CN103076027B - Based on the resonant mode intelligent sensor lag error compensation method of geneva core FCM algorithm - Google Patents
Based on the resonant mode intelligent sensor lag error compensation method of geneva core FCM algorithm Download PDFInfo
- Publication number
- CN103076027B CN103076027B CN201210441964.0A CN201210441964A CN103076027B CN 103076027 B CN103076027 B CN 103076027B CN 201210441964 A CN201210441964 A CN 201210441964A CN 103076027 B CN103076027 B CN 103076027B
- Authority
- CN
- China
- Prior art keywords
- geneva
- network
- core
- matrix
- resonant mode
- 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.)
- Expired - Fee Related
Links
Abstract
The invention discloses a kind of resonant mode intelligent sensor lag error compensation method based on geneva core FCM algorithm, first data acquisition is carried out to the input and output of resonant mode intelligent sensor, then according to accuracy requirement setting neural network correlation parameter, ensure network calculations speed and precision, finally data are sent into network, completed the calculating of network node central value and geneva matrix by the fuzzy clustering method of mahalanobis distance, and complete weight computing, finally obtain network architecture parameters.The invention has the advantages that and use less neural network structure to realize high precision error compensation, overcome traditional algorithm precision low, the shortcoming that network structure is huge.
Description
Technical field
The present invention relates to the compensation method of a kind of intelligent sensor lag error, particularly a kind of resonant mode intelligent sensor lag error compensation technique based on geneva core FCM algorithm, can adaptive alignment error corrective network parameter.
Background technology
The features such as resonant mode intelligent sensor, as a kind of sensors, has precision high, stable performance, and antijamming capability is strong.But the nonlinearity erron caused due to material behavior etc. is the large obstacle affecting resonant mode intelligent sensor precision always.This one of them important component part is exactly lag error.Existing lag error disposal route generally adopts least square method: namely by the non-linear retardant curve of fitting of a polynomial, make following formula minimum
E
2=∑[F(x
i)-Yi
] 2
Complete the compensation to lag error.Although said method calculated amount is little, computing velocity is high, and due to the limitation of least square method algorithm, arithmetic accuracy is not high.Or depend on certain Hysteresis Model, as Preisach model or JA model, Bouce-Wen model etc. carries out computing, as shown in the formula the lag error computing formula being Preisach model
By above formula, (see Fig. 1) is described to retardant curve visible, wherein α is current hysteresis loop maximal value, β is current hysteresis loop minimum value, u (t) is current input, μ (α, β, u (t)), the weight function that v (α) is corresponding hysteresis loop
or 1, can find out that employing utilizes Hysteresis Model to compensate error by above-mentioned formula, if be required to meet system output accuracy, than introducing too much parameter, thus cause computing complicated, system storage scale is large, the features such as poor real, and the determination of parameter also requires great many of experiments.
For high precision resonant mode intelligent sensor, to sensor output accuracy and sensor real-time, all there is very high requirement, as can be seen from above-mentioned two kinds of algorithms, under the condition not increasing existing hardware system cost, existing algorithm is difficult to obtain good balance in arithmetic speed and algorithm real-time, meeting the requirement of precision and arithmetic speed during to seek common ground, only having and increasing hardware system processing speed, bringing the lifting of whole Intelligent Sensorsystem cost.
Summary of the invention
The object of the invention is to the shortcoming overcoming above-mentioned prior art, a kind of resonant mode intelligent sensor lag error compensation method based on geneva core FCM algorithm is provided.Have network size little, parameter is few, and computing velocity is fast, and precision is high, the advantage that extensibility is strong.This algorithm, to the method for each input quantity nonopiate expansion in higher dimensional space, is especially applicable to carrying out error compensation to intelligent sensor.
For achieving the above object, the present invention is by the following technical solutions:
Based on a resonant mode intelligent sensor lag error compensation method for geneva core FCM algorithm, (1) inputoutput data to resonant transducer gathers, and forms input and output training dataset; (2) initialization network: select network parameter according to data set quantity and compensation precision, compensation speed, comprise Network Central Node number, and to geneva matrix initialize; (3) input vector is projected to mahalanobis space feeding network and carry out fuzzy clustering until the cluster of adjacent twice is identical; (4) data by completing cluster carry out output layer weight computing, calculate output error, if do not meet accuracy requirement, then return step (2), geneva matrix is adjusted, repeat above step, stop accuracy requirement until meet, circulation stops, and finally carries out network and exports weight computing.
As the preferred embodiments of the present invention, the data centralization in step (1) at least comprises the interior maximum value of loading zone and minimal value;
As the preferred embodiments of the present invention, in step (3), cluster is carried out according to following steps:
3.1) data set in step (1) is sent in network;
3.2) according to following formula, the central point initial value C of geneva core is calculated to the mahalanobis distance r of data set x:
Wherein, M is geneva matrix, and T is matrix transpose symbol;
3.3) according to step 3.2) the distance r that obtains calculates degree of membership μ
ik:
Wherein, j is kernel function numbering, i.e. a jth kernel function, and K is kernel function number, and m is fuzzy clustering parameter;
3.4) according to degree of membership, geneva core center is adjusted:
Wherein, c
ifor the centerpoint value of geneva core, N is the data amount check that training data is concentrated;
3.5) judge whether cluster completes, if do not complete, repeat step 3.1) to 3.4), until cluster completes;
In step (4), described output error is the error J that network exports between F (x) and training objective y, and it is according to following formulae discovery:
If output error does not meet accuracy requirement, then calculate the adjustment amount of geneva matrix according to backpropagation, then adjust geneva matrix, wherein, the adjustment amount of geneva matrix is according to following formulae discovery:
σM
i=η*e*w
i*(-1)*2*M*(x-c
i)*(x-c
i)
T,
Wherein, η is learning rate, w
ifor the network before geneva adjustment of matrix exports weights;
Geneva matrix M after adjustment
jfor, M
j+1=M
j+ η * σ M
i, wherein, j is error back propagation number of times.
Core of the present invention achieves the clustering method based on geneva Kernel fuzzy clustering, high-precision nonlinear error compensation can be realized by less network size, thus reduce the storage size of systematic parameter, achieve the compensation effect reaching degree of precision with less calculation cost.Overcome least square method precision low, modelling storage size is large, the shortcoming of poor real.
Accompanying drawing explanation
Fig. 1 is lag error curve synoptic diagram
Fig. 2 is backoff algorithm network structure of the present invention;
Fig. 3 is the lag error curve map of not compensated in the present invention;
Fig. 4 is algorithm flow chart in the present invention;
Fig. 5 is algorithm and existing arithmetic accuracy comparison diagram in the present invention;
Fig. 6 calculates 12 data to export used time contrast under different IPs function number in the present invention.
Embodiment:
Technical scheme of the present invention is carried out in accordance with the following steps:
(1) the sensor output that first testing standard input is lower, obtains a series of inputoutput data collection;
(2) initialization network, according to data set scale and accuracy requirement, the basic parameter of setting network, comprises input vector dimension, hides into neuron number etc., and to geneva matrix initialize;
(3) send into network after input vector being projected to mahalanobis space and carry out fuzzy clustering;
(4) data by completing cluster carry out output layer weight computing, and calculate output error, if preset accuracy requirement, then return 2, adjust geneva matrix, repeat above step, until reach accuracy requirement.
Inputoutput data collection in described step (1) refers to:
Table 1 measurement data
Because this technology have employed the compensation method according to Preisach Hysteresis Model, this model is a kind of model based on phenomenon, and therefore our image data is as shown in table 1, and a point different interval gathers data inputoutput data.
In step (3), input vector is projected mahalanobis space and processes by this technology.Algorithm steps is as follows:
In step (4), the data projecting mahalanobis space are sent into kernel function and carries out distance calculating, complete fuzzy clustering.Main algorithm is as follows:
1) the distance r obtained in formula (1) is sent into each kernel function, calculate the output of core, this output is the degree of membership to each core;
2) according to degree of membership to kernel function center c
iadjust;
3) judge whether cluster completes, if do not complete, be back to 1) cluster again;
4), after cluster completes, to calculate output layer weights such as formula (4) by RLS algorithm and calculate output error as formula (5), judge whether to meet accuracy requirement, if do not meet, error back propagation, such as formula (6), adjusts geneva matrix, cluster again, until meet accuracy requirement;
ω=YΦ
-1(4)
FCM algorithm is a class intelligent algorithm, and by calculating the degree of membership of kernel function each input point, obtain input to output fuzzy clustering, iteration reaches optimum.
Below in conjunction with accompanying drawing, the present invention is described in further detail:
See Fig. 3, it is certain resonant transducer curve of output, obvious lag error as we can see from the figure, according to the requirement of the present invention to input, first image data is as table 1, and is x=[α by data preparation, β, u], be respectively the maximal value of current hysteresis loop, minimum value and current input value.
According to image data scale and accuracy requirement, determine hidden layer kernel function number K=45.Setting geneva matrix initial value is 3*3 unit matrix M
0=I
3.Data set is sent in network and calculate mahalanobis distance according to formula (1).Obtain mahalanobis distance R=[r
1, r
2r
ir
n], n is inputoutput data number.
The mahalanobis distance obtained is sent into fuzzy clustering core and calculates output respectively, obtain hidden layer output matrix
This matrix is substituted into formula (4), according to inputoutput pair
(d is the standard value in measuring process), calculates and exports weights, obtain weight vector ω=[ω
1, ω
2..., ω
45].Thus obtain network output y=ω * Φ.
Export y by computational grid and obtain error vector e=[e
1, e
2..., e
90].The cost function of computational grid is as the output error of network.
If result does not now meet preset accuracy requirement, then this function result is passed through formula (6) backpropagation, obtain the adjustment amount σ Mi of Metzler matrix
,m
j+1=M
j+ η * σ M
i, wherein j is error back propagation number of times, and η is learning rate.After each backpropagation of error, the Metzler matrix of each kernel function is adjusted, until the output error of network (formula (7)) meets the accuracy requirement preset.
After completing above-mentioned steps, obtain final hidden layer and export
by Φ
k,ncalculate final network output layer weights, complete network training.
As can be seen from computational accuracy comparison diagram 5 and calculating used time comparison diagram 6, in the present invention, technology is not when substantially increasing the calculating used time, can reach the precision being far more than existing algorithm.
Can obtain thus, the intelligent compensation of resonant transducer lag error can be realized based on geneva core FCM algorithm algorithm, under lower hardware requirement, can degree of precision be reached, and sensor real-time is had an impact hardly.And based on the training method of artificial intelligence also without the need to manual operation and calculating, setup parameter.
Above content is in conjunction with concrete preferred implementation further description made for the present invention; can not assert that the specific embodiment of the present invention is only limitted to this; for general technical staff of the technical field of the invention; without departing from the inventive concept of the premise; some simple deduction or replace can also be made, all should be considered as belonging to the present invention by submitted to claims determination scope of patent protection.
Claims (4)
1., based on a resonant mode intelligent sensor lag error compensation method for geneva core FCM algorithm, it is characterized in that: comprise the following steps:
(1) inputoutput data of resonant transducer is gathered, form input and output training dataset;
(2) initialization network: select network parameter according to data set quantity and compensation precision, compensation speed, comprise Network Central Node number, and to geneva matrix initialize;
(3) input vector is projected to mahalanobis space, send into network, carry out fuzzy clustering, until the cluster of adjacent twice is identical;
(4) data by completing cluster carry out output layer weight computing, calculate output error, if do not meet accuracy requirement, then return step (2), geneva matrix is adjusted, repeats above step, stop accuracy requirement until meet, circulation stops, and finally carries out network and exports weight computing; Described output error is the error J that network exports between F (x) and training objective y, and it is according to following formulae discovery:
2. as claimed in claim 1 based on the resonant mode intelligent sensor lag error compensation method of geneva core FCM algorithm, it is characterized in that, the data centralization in step (1) at least comprises the interior maximum value of loading zone and minimal value.
3. as claimed in claim 1 based on the resonant mode intelligent sensor lag error compensation method of geneva core FCM algorithm, it is characterized in that, in step (3), cluster is carried out according to following steps:
3.1) data set in step (1) is sent in network;
3.2) according to following formula, the central point initial value C of geneva core is calculated to the mahalanobis distance r of data set x:
Wherein, M is geneva matrix, and T is matrix transpose symbol;
3.3) according to step 3.2) the mahalanobis distance r that obtains calculates degree of membership μ
ik:
Wherein, j is kernel function numbering, i.e. a jth kernel function, and K is kernel function number, and m is fuzzy clustering parameter;
3.4) according to degree of membership, geneva core center is adjusted:
Wherein, c
ifor the centerpoint value of geneva core, N is the data amount check that training data is concentrated;
3.5) judge whether cluster completes, if do not complete, repeat step 3.1) to 3.4), until cluster completes.
4. according to claim 3 based on the resonant mode intelligent sensor lag error compensation method of geneva core FCM algorithm, it is characterized in that, if output error does not meet accuracy requirement, the adjustment amount of geneva matrix is then calculated according to backpropagation, then geneva matrix is adjusted, wherein, the adjustment amount of geneva matrix is according to following formulae discovery:
σM
i=η*e*w
i*(-1)*2*M*(x-c
i)*(x-c
i)
T,
Wherein, η is learning rate, w
ifor the network before geneva adjustment of matrix exports weights;
Geneva matrix M after adjustment
jfor, M
j+1=M
j+ η * σ M
i, wherein, j is error back propagation number of times.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201210441964.0A CN103076027B (en) | 2012-11-08 | 2012-11-08 | Based on the resonant mode intelligent sensor lag error compensation method of geneva core FCM algorithm |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201210441964.0A CN103076027B (en) | 2012-11-08 | 2012-11-08 | Based on the resonant mode intelligent sensor lag error compensation method of geneva core FCM algorithm |
Publications (2)
Publication Number | Publication Date |
---|---|
CN103076027A CN103076027A (en) | 2013-05-01 |
CN103076027B true CN103076027B (en) | 2015-08-05 |
Family
ID=48152638
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201210441964.0A Expired - Fee Related CN103076027B (en) | 2012-11-08 | 2012-11-08 | Based on the resonant mode intelligent sensor lag error compensation method of geneva core FCM algorithm |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN103076027B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109596040A (en) * | 2018-11-15 | 2019-04-09 | 天津大学 | Flexible sensor hysteresis compensation method based on BP neural network |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6067515A (en) * | 1997-10-27 | 2000-05-23 | Advanced Micro Devices, Inc. | Split matrix quantization with split vector quantization error compensation and selective enhanced processing for robust speech recognition |
CN101814160A (en) * | 2010-03-08 | 2010-08-25 | 清华大学 | RBF neural network modeling method based on feature clustering |
-
2012
- 2012-11-08 CN CN201210441964.0A patent/CN103076027B/en not_active Expired - Fee Related
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6067515A (en) * | 1997-10-27 | 2000-05-23 | Advanced Micro Devices, Inc. | Split matrix quantization with split vector quantization error compensation and selective enhanced processing for robust speech recognition |
CN101814160A (en) * | 2010-03-08 | 2010-08-25 | 清华大学 | RBF neural network modeling method based on feature clustering |
Non-Patent Citations (3)
Title |
---|
FCM与马氏空间约束条件下的快速图像分割技术研究;刘思远等;《计算机应用研究》;20070831;第220-222页 * |
基于核函数的模糊聚类算法研究;程可嘉;《中国优秀硕士学位论文全文数据库 信息科技辑》;20091215;第33页倒数第1段,第45页倒数第1段至第48页倒数第1段及图4-1 * |
基于模糊C均值与Markov随机场的图像分割;蔡涛等;《计算机工程》;20071031;第33卷(第20期);第34-36、39页 * |
Also Published As
Publication number | Publication date |
---|---|
CN103076027A (en) | 2013-05-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Wang et al. | Temperature error correction based on BP neural network in meteorological wireless sensor network | |
El-Keib et al. | Application of artificial neural networks in voltage stability assessment | |
CN105023043A (en) | AOD-based PM2.5 inversion model for Hangzhou region | |
CN102393881B (en) | A kind of high-precision detecting method of real-time many sensing temperatures data fusion | |
Niu et al. | Model turbine heat rate by fast learning network with tuning based on ameliorated krill herd algorithm | |
CN103106535A (en) | Method for solving collaborative filtering recommendation data sparsity based on neural network | |
CN105260786A (en) | Comprehensive optimization method of simulation credibility evaluation model of electric propulsion system | |
CN102663495B (en) | Neural net data generation method for nonlinear device modeling | |
CN106679880A (en) | Pressure sensor temperature compensating method based on FOA-optimized SOM-RBF | |
Kumar et al. | Parametric studies of a simple direct expansion solar assisted heat pump using ANN and GA | |
CN102645894B (en) | Fuzzy adaptive dynamic programming method | |
CN107966600A (en) | A kind of electricity anti-theft system and its electricity anti-theft method based on deep learning algorithm | |
CN102831476B (en) | Pattern detecting device and pattern detecting method for pulse neural network | |
CN103106305A (en) | Space grid structure model step-by-step correction method based on actual measurement mode | |
CN103714382A (en) | Multi-index comprehensive evaluation method for reliability of urban rail train security detection sensor network | |
CN103674778A (en) | Industrial melt index soft measuring meter and method based on RBF (radial basis function) particle swarm optimization | |
CN103559542A (en) | Extension neural network pattern recognition method based on priori knowledge | |
CN109916494A (en) | Weighing-appliance scaling method and device | |
CN103076027B (en) | Based on the resonant mode intelligent sensor lag error compensation method of geneva core FCM algorithm | |
CN103559541A (en) | Back propagation method for out-of-order data stream in big data | |
CN103279030A (en) | Bayesian framework-based dynamic soft measurement modeling method and device | |
CN104678989A (en) | State perception optimization method for improving fault diagnosability in noise environment | |
CN104899642A (en) | Gross distortion flexible body dynamic stress compensation method based on mixing nerve network model | |
Jafarkazemi et al. | Performance prediction of flat-plate solar collectors using MLP and ANFIS | |
Zhangang et al. | Genetic algorithm-based RBF neural network load forecasting model |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
C14 | Grant of patent or utility model | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20150805 Termination date: 20181108 |