CA2203156A1 - Hybrid filter bank analog/digital converter - Google Patents
Hybrid filter bank analog/digital converterInfo
- Publication number
- CA2203156A1 CA2203156A1 CA002203156A CA2203156A CA2203156A1 CA 2203156 A1 CA2203156 A1 CA 2203156A1 CA 002203156 A CA002203156 A CA 002203156A CA 2203156 A CA2203156 A CA 2203156A CA 2203156 A1 CA2203156 A1 CA 2203156A1
- Authority
- CA
- Canada
- Prior art keywords
- time
- subband signals
- analog
- bank
- filters
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Abandoned
Links
Classifications
-
- H—ELECTRICITY
- H03—ELECTRONIC CIRCUITRY
- H03M—CODING; DECODING; CODE CONVERSION IN GENERAL
- H03M1/00—Analogue/digital conversion; Digital/analogue conversion
- H03M1/12—Analogue/digital converters
- H03M1/1205—Multiplexed conversion systems
- H03M1/121—Interleaved, i.e. using multiple converters or converter parts for one channel
-
- H—ELECTRICITY
- H03—ELECTRONIC CIRCUITRY
- H03H—IMPEDANCE NETWORKS, e.g. RESONANT CIRCUITS; RESONATORS
- H03H17/00—Networks using digital techniques
- H03H17/02—Frequency selective networks
- H03H17/0248—Filters characterised by a particular frequency response or filtering method
- H03H17/0264—Filter sets with mutual related characteristics
- H03H17/0266—Filter banks
Abstract
A hybrid filter bank analog-to-digital converter includes continuous-time analysis filters and discrete-time synthesis filters. The continuous-time analysis filters partition a continuous-time wibedand input signal into continuous-time subband signals. An analog-to-digital converter bank quantizes the subband signals at a low data rate. A bank of upsamplers increases the data rate of the quantized subband signals. A bank of discrete-time synthesis filters processes the upsampled subband signals, generating signals which are the discrete-time approximation of the continuous-time subband signals. The subband signals may be recombined into a discrete-time wideband signal which is the discrete-time approximation of the continuous-time wideband input signal. The linearity errors, analog-to-digital converter mismatches and quantization noise are not compounded between the frequency bands, thereby increasing resolution. The system may also include compensators for compensating for non-linear behavior in the analog-to-digital converters and may include application-specific processors for further processing of the compensated subband signals. An alternative embodiment employs discrete-time analysis filters and continuous-time synthesis filters for digital-to-analog conversion.
Description
CA 022031~6 1997-04-18 W O 96/13097 PC~rAUS95/13916 HYBRID FILTER BANK ANALOG/DIGITAL CONVERTER
Backqround of the Invention The rapid expansion of digital signal processing hardware and software has dramatically increased the demand for high-speed and high-resolution analog-to-digital converters. The most popular techniques foranalog-to-digital conversion to date include successive approximation, sigma-delta, sub-ranging, flash, and time-interleaving.
The advancement of analog-to-digital converter technology involves a trade-off between sampling rate in samples per second (Sa/s) and resolution in bits.
Sigma-delta converters currently provide high resolution (20 bits) at relatively low sampling rates (10 kSa/s) while time-interleaved converters provide low resolution (8 bits) at high sampling rates (8 GSa/s). The remaining techniques including successive approximation (100 kSa/s at 16 bits resolution), subranging (1 MSa/s at 14 bits resolution), and flash (100 MSa/s at 10 bits resolution), span the range of sampling rate, resolution, and cost between the time-interleaving and sigma-delta techniques.
Recently, manufacturers began studying and perfecting time-interleaved analog-to-digital conversion. In this technique, a bank of time-multiplexed analog-to-digital converters with low sampling rates are combined to realize a single high-speed, high-resolution analog-to-digital converter.
The primary drawback of time-interleaved analog-to-digital conversion is that linearity errors and mismatches between each of the converters in the bank are compounded across the full bandwidth of the device, CA 022031~6 1997-04-18 limiting the resolution of the system. Time-interleaving is also prone to timing errors because it requires very accurate interleaved clock signals which limits the speed and resolution of the system and which introduces harmonic distortion. In addition, a voltage offset between two converters in the bank may cause the converters to digitize the same input voltage to different codes. Since the converters output data sequentially, the output exhibits an error signal having a period equivalent to the sample period of a single converter, but the frequency content can increase because each cycle of the error contains a number of offset errors equivalent to the number of converters in the bank. Therefore, this source of error introduces harmonic distortion spurs that limit the effective resolution of the system. A difference in voltage gain between two converters also causes each converter to digitize the same input voltage to a different code. The magnitude of the error grows as the input voltage grows. Overcoming these fundamental limitations is difficult because it requires reduction of error for each converter. The gain and offset of an individual converter can be adjusted with external resistors and the remaining harmonic distortion can be reduced with dynamic compensation techniques. However, the remaining linearity errors of each converter are still compounded across the full bandwidth of the system.
Discrete-time Quaudrature Mirror Filter (QMF) banks have been employed in analog-to-digital conversion techniques as described in:
CA 022031~6 1997-04-18 W O 96/13097 PC~rrUS95/13916 Antonio Petraglia and Sanjit K.
Mitra, "High-Speed A/D Conversion Incorporating a QMF Bank," IEEE
Transac ti ons on Ins trumen ta ti on and Measurement, 41 (3) :42 7-43 1 (June, 1992 ) .
In this system, discrete-time switched-capacitor analysis filters decompose the wideband input signal into several contiguous frequency subbands. An individual analog-to-digital converter is assigned to each subband. All converters are driven by a common clock. The subbands are allocated certain quantization bits thereby allowing separate control of resolution in each subband. Discrete-time finite impulse response 15 synthesis filters reconstruct each subband, cancelling the aliasing caused by other subbands. Therefore, the errors associated with a particular subband are not compounded with errors of other subbands as in time-interleaved analog-to-digital conversion. In addition, the discrete-time QMF bank approach does not require extremely accurate time-skewed clock signals.
The primary disadvantage of the discrete-time QMF
bank analog-to-digital converter is that the switched-capacitors introduce switching noise that can limit the signal-to-noise ratio and thereby limit the resolution and speed of the system. Typical switched-capacitor filters have a signal-to-noise ratio of about 85 dB and are limited to sampling rates of about 150 kSa/sec.
This limits the bandwidth of the discrete-time QMF
bank.
Summary of the Invention The present invention is directed to an apparatus and method for high-speed high-resolution analog/digital conversion. Note that the term "analog/digital" as used herein includes both analog-CA 022031~6 1997-04-18 W O 96/13097 PCTrUS95/13916 to-digital and digital-to-analog conversion. In the present invention, converter quantization noise and linearity errors are confined to each individual subband and are therefore not compounded. The converter bank is driven by a single clock, simplifying timing constraints. Synthesis filters compensate for remalnlng tlmlng errors.
For subband coding applications such as compression or adaptive array processing, the present invention is capable of direct channelization without the requirement of additional hardware. The present lnvention is amenable to non-uniform channel bandwidths and can be implemented in very large scale integration (VLSI). The invention overcomes the limitations and disadvantages of the conventional discrete-time QMF
array architectures.
In accordance with the present invention, a converter bank is incorporated into a hybrid system having, in the case of an analog-to-digital converter, continuous-time analysis filters and discrete-time synthesis filters. Integration of continuous-time filters and discrete-time filters does present design problems, but they have been overcome with this invention. Continuous-time analysis filters improve system speed and resolution by overcoming the limitations of switched-capacitor filters used in the prior art. The continuous-time analysis filters partition a continuous-time wideband input signal into continuous-time subband signals. In a preferred embodiment, an analog-to-digital converter, a rate changer, and a discrete-time synthesis filter are provided for each subband. The individual analog-to-digital converters in the array quantize the subband signals into binary bits. The rate changer, in the form of an upsampler or downsampler increases or CA 022031~6 1997-04-18 WO 96/13097 PCI~/US95/13916 decreases the data rate of the quantized subband signal. The discrete-time synthesis filters process the subband signal, generating a signal which is the discrete-time approximation of the continuous-time - 5 subband signal.
An adder may be provided for recombining the discrete-time subband signals into a discrete-time wideband signal which is the discrete-time approximation of the continuous-time wideband input signal. A discrete-time compensator may compensate for non-linear behavior in the analog-to-digital converters. The synthesis filters compensate for gain error, aliasing error and non-linear phase in the continuous-time analysis filters. Application-specific processors may be included for further processing of the subband signals either before or after conversion to discrete-time. Non-uniform subband channel bandwidths may be employed as well as uniform subband channel bandwidths. A tree-structure architecture may be implemented to simplify the multiple-channel parallel architecture into a hierarchy of multiple-channel systems. A wavelet-type structure may be implemented for systems requiring non-uniform channel bandwidths.
An alternative embodiment for digital-to-analog conversion employs discrete-time analysis filters to partition a wideband discrete-time input signal into subbands. Each subband is assigned its own digital-to-analog converter and subband processor. Continuous-time synthesis filters process the subband signal. The subband signals may be recombined to form a wideband, continuous-time output signal.
In a preferred embodiment of the invention, the hybrid system comprises an M-channel, maximally-decimated, parallel, uniform bandwidth architecture, CA 022031~6 1997-04-18 W O 96/13097 PC~rtUS9Stl3916 where M represents an integer. In a maximally-decimated embodiment, each subband converter samples at 1/M the effective sampling rate of the system. The wideband input signal is partitioned into M subband signals by M analysis filters. For the digital-to-analog conversion case, each subband signal is then downsampled by a factor of M. Each subband signal is then passed through a converter, either analog-to-digital or digital-to-analog and processed. For the analog-to-digital case, the subband signal is then upsampled by a factor of M. Next, the subband signal is processed in a bank of M synthesis filters and recombined to form a wideband output signal.
A more efficient implementation employs polyphase decomposition to interchange the order of analysis filtering and downsampling in the digital-to-analog case, or upsampling and synthesis filtering in the analog-to-digital case. This allows for filtering at the subband data rate. To reduce the effects of aliasing in the maximally-decimated system and to improve resolution, oversampling may be used. A subset of the analog/digital converters in the bank may operate at higher resolution than the remaining analog/digital converters for applications requiring greater dynamic range at select subbands. The subbands may be mixed down to baseband or up to higher frequencies to alleviate the requirement for high-speed, high-precision sample and hold circuitry in the analog/digital converters.
An optimization algorithm determines the analysis filters and the synthesis filters in an iterative process which incrementally adjusts the analysis and synthesis filters such that the system distortion function converges on a perfect delay and such that the system aliasing functions converge on zero.
CA 022031~6 1997-04-18 WO96/13097 PCT~S9S/13916 Brief Description of the Drawings FIG. 1 is a block diagram of a general hybrid filter bank for analog-to-digital or digital-to-analog conversion in accordance with the present invention.
FIG. 2 is a block diagram of a hybrid filter bank analog-to-digital converter having subband processor banks and a discrete-time compensation bank for compensating for non-ideal analog behavior in accordance with the present invention.
FIG. 3 is a block diagram of an M-channel, maximally-decimated, parallel, uniform channel width, hybrid filter bank analog-to-digital converter in accordance with the present invention.
FIGs. 4A and 4B are block diagrams of equivalent methods for filtering. FIG. 4A is a representation of filtering following upsampling. FIG. 4B shows an equivalent structure using polyphase decomposition to enable efficient filtering at the lower data rate prior to upsampling. In the analog-to-digital case, this can be used to enable efficient synthesis filtering at the lower data rate.
FIGS. 4C and 4D are block diagrams of equivalent methods for filtering. FIG. 4C is a representation of filtering followed by downsampling. FIG. 4D shows an equivalent structure using polyphase decomposition to enable efficient filtering at the lower data rate after downsampling. In the digital-to-analog case, this can be used to enable efficient analysis filtering at the lower data rate.
FIG. 5A is a block diagram of a tree-structured hybrid filter bank analog-to-digital converter. FIG.
5B is a block diagram of a wavelet-type hybrid filter bank analog-to-digital converter.
CA 02203l~6 1997-04-l8 WO96/13097 PCT~S9S/13916 FIG. 6 is a block diagram of a hybrid filter bank digital-to-analog converter in accordance with the present invention.
FIG. 7 is a flow diagram for an optimization algorithm for calculating analysis filters Hk (5), and then for calculating discrete time synthesis filters Fk (Z) based on the analysis filters for an analog-to-digital converter in accordance with the present lnvent lon .
Detailed Description of the Preferred Embodiments The hybrid filter bank of the present invention is a novel class of filter bank which is less understood than the all-discrete-time class. The complexity of the hybrid filter bank lies in the relationship between the continuous-time analysis filters and the discrete-time synthesis filters.
FIG. 1 is a block diagram of a general hybrid filter bank analog/digital converter 11. An analysis filter bank 12 partitions a wideband input signal 10 into subband signals 12A. Each subband signal 12A is assigned its own analog/digital converter and subband processor 14 where the subband signal 12A undergoes conversion and processing. A synthesis filter bank 16 further processes the converted and processed subband slgnals 14A. The synthesized subband signals 16A are combined at an optional adder 18 to form a wideband output signal 20. For analog-to-digital conversion, continuous-time analysis fiIters 12 and discrete-time synthesis filters 16 are employed and the synthesis bank 16 may include data rate changers to increase or decrease the data rate. For digital-to-analog conversion, discrete-time analysis filters 12 and continuous-time synthesis filters 16 are employed and the analysis bank 12 may include data rate changers for CA 02203l~6 1997-04-l8 WO96/13097 PCT~S95/13916 increasing or decreasing the data rate. The data rate changers are either upsamplers, which increase the data rate by inserting zeros between the samples, or downsamplers, which decrease the sampling rate by periodically discarding samples.
The embodiment of FIG. 1 provides for subband channelization of the wideband input signal 10 at signals 16A or 12A without the requirement of additional hardware. This is particularly useful in applications such as adaptive array processing and compression techniques that perform subband coding.
A hybrid filter bank analog-to-digital converter with discrete-time compensation is shown in the block diagram of FIG. 2. Rather than filtering the continuous-time wideband input signal 22 with discrete-time switched-capacitors before analysis filtering in discrete-time as does the prior art, the embodiment of FIG. 2 overcomes the limitations of the prior art by using continuous-time analysis filters 24.
A continuous-time wideband input signal 22 is partitioned into continuous-time subband signals 24A by the continuous-time analysis filter bank 24. A bank of analog-to-digital converters 26, one converter for each subband, converts the continuous-time subband signals 24A into discrete-time subband signals 26A. An optional discrete-time compensator bank 28 may be employed for compensating for non-linear behavior in the analog-to-digital converters 26. The compensated discrete-time subband signals 28A may be further processed by an optional discrete-time subband processor bank 30 for frequency-dependent processing of individual subband data streams, for example, adaptive-array processing and compression. Similarly, the subband processors 25 may optionally be employed for processing the continuous-time subband signals 24A
CA 022031~6 1997-04-18 W O 96/13097 PC~rrUS95/13916 before quantization by the analog-to-digital converters 26. The resultant signals 30A are further processed by the discrete-time synthesis filter bank 32 which may include data rate changers for increasing or decreasing the data rate. The synthesis filters 32 compensate for non-linear phase and gain errors in the analysis filters 24 and also cancel aliasing.
The output 32A of the discrete-time synthesis filter bank 32 is the discrete-time approximation of the continuous-time subband inputs 24A. The synthesis filter bank 32 processes the signals so that upon recombination at an optional adder 34, aliasing is cancelled and the subband signals 32A are combined into a wideband discrete-time output signal 40 which is the discrete-time approximation of the wideband continuous-time input signal 22.
The subbands 24A may be mixed down to baseband with mixers 27 before sampling at the analog-to-digital converters 26 to alleviate the requirement for high-speed, high-precision sample-and-hold circuitry in the analog-to-digital converters 26. This allows for analog-to-digital conversion at a lower sampling rate, and increases the dynamic range of the system by alleviating the wide analog bandwidth constraints of the converter bank. An analogous structure may be used to mix the subband signals up to higher frequencies after conversion for the digital-to-analog case.
FIG. 3 is a block diagram representing an M-channel, maximally-decimated, parallel, uniform channel, hybrid filter bank analog-to-digital converter. Continuous-time analysis filters Hk (5) partition a wideband continuous-time input u(t) into M
consecutive, uniform bandwidth subband signals xk(t).
Each continuous-time subband signal xk(t) is converted to discrete-time by an analog-to-digital converter 42 CA 022031~6 1997-04-18 W O 96/13097 PC~rAUS95/13916 sampling at a subband rate of l/M multiplied by the effective sample rate of the system. Each discrete-time converted subband signal xk(n)is upsampled 44 by a factor of M. Each upsampled discrete-time subband signal vk(n) is processed by a discrete-time synthesis filter Fk (Z)- The resulting signals Yk(n) are optionally combined at an adder 46 to form a wideband discrete-time output signal y(n).
A mathematical derivation of the properties of the hybrid filter bank analog-to-digital converter shown in FIG. 3 will now be described. The first goal is to determine the input/output function for this architecture which is derived below in equations 1 through 9. The second goal is to find expressions for the Fourier transform of the synthesis filters Fk (eiW) that provide perfect reconstruction given a set of analysis filters Hk (iQ) which are defined in equations 10 through 16, with an example given for M=2 at equations 17-22.
In the frequency domain, the result of filtering the input U(jQ), bandlimited to Qn rad/sec, through each analysis filter Hk (iQ) is Xk(jQ), Xk( jQ) = Ut jQ) Hk( jQ) (1) In the time-domain, analog-to-digital converter 42 errors can be modeled as gain errors ak and DC offset errors bk, Xk[n]=(l+ak)Xk[n] + bk (2) The effects of analog-to-digital converter errors are seen in the frequency domain along with aliasing error WO96/13097 PCT~S95/13916 The effects of analog-to-digital converter errors are seen in the frequency domain along with aliasing error due to undersampling, where T is the sampling period, T=M7r/n", xk(ei~)= T ~0 Xk( T T (3) kl _ o 2 r~ ~ ( ~ - 2 ~ 1 ), 0 < ~1) < 2 Substituting for X~(jQ), 1+a M-1 ~ j2 1 J ;Zn1 + bk 2 2~ (~-2~1), O<~(M-1) 2~ .
The result of upsampling by a factor of M is represented as V~.(~).
Vk( e j"') =Xk( e j~M) =--T k 2 U( j T ~ j T ) Hk( j T ~ j T
+bk 2 2~0 (~M-2r~1), 0<~ (M 1) 2J~.
The result of filtering Vk(ei~) through synthesis filters Fk(~J is represented as Yk(~J, Yk(e ) =Fk(ej ) Vk(ej~) Tk ~ Fk(ej~) U( j(T)M- j2Tl )Hk( jtM-~) (6) M-l +b 2 F (e~) 2~ (~M-2J~1) W O 96/13097 PC~r~US95/13916 M-l .. Y(ej~) = 2 Yk(ei~) . (7) Substituting for Yk(ei~), the reconstructed wideband output signal Y(eiW) is Y(ei'~) = 2 T k 2 Fk (ei'~) U( j T - j T
. (H ( j(~M j2 1~) (8) M-1 M-l + ~ bk ~ Fk(ej~) 27~ M-2 k=O 1=0 Rearranging the summation gives the output Y(ei~), in terms of the input U(jQ), the analysis filters Hk (ei~) and the synthesis filters Fk (ei~) .
y(ej~) = 2 U( jT - j T ) ¦~ 2 (1 ak) Fk(ej~.,)Hk( ~ M_ j2 1~¦ (9) M-1 M-l + 2 bk 2 Fk(ej(i') 2rc8 ((~M-2t~1) k=O 1=0 This expression is simplified by introducing functions Tp (ei~) and ~, Y(ej~) = 2 U( jTM- j2Tp) Tp(ei'~') + ~ , (10) where CA 022031~6 1997-04-18 W O 96/13097 PC~rAUS95/13916 T (ej~) = 2 k Fk(ej~)Hk( jT - i T ) ~ (11) O < ~ (MMl) and M-l M-l = 2~ 2 bk Fk(ej~ (~M-2~p) . (12) k=O p=O
Equations 10, 11, and 12 describe the mathematical input/output function for the hybrid filter bank analog-to-digital converter on the interval Oc~5 (M-1)2~
M
Perfect reconstruction occurs when the sole effect of the filter bank architecture is to introduce a delay, Y(ej~) = Ae~j~dU(ej~) . (13) Therefore, the distortion function To (ei~) should correspond to a perfect delay, d, To(e j~) = Ae -j~d . (14) Also, the aliasing functions, Tp (ei~) where lspsM-l, should be equal to zero, CA 022031~6 1997-04-18 WO96/13097 pcT~sssll39l6 Tp(ej~) = 0 . (15) ~ corresponds to DC offset errors in the analog-to-digital converters and should be zeroed by adjusting each analog-to-digital converter offset bk to zero.
~ = o . (16) Using these constraints, equation 11 can be solved as a system of simultaneous linear equations for the Fourier transform of the desired synthesis filters Fk(ej~) that provide perfect reconstruction given a set of analysis filters Hk (eiW) .
As an example, for M=2, the input/output function 10 is Y(ej'l') =U ( j2 ) Tc (ei'l')+ U ( j2w_ j2 )Tl(ei~')+~
For perfect reconstruction, the distortion function should be equal to a perfect delay, TO~ei )=--T (l+aO)Fo(ei )Ho( T ) (18) + T(l+al)Fl(ei)Hl( T )=e i , Oc~s~, and the aliasing function should be equal to zero, WO96/13097 PCT~S95/13916 and the aliasing function should be equal to zero, Tl(ej~)= T(l+aO)FO(e )HJ~ T T ) (19) -T(1+a~)FI(ej )Hl( jT - jT )=' <~<~ .
Also, the DC offset errors, ~, should be set to zero ,/3=2~, b,~ F, (e~ (~) +~ -7r) ) = (20) 1:=0 Solving the system of simultaneous equations above for F0(ei~J and F~(ej~J, the Fourier transform of the synthesis filters that provide perfect reconstruction gives Fo ( e i~ ) = A e ~ j~ d T
( T - jT ) (21) ~ Ho( T )H1( T - jT ) -Hl ( i T ) Ho( j2 _ j2 and Discrete-time synthesis filters, Fk(Z), are determined by calculating a system function that approximates the desired frequency responses, F~
using the optimization described below. The order of CA 022031~6 1997-04-18 W O 96/13097 PC~rrUS95/13916 using the optimization described below. The order of the synthesis filter is determined by the level of error that can be tolerated in the system. High-order synthesis filters reduce reconstruction errors because they more closely approximate the desired frequency responses than do low-order synthesis filters.
The reverse structure of the hybrid filter bank analog-to-digital converter is the hybrid filter bank digital-to-analog converter which uses discrete-time analysis filters, downsamplers, an array of digital-to-analog converts and continuous-time synthesis filters.
It can be shown that using the continuous-time and discrete-time filters designed for the hybrid filter bank analog-to-digital converter will result in the same reconstruction error as the hybrid filter bank digital-to-analog converter. The digital-to-analog converter is described below in conjunction with FIG.
6.
The optimization algorithm described below in conjunction with FIG. 7 calculates the optimal system delay d and formulates suitable analysis and synthesis filters which enable accurate reconstruction. A near-perfect reconstruction solution can be designed by approximating a discrete-time perfect reconstruction solution, but good approximations require high-order filters. The optimization algorithm described below calculates near-perfect-reconstruction hybrid filter banks that are not necessarily hardware intensive.
The algorithm has two sections, the analysis filter optimization, followed by the synthesis filter optimization. From a given lowpass continuous-time analysis filter, Ho(sJl the analysis filter optimization calculates the remaining bandpass analysis filters Hk(s), lsksM-l, and the system delay, d, to minimize reconstruction error. The optimization CA 022031~6 1997-04-18 calculates samples of the functions Tp (ej~), OspsM-l, and iteratively adjusts the analysis filters Hk (s), and system delay, d, such that the distortion function, T0(ej~) is as close to a perfect delay, e~j~d, as possible and the aliasing functions Tp(ej~), for lsp~M-1 are as close to zero as possible. The algorithm uses a least-squares error criterion with the Matlab multi-variable minimization algorithm fmins.
Given analysis filters Hk (s) and system delay d, calculated above, the synthesis filter optimization calculates the synthesis filters Fk (Z) that minimize the reconstruction error. One method is the Matlab function invfreqz which calculates system functions Fk(z) whose Fourier transforms approximate the ideal Fourier transforms. Such an algorithm is quick enough to provide a rough solution, but may not provide the highest accuracy reconstruction possible. This algorithm is used in the analysis filter optimization above to evaluate the aliasing functions Tp(e~
Similar to the analysis filter optimization, the synthesis filter optimization iteratively adjusts the synthesis filters, Fk(z), to minimize a least-squares error criterion in Tp(ej~) with the Matlab multi-variable minimization algorithm fmins. The best results are obtained when the algorithm uses the analysis filters calculated in the analysis filter optimization above, but it also works for any set of analysis filters, enabling the use of standard filter types (such as Butterworth, Chebyshev, or Elliptic filters). The synthesis filter optimization also enables accurate calibration of the hardware system by optimizing the synthesis filters based on measurement of the hardware analysis filters.
The technique of FIG. 3 requires synthesis filters Fk (Z) operating at the effective sample rate of the CA 022031 j6 1997-04-18 WO 96/13097 PCT/US9!i/13916 system, which can be extremely fast. The upsampler 48 and synthesis filter Fk (ZJ combination from FIG. 3 is reproduced in FIG. 4A. FIG. 4B shows a polyphase decomposition technique which is the equivalent of the combination shown in FIG. 4A. Polyphase decomposition rearranges the order so that synthesis filtering Rk (Z) is performed before upsampling. Synthesis filtering Rk (Z) is performed at the lower sample rate of the analog-to-digital converters, which is 1/M multiplied by the effective sample rate of the system. Polyphase decomposition replaces each upsampler 44 and length L
FIR synthesis filter Fk with M length L/M filters Rk (Z) followed by M upsamplers 50. The process requires M
filters for each subband, but each filter has fewer coefficients. A similar polyphase decomposition structure can be used to enable analysis filtering at the subband sample rate in the digital-to-analog case.
In an oversampling system where the analog-to-digital converters are sampling at a rate higher than the effective sampling rate of the hybrid filter bank, downsampling may be needed after filtering and conversion as shown in FIG. 4C. FIG. 4D shows an embodiment for polyphase decomposition wherein the converted signal Xk(~) is first downsampled 51 to lower the data rate, and then filtered by synthesis filters Rk tZ)- In this embodiment, the filtering Rk (Z) is performed at the lower data rate. In general, the polyphase filtering methods in FIGS. 4B and 4D may be used to enable filtering at a lower data rate by interchanging the order of data rate changing and filtering.
One technique for reducing aliasing is to increase analysis filter cut-off slope and stopband attenuation.
However, this technique increases hardware complexity.
CA 022031~6 1997-04-18 Another improved technique is oversampling. The maximally-decimated architecture shown in FIG. 3 is, by definition, operating at the theoretical limit Nyquist sampling rate with no leeway for aliasing.
oversampling would widen the frequency channels allocated to each converter in the array and thus the channels would overlap. This places less emphasis on the analysis filters Hk (s) to reduce aliasing, and relieves the sharp cut-off slope and high stopband attenuation constraints, thereby allowing lower order analysis filter Hk (s) implementations with reduced hardware requirements. Oversampling also separates quantization noise from the signal and the noise is filtered out by the synthesis filters, reducing system noise and improving system resolution. Each factor of four in sampling speed improves resolution by 1 effective bit. In the analog-to-digital converter case, where oversampling is used, it may be necessary to downsample the converted subband signal after synthesis filtering because, in the oversampled system, the individual converters may be operating at a rate higher than the effective sampling rate of the entire system.
The M-channel hybrid filter bank can be simplified to form a series of multiple-channel blocks arranged in a tree structure. This simplifies the optimization algorithm and design procedure. FIG. 5A is a block diagram of a two-channel tree structure for analog-to-digital conversion. The analysis filter bank 52 comprises hierarchical pairs of continuous-time analysis filters Ho(s) and H1(s) which partition the input signal u(t) into subbands, the subbands being further partitioned by continuous-time analysis filters Hoo(S)~ Ho1(S)l H1o(S)I H1l(s). Each subband signal is converted to discrete-time 53, upsampled 54, and then CA 022031~6 1997-04-18 processed and recombined in a hierarchical discrete time synthesis filter bank 55. Rather than optimizing an M-channel system, which can be prohibitively time-consuming, the optimization is reduced to a series of two-channel optimizations. This embodiment is not limited to two-channel blocks, nor is it limited to two levels of hierarchy. This technique may also be applied to digital-to-analog conversion.
The invention may also be implemented in a wavelet-type structure as shown in FIG. 5B. The continuous-time input signal u(t) is partitioned into subband signals by a continuous-time analysis filter Ho(s)~ One of the subbands u1(t) is further partitioned into smaller subbands u2(t) by an analysis filter H1(s) and each subband signal is converted to discrete-time 56 and upsampled 57. The subbands are recombined in a wavelet-structured synthesis filter bank 58. This structure is applicable in systems requiring non-uniform channel bandwidths for multi-resolution or wavelet analysis. This technique isappropriate for example, in audio applications, which require a wider bandwidth at lower frequencies.
The invention may also include higher resolution analog-to-digital converters in a subset of the subbands for applications requiring additional dynamic range at select frequency bands. For example, certain subbands may employ a converter with a higher number of quantization bits. Alternatively, certain subbands may be sampled at higher rates.
FIG. 6 is a block diagram of a M-channel maximally-decimated, parallel, uniform channel, hybrid filter bank digital-to-analog converter in accordance with the present invention. A wideband discrete-time input signal u(n) is partitioned into discrete-time subband signals uk(n) by discrete-time analysis filters CA 022031~6 1997-04-18 W O 96/13097 PC~rrUS95/13916 Hk (Z) . The subband signals uk(n) are downsampled 60 and the resulting downsampled signals xk(n) are converted from discrete-time to continuous-time by a bank of digital-to-analog converters 62. The resulting continuous-time subband signals xk(t) are further processed by continuous-time synthesis filters Fk (s) which generate signals Yk (t) which are the continuous-time approximation of the discrete-time subband signals uk(n). The continuous-time subband signals Yk (t) may be combined at an adder 64 to form a wideband continuous-time output signal y(t) which is the continuous-time approximation of the discrete-time wideband input signal u(n). Note that the digital-to-analog conversion technique shown in FIG. 6 may be implemented with any of the embodiments described herein, including, but not limited to polyphase decomposition, tree-structure, wavelet structure, subband processing, subband compensation and oversampllng.
FIG. 7 is a flow diagram of an algorithm for optimizing the analysis filters Hk(s) and synthesis filters Fk(z). The algorithm is broken down into two stages 92, 94. In the first stage 92, analysis filters Hk(s) are calculated and synthesis filters Fk(z) are approximated to minimize reconstruction error. In the second stage 94, the best synthesis filters are calculated to minimize reconstruction error. In each stage, the resulting input/output function for the system is evaluated, assuming that the DC offset errors ~ in the analog-to-digital converters have been zeroed, and then an error function is calculated. The error function quantifies the difference between the resulting input/output function and an ideal delay.
This data is used to make another guess at the filters CA 022031~6 1997-04-18 to reduce reconstruction error. This process repeats until the error is below a prescribed threshold in each stage.
In the first stage 92, a guess is made with a 5 standard multi-variable minimization routine, such as the Matlab routine fmins, for the analysis filters Hk(s) and the system delay d, 80. The synthesis filters Fk (Z) are then roughly approximated by fitting the frequency response curve to an optimal frequency 10 response 82 using standard filter approximation routines such as the Matlab routine invfreqz. The errors in Tk(eiW) are then evaluated 84 and a determination is made as to whether the error in Tk(eiW) is below a threshold le~el 86. If the error is 15 not below threshold 86, then another iteration of the first stage 92 is required and the process continues to adjust the analysis filters Hk(s) with a standard minimization algorithm until the error is below threshold. If the error is below threshold, then the 20 algorithm proceeds to the second stage 94.
In the second stage 94, given analysis filters Hk(s), synthesis filters Fk(z) are calculated which minimize reconstruction error. A measurement of the response of the analysis filters Hk(s) is made 87.
25 Next, the errors in Tk(eiW) are evaluated 88. A
determination is then made as to whether the error in Tk(eiW) is below a prescribed threshold 90. If the error is not below threshold, then the synthesis filter Fk(z) is adjusted with a standard minimization 30 algorithm to minimize the error 96 and another iteration of the second stage 94 is required. If the error is below threshold 90, then the analysis filters Hk(s) and synthesis filters Fk(z) are determined.
In practice, the frequency response of the 35 analysis filters Hk(s) must be measured 87 to determine CA 022031~6 1997-04-18 WO96/13097 PCT~S95/13916 their actual transfer functions, since the analysis filters calculated in the first stage 92 of the optimization can never be ideally implemented in hardware. The error function in the optimization algorithm requires accurate analysis filter amplitude and phase data at g equally-spaced frequency points.
One approach involves directly measuring the frequency response of the analysis filters at g equally-spaced frequencies. Accurate measurements are required or reconstruction is degraded.
The order of the synthesis filters will have a direct impact on how accurately the ideal solutions can be approximated. For FIR synthesis filters, once the length L has been decided, the initial guess of the optimal delay d of the system can be set to L/2. The rationale for this choice is that the magnitude of the impulse response coefficients is greatest in the vicinity of the dth coefficient and the magnitude tapers off on either side. Choosing d to be one-half of the length of the impulse response assures that the most significant coefficients are not truncated by the finite length of the impulse response.
The analog-to-digital converter gain errors ak and offset errors bk must also be measured 87. By measuring the amplitude of the harmonics that the analog-to-digital converter produces when sampling a pure sinusoid, an analog-to-digital converter non-linearity curve can be deduced, where the zero order term corresponds to offset errors bk which should be nulled to zero, and the first order term corresponds to gain errors ak.
The error function Tk(ei~) to be minimized must be calculated. As stated above, perfect reconstruction occurs when the only effect of the filter bank architecture is to introduce a delay. Thus, the CA 022031~6 1997-04-18 W O 96/13097 PC~rrUS95/13916 distortion function T0(ej~) should be as close to perfect delay Ae~iWd as possible, and the aliasing functions Tp(ej~), lspsM-1, should be as close to zero as possible. o(~) is a measurement at a specific frequency ~1~ of the squared difference between the distortion function T0(ej~) and an ideal delay Ae~j~d (both real and imaginary parts), Eo((I)) = [~{To (e j~) } --~R{Ae j~d} ] 2 (23) + [~{To (ej~) } -~{Ae~i~d} ] 2 p (w) is a measurement at a specific frequency ~ of the squared difference between the aliasing function Tp (ej~) and zero (both real and imaginary parts), Ep(~) = ~{TP(e j~) }2 + ~{TP(e j~) }2, l~p~M-1 . (24) Therefore, if the system is a perfect delay at frequency ~, both o(~) and p(~) will be zero. The system should be as close to a perfect delay as possible at all frequencies, so total error function should sample o(~) and p(~) at g equally-spaced frequencies to assure that optimization occurs across the entire frequency band. The total error function that the algorithm minimizes is = ~ EPEP ( ) . ( 25) where EP are weighting factors that can be used to favor reduction of either the aliasing error or the distortion error.
The number of frequency points g to be examined in the error function must be determined. If L is the CA 022031~6 1997-04-18 W O 96/13097 PCTrUS95/13916 length of each synthesis filter, then the minimization program will be operating on a transfer function with ML variables, the L coefficients of the M finite impulse response synthesis filters. If g is chosen as ML, then the input/output function will be relatively smooth between the frequency response measurement points. In other words, the algorithm is reasonably assured of optimizing the response over the entire frequency band, not just at the g frequencies that the algorithm uses. For example, in an M=2 system, if the length of the synthesis filters is chosen to be L=64, then evaluating the error function at g=128 equally-spaced frequencies is the minimum recommended.
The total error function is a function of analysis filters Hkfs), gain errors ak, and synthesis filters Fk(z). Hk(s) and ak are first measured, and an initial guess at Fk(z) is made. The minimization of in both the first stage 92 and the second stage 94 is done numerically using a routine such as the MatlabTM
function fmins, which uses a Nelder-Mead simplex search to iteratively adjust filters Fk(z), Hk(z), or system delay d to minimize the error . The algorithm is a direct search that does not require derivative information. Let C be the number of variables the algorithm is manipulating to minimize the error. A
simplex in C-dimensional space is characterized by the C+1 distinct vectors which are its vertices. In two-space, a simplex is a triangle; in three-space, it is a pyramid. At each step of the search, a new point in or near the current simplex is generated. Error evaluated at the new point 84, 88 is compared with the error value at the vertices of the simplex, and usually, one of the vertices is replaced by a new point, giving a new simplex 80, 96. This step is CA 022031~6 1997-04-18 W O 96/13097 PC~rAUS95/13916 repeated until the diameter of the simplex is less than the specified tolerance 86, 90.
An implementation of the Hybrid Filter Bank analog-to-digital converter preferably includes an efficient automatic calibration routine that can be performed on a regular basis. One method of calibration includes injection of a known wideband analog signal into the continuous-time analysis filters, followed by the optimization of the synthesis filters to provide the best mathematical reconstruction of the original continuous-time signal. Measurement and modeling errors will be bypassed using this method because the process is directly optimizing the system based on the performance of the analysis filters.
Although the process is computationally intensive, once the initial calibration has been done, re-calibration could be accomplished with a reasonable amount of computation on a regular basis. The computational requirements of the M-channel case can be reduced by 2 0 simplifying the architecture to form a series of recursive two channel blocks arranged in a tree structure as shown in FIG. 5A.
The analysis filter type affects the accuracy of the synthesis filter reconstruction. Reconstruction accuracy is also a function of the order, ripple, and group delay of the analysis filters. In addition to analysis filters designed by the first stage 92 of the optimization algorithm, experiments have been conducted on several standard types of analysis filters including Elliptic, Chebyshev I, and Butterworth. The experiments demonstrated that reconstruction accuracy is nearly independent of analysis filter order.
However, low order analysis filters with greater stopband attenuation generated greater out-of-band attenuation of analog-to-digital converter linearity CA 022031~6 1997-04-18 WO 96/13097 PCI'/US95/13916 errors and quantization errors. These errors are compounded across the subband channels similar to the compounding of errors in time-interleaved analog-to-digital converters, thereby increasing overall distortion. Note that this analog-to-digital converter error should not be confused with the mathematical reconstruction error of the hybrid bank structure discussed above. Filters with passband ripple and non-constant group delay, such as Chebyshev and Elliptic filters, have lower reconstruction accuracy. Better reconstruction was achieved with Butterworth filters since they have no passband ripple and since group delay is relatively constant. The best analysis filters were designed with the first stage 92 of the optimization algorithm.
The synthesis filters reconstruct the signal by suppressing upsampled images, cancelling aliasing and compensating for the analysis filter ripple and nonlinear phase. The order of the synthesis filter determines how accurately the objectives are met.
Increasing the analysis filter cut-off slope and stopband attenuation decreases the amount of aliasing.
This alleviates the requirements imposed on synthesis filters to cancel aliasing, thereby improving reconstruction accuracy. This also increases out-of-band attenuation, which suppresses linearity errors and quantization noise between the frequency bands.
A two channel hybrid filter bank system employing seventh order analog Butterworth analysis filters and length 64 finite impulse response 16-bit digital synthesis filters optimized by the second stage 94 of the optimization algorithm was simulated. The system exhibited 0.113 dB average deviation from 0 dB
distortion (3.13 dB peak distortion at w=~ rad/s), -81.5 dB average aliasing (-42.1 dB peak aliasing at ~=0 CA 02203l~6 1997-04-l8 WO96tl3097 PCTtUS95/13916 rad/s), and almost constant group delay of 32 samples.
Errors were concentrated near ~=0 and ~=~ and did not severely limit the useful bandwidth of the system.
Aliasing was cancelled below -68 dB over 93~ of the spectrum. The signal-to-noise ratio of an ideal n-bit analog-to-digital converter is 6.02n + 1.76 dB, so the system is capable of analog-to-digital conversion at approximately 11 bits of resolution because the aliasing errors would be at the noise floor of the system.
A system with seventh-order analysis filters optimized by the first stage 92 and length 64, 16-bit coefficient FIR digital synthesis filters optimized by the second stage 94 was simulated. The optimized system demonstrated 0.00095 dB average deviation from 0 dB distortion (0.0045 dB peak distortion), -108 dB
average aliasing (-42.3 dB peak aliasing at ~=0 rad/s), and an almost constant group delay of 34 samples.
Aliasing was cancelled below -74 dB over 95~ of the spectrum, so the optimized hybrid filter bank is capable of analog-to-digital conversion at approximately 12 bits of resolution. Results from the filter design algorithm indicate that the mean-squared reconstruction error for the Butterworth hybrid filter bank is 46 times larger than that of the optimized filter bank.
Results from a simulation with wideband input reveal approximately 30 dB out-of-band attenuation.
This illustrates how the linearity errors and analog-to-digital converter mismatches of each of the analog-to-digital converters in the array are suppressed by about 30 dB between the channels and are therefore not compounded as in time-interleaved analog-to-digital conversion. For example, a 50 MSa/s time-interleaved system using two 12-bit 25 MSa/s analog-to-digital CA 022031~6 1997-04-18 W O 96/13097 PC~rAUS95/13916 converters was evaluated. Because of mismatches between the analog-to-digital converters, the system's spurious-free dynamic range was only 56.7 dB, or 9.1 bits. The corresponding two-channel 50 MSa/s hybrid filter bank analog-to-digital converter does not compound the mismatches, so the resolution was limited by the worst analog-to-digital converter in the array, which in this case had a spurious-free dynamic range of 82.2 dB, or 13.4 bits.
It is expected that a fully-developed hybrid filter bank analog-to-digital converter would exceed 8 GSa/s sampling rate at 8 bits resolution, and would exceed 2 MSa/s sampling rate at 16 bits resolution.
The filter design algorithm optimizes the analog analysis filters and digital synthesis filters to cancel aliasing and compensate for ripple, non-linear phase, and imperfect implementations of the analog analysis filters to form a near-perfect-reconstruction hybrid filter bank. The hybrid filter bank isolates the analog-to-digital converters in the array so that analog-to-digital errors are not compounded between the channels, thereby improving speed and resolution over conventional array architectures. Although the experiment focused on an analog-to-digital system, the optimization algorithm is equally applicable to the digital-to-analog case.
Further details of an implementation of the invention can be found in the Massachusetts Institute of Technology M.S. Thesis of Scott Richard Velazquez, 1994, which is incorporated herein by reference.
While this invention has been particularly shown and describe with references to preferred embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the scope of the invention as defined by the appended claims.
A~AENDED
Backqround of the Invention The rapid expansion of digital signal processing hardware and software has dramatically increased the demand for high-speed and high-resolution analog-to-digital converters. The most popular techniques foranalog-to-digital conversion to date include successive approximation, sigma-delta, sub-ranging, flash, and time-interleaving.
The advancement of analog-to-digital converter technology involves a trade-off between sampling rate in samples per second (Sa/s) and resolution in bits.
Sigma-delta converters currently provide high resolution (20 bits) at relatively low sampling rates (10 kSa/s) while time-interleaved converters provide low resolution (8 bits) at high sampling rates (8 GSa/s). The remaining techniques including successive approximation (100 kSa/s at 16 bits resolution), subranging (1 MSa/s at 14 bits resolution), and flash (100 MSa/s at 10 bits resolution), span the range of sampling rate, resolution, and cost between the time-interleaving and sigma-delta techniques.
Recently, manufacturers began studying and perfecting time-interleaved analog-to-digital conversion. In this technique, a bank of time-multiplexed analog-to-digital converters with low sampling rates are combined to realize a single high-speed, high-resolution analog-to-digital converter.
The primary drawback of time-interleaved analog-to-digital conversion is that linearity errors and mismatches between each of the converters in the bank are compounded across the full bandwidth of the device, CA 022031~6 1997-04-18 limiting the resolution of the system. Time-interleaving is also prone to timing errors because it requires very accurate interleaved clock signals which limits the speed and resolution of the system and which introduces harmonic distortion. In addition, a voltage offset between two converters in the bank may cause the converters to digitize the same input voltage to different codes. Since the converters output data sequentially, the output exhibits an error signal having a period equivalent to the sample period of a single converter, but the frequency content can increase because each cycle of the error contains a number of offset errors equivalent to the number of converters in the bank. Therefore, this source of error introduces harmonic distortion spurs that limit the effective resolution of the system. A difference in voltage gain between two converters also causes each converter to digitize the same input voltage to a different code. The magnitude of the error grows as the input voltage grows. Overcoming these fundamental limitations is difficult because it requires reduction of error for each converter. The gain and offset of an individual converter can be adjusted with external resistors and the remaining harmonic distortion can be reduced with dynamic compensation techniques. However, the remaining linearity errors of each converter are still compounded across the full bandwidth of the system.
Discrete-time Quaudrature Mirror Filter (QMF) banks have been employed in analog-to-digital conversion techniques as described in:
CA 022031~6 1997-04-18 W O 96/13097 PC~rrUS95/13916 Antonio Petraglia and Sanjit K.
Mitra, "High-Speed A/D Conversion Incorporating a QMF Bank," IEEE
Transac ti ons on Ins trumen ta ti on and Measurement, 41 (3) :42 7-43 1 (June, 1992 ) .
In this system, discrete-time switched-capacitor analysis filters decompose the wideband input signal into several contiguous frequency subbands. An individual analog-to-digital converter is assigned to each subband. All converters are driven by a common clock. The subbands are allocated certain quantization bits thereby allowing separate control of resolution in each subband. Discrete-time finite impulse response 15 synthesis filters reconstruct each subband, cancelling the aliasing caused by other subbands. Therefore, the errors associated with a particular subband are not compounded with errors of other subbands as in time-interleaved analog-to-digital conversion. In addition, the discrete-time QMF bank approach does not require extremely accurate time-skewed clock signals.
The primary disadvantage of the discrete-time QMF
bank analog-to-digital converter is that the switched-capacitors introduce switching noise that can limit the signal-to-noise ratio and thereby limit the resolution and speed of the system. Typical switched-capacitor filters have a signal-to-noise ratio of about 85 dB and are limited to sampling rates of about 150 kSa/sec.
This limits the bandwidth of the discrete-time QMF
bank.
Summary of the Invention The present invention is directed to an apparatus and method for high-speed high-resolution analog/digital conversion. Note that the term "analog/digital" as used herein includes both analog-CA 022031~6 1997-04-18 W O 96/13097 PCTrUS95/13916 to-digital and digital-to-analog conversion. In the present invention, converter quantization noise and linearity errors are confined to each individual subband and are therefore not compounded. The converter bank is driven by a single clock, simplifying timing constraints. Synthesis filters compensate for remalnlng tlmlng errors.
For subband coding applications such as compression or adaptive array processing, the present invention is capable of direct channelization without the requirement of additional hardware. The present lnvention is amenable to non-uniform channel bandwidths and can be implemented in very large scale integration (VLSI). The invention overcomes the limitations and disadvantages of the conventional discrete-time QMF
array architectures.
In accordance with the present invention, a converter bank is incorporated into a hybrid system having, in the case of an analog-to-digital converter, continuous-time analysis filters and discrete-time synthesis filters. Integration of continuous-time filters and discrete-time filters does present design problems, but they have been overcome with this invention. Continuous-time analysis filters improve system speed and resolution by overcoming the limitations of switched-capacitor filters used in the prior art. The continuous-time analysis filters partition a continuous-time wideband input signal into continuous-time subband signals. In a preferred embodiment, an analog-to-digital converter, a rate changer, and a discrete-time synthesis filter are provided for each subband. The individual analog-to-digital converters in the array quantize the subband signals into binary bits. The rate changer, in the form of an upsampler or downsampler increases or CA 022031~6 1997-04-18 WO 96/13097 PCI~/US95/13916 decreases the data rate of the quantized subband signal. The discrete-time synthesis filters process the subband signal, generating a signal which is the discrete-time approximation of the continuous-time - 5 subband signal.
An adder may be provided for recombining the discrete-time subband signals into a discrete-time wideband signal which is the discrete-time approximation of the continuous-time wideband input signal. A discrete-time compensator may compensate for non-linear behavior in the analog-to-digital converters. The synthesis filters compensate for gain error, aliasing error and non-linear phase in the continuous-time analysis filters. Application-specific processors may be included for further processing of the subband signals either before or after conversion to discrete-time. Non-uniform subband channel bandwidths may be employed as well as uniform subband channel bandwidths. A tree-structure architecture may be implemented to simplify the multiple-channel parallel architecture into a hierarchy of multiple-channel systems. A wavelet-type structure may be implemented for systems requiring non-uniform channel bandwidths.
An alternative embodiment for digital-to-analog conversion employs discrete-time analysis filters to partition a wideband discrete-time input signal into subbands. Each subband is assigned its own digital-to-analog converter and subband processor. Continuous-time synthesis filters process the subband signal. The subband signals may be recombined to form a wideband, continuous-time output signal.
In a preferred embodiment of the invention, the hybrid system comprises an M-channel, maximally-decimated, parallel, uniform bandwidth architecture, CA 022031~6 1997-04-18 W O 96/13097 PC~rtUS9Stl3916 where M represents an integer. In a maximally-decimated embodiment, each subband converter samples at 1/M the effective sampling rate of the system. The wideband input signal is partitioned into M subband signals by M analysis filters. For the digital-to-analog conversion case, each subband signal is then downsampled by a factor of M. Each subband signal is then passed through a converter, either analog-to-digital or digital-to-analog and processed. For the analog-to-digital case, the subband signal is then upsampled by a factor of M. Next, the subband signal is processed in a bank of M synthesis filters and recombined to form a wideband output signal.
A more efficient implementation employs polyphase decomposition to interchange the order of analysis filtering and downsampling in the digital-to-analog case, or upsampling and synthesis filtering in the analog-to-digital case. This allows for filtering at the subband data rate. To reduce the effects of aliasing in the maximally-decimated system and to improve resolution, oversampling may be used. A subset of the analog/digital converters in the bank may operate at higher resolution than the remaining analog/digital converters for applications requiring greater dynamic range at select subbands. The subbands may be mixed down to baseband or up to higher frequencies to alleviate the requirement for high-speed, high-precision sample and hold circuitry in the analog/digital converters.
An optimization algorithm determines the analysis filters and the synthesis filters in an iterative process which incrementally adjusts the analysis and synthesis filters such that the system distortion function converges on a perfect delay and such that the system aliasing functions converge on zero.
CA 022031~6 1997-04-18 WO96/13097 PCT~S9S/13916 Brief Description of the Drawings FIG. 1 is a block diagram of a general hybrid filter bank for analog-to-digital or digital-to-analog conversion in accordance with the present invention.
FIG. 2 is a block diagram of a hybrid filter bank analog-to-digital converter having subband processor banks and a discrete-time compensation bank for compensating for non-ideal analog behavior in accordance with the present invention.
FIG. 3 is a block diagram of an M-channel, maximally-decimated, parallel, uniform channel width, hybrid filter bank analog-to-digital converter in accordance with the present invention.
FIGs. 4A and 4B are block diagrams of equivalent methods for filtering. FIG. 4A is a representation of filtering following upsampling. FIG. 4B shows an equivalent structure using polyphase decomposition to enable efficient filtering at the lower data rate prior to upsampling. In the analog-to-digital case, this can be used to enable efficient synthesis filtering at the lower data rate.
FIGS. 4C and 4D are block diagrams of equivalent methods for filtering. FIG. 4C is a representation of filtering followed by downsampling. FIG. 4D shows an equivalent structure using polyphase decomposition to enable efficient filtering at the lower data rate after downsampling. In the digital-to-analog case, this can be used to enable efficient analysis filtering at the lower data rate.
FIG. 5A is a block diagram of a tree-structured hybrid filter bank analog-to-digital converter. FIG.
5B is a block diagram of a wavelet-type hybrid filter bank analog-to-digital converter.
CA 02203l~6 1997-04-l8 WO96/13097 PCT~S9S/13916 FIG. 6 is a block diagram of a hybrid filter bank digital-to-analog converter in accordance with the present invention.
FIG. 7 is a flow diagram for an optimization algorithm for calculating analysis filters Hk (5), and then for calculating discrete time synthesis filters Fk (Z) based on the analysis filters for an analog-to-digital converter in accordance with the present lnvent lon .
Detailed Description of the Preferred Embodiments The hybrid filter bank of the present invention is a novel class of filter bank which is less understood than the all-discrete-time class. The complexity of the hybrid filter bank lies in the relationship between the continuous-time analysis filters and the discrete-time synthesis filters.
FIG. 1 is a block diagram of a general hybrid filter bank analog/digital converter 11. An analysis filter bank 12 partitions a wideband input signal 10 into subband signals 12A. Each subband signal 12A is assigned its own analog/digital converter and subband processor 14 where the subband signal 12A undergoes conversion and processing. A synthesis filter bank 16 further processes the converted and processed subband slgnals 14A. The synthesized subband signals 16A are combined at an optional adder 18 to form a wideband output signal 20. For analog-to-digital conversion, continuous-time analysis fiIters 12 and discrete-time synthesis filters 16 are employed and the synthesis bank 16 may include data rate changers to increase or decrease the data rate. For digital-to-analog conversion, discrete-time analysis filters 12 and continuous-time synthesis filters 16 are employed and the analysis bank 12 may include data rate changers for CA 02203l~6 1997-04-l8 WO96/13097 PCT~S95/13916 increasing or decreasing the data rate. The data rate changers are either upsamplers, which increase the data rate by inserting zeros between the samples, or downsamplers, which decrease the sampling rate by periodically discarding samples.
The embodiment of FIG. 1 provides for subband channelization of the wideband input signal 10 at signals 16A or 12A without the requirement of additional hardware. This is particularly useful in applications such as adaptive array processing and compression techniques that perform subband coding.
A hybrid filter bank analog-to-digital converter with discrete-time compensation is shown in the block diagram of FIG. 2. Rather than filtering the continuous-time wideband input signal 22 with discrete-time switched-capacitors before analysis filtering in discrete-time as does the prior art, the embodiment of FIG. 2 overcomes the limitations of the prior art by using continuous-time analysis filters 24.
A continuous-time wideband input signal 22 is partitioned into continuous-time subband signals 24A by the continuous-time analysis filter bank 24. A bank of analog-to-digital converters 26, one converter for each subband, converts the continuous-time subband signals 24A into discrete-time subband signals 26A. An optional discrete-time compensator bank 28 may be employed for compensating for non-linear behavior in the analog-to-digital converters 26. The compensated discrete-time subband signals 28A may be further processed by an optional discrete-time subband processor bank 30 for frequency-dependent processing of individual subband data streams, for example, adaptive-array processing and compression. Similarly, the subband processors 25 may optionally be employed for processing the continuous-time subband signals 24A
CA 022031~6 1997-04-18 W O 96/13097 PC~rrUS95/13916 before quantization by the analog-to-digital converters 26. The resultant signals 30A are further processed by the discrete-time synthesis filter bank 32 which may include data rate changers for increasing or decreasing the data rate. The synthesis filters 32 compensate for non-linear phase and gain errors in the analysis filters 24 and also cancel aliasing.
The output 32A of the discrete-time synthesis filter bank 32 is the discrete-time approximation of the continuous-time subband inputs 24A. The synthesis filter bank 32 processes the signals so that upon recombination at an optional adder 34, aliasing is cancelled and the subband signals 32A are combined into a wideband discrete-time output signal 40 which is the discrete-time approximation of the wideband continuous-time input signal 22.
The subbands 24A may be mixed down to baseband with mixers 27 before sampling at the analog-to-digital converters 26 to alleviate the requirement for high-speed, high-precision sample-and-hold circuitry in the analog-to-digital converters 26. This allows for analog-to-digital conversion at a lower sampling rate, and increases the dynamic range of the system by alleviating the wide analog bandwidth constraints of the converter bank. An analogous structure may be used to mix the subband signals up to higher frequencies after conversion for the digital-to-analog case.
FIG. 3 is a block diagram representing an M-channel, maximally-decimated, parallel, uniform channel, hybrid filter bank analog-to-digital converter. Continuous-time analysis filters Hk (5) partition a wideband continuous-time input u(t) into M
consecutive, uniform bandwidth subband signals xk(t).
Each continuous-time subband signal xk(t) is converted to discrete-time by an analog-to-digital converter 42 CA 022031~6 1997-04-18 W O 96/13097 PC~rAUS95/13916 sampling at a subband rate of l/M multiplied by the effective sample rate of the system. Each discrete-time converted subband signal xk(n)is upsampled 44 by a factor of M. Each upsampled discrete-time subband signal vk(n) is processed by a discrete-time synthesis filter Fk (Z)- The resulting signals Yk(n) are optionally combined at an adder 46 to form a wideband discrete-time output signal y(n).
A mathematical derivation of the properties of the hybrid filter bank analog-to-digital converter shown in FIG. 3 will now be described. The first goal is to determine the input/output function for this architecture which is derived below in equations 1 through 9. The second goal is to find expressions for the Fourier transform of the synthesis filters Fk (eiW) that provide perfect reconstruction given a set of analysis filters Hk (iQ) which are defined in equations 10 through 16, with an example given for M=2 at equations 17-22.
In the frequency domain, the result of filtering the input U(jQ), bandlimited to Qn rad/sec, through each analysis filter Hk (iQ) is Xk(jQ), Xk( jQ) = Ut jQ) Hk( jQ) (1) In the time-domain, analog-to-digital converter 42 errors can be modeled as gain errors ak and DC offset errors bk, Xk[n]=(l+ak)Xk[n] + bk (2) The effects of analog-to-digital converter errors are seen in the frequency domain along with aliasing error WO96/13097 PCT~S95/13916 The effects of analog-to-digital converter errors are seen in the frequency domain along with aliasing error due to undersampling, where T is the sampling period, T=M7r/n", xk(ei~)= T ~0 Xk( T T (3) kl _ o 2 r~ ~ ( ~ - 2 ~ 1 ), 0 < ~1) < 2 Substituting for X~(jQ), 1+a M-1 ~ j2 1 J ;Zn1 + bk 2 2~ (~-2~1), O<~(M-1) 2~ .
The result of upsampling by a factor of M is represented as V~.(~).
Vk( e j"') =Xk( e j~M) =--T k 2 U( j T ~ j T ) Hk( j T ~ j T
+bk 2 2~0 (~M-2r~1), 0<~ (M 1) 2J~.
The result of filtering Vk(ei~) through synthesis filters Fk(~J is represented as Yk(~J, Yk(e ) =Fk(ej ) Vk(ej~) Tk ~ Fk(ej~) U( j(T)M- j2Tl )Hk( jtM-~) (6) M-l +b 2 F (e~) 2~ (~M-2J~1) W O 96/13097 PC~r~US95/13916 M-l .. Y(ej~) = 2 Yk(ei~) . (7) Substituting for Yk(ei~), the reconstructed wideband output signal Y(eiW) is Y(ei'~) = 2 T k 2 Fk (ei'~) U( j T - j T
. (H ( j(~M j2 1~) (8) M-1 M-l + ~ bk ~ Fk(ej~) 27~ M-2 k=O 1=0 Rearranging the summation gives the output Y(ei~), in terms of the input U(jQ), the analysis filters Hk (ei~) and the synthesis filters Fk (ei~) .
y(ej~) = 2 U( jT - j T ) ¦~ 2 (1 ak) Fk(ej~.,)Hk( ~ M_ j2 1~¦ (9) M-1 M-l + 2 bk 2 Fk(ej(i') 2rc8 ((~M-2t~1) k=O 1=0 This expression is simplified by introducing functions Tp (ei~) and ~, Y(ej~) = 2 U( jTM- j2Tp) Tp(ei'~') + ~ , (10) where CA 022031~6 1997-04-18 W O 96/13097 PC~rAUS95/13916 T (ej~) = 2 k Fk(ej~)Hk( jT - i T ) ~ (11) O < ~ (MMl) and M-l M-l = 2~ 2 bk Fk(ej~ (~M-2~p) . (12) k=O p=O
Equations 10, 11, and 12 describe the mathematical input/output function for the hybrid filter bank analog-to-digital converter on the interval Oc~5 (M-1)2~
M
Perfect reconstruction occurs when the sole effect of the filter bank architecture is to introduce a delay, Y(ej~) = Ae~j~dU(ej~) . (13) Therefore, the distortion function To (ei~) should correspond to a perfect delay, d, To(e j~) = Ae -j~d . (14) Also, the aliasing functions, Tp (ei~) where lspsM-l, should be equal to zero, CA 022031~6 1997-04-18 WO96/13097 pcT~sssll39l6 Tp(ej~) = 0 . (15) ~ corresponds to DC offset errors in the analog-to-digital converters and should be zeroed by adjusting each analog-to-digital converter offset bk to zero.
~ = o . (16) Using these constraints, equation 11 can be solved as a system of simultaneous linear equations for the Fourier transform of the desired synthesis filters Fk(ej~) that provide perfect reconstruction given a set of analysis filters Hk (eiW) .
As an example, for M=2, the input/output function 10 is Y(ej'l') =U ( j2 ) Tc (ei'l')+ U ( j2w_ j2 )Tl(ei~')+~
For perfect reconstruction, the distortion function should be equal to a perfect delay, TO~ei )=--T (l+aO)Fo(ei )Ho( T ) (18) + T(l+al)Fl(ei)Hl( T )=e i , Oc~s~, and the aliasing function should be equal to zero, WO96/13097 PCT~S95/13916 and the aliasing function should be equal to zero, Tl(ej~)= T(l+aO)FO(e )HJ~ T T ) (19) -T(1+a~)FI(ej )Hl( jT - jT )=' <~<~ .
Also, the DC offset errors, ~, should be set to zero ,/3=2~, b,~ F, (e~ (~) +~ -7r) ) = (20) 1:=0 Solving the system of simultaneous equations above for F0(ei~J and F~(ej~J, the Fourier transform of the synthesis filters that provide perfect reconstruction gives Fo ( e i~ ) = A e ~ j~ d T
( T - jT ) (21) ~ Ho( T )H1( T - jT ) -Hl ( i T ) Ho( j2 _ j2 and Discrete-time synthesis filters, Fk(Z), are determined by calculating a system function that approximates the desired frequency responses, F~
using the optimization described below. The order of CA 022031~6 1997-04-18 W O 96/13097 PC~rrUS95/13916 using the optimization described below. The order of the synthesis filter is determined by the level of error that can be tolerated in the system. High-order synthesis filters reduce reconstruction errors because they more closely approximate the desired frequency responses than do low-order synthesis filters.
The reverse structure of the hybrid filter bank analog-to-digital converter is the hybrid filter bank digital-to-analog converter which uses discrete-time analysis filters, downsamplers, an array of digital-to-analog converts and continuous-time synthesis filters.
It can be shown that using the continuous-time and discrete-time filters designed for the hybrid filter bank analog-to-digital converter will result in the same reconstruction error as the hybrid filter bank digital-to-analog converter. The digital-to-analog converter is described below in conjunction with FIG.
6.
The optimization algorithm described below in conjunction with FIG. 7 calculates the optimal system delay d and formulates suitable analysis and synthesis filters which enable accurate reconstruction. A near-perfect reconstruction solution can be designed by approximating a discrete-time perfect reconstruction solution, but good approximations require high-order filters. The optimization algorithm described below calculates near-perfect-reconstruction hybrid filter banks that are not necessarily hardware intensive.
The algorithm has two sections, the analysis filter optimization, followed by the synthesis filter optimization. From a given lowpass continuous-time analysis filter, Ho(sJl the analysis filter optimization calculates the remaining bandpass analysis filters Hk(s), lsksM-l, and the system delay, d, to minimize reconstruction error. The optimization CA 022031~6 1997-04-18 calculates samples of the functions Tp (ej~), OspsM-l, and iteratively adjusts the analysis filters Hk (s), and system delay, d, such that the distortion function, T0(ej~) is as close to a perfect delay, e~j~d, as possible and the aliasing functions Tp(ej~), for lsp~M-1 are as close to zero as possible. The algorithm uses a least-squares error criterion with the Matlab multi-variable minimization algorithm fmins.
Given analysis filters Hk (s) and system delay d, calculated above, the synthesis filter optimization calculates the synthesis filters Fk (Z) that minimize the reconstruction error. One method is the Matlab function invfreqz which calculates system functions Fk(z) whose Fourier transforms approximate the ideal Fourier transforms. Such an algorithm is quick enough to provide a rough solution, but may not provide the highest accuracy reconstruction possible. This algorithm is used in the analysis filter optimization above to evaluate the aliasing functions Tp(e~
Similar to the analysis filter optimization, the synthesis filter optimization iteratively adjusts the synthesis filters, Fk(z), to minimize a least-squares error criterion in Tp(ej~) with the Matlab multi-variable minimization algorithm fmins. The best results are obtained when the algorithm uses the analysis filters calculated in the analysis filter optimization above, but it also works for any set of analysis filters, enabling the use of standard filter types (such as Butterworth, Chebyshev, or Elliptic filters). The synthesis filter optimization also enables accurate calibration of the hardware system by optimizing the synthesis filters based on measurement of the hardware analysis filters.
The technique of FIG. 3 requires synthesis filters Fk (Z) operating at the effective sample rate of the CA 022031 j6 1997-04-18 WO 96/13097 PCT/US9!i/13916 system, which can be extremely fast. The upsampler 48 and synthesis filter Fk (ZJ combination from FIG. 3 is reproduced in FIG. 4A. FIG. 4B shows a polyphase decomposition technique which is the equivalent of the combination shown in FIG. 4A. Polyphase decomposition rearranges the order so that synthesis filtering Rk (Z) is performed before upsampling. Synthesis filtering Rk (Z) is performed at the lower sample rate of the analog-to-digital converters, which is 1/M multiplied by the effective sample rate of the system. Polyphase decomposition replaces each upsampler 44 and length L
FIR synthesis filter Fk with M length L/M filters Rk (Z) followed by M upsamplers 50. The process requires M
filters for each subband, but each filter has fewer coefficients. A similar polyphase decomposition structure can be used to enable analysis filtering at the subband sample rate in the digital-to-analog case.
In an oversampling system where the analog-to-digital converters are sampling at a rate higher than the effective sampling rate of the hybrid filter bank, downsampling may be needed after filtering and conversion as shown in FIG. 4C. FIG. 4D shows an embodiment for polyphase decomposition wherein the converted signal Xk(~) is first downsampled 51 to lower the data rate, and then filtered by synthesis filters Rk tZ)- In this embodiment, the filtering Rk (Z) is performed at the lower data rate. In general, the polyphase filtering methods in FIGS. 4B and 4D may be used to enable filtering at a lower data rate by interchanging the order of data rate changing and filtering.
One technique for reducing aliasing is to increase analysis filter cut-off slope and stopband attenuation.
However, this technique increases hardware complexity.
CA 022031~6 1997-04-18 Another improved technique is oversampling. The maximally-decimated architecture shown in FIG. 3 is, by definition, operating at the theoretical limit Nyquist sampling rate with no leeway for aliasing.
oversampling would widen the frequency channels allocated to each converter in the array and thus the channels would overlap. This places less emphasis on the analysis filters Hk (s) to reduce aliasing, and relieves the sharp cut-off slope and high stopband attenuation constraints, thereby allowing lower order analysis filter Hk (s) implementations with reduced hardware requirements. Oversampling also separates quantization noise from the signal and the noise is filtered out by the synthesis filters, reducing system noise and improving system resolution. Each factor of four in sampling speed improves resolution by 1 effective bit. In the analog-to-digital converter case, where oversampling is used, it may be necessary to downsample the converted subband signal after synthesis filtering because, in the oversampled system, the individual converters may be operating at a rate higher than the effective sampling rate of the entire system.
The M-channel hybrid filter bank can be simplified to form a series of multiple-channel blocks arranged in a tree structure. This simplifies the optimization algorithm and design procedure. FIG. 5A is a block diagram of a two-channel tree structure for analog-to-digital conversion. The analysis filter bank 52 comprises hierarchical pairs of continuous-time analysis filters Ho(s) and H1(s) which partition the input signal u(t) into subbands, the subbands being further partitioned by continuous-time analysis filters Hoo(S)~ Ho1(S)l H1o(S)I H1l(s). Each subband signal is converted to discrete-time 53, upsampled 54, and then CA 022031~6 1997-04-18 processed and recombined in a hierarchical discrete time synthesis filter bank 55. Rather than optimizing an M-channel system, which can be prohibitively time-consuming, the optimization is reduced to a series of two-channel optimizations. This embodiment is not limited to two-channel blocks, nor is it limited to two levels of hierarchy. This technique may also be applied to digital-to-analog conversion.
The invention may also be implemented in a wavelet-type structure as shown in FIG. 5B. The continuous-time input signal u(t) is partitioned into subband signals by a continuous-time analysis filter Ho(s)~ One of the subbands u1(t) is further partitioned into smaller subbands u2(t) by an analysis filter H1(s) and each subband signal is converted to discrete-time 56 and upsampled 57. The subbands are recombined in a wavelet-structured synthesis filter bank 58. This structure is applicable in systems requiring non-uniform channel bandwidths for multi-resolution or wavelet analysis. This technique isappropriate for example, in audio applications, which require a wider bandwidth at lower frequencies.
The invention may also include higher resolution analog-to-digital converters in a subset of the subbands for applications requiring additional dynamic range at select frequency bands. For example, certain subbands may employ a converter with a higher number of quantization bits. Alternatively, certain subbands may be sampled at higher rates.
FIG. 6 is a block diagram of a M-channel maximally-decimated, parallel, uniform channel, hybrid filter bank digital-to-analog converter in accordance with the present invention. A wideband discrete-time input signal u(n) is partitioned into discrete-time subband signals uk(n) by discrete-time analysis filters CA 022031~6 1997-04-18 W O 96/13097 PC~rrUS95/13916 Hk (Z) . The subband signals uk(n) are downsampled 60 and the resulting downsampled signals xk(n) are converted from discrete-time to continuous-time by a bank of digital-to-analog converters 62. The resulting continuous-time subband signals xk(t) are further processed by continuous-time synthesis filters Fk (s) which generate signals Yk (t) which are the continuous-time approximation of the discrete-time subband signals uk(n). The continuous-time subband signals Yk (t) may be combined at an adder 64 to form a wideband continuous-time output signal y(t) which is the continuous-time approximation of the discrete-time wideband input signal u(n). Note that the digital-to-analog conversion technique shown in FIG. 6 may be implemented with any of the embodiments described herein, including, but not limited to polyphase decomposition, tree-structure, wavelet structure, subband processing, subband compensation and oversampllng.
FIG. 7 is a flow diagram of an algorithm for optimizing the analysis filters Hk(s) and synthesis filters Fk(z). The algorithm is broken down into two stages 92, 94. In the first stage 92, analysis filters Hk(s) are calculated and synthesis filters Fk(z) are approximated to minimize reconstruction error. In the second stage 94, the best synthesis filters are calculated to minimize reconstruction error. In each stage, the resulting input/output function for the system is evaluated, assuming that the DC offset errors ~ in the analog-to-digital converters have been zeroed, and then an error function is calculated. The error function quantifies the difference between the resulting input/output function and an ideal delay.
This data is used to make another guess at the filters CA 022031~6 1997-04-18 to reduce reconstruction error. This process repeats until the error is below a prescribed threshold in each stage.
In the first stage 92, a guess is made with a 5 standard multi-variable minimization routine, such as the Matlab routine fmins, for the analysis filters Hk(s) and the system delay d, 80. The synthesis filters Fk (Z) are then roughly approximated by fitting the frequency response curve to an optimal frequency 10 response 82 using standard filter approximation routines such as the Matlab routine invfreqz. The errors in Tk(eiW) are then evaluated 84 and a determination is made as to whether the error in Tk(eiW) is below a threshold le~el 86. If the error is 15 not below threshold 86, then another iteration of the first stage 92 is required and the process continues to adjust the analysis filters Hk(s) with a standard minimization algorithm until the error is below threshold. If the error is below threshold, then the 20 algorithm proceeds to the second stage 94.
In the second stage 94, given analysis filters Hk(s), synthesis filters Fk(z) are calculated which minimize reconstruction error. A measurement of the response of the analysis filters Hk(s) is made 87.
25 Next, the errors in Tk(eiW) are evaluated 88. A
determination is then made as to whether the error in Tk(eiW) is below a prescribed threshold 90. If the error is not below threshold, then the synthesis filter Fk(z) is adjusted with a standard minimization 30 algorithm to minimize the error 96 and another iteration of the second stage 94 is required. If the error is below threshold 90, then the analysis filters Hk(s) and synthesis filters Fk(z) are determined.
In practice, the frequency response of the 35 analysis filters Hk(s) must be measured 87 to determine CA 022031~6 1997-04-18 WO96/13097 PCT~S95/13916 their actual transfer functions, since the analysis filters calculated in the first stage 92 of the optimization can never be ideally implemented in hardware. The error function in the optimization algorithm requires accurate analysis filter amplitude and phase data at g equally-spaced frequency points.
One approach involves directly measuring the frequency response of the analysis filters at g equally-spaced frequencies. Accurate measurements are required or reconstruction is degraded.
The order of the synthesis filters will have a direct impact on how accurately the ideal solutions can be approximated. For FIR synthesis filters, once the length L has been decided, the initial guess of the optimal delay d of the system can be set to L/2. The rationale for this choice is that the magnitude of the impulse response coefficients is greatest in the vicinity of the dth coefficient and the magnitude tapers off on either side. Choosing d to be one-half of the length of the impulse response assures that the most significant coefficients are not truncated by the finite length of the impulse response.
The analog-to-digital converter gain errors ak and offset errors bk must also be measured 87. By measuring the amplitude of the harmonics that the analog-to-digital converter produces when sampling a pure sinusoid, an analog-to-digital converter non-linearity curve can be deduced, where the zero order term corresponds to offset errors bk which should be nulled to zero, and the first order term corresponds to gain errors ak.
The error function Tk(ei~) to be minimized must be calculated. As stated above, perfect reconstruction occurs when the only effect of the filter bank architecture is to introduce a delay. Thus, the CA 022031~6 1997-04-18 W O 96/13097 PC~rrUS95/13916 distortion function T0(ej~) should be as close to perfect delay Ae~iWd as possible, and the aliasing functions Tp(ej~), lspsM-1, should be as close to zero as possible. o(~) is a measurement at a specific frequency ~1~ of the squared difference between the distortion function T0(ej~) and an ideal delay Ae~j~d (both real and imaginary parts), Eo((I)) = [~{To (e j~) } --~R{Ae j~d} ] 2 (23) + [~{To (ej~) } -~{Ae~i~d} ] 2 p (w) is a measurement at a specific frequency ~ of the squared difference between the aliasing function Tp (ej~) and zero (both real and imaginary parts), Ep(~) = ~{TP(e j~) }2 + ~{TP(e j~) }2, l~p~M-1 . (24) Therefore, if the system is a perfect delay at frequency ~, both o(~) and p(~) will be zero. The system should be as close to a perfect delay as possible at all frequencies, so total error function should sample o(~) and p(~) at g equally-spaced frequencies to assure that optimization occurs across the entire frequency band. The total error function that the algorithm minimizes is = ~ EPEP ( ) . ( 25) where EP are weighting factors that can be used to favor reduction of either the aliasing error or the distortion error.
The number of frequency points g to be examined in the error function must be determined. If L is the CA 022031~6 1997-04-18 W O 96/13097 PCTrUS95/13916 length of each synthesis filter, then the minimization program will be operating on a transfer function with ML variables, the L coefficients of the M finite impulse response synthesis filters. If g is chosen as ML, then the input/output function will be relatively smooth between the frequency response measurement points. In other words, the algorithm is reasonably assured of optimizing the response over the entire frequency band, not just at the g frequencies that the algorithm uses. For example, in an M=2 system, if the length of the synthesis filters is chosen to be L=64, then evaluating the error function at g=128 equally-spaced frequencies is the minimum recommended.
The total error function is a function of analysis filters Hkfs), gain errors ak, and synthesis filters Fk(z). Hk(s) and ak are first measured, and an initial guess at Fk(z) is made. The minimization of in both the first stage 92 and the second stage 94 is done numerically using a routine such as the MatlabTM
function fmins, which uses a Nelder-Mead simplex search to iteratively adjust filters Fk(z), Hk(z), or system delay d to minimize the error . The algorithm is a direct search that does not require derivative information. Let C be the number of variables the algorithm is manipulating to minimize the error. A
simplex in C-dimensional space is characterized by the C+1 distinct vectors which are its vertices. In two-space, a simplex is a triangle; in three-space, it is a pyramid. At each step of the search, a new point in or near the current simplex is generated. Error evaluated at the new point 84, 88 is compared with the error value at the vertices of the simplex, and usually, one of the vertices is replaced by a new point, giving a new simplex 80, 96. This step is CA 022031~6 1997-04-18 W O 96/13097 PC~rAUS95/13916 repeated until the diameter of the simplex is less than the specified tolerance 86, 90.
An implementation of the Hybrid Filter Bank analog-to-digital converter preferably includes an efficient automatic calibration routine that can be performed on a regular basis. One method of calibration includes injection of a known wideband analog signal into the continuous-time analysis filters, followed by the optimization of the synthesis filters to provide the best mathematical reconstruction of the original continuous-time signal. Measurement and modeling errors will be bypassed using this method because the process is directly optimizing the system based on the performance of the analysis filters.
Although the process is computationally intensive, once the initial calibration has been done, re-calibration could be accomplished with a reasonable amount of computation on a regular basis. The computational requirements of the M-channel case can be reduced by 2 0 simplifying the architecture to form a series of recursive two channel blocks arranged in a tree structure as shown in FIG. 5A.
The analysis filter type affects the accuracy of the synthesis filter reconstruction. Reconstruction accuracy is also a function of the order, ripple, and group delay of the analysis filters. In addition to analysis filters designed by the first stage 92 of the optimization algorithm, experiments have been conducted on several standard types of analysis filters including Elliptic, Chebyshev I, and Butterworth. The experiments demonstrated that reconstruction accuracy is nearly independent of analysis filter order.
However, low order analysis filters with greater stopband attenuation generated greater out-of-band attenuation of analog-to-digital converter linearity CA 022031~6 1997-04-18 WO 96/13097 PCI'/US95/13916 errors and quantization errors. These errors are compounded across the subband channels similar to the compounding of errors in time-interleaved analog-to-digital converters, thereby increasing overall distortion. Note that this analog-to-digital converter error should not be confused with the mathematical reconstruction error of the hybrid bank structure discussed above. Filters with passband ripple and non-constant group delay, such as Chebyshev and Elliptic filters, have lower reconstruction accuracy. Better reconstruction was achieved with Butterworth filters since they have no passband ripple and since group delay is relatively constant. The best analysis filters were designed with the first stage 92 of the optimization algorithm.
The synthesis filters reconstruct the signal by suppressing upsampled images, cancelling aliasing and compensating for the analysis filter ripple and nonlinear phase. The order of the synthesis filter determines how accurately the objectives are met.
Increasing the analysis filter cut-off slope and stopband attenuation decreases the amount of aliasing.
This alleviates the requirements imposed on synthesis filters to cancel aliasing, thereby improving reconstruction accuracy. This also increases out-of-band attenuation, which suppresses linearity errors and quantization noise between the frequency bands.
A two channel hybrid filter bank system employing seventh order analog Butterworth analysis filters and length 64 finite impulse response 16-bit digital synthesis filters optimized by the second stage 94 of the optimization algorithm was simulated. The system exhibited 0.113 dB average deviation from 0 dB
distortion (3.13 dB peak distortion at w=~ rad/s), -81.5 dB average aliasing (-42.1 dB peak aliasing at ~=0 CA 02203l~6 1997-04-l8 WO96tl3097 PCTtUS95/13916 rad/s), and almost constant group delay of 32 samples.
Errors were concentrated near ~=0 and ~=~ and did not severely limit the useful bandwidth of the system.
Aliasing was cancelled below -68 dB over 93~ of the spectrum. The signal-to-noise ratio of an ideal n-bit analog-to-digital converter is 6.02n + 1.76 dB, so the system is capable of analog-to-digital conversion at approximately 11 bits of resolution because the aliasing errors would be at the noise floor of the system.
A system with seventh-order analysis filters optimized by the first stage 92 and length 64, 16-bit coefficient FIR digital synthesis filters optimized by the second stage 94 was simulated. The optimized system demonstrated 0.00095 dB average deviation from 0 dB distortion (0.0045 dB peak distortion), -108 dB
average aliasing (-42.3 dB peak aliasing at ~=0 rad/s), and an almost constant group delay of 34 samples.
Aliasing was cancelled below -74 dB over 95~ of the spectrum, so the optimized hybrid filter bank is capable of analog-to-digital conversion at approximately 12 bits of resolution. Results from the filter design algorithm indicate that the mean-squared reconstruction error for the Butterworth hybrid filter bank is 46 times larger than that of the optimized filter bank.
Results from a simulation with wideband input reveal approximately 30 dB out-of-band attenuation.
This illustrates how the linearity errors and analog-to-digital converter mismatches of each of the analog-to-digital converters in the array are suppressed by about 30 dB between the channels and are therefore not compounded as in time-interleaved analog-to-digital conversion. For example, a 50 MSa/s time-interleaved system using two 12-bit 25 MSa/s analog-to-digital CA 022031~6 1997-04-18 W O 96/13097 PC~rAUS95/13916 converters was evaluated. Because of mismatches between the analog-to-digital converters, the system's spurious-free dynamic range was only 56.7 dB, or 9.1 bits. The corresponding two-channel 50 MSa/s hybrid filter bank analog-to-digital converter does not compound the mismatches, so the resolution was limited by the worst analog-to-digital converter in the array, which in this case had a spurious-free dynamic range of 82.2 dB, or 13.4 bits.
It is expected that a fully-developed hybrid filter bank analog-to-digital converter would exceed 8 GSa/s sampling rate at 8 bits resolution, and would exceed 2 MSa/s sampling rate at 16 bits resolution.
The filter design algorithm optimizes the analog analysis filters and digital synthesis filters to cancel aliasing and compensate for ripple, non-linear phase, and imperfect implementations of the analog analysis filters to form a near-perfect-reconstruction hybrid filter bank. The hybrid filter bank isolates the analog-to-digital converters in the array so that analog-to-digital errors are not compounded between the channels, thereby improving speed and resolution over conventional array architectures. Although the experiment focused on an analog-to-digital system, the optimization algorithm is equally applicable to the digital-to-analog case.
Further details of an implementation of the invention can be found in the Massachusetts Institute of Technology M.S. Thesis of Scott Richard Velazquez, 1994, which is incorporated herein by reference.
While this invention has been particularly shown and describe with references to preferred embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the scope of the invention as defined by the appended claims.
A~AENDED
Claims (17)
1. A system for converting a signal between continuous-time and discrete-time comprising:
an analysis filter bank (12) for partitioning a wideband input signal (10) into first subband signals (12A);
a converter bank (14) of analog/digital converters for converting the first subband signals (12A) to second subband signals (14A); and a synthesis filter bank (16) for reconstructing the second subband signals (14A) such that third subband signals (16A) are generated which are the converted approximation of the first subband signals (12A); characterized in that:
one of the analysis bank (12) and synthesis bank (16) comprises continuous-time analog filters and the other comprises discrete-time digital filters.
an analysis filter bank (12) for partitioning a wideband input signal (10) into first subband signals (12A);
a converter bank (14) of analog/digital converters for converting the first subband signals (12A) to second subband signals (14A); and a synthesis filter bank (16) for reconstructing the second subband signals (14A) such that third subband signals (16A) are generated which are the converted approximation of the first subband signals (12A); characterized in that:
one of the analysis bank (12) and synthesis bank (16) comprises continuous-time analog filters and the other comprises discrete-time digital filters.
2. The system of Claim 1 for converting a continuous-time signal to a discrete-time signal comprising:
a bank of continuous-time analog analysis filters (24) for partitioning a continuous-time wideband input signal (22) into first subband signals (24A);
a bank of analog-to-digital converters (26) for quantizing the first subband signals (24A) into second subband signals (26A);
a bank of rate changers (44) for adjusting the data rate of the second subband signals (26A);
and a bank of discrete-time digital synthesis filters (32) for reconstructing the second subband signals (26A) such that third subband signals (32A) are generated which are the discrete-time approximation of the continuous-time first subband signals (24A).
a bank of continuous-time analog analysis filters (24) for partitioning a continuous-time wideband input signal (22) into first subband signals (24A);
a bank of analog-to-digital converters (26) for quantizing the first subband signals (24A) into second subband signals (26A);
a bank of rate changers (44) for adjusting the data rate of the second subband signals (26A);
and a bank of discrete-time digital synthesis filters (32) for reconstructing the second subband signals (26A) such that third subband signals (32A) are generated which are the discrete-time approximation of the continuous-time first subband signals (24A).
3. The system of Claim 2 further comprising a bank of discrete-time compensators (28) for compensating for non-ideal analog behavior in the analog-to-digital converters (26).
4. The system of any one of preceding Claims 2 or 3 wherein subband processing (25) is performed before or after or both before and after the analog-to-digital converters (26) quantize the first subband signals (24A).
5. The system of any preceding claim further comprising an adder (18,34,64) for recombining the third subband signals (16A,32A) into a wideband output signal (20,40) which is the converted approximation of the wideband input signal (10,22).
6. The system of any preceding claim wherein the analysis filters (12,24) and the synthesis filters (16,32) are determined by an optimization technique which iteratively adjusts the analysis and synthesis filters such that the system distortion function converges on a perfect delay and such that the system aliasing functions converge on zero.
7. The system of any preceding claim wherein polyphase decomposition is employed such that filtering is accomplished at the lowest data rate.
8. The system of any preceding claim wherein the bank of analog/digital converters (14,26) includes a subset of analog/digital converters operating at a higher resolution than the remaining analog/digital converters.
9. The system of any preceding claim wherein the bandwidth of the subbands is non-uniform.
10. The system of any preceding claim wherein the first subband signals (12A,24A) are oversampled by the analog/digital converters for increasing resolution.
11. The system of any one of preceding Claims 1 and 5-10 wherein the analysis filter bank (12) comprises discrete-time digital filters, the converter bank (14) comprises digital-to-analog converters, and the synthesis filter bank (16) comprises continuous-time analog filters.
12. The system of any preceding claim further comprising a mixing circuit (27) for mixing the first subband signals down to baseband or up to higher frequencies.
13. The system of any preceding claim further comprising a bank of application specific processors (30) for further processing of the subband signals.
14. A method for converting a signal between continuous-time and discrete-time comprising:
partitioning a wideband input signal (10) into first subband signals (12A) at an analysis filter bank (12);
converting the first subband signals (12A) to.
second subband signals (14A) at a bank of analog/digital converters (14); and reconstructing the second subband signals (14A) at a synthesis filter bank (16) such that third subband signals (16A) are generated which are the converted approximation of the first subband signals (12A); characterized in that:
one of the analysis bank (12) and synthesis bank (14) comprises continuous-time analog filters and the other comprises discrete-time digital filters.
partitioning a wideband input signal (10) into first subband signals (12A) at an analysis filter bank (12);
converting the first subband signals (12A) to.
second subband signals (14A) at a bank of analog/digital converters (14); and reconstructing the second subband signals (14A) at a synthesis filter bank (16) such that third subband signals (16A) are generated which are the converted approximation of the first subband signals (12A); characterized in that:
one of the analysis bank (12) and synthesis bank (14) comprises continuous-time analog filters and the other comprises discrete-time digital filters.
15. The method of Claim 14 for converting a continuous-time signal to a discrete-time signal comprising the steps of:
partitioning a continuous-time wideband input signal (22) into first subband signals (24A) at a bank of continuous-time analog analysis filters (24);
quantizing the first subband signals (24A) into second subband signals (26A) at a bank of analog-to-digital converters (26);
adjusting the data rate of the second subband signals (26A) at a bank of rate changers (44); and reconstructing the second subband signals (26A) at a bank of discrete-time digital synthesis filters (32) for generating third subband signals (32A) which are the discrete-time approximation of the continuous-time first subband signals (24A).
partitioning a continuous-time wideband input signal (22) into first subband signals (24A) at a bank of continuous-time analog analysis filters (24);
quantizing the first subband signals (24A) into second subband signals (26A) at a bank of analog-to-digital converters (26);
adjusting the data rate of the second subband signals (26A) at a bank of rate changers (44); and reconstructing the second subband signals (26A) at a bank of discrete-time digital synthesis filters (32) for generating third subband signals (32A) which are the discrete-time approximation of the continuous-time first subband signals (24A).
16. The method of any one of preceding Claims 14-15 further comprising the step of recombining (18,34,64) the third subband signals (16A,32A) into a wideband output signal (20,40) which is the converted approximation of the wideband input signal (10,22).
17. The method of any one of preceding Claims 14-16 further comprising the step of optimizing the analysis filters and the synthesis filters by iteratively adjusting the analysis and synthesis filters such that the system distortion function converges on a perfect delay and such that the system aliasing functions converge on zero.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| US08/326,474 US5568142A (en) | 1994-10-20 | 1994-10-20 | Hybrid filter bank analog/digital converter |
| US08/326,474 | 1994-10-20 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CA2203156A1 true CA2203156A1 (en) | 1996-05-02 |
Family
ID=23272365
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CA002203156A Abandoned CA2203156A1 (en) | 1994-10-20 | 1995-10-19 | Hybrid filter bank analog/digital converter |
Country Status (5)
| Country | Link |
|---|---|
| US (1) | US5568142A (en) |
| EP (1) | EP0787385A1 (en) |
| JP (1) | JPH10507891A (en) |
| CA (1) | CA2203156A1 (en) |
| WO (1) | WO1996013097A1 (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8583390B2 (en) | 2002-10-24 | 2013-11-12 | Teledyne Lecroy, Inc. | High bandwidth oscilloscope for digitizing an analog signal having a bandwidth greater than the bandwidth of digitizing components of the oscilloscope |
| US10659071B2 (en) | 2002-10-24 | 2020-05-19 | Teledyne Lecroy, Inc. | High bandwidth oscilloscope |
Families Citing this family (104)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO1997011527A1 (en) * | 1995-09-22 | 1997-03-27 | Philips Electronics N.V. | Transmission system using time dependent filter banks |
| US5608665A (en) * | 1995-10-03 | 1997-03-04 | Wyszynski; Adam S. | Self-tuned, continuous-time active filter |
| JP3327114B2 (en) * | 1996-04-24 | 2002-09-24 | ソニー株式会社 | Signal processing device, signal recording device, and signal reproducing device |
| US5771299A (en) * | 1996-06-20 | 1998-06-23 | Audiologic, Inc. | Spectral transposition of a digital audio signal |
| US6252535B1 (en) | 1997-08-21 | 2001-06-26 | Data Fusion Corporation | Method and apparatus for acquiring wide-band pseudorandom noise encoded waveforms |
| EP0897222A4 (en) * | 1997-02-10 | 2007-07-25 | Sony Corp | Method and equipment for processing data |
| US6075820A (en) * | 1997-05-28 | 2000-06-13 | Lucent Technologies Inc. | Sampling receiver with multi-branch sigma-delta modulators and digital channel mismatch correction |
| US6018753A (en) * | 1997-07-29 | 2000-01-25 | Lucent Technologies Inc. | Interpolating filter banks in arbitrary dimensions |
| SE513044C2 (en) * | 1997-12-29 | 2000-06-26 | Ericsson Telefon Ab L M | Analog-digital converter with global clock and global switch |
| US5973470A (en) * | 1998-03-25 | 1999-10-26 | Mcdonnell Douglas Corp. | Compensation for generic servoamplifier usage with high performance direct drive valves |
| WO1999050678A2 (en) * | 1998-03-31 | 1999-10-07 | Siemens Aktiengesellschaft | Method and configuration for processing at least one analog signal comprising several frequency ranges |
| JP3484980B2 (en) * | 1998-06-23 | 2004-01-06 | 日本電気株式会社 | Wireless receiver |
| US6182031B1 (en) * | 1998-09-15 | 2001-01-30 | Intel Corp. | Scalable audio coding system |
| US6177893B1 (en) | 1998-09-15 | 2001-01-23 | Scott R. Velazquez | Parallel processing analog and digital converter |
| CA2256779A1 (en) * | 1998-12-21 | 2000-06-21 | Tet Hin Yeap | High speed analog-to-digital converter and digital-to-analog converter |
| US6757343B1 (en) * | 1999-06-16 | 2004-06-29 | University Of Southern California | Discrete wavelet transform system architecture design using filterbank factorization |
| AUPQ381499A0 (en) * | 1999-11-02 | 1999-11-25 | Commonwealth Of Australia, The | Enhanced direct digitising array arrangement |
| US6947509B1 (en) | 1999-11-30 | 2005-09-20 | Verance Corporation | Oversampled filter bank for subband processing |
| WO2001041320A1 (en) * | 1999-11-30 | 2001-06-07 | Verance Corporation | Oversampled filter bank for subband processing |
| DE10008699C1 (en) | 2000-02-24 | 2001-05-23 | Daimler Chrysler Ag | Analogue-digital signal conversion method has input signal transformed via linear function with resulting coefficients used for retransformation into digital region via second linear function |
| US7933341B2 (en) * | 2000-02-28 | 2011-04-26 | Broadcom Corporation | System and method for high speed communications using digital signal processing |
| USRE41831E1 (en) | 2000-05-23 | 2010-10-19 | Marvell International Ltd. | Class B driver |
| US7194037B1 (en) | 2000-05-23 | 2007-03-20 | Marvell International Ltd. | Active replica transformer hybrid |
| US7312739B1 (en) | 2000-05-23 | 2007-12-25 | Marvell International Ltd. | Communication driver |
| US6775529B1 (en) | 2000-07-31 | 2004-08-10 | Marvell International Ltd. | Active resistive summer for a transformer hybrid |
| US7433665B1 (en) | 2000-07-31 | 2008-10-07 | Marvell International Ltd. | Apparatus and method for converting single-ended signals to a differential signal, and transceiver employing same |
| JP4454109B2 (en) * | 2000-06-14 | 2010-04-21 | 日本テキサス・インスツルメンツ株式会社 | Method and apparatus for improving signal-to-noise ratio in digital-to-analog conversion processing of a pulse density modulation signal (PDM) |
| DE10028593C1 (en) * | 2000-06-14 | 2001-10-18 | Daimler Chrysler Ag | Digital/analogue signal conversion method uses transformation with orthogonal functions and determination of coefficients for re-conversion into analogue range |
| US7606547B1 (en) | 2000-07-31 | 2009-10-20 | Marvell International Ltd. | Active resistance summer for a transformer hybrid |
| US6339390B1 (en) * | 2000-10-04 | 2002-01-15 | Scott R. Velazquez | Adaptive parallel processing analog and digital converter |
| US6996198B2 (en) * | 2000-10-27 | 2006-02-07 | At&T Corp. | Nonuniform oversampled filter banks for audio signal processing |
| EP1360761A4 (en) * | 2001-01-05 | 2006-05-10 | Philip Druck | N dimensional, non-linear, static, adaptive, digital filter design using d scale non-uniform sampling |
| US6701297B2 (en) | 2001-03-02 | 2004-03-02 | Geoffrey Layton Main | Direct intermediate frequency sampling wavelet-based analog-to-digital and digital-to-analog converter |
| SE0101175D0 (en) * | 2001-04-02 | 2001-04-02 | Coding Technologies Sweden Ab | Aliasing reduction using complex-exponential-modulated filter banks |
| US6473013B1 (en) * | 2001-06-20 | 2002-10-29 | Scott R. Velazquez | Parallel processing analog and digital converter |
| US7206459B2 (en) * | 2001-07-31 | 2007-04-17 | Ricoh Co., Ltd. | Enhancement of compressed images |
| US20030058148A1 (en) * | 2001-09-21 | 2003-03-27 | Sheen Timothy W. | Multiple a-to-d converter scheme employing digital crossover filter |
| JP4222816B2 (en) * | 2001-12-06 | 2009-02-12 | 本田技研工業株式会社 | Plant control system using frequency shaping response assignment control |
| US6567030B1 (en) * | 2002-02-27 | 2003-05-20 | Lecroy Corporation | Sample synthesis for matching digitizers in interleaved systems |
| CN1663257A (en) * | 2002-04-19 | 2005-08-31 | 德国普莱特科技公司 | Wavelet transform system, method and computer program product |
| JP3956847B2 (en) * | 2002-04-24 | 2007-08-08 | 株式会社デンソー | A / D conversion method and apparatus |
| WO2004039021A1 (en) * | 2002-10-25 | 2004-05-06 | The Trustees Of Columbia University In The City Of New York | Time encoding and decoding of a signal |
| US7271756B2 (en) * | 2002-12-04 | 2007-09-18 | Nxp B.V. | Non-linear distribution of voltage steps in flash-type A/D converters |
| JP3851870B2 (en) * | 2002-12-27 | 2006-11-29 | 株式会社東芝 | Variable resolution A / D converter |
| US7324036B2 (en) * | 2003-05-12 | 2008-01-29 | Hrl Laboratories, Llc | Adaptive, intelligent transform-based analog to information converter method and system |
| WO2004112298A2 (en) * | 2003-05-27 | 2004-12-23 | The Trustees Of Columbia University In The City Of New York | Multichannel time encoding and decoding of a signal |
| US20050018796A1 (en) * | 2003-07-07 | 2005-01-27 | Sande Ravindra Kumar | Method of combining an analysis filter bank following a synthesis filter bank and structure therefor |
| DE10343346B4 (en) * | 2003-09-12 | 2011-01-27 | Qimonda Ag | Method for testing an electrical circuit and device for carrying out the method |
| JP2005217837A (en) * | 2004-01-30 | 2005-08-11 | Sony Corp | Sampling rate conversion apparatus and method thereof, and audio apparatus |
| US7715928B1 (en) | 2004-05-17 | 2010-05-11 | University Of Southern California | Efficient communications and data compression algorithms in sensor networks using discrete wavelet transform |
| US7280059B1 (en) * | 2004-05-20 | 2007-10-09 | The Trustees Of Columbia University In The City Of New York | Systems and methods for mixing domains in signal processing |
| JP3947185B2 (en) * | 2004-06-01 | 2007-07-18 | 株式会社アドバンテスト | Analog-digital conversion method and analog-digital conversion apparatus |
| US7253761B1 (en) * | 2004-11-08 | 2007-08-07 | United States Of America As Represented By The Secretary Of The Army | Analog to digital conversion with signal expansion |
| EP1844550B1 (en) * | 2005-02-04 | 2010-12-08 | Signal Processing Devices Sweden AB | Estimation of timing errors in a time-interleaved analog to digital converter system |
| US7138933B2 (en) * | 2005-04-26 | 2006-11-21 | Analog Devices, Inc. | Time-interleaved signal converter systems with reduced timing skews |
| US7119724B1 (en) * | 2005-05-25 | 2006-10-10 | Advantest Corporation | Analog digital converter and program therefor |
| KR100749074B1 (en) * | 2005-05-25 | 2007-08-13 | (주)에이스딕시오 | AD Converting Device |
| US7312662B1 (en) | 2005-08-09 | 2007-12-25 | Marvell International Ltd. | Cascode gain boosting system and method for a transmitter |
| US7577892B1 (en) | 2005-08-25 | 2009-08-18 | Marvell International Ltd | High speed iterative decoder |
| KR101184323B1 (en) * | 2005-11-03 | 2012-09-19 | 삼성전자주식회사 | Analog to digital conversion method and apparatus of receiver supporting software defined multi-standard radios |
| US7196650B1 (en) * | 2006-01-27 | 2007-03-27 | Analog Devices, Inc. | Signal converter systems and methods with enhanced signal-to-noise ratios |
| US7304597B1 (en) * | 2006-05-26 | 2007-12-04 | Lecroy Corporation | Adaptive interpolation for use in reducing signal spurs |
| ES2438176T3 (en) * | 2006-07-04 | 2014-01-16 | Electronics And Telecommunications Research Institute | Method to restore a multi-channel audio signal using a HE-AAC decoder and a MPEG surround decoder |
| US7346137B1 (en) * | 2006-09-22 | 2008-03-18 | At&T Corp. | Nonuniform oversampled filter banks for audio signal processing |
| US7782235B1 (en) | 2007-04-30 | 2010-08-24 | V Corp Technologies, Inc. | Adaptive mismatch compensators and methods for mismatch compensation |
| US20080279311A1 (en) * | 2007-05-07 | 2008-11-13 | Roger Lee Jungerman | AD Converter Bandwidth Enhancement Using An IQ Demodulator And Low Frequency Cross-Over Network |
| WO2008151137A2 (en) * | 2007-06-01 | 2008-12-11 | The Trustees Of Columbia University In The City Of New York | Real-time time encoding and decoding machines |
| WO2009006405A1 (en) * | 2007-06-28 | 2009-01-08 | The Trustees Of Columbia University In The City Of New York | Multi-input multi-output time encoding and decoding machines |
| US8238500B2 (en) | 2007-10-02 | 2012-08-07 | Seagate Technology Llc | Channel architecture with multiple signal processing branches for a given physical channel |
| JP4470996B2 (en) * | 2007-12-25 | 2010-06-02 | セイコーエプソン株式会社 | A / D conversion circuit and electronic device |
| US8145068B2 (en) * | 2008-03-17 | 2012-03-27 | Arris Group, Inc. | Dual use of test point power monitor |
| TW200943732A (en) * | 2008-04-15 | 2009-10-16 | Novatek Microelectronics Corp | Time interleaved analog to digital convert apparatus |
| US7649485B1 (en) | 2008-05-28 | 2010-01-19 | Hrl Laboratories, Llc | Multi-rate analog to digital converter with proportional filter bank |
| US7847719B2 (en) * | 2008-06-11 | 2010-12-07 | The Board Of Trustees Of The University Of Illinois | Signal processors, signal processing methods, and digital filter configuration methods |
| JP5215477B2 (en) * | 2008-12-16 | 2013-06-19 | シグナル・プロセシング・デバイシーズ・スウェーデン・エービー | Method and apparatus for evaluating and compensating for non-linearity errors |
| CN102414988B (en) * | 2009-04-29 | 2014-03-19 | 香港大学 | Methods or structures for reconstruction of substantially uniform samples from substantially nonuniform samples |
| US9419637B1 (en) | 2010-01-05 | 2016-08-16 | Syntropy Systems, Llc | Multi-mode sampling/quantization converters |
| US8089382B2 (en) * | 2009-06-26 | 2012-01-03 | Syntropy Systems, Llc | Sampling/quantization converters |
| US8416111B2 (en) * | 2010-01-05 | 2013-04-09 | Sytropy Systems, LLC | Multimode sampling/quantization converters |
| US9130584B1 (en) | 2010-01-05 | 2015-09-08 | Syntropy Systems, Llc | Multi-mode sampling/quantization converters |
| US9209829B2 (en) | 2009-06-26 | 2015-12-08 | Syntropy Systems, Llc | Sampling/quantization converters |
| US8917198B2 (en) | 2010-01-05 | 2014-12-23 | Syntropy Systems, Llc | Multi-mode sampling/quantization converters |
| US9680498B2 (en) | 2009-06-26 | 2017-06-13 | Syntropy Systems, Llc | Sampling/quantization converters |
| US8943112B2 (en) | 2009-06-26 | 2015-01-27 | Syntropy Systems, Llc | Sampling/quantization converters |
| US8299947B2 (en) * | 2009-06-26 | 2012-10-30 | Syntropy Systems, Llc | Sampling/quantization converters |
| US9654128B2 (en) | 2010-01-05 | 2017-05-16 | Syntropy Systems, Llc | Multi-mode sampling/quantization converters |
| US8855011B2 (en) | 2010-07-12 | 2014-10-07 | University Of Southern California | Distributed transforms for efficient data gathering in sensor networks |
| WO2012109407A1 (en) | 2011-02-09 | 2012-08-16 | The Trustees Of Columbia University In The City Of New York | Encoding and decoding machine with recurrent neural networks |
| US9306590B2 (en) | 2011-05-26 | 2016-04-05 | Tektronix, Inc. | Test and measurement instrument including asynchronous time-interleaved digitizer using harmonic mixing |
| US9568503B2 (en) | 2011-05-26 | 2017-02-14 | Tektronix, Inc. | Calibration for test and measurement instrument including asynchronous time-interleaved digitizer using harmonic mixing |
| US9432042B2 (en) | 2011-05-26 | 2016-08-30 | Tektronix, Inc. | Test and measurement instrument including asynchronous time-interleaved digitizer using harmonic mixing |
| US8742749B2 (en) | 2011-05-26 | 2014-06-03 | Tektronix, Inc. | Test and measurement instrument including asynchronous time-interleaved digitizer using harmonic mixing |
| US8849602B2 (en) | 2011-10-27 | 2014-09-30 | Agilent Technologies, Inc. | Calibrating reconstructed signal using multi-tone calibration signal |
| US20140163940A1 (en) * | 2012-12-11 | 2014-06-12 | David E. Erisman | Method and system for modeling rf emissions occurring in a radio frequency band |
| WO2014125600A1 (en) * | 2013-02-14 | 2014-08-21 | 三菱電機株式会社 | Demultiplexing device, multiplexing device, and relay device |
| US8928514B1 (en) | 2013-09-13 | 2015-01-06 | Tektronix, Inc. | Harmonic time domain interleave to extend oscilloscope bandwidth and sample rate |
| US9250313B2 (en) * | 2013-12-04 | 2016-02-02 | Raytheon Company | Electronically reconfigurable bandwidth and channel number analog-to-digital converter circuit for radar systems |
| US20150372846A1 (en) * | 2014-06-20 | 2015-12-24 | Ikanos Communications, Inc. | Dual band analog front end for high speed data transmissions in dmt systems |
| US9621175B2 (en) | 2015-02-11 | 2017-04-11 | Syntropy Systems, Llc | Sampling/quantization converters |
| US9231608B1 (en) * | 2015-03-19 | 2016-01-05 | Teledyne Lecroy, Inc. | Method and apparatus for correction of time interleaved ADCs |
| US9935604B2 (en) * | 2015-07-06 | 2018-04-03 | Xilinx, Inc. | Variable bandwidth filtering |
| DE102016014795A1 (en) * | 2016-12-12 | 2018-06-14 | Rosenberger Hochfrequenztechnik Gmbh & Co. Kg | Method and apparatus for transmitting or receiving at least one high frequency signal with parallel and subsampled baseband signal processing |
| US10135475B1 (en) * | 2017-10-11 | 2018-11-20 | The Boeing Company | Dynamic low-latency processing circuits using interleaving |
| US10601434B1 (en) * | 2019-03-29 | 2020-03-24 | Intel Corporation | Apparatus for calibrating a time-interleaved analog-to-digital converter |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5191305A (en) * | 1991-07-02 | 1993-03-02 | Interstate Electronics Corporation | Multiple bandpass filter |
| US5412690A (en) * | 1993-03-08 | 1995-05-02 | Motorola, Inc. | Method and apparatus for receiving electromagnetic radiation within a frequency band |
| US5392044A (en) * | 1993-03-08 | 1995-02-21 | Motorola, Inc. | Method and apparatus for digitizing a wide frequency bandwidth signal |
| US5465072A (en) * | 1994-08-05 | 1995-11-07 | Linear Technology Corporation | Tunable operational transcondunctance amplifier having high linearity |
-
1994
- 1994-10-20 US US08/326,474 patent/US5568142A/en not_active Expired - Fee Related
-
1995
- 1995-10-19 WO PCT/US1995/013916 patent/WO1996013097A1/en not_active Application Discontinuation
- 1995-10-19 JP JP8514143A patent/JPH10507891A/en not_active Ceased
- 1995-10-19 EP EP95938940A patent/EP0787385A1/en not_active Withdrawn
- 1995-10-19 CA CA002203156A patent/CA2203156A1/en not_active Abandoned
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8583390B2 (en) | 2002-10-24 | 2013-11-12 | Teledyne Lecroy, Inc. | High bandwidth oscilloscope for digitizing an analog signal having a bandwidth greater than the bandwidth of digitizing components of the oscilloscope |
| US9325342B2 (en) | 2002-10-24 | 2016-04-26 | Teledyne Lecroy, Inc. | High bandwidth oscilloscope |
| US9660661B2 (en) | 2002-10-24 | 2017-05-23 | Teledyne Lecroy, Inc. | High bandwidth oscilloscope |
| US10135456B2 (en) | 2002-10-24 | 2018-11-20 | Teledyne Lecroy, Inc. | High bandwidth oscilloscope |
| US10333540B2 (en) | 2002-10-24 | 2019-06-25 | Teledyne Lecroy, Inc. | High bandwidth oscilloscope |
| US10659071B2 (en) | 2002-10-24 | 2020-05-19 | Teledyne Lecroy, Inc. | High bandwidth oscilloscope |
Also Published As
| Publication number | Publication date |
|---|---|
| JPH10507891A (en) | 1998-07-28 |
| US5568142A (en) | 1996-10-22 |
| EP0787385A1 (en) | 1997-08-06 |
| WO1996013097A1 (en) | 1996-05-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US5568142A (en) | Hybrid filter bank analog/digital converter | |
| Velazquez et al. | A hybrid filter bank approach to analog-to-digital conversion | |
| US6473013B1 (en) | Parallel processing analog and digital converter | |
| US6177893B1 (en) | Parallel processing analog and digital converter | |
| US6339390B1 (en) | Adaptive parallel processing analog and digital converter | |
| US9680497B2 (en) | Conversion of a discrete-time quantized signal into a continuous-time, continuously variable signal | |
| US9048865B2 (en) | Conversion of a discrete time quantized signal into a continuous time, continuously variable signal | |
| Kester | Mixed-signal and DSP design techniques | |
| US5392042A (en) | Sigma-delta analog-to-digital converter with filtration having controlled pole-zero locations, and apparatus therefor | |
| US4890106A (en) | Apparatus and methods for digital-to-analog conversion using modified LSB switching | |
| EP0454406B1 (en) | Multi-stage sigma-delta analog-to-digital converter | |
| EP2514098B1 (en) | Conversion of a discrete-time quantized signal into a continuous-time, continuously variable signal | |
| Tsai et al. | Bandwidth mismatch and its correction in time-interleaved analog-to-digital converters | |
| US5838272A (en) | Error correcting sigma-delta modulation decoding | |
| US6771198B2 (en) | Method and apparatus for converting between analog and digital domains using frequency interleaving | |
| AU4019899A (en) | Flexible and programmable delta-sigma analog signal converter | |
| CA2689287A1 (en) | Compensation of mismatch errors in a time-interleaved analog-to-digital converter | |
| JPS61177819A (en) | Oversampling type digital/analog converter | |
| Pinheiro et al. | Improving the near-perfect hybrid filter bank performance in the presence of realization errors | |
| Bunton | An improved FX correlator | |
| Petraglia et al. | QMF-based A/D converters: Overview and new results | |
| Petrescu et al. | Sensitivity of hybrid filter banks A/D converters to analog realization errors and finite word length | |
| Löwenborg et al. | A survey of filter bank A/D converters | |
| Khakpour et al. | Adaptive noise cancellation based architecture for correction of gain and offset mismatch errors in time-interleaved ADC | |
| USRE34660E (en) | Apparatus and methods for digital-to-analog conversion using modified LSB switching |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| FZDE | Discontinued |