DE102006014475A1 - Procedure for controlling a spinning preparation machine e.g. carding engine, drawing frame/rotor spinning machine, by determining input variables of a control device of the spinning machine so that parameter of the machine is optimized - Google Patents

Abstract

The procedure for controlling a spinning preparation machine e.g. a carding engine, drawing frame or a rotor spinning machine, comprises determining input variables (REP, RI) of a control device of the spinning machine so that parameter (CV%) of the textile goods produced by the spinning machine is optimized. The implied input variables are determined by a neuronal network in dependence of influencing variables (EG, VE, VV, VVD, HVD), which influence the effect of the implied input variables on implied parameter. The procedure for controlling a spinning preparation machine e.g. a carding engine, drawing frame or a rotor spinning machine, comprises determining input variables (REP, RI) of a control device of the spinning machine so that parameter (CV%) of the textile goods produced by the spinning machine is optimized. The implied input variables are determined by a neuronal network in dependence of influencing variables (EG, VE, VV, VVD, HVD), which influence the effect of the implied input variables on implied parameter. Each implied influencing variable of an assigned neuron is fed to an input layer of the neuronal network as input neuron variables. The implied influencing variable is standardized before it is brought to the assigned neurons of the input layer. Each implied input variables is delivered in standardized form by the assigned neurons to input layer of the neuronal network as output neuron variables. The implied neuronal network has a covered layer with covered neurons. In the neuronal network, each neuron of a layer is connected with each neuron of an adjacent layer and all connections are arranged between the neurons of a variable weight. The implied input variables are determined periodically or continuously through foreword propagation of the influencing variable in an operational phase of the spinning machine during varying the implied influencing variables. The implied input variables are automatically transmitted after its determination for control device of the spinning machine. Records are produced in an external learning phase. Each record consists of values of implied influencing variable and/or empirically determined values of the implied input variables, which lead in each value of the implied influencing variable to an optimal value of the parameter. The neuronal network is formed in an externally operating learning phase through reverse propagation. In the learning phase, the variable weight of the connection between the neurons is adjusted. The neuronal network is worked in the learning phase through supervised learning. In an externally operating test phase, the unit of the neuronal network determines and is verified the value of the implied input variables consulted for controlling the spinning machine. In the value of the implied input variables, different test values are used for the implied input variables to receive a measurement result of a parameter of the produced textile goods. The records are produced for the training of the neuronal network. The neuronal network is manually or automatically given for the determination of implied input variables and influencing variable through a sensor device and/or a computing device of the spinning machine or a system control. The implied influencing variable is a material variable that represents the material properties submitting the spinning machine e.g. thread type, thread length, a quality significant parameter and/or the type of pretreatment, is an environment variable that represents the characteristics of the environment e.g. ambient temperature and/or ambient humidity, and is a process parameter that represents the characteristics of the textile machine e.g. drawing-in speed, web tension, predistortion, main draft distance of a stretching unit of the textile machine and/or position of band redirection poles attached to the stretching unit. The implied input variable is standard trigger point, regulation intensity or a contact pressure of an upper roll or a roll pair of the stretching unit. The implied parameter is a quality significant parameter e.g. a parameter value or a thread volume drawn by the stretching unit. The input variable is determined by the neural network, which is separately realized by the textile machine or integrated in the textile machine and which is formed as a software or hardware control evaluation device or its part. Independent claims are included for: (1) a device for controlling a spinning preparation machine; and (2) a spinning preparation machine. .

Description

The The present invention relates to a method for controlling a Textile machine, in particular a spinning preparation machine, for example, a card or track, or a spinning machine, For example, a rotor spinning machine, in which one or more Input variables of a Control device of the textile machine to be determined so that one or more characteristics of the Textile goods produced with the textile machine can be optimized.

Farther the invention relates to a device for carrying out the Procedure, with an output-side interface to the output one or more input variables of a Control device of a textile machine, wherein said Influencing factors at least a characteristic of the the textile machine produced textile material can be optimized.

As well The present invention relates to a textile machine, in particular a spinning preparation machine, for example a card or Track, or a spinning machine, such as a rotor spinning machine.

The Properties of a textile product produced with a textile machine hang of a large number of influencing variables, which in the control of the Textile machine to consider are off. This results in many cases the task of an input of a Control device of a textile machine in number so too determine that a characteristic of the with the textile machine produced textile material is optimized. in this connection is common the exact relationship between the sought input and the resulting characteristic of the textile material unknown or at least not analytically writable.

in this connection occurs frequently added that the influence of the input variable on the characteristic of the textile material may be dependent on a variety of constraints. That is, the concrete value of an input which in the one case to an optimum value of the characteristic of the textile material leads, can be a completely unsatisfactory result under other boundary conditions to lead. The to be considered Boundary conditions, in the following influencing factors, which the effect of the searched input on the influence the resulting parameter are usually known as such, but it is itself in numerically known Influencing variables as a rule not possible, the required input value analytically to determine in order to achieve the optimum value of the characteristic of the textile material.

In practice, it is therefore common the requested input quantity non-company Attempts to determine. It is based on experience, a certain value for set the input, produces a certain amount of textile material and the resulting Characteristic of the textile material determined. Then the input quantity becomes a certain value changed in turn generates textile material and determines the now resulting value of the characteristic. This procedure is then repeated many times, for example twenty times, repeatedly and ultimately that value of the input selected, the led to the best value of the characteristic. The so determined input size can then in a subsequent production plant of the textile machine be used.

adversely in such a method, however, it is necessary a considerable amount, sometimes an intolerable amount Committee is produced. It is also disadvantageous that the Influencing factors which the effect of the input quantity on the Influencing the parameter not directly considered become. Changes So only one of the factors in a significant way, so the input size must be again be determined by means of the described test method.

If, which is common in practice occurs to determine several input variables are and / or several characteristics of the to be optimized with the textile machine textile goods, so reinforced continue the previously identified problem. In particular, be the required test series much more expensive.

The problem outlined will be illustrated once more by way of example of a method for controlling the distortion of a default field of a textile machine. In the control of the delay of a default field of a textile machine, it is necessary to determine the position of the so-called Regeleinsatzpunktes. The Regeleinsatzpunkt is the location at which an upstream of the default field measured in terms of its length-specific mass portion of a Fasergemenges, such as a sliver or a nonwoven fabric is located when a required to equalize the length-specific mass control intervention in the delay of the default field. The location of the Regeleinsatzpunktes can be specified, for example, as the distance from the measuring point. Such an indication ultimately indicates that distance which a certain section of the fiber amount travels from the measuring location to the point of delay. Alternatively, the location of the Regeleinsatzpunktes as the running time, which requires a certain portion of the fiber amount from the site to the location of default, could be specified. Both details are synonymous technically equivalent. For conversion, only the speed of Fasergemenges be can be known.

Of the Rule point should now be determined and set so that the drafting leaving fiber amount in its longitudinal course one possible uniform cross section having. The so-called CV% value is a parameter which this uniformity expresses. The lower the CV% value, the smoother it is distributed the mass of Fasergemenges. The location of the rule entry point, which to the best possible, so at least, CV% value leads, depends on a large number of factors. Such factors are Among other things, the settings of the drafting of the textile machine. Nevertheless, the rule insert point can not be determined analytically if these setting values and the further influencing variables are known are. It is therefore common the rule usage point by means of more or less automated To determine the test procedure.

So is out of the DE 10041892 A1 a method for controlling the delay of a default field on a Regulierstrecke known, the Regeleinsatzpunkt is determined in a pre-operational test or adjustment run. For this purpose, various control application points are set in succession on a trial basis, with several CV% values having a different reference length of the fiber amount leaving the default field being determined at each trial set point of use. The CV% values determined at a certain control point are added together to obtain a quality index. The quality indices determined at the different experimentally set control points are then used to form a polynomial of the second degree, the minimum of which is determined by means of numerical methods and considered as an optimal control point of use.

A disadvantage of the by the DE 10041892 A1 The disclosed method is that when it is carried out, typically several thousand meters of rejects are produced. It is equally disadvantageous that the method described above is relatively time-consuming, so that the productivity of the textile machine is reduced.

As well It is disadvantageous that the described process is repeated in its entirety must, if only one of the aforementioned influencing factors changes. This leads to another production shutdown and another rejection.

task Thus, the present invention is a method, an apparatus to carry out of the method and to provide a textile machine, which the Avoid mentioned disadvantages.

The Task is solved by a method, a device and a textile machine with the characteristics of the independent Claims.

at the method according to the invention the required input quantity or the sought input variables are dependent of one or more influencing factors which influence the effect of said input variables influence said parameters, determined by means of a neural network.

Provided, that a suitable and appropriately trained neural network is used, it is possible one or even several input variables for arbitrary values of the influencing variables without execution of test or set runs so that one or more resulting characteristics of the textile machine produced textile good to be optimized. To determine the input variables only the concrete numerical values of the influencing variables are determined. If determines these factors are, can the respective input values practically be determined without any delay. A committee will be there not produced.

Under a neural network should be understood as a computing device become a variety of artificial Neurons encompass each other for information processing are connected. The term "neuronal Networks "refers to the present application, therefore, exclusively on artificial neural networks. such neural networks allow the execution of information processing processes, their structure and functioning live on the nervous networks Orient organisms. This is also the case with the ones considered here artificial neural networks, the input signals through multiple layers of Neurons, where the neurons of a layer are the respectively supplied signals modify and over divergent connections to usually several neurons one send next layer. The central aspect here is the parallel one Information processing, which in contrast to the serial concepts classical computer systems stands.

Since neural networks are capable of learning, it is not necessary to know the exact relationship between the influencing variables considered and the input variables to be determined. All that is required is that the influencing factors are known and quantifiable as such. Using the example of the rule point search in a drafting system, this means that it first has to be determined which settings on the drafting system have any effect on the best possible control point of use. Then the numerical values of the identified influencing variables must be determined and the Neural network are supplied, so that the respective input variables can be determined by means of the neural network.

to Determination of the sought input variables is preferred each said Influence size one respectively associated neuron of an input layer of the neural Network supplied as a neuron input. hereby it is ensured that each of the determined influencing factors contributes the determination of the respective input variables is taken into account. alternative would it be conceivable, derived from various of the mentioned parameters Sizes to the To transmit neurons to the input layer of the neural network. This would the complexity of the neural network, but then the interactions would have to decrease the summarized influencing factors exactly be known.

Prefers is at least one said influencing variable, preferably each said Normalized influence variable, before being supplied to the respectively assigned neuron of the input layer. This can be taken into account become that the value ranges of the influencing variables strongly can distinguish. The value ranges of the influencing variables can be, for example to the interval [-1, +1] or the interval [0, 1]. In practical However, experiments have shown that the use of the latter Intervals often leads to an improved trainability of the neural network. Provided the values of an influencing variable as nominal values exist, that is, for example, in the designation of a type of fiber, so can the Materials are assigned numbers and these are also normalized.

advantageously, is at least one of the sought inputs, preferably each of the sought Input variables, from an associated neuron of an output layer of the neural Network delivered as a neuron output. In this case, the sought input variables become direct calculated using the neural network. Alternatively, the Input variables, however also indirectly from one or more neuron output quantities of the neural network, which is an additional one Computing would result.

Prefers is at least one said input, preferably each said Input quantity in standardized Form delivered. This ensures that within the neural network exclusively processed normalized sizes so that the neural network is easy to train. Provided to control the textile machine the input in denormierter form is needed denormation can be carried out by means of a separate denormation stage, which is connected downstream of the neural network.

advantageously, a neural network is used which has one or more hidden layers with hidden neurons. By such an increase of the Number of neurons contained in the neural network can increase the learning ability of the network, so also the accuracy of the thus determined input variables can be improved.

advantageously, Such a neural network is used, in which every neuron one layer with each neuron of an adjacent layer, preferably is associated with each neuron of each adjacent layer. hereby The number of connections between neurons will be different Layers maximized given the number of neurons per layer. This increases on the one hand, the performance of the network, on the other hand, there is also some redundancy, In this case, the failure of a connection is less noticeable power.

there It is preferred to use a neural network in which all connections between the neurons each assigned a variable weight is. The weight of a connection indicates with what factor Output signal of a neuron is multiplied before it from the next Neuron is processed. If any existing connection is a variable Weight, so is the performance of the neural network maximized, since then every single connection for information storage can contribute.

Especially said input variables are preferred by forward propagation the underlying factors. Vorwärtspropagierung means that the flow of information only towards the output layer takes place. Such a method can easily with a forward calculation Net (feed forward net) become. in principle could However, also be provided that at least part of the flow of information in the neural network flows in the direction of the input side. For this would be a relatively complex recurring network required.

In an embodiment the method according to the invention are said input quantities before an operational phase of the textile machine determined. The thus determined Input size can then in the operational phase of the textile machine for control be used. Such a procedure always makes sense if it is to be expected that the relevant influencing factors have a longer Period during operation of the textile machine remain constant.

In many cases However, it makes sense if the sought input variables in the operational phase of the textile machine can be repeatedly determined. A repetition may be, for example, periodically or on entry of a predefined event. An initiation a repetition can be provided in particular if a relevant factor changes. Intended However, it can also be that the sought input variables continuously or quasi-continuously determined over time. From one continuous or quasi-continuous determination of the input variable can be spoken when the repetition rate of the calculation only is limited by the duration of the calculation itself, so no breaks are provided. Due to the repeated determination of the input variables ensuring that the value of the input variable is available at all times, which leads to the optimization of the relevant characteristic of the textile material.

Especially Preferably, the determined input variable is automatically for control the textile machine used. This is especially useful if the input variables with a high repetition rate can be determined. Alternatively, however, could also be provided that a newly determined value of the input size only after confirmation by an operator is used to control the textile machine.

Prefers be in an off-farm Trial phase records Each record contains values of said influencing variables as well as empirically determined values of said input variables, which at the respective values of the influencing variables to an optimal value of the parameters. With In other words, there will be records generates the input variables of the neural network and the corresponding nominal values of the Output variables of the neural network. Such records allow a targeted Training the neural network.

For this it is beneficial if the neural network is in an off-site Learning phase by reverse propagation is trained. backpropagation means that the flow of information from the output layer to the input layer. When the training of the neural network into an external company Learning phase may be an undesirable influence on the production process the textile machine can be avoided.

advantageously, become in the learning phase the variable weights of the connections adjusted between the neurons. Alternatively or additionally could also a parameter of information transfer be adapted within a neuron. In particular, the so-called activation function or de output function can be adjusted.

Especially The neural network is preferably monitored during the learning phase Learning exercises. monitored Learning means that the neural network has both the neuron input quantities, here So the influencing factors, as also the neuron output quantities, here So the input variables, given where, after each learning step, the deviation between a predetermined and a calculated neuron output quantity determined and the network is specifically optimized by a learning rule.

In an advantageous embodiment of the method according to the invention It is envisaged that in an external test phase, the determined by the neural network and the control the textile machine used values of said input variables are checked, in which a plurality of different test values for said Input variables used is at least one measurement result of at least one Characteristic of the textile material to obtain. From the different test values can then the test values are selected which are the best value of each Characteristics of the Textile good led. If this deviations between the original means of the neural network and the input during the test phase determined input variables, can further records be generated for training the neural network.

In an embodiment of the method is said to the neural network for the determination Input variables at least a said influencing variable by set an operator manually. This is particularly advantageous if the manually specified influencing variable does not change over time or only negligible little changed.

In many cases However, it is advantageous if the neural network for determination at least a said influencing variable by a sensor device and / or a computing device of Textile machine or a plant control automatically specified becomes. This is particularly desirable when the automatically predetermined influencing variable in the operation of the textile machine often or even changed constantly. In this way it is possible that the determined input quantities always are adapted to the currently prevailing influencing variables.

Preferably, at least one said influencing variable, a material size, which represents one or more properties of the material presented to the textile machine. The material size In particular, the length-specific weight of the band (s) presented, the type of fiber, for example cotton, the fiber length, a quality-indicative quantity, for example a CV% value with a certain cut length or a size derived from one or more CV% values, and / or or the type of pretreatment of the material submitted. The type of pretreatment is understood to mean, for example, the number of drafting units via which a sliver was guided before it is presented to the textile machine to be controlled.

In many cases it is also advantageous if an environmental quantity is taken into account as an influencing variable. An environmental size represents one or more characteristics of the environment of the textile machine, for example the ambient temperature and / or the ambient air humidity.

In many cases It is advantageous if, alternatively or additionally, operating variables of Textile machine considered as influencing factors become. Represent operating variables one or more operational characteristics of the textile machine. The term farm size includes in particular setting values, values derived therefrom, constructive predetermined values and / or the configuration of the textile machine. At a track can For example, the intake speed, the intake tension, the Pre-delay, the total delay, the pre-delay distance, the main draft distance and / or the position of the Bandumlenkstäbe the drafting of the track considered as farm sizes become.

Prefers is a said input is the rule point, the rule intensity or the contact pressure of an upper roller of a pair of rollers of a drafting system of the textile machine. Just the above sizes affect the properties a given by a drafting textile material in essential Wise. However, their determination is usually analytical not possible. This is especially because the effect of these variables on The textile is dependent on a number of factors, including the exact Type of dependency usually not analytically describable.

The can be optimized characteristic in particular a quality-characterizing Size, for example the CV% value or a quantity derived therefrom, that of the textile machine be delivered textile good. In particular, the to be optimized Parameter of the CV% value or a derived quantity of one be from a drafting of the textile machine discharged sliver.

If used to determine the desired inputs a neural network is, which is realized separately from the textile machine, so can with one and the same neural network input variables for control devices of various Textile machines are determined. For example, one can do this suitable neural network in a plant control or in a be integrated portable device. The input variables can then manually or preferably automatically transferred to the respective textile machine become.

If however, the input size of a Control device of a textile machine by means of a neural Net is integrated, which integrates into this textile machine is, so in many cases lead to a more precise determination of the input variables, since the training of the neural network the individual conditions of use of each Textile machine implicitly be taken into account. For example, production-related deviations of the textile machine automatically taken into account by the standard. It is preferable that all empirical values used to train the neural network were determined directly with the textile machine to be controlled.

Prefers is used to determine the input variable or the input variables neural network used, which as a software-driven Be calculating device or formed as part of such. Such a software implementation The neural network can usually be accomplished inexpensively.

In some cases However, it can also be advantageous if a neural network is used, which as a hardware-controlled calculation device or as part such is formed. The advantages arise in particular for more complex problems, as hardware implemented neural networks with higher Speed can work.

A Apparatus for carrying out a method according to the invention is characterized by an input side interface to Input of one or more influencing factors, which the effect of the said input variables affecting at least one parameter. In this case, the input-side interface is so with a neural Connected network that said factors of an input layer of the Neural network feedable are. Furthermore, the output side interface with a Output layer of the neural network so connected, so that said input variables in dependence can be output from said influencing variables are. With such a device, the advantages of the invention be realized.

advantageously, the input-side interface is designed for normalization of said influencing variables. On separate standardization levels can then be waived.

As well it is advantageous if the output-side interface for denormalization formed of these input variables is. In this case, you can do without a special denormalization stage become.

Around a powerful Device can be provided that the neural network has at least one hidden layer with hidden neurons.

To For the same purpose it can be provided that each neuron of a layer with each neuron of an adjacent layer, preferably with each Neuron is connected to each adjacent layer.

Also can be provided that all connections between the neurons each associated with a variable weight.

Prefers the neural network is designed so that said input quantities through Vorwärtspropagierung said influencing variables can be determined.

Especially Preferably, the neural network is configured to be by reverse propagation is trainable.

Prefers the device is separate from the textile machine, for example as part of a plant control or as a standalone, portable device, realized, wherein preferably said input variables means an at least temporarily producible data connection automatically transferable to the textile machine are.

alternative The device can be an integral part of the textile machine be educated.

In a preferred embodiment is the neural network as a software-driven computing device realized.

In a further embodiment is the neural network as a hardware-driven computing device realized.

A textile machine according to the invention is characterized in that a device for carrying out the provided inventive method is. This results in the advantages according to the invention.

Further Advantages of the invention are in the following embodiments described. Show it:

1 a route as an example of a textile machine according to the prior art;

2 a route according to the invention;

3 a device according to the invention with a neural network;

4 an enlarged view of a neuron of the neural network; and

5 a flow diagram of a method according to the invention.

1 shows a trained according to the prior art route 1 , as an example of a textile machine 1 , That of the track 1 submitted material FB _to be in the direction of LR via a tape feed 2 , an inlet sensor unit 3 , a diverter 4 , a drafting system 5 , an exit guide 6 and about a tape storage 7 guided.

The only sketchy tape feed 2 has a deflection tube 8th which is arranged so that a presented sliver FB _to from one of the route 1 provided spinning can 9 can be removed.

The tape feeder 2 However, could also be designed so that they can take over a submitted sliver FB _to directly from a running card. Also, and in practice this is the most common case, could be the tape feed 2 for the simultaneous removal of several submitted slivers _to FB from various-supplied sliver cans 9 be educated.

The submitted sliver FB _to or the group of slivers FB presented _to be from the tape feeder 2 to the inlet sensor unit 3 transported. This is _for the section-wise detection of the length specific mass of the fiber sliver guided through it FB. Is a group of fiber ribbons _to FB presented, is through the inlet sensor unit 3 _to detected the total mass of fiber ribbons contained therein FB. The individual measured sections AB usually have a length of a few millimeters. For each measured section AB is through the inlet sensor unit 3 generates a measured value MW. For this purpose, the inlet sensor unit 3 as indicated by a driven feeler roller pair 3a . 3a ' include. However, they are also inlet sensor units 3 known which work according to other physical principles.

Below slivers If FB is _to speak, should thus be possible that only a sliver FB is _{to be} meant. The deflection unit 4 with the Bandumlenkstäben 4a . 4b . 4c and 4d serves to that of the inlet sensor unit 3 measured fiber ribbons FB _to transverse to the direction evenly spread and in the area of the inlet sensor unit 3 Pressing the fiber slivers FB _to dissolve. In order _to achieve such a uniform spread even with different strength and number of slivers presented FB, the position of the Bandumlenkstäbe 4a . 4b . 4c . 4d adjustable. By adjusting the height of the Bandumlenkstäbe both too wide and too narrow spread of the fiber ribbons FB _{can be} corrected.

When transporting the slivers FB _to from the inlet sensor unit 3 to the drafting system 5 the fiber ribbons FB are subjected _to a pull-in voltage VE, which by a different peripheral speed of the Tastwalzenpaares 3a . 3a ' and the feed roller pair 5a . 5a ' is effected. The pull-in voltage VE can be set, for example, via change gears (not shown), typically in the range from 0.95 to 1.05. In this case, the determination of the pull-in voltage VE essentially takes place as a function of the distance 1 as fiber slivers FB _to presented fiber material. The pull-in voltage VE is usually adjusted only during a batch change and otherwise during operation of the track 1 kept constant.

The drafting system 5 includes the already mentioned input roller pair 5a . 5a ' and a middle roll pair 5b . 5b ' and a pair of delivery rollers 5c . 5c ' , The roller pairs 5a . 5a '; 5b . 5b '; 5c . 5c '; are driven so that the speed increases from pair of rollers to pair of rollers in the running direction. As a result, the fiber ribbons FB _to both the Vorverzugsfeld VVF, which between the pair of input rollers 5a . 5a ' and the middle roll pair 5b . 5b ' is formed, as well as in the main drafting field VF, which between the middle roller pair 5b . 5b ' and the pair of delivery rollers 5c . 5c ' is formed, warped.

The pre-delay distance VVD indicates the distance between the nip lines of the feed roller pair 5a . 5a ' and the center roller pair 5b . 5b ' , Furthermore, the main drafting distance HVD corresponds to the distance of the clamping lines of the middle roller pair 5b . 5b ' and the delivery roller pair 5c . 5c ' , Both the pre-drafting distance VVD and the main drafting distance HVD are for adjusting the drafting system 5 adjustable to the fiber sliver FB _to be forgiven.

The lower rollers 5a . 5d . 5c of the drafting system 5 are stationary, however, are the top rollers 5a ' . 5b ' 5c ' movably mounted and are not shown by means of stress against the lower rollers 5a . 5b . 5c pressed so that _to results in a secure clamping of the slivers FB.

The exit guide 6 includes a measuring funnel 10 as well as take-off rolls 11 , The measuring funnel 10 initially serves to compress the warped slivers FB _vz so that a single compact fiber sliver FB _ab originates. For this purpose, it has a funnel-shaped section in its upstream part. In addition, the measuring funnel generates 10 by means of a downstream of the funnel-shaped portion arranged outlet sensor, a signal S, which corresponds to the length-specific mass of the discharged sliver FB _from . The extraction rollers 11 serve _from the removal of the fiber band FB from the measuring hoppers 10 and the further compaction of the fiber band FB _from.

The tape storage 7 , which is not explained in detail here, serves the orderly filing of the means of the track 1 fiber band FB generated _from a sliver can 12 ,

The drafting system 5 is via a control device 13 controllable. In this case, the control device 13 an operating unit 14 which allows an operator to specify settings, which then as an input to the control device 13 be transmitted. Such input variables which can be predetermined by the operator are the delivery speed LG, the band weight BG, the control application point REP and the rule intensity RI. Further input of the control device 13 is the current measured value MW, that of the inlet sensor unit 3 automatically to the control device 13 is transmitted. Shown are only selected input variables of the control device 13 presented and discussed. In practice, the control device 13 fed to further input variables.

The control device 13 is designed so that it depends on their input variables, the speed of the input roller pair 5a . 5a ' , the speed of the middle roller pair 5b . 5b ' and the speed of the delivery roller pair 5c . 5c ' controlled by acting on not shown drive means. In unregulated operation, a fixed preliminary delay VV is determined for the default draft field VF and a fixed HV for the main drafting field VF. In regulated operation, however, control interventions take place in the default of at least one default field.

The speeds of the roller pairs 5a . 5a '; 5b . 5b '; 5c . 5c ' of the drafting system 5 be if no control intervention takes place, especially in response to the predetermined delivery speed LG, the predetermined band BG weight of the dispensed fiber band FB _from and the added mass of the led sliver FB _to determined.

The delivery speed LG is the speed with which the warped fiber slivers FB the drafting system 5 leave. The predetermined by the operator delivery speed LG is the basis for the control of the speed of the pair of delivery rollers 5c . 5c ' ,

The band weight BG describes the average length-specific mass of the track 1 discharged sliver FB _from . From this size, in conjunction with the mass of the fed sliver FB _to the necessary total draft of the drafting system 5 be determined. The distribution of the total delay on the default draft field VVF and on the main drafting field VF is based on empirical values. From this can now the speed of the pair of input rollers 5a . 5a ' and the number of revolutions of the middle roller pair 5b . 5b ' be determined.

In regulated operation, distortion changes, also called control interventions, are provided around the drafting system 5 To fed to fiber slices FB _too uniform. In the usual way, in such a control intervention in the embodiment of 1 the speed of the delivery roller pair 5c . 5c ' kept constant, whereas the speed of the input roller pair 5a . 5a ' and the number of revolutions of the middle roller pair 5b . 5b ' each increased or decreased by the same percentage. Therefore, the delay VV in Vorverzugsfeld VVF is constant, while the default HV of the main drafting field VF is changeable.

Control interventions are carried out on the basis of the measurements of the inlet sensor unit 3 , A control in which the measuring location in front of the drafting system 5 is located, so an inlet sensor unit 3 is commonly called open-loop control. In such a control, the running distance or the running time of a section AB of the supplied fiber ribbons FB is _{to be considered} up to the point REP at which the control intervention is to take place. Running distance and running time are linked by the speed of the drafting system. In this case, the determination of the regular use point REP attains an outstanding importance. This is explained below:
The drafting zone the fiber mixture FV supplied _to FG consists of successively arranged portions. The reference AB _n denotes the section which is measured by the sensor device in the moment shown. Downstream of section AB _n is section AB _n-1 , upstream section AB _{n + 1} . For reasons of simplification, the further sections are not designated by reference symbols. For each of the sections, at least one measured value MW is determined, which corresponds to the length-specific mass of the respective section and to the control device 13 is transmitted.

When the measured section AB _{n reaches} the control point REP, that is, the position designated AB ' _n , is determined by the controller 13 causes a corresponding control intervention. If the section AB _n has a length-specific mass lying above the average, for example, then an increase in the delay to equalize the fiber quantity FG is initiated.

The location of the control point of application REP is usually referred to as the distance A of the Regeleinsatzpunktes of the sensor device 11 specified. The control device 10 is formed so that the distance A in a certain, indicated by a double arrow area, can be predetermined. The exact location of the control point REP can by the operator by means of the control unit 14 to the control device 13 be transmitted. The predetermined position of the control application point REP or the predetermined distance A is critically _on the quality of the discharged from the drafting zone VF fiber mixture FG. The optimal position of the control point REP, in short the optimal control point REP, depends on a large number of factors and can not be determined analytically. These influencing variables can also occur during the production phase of the textile machine 1 change.

There the optimal position of the control point REP not with sufficient Accuracy can be determined analytically, according to the state of the art Technology is the control point REP in a pre-operational setting or test run and for a longer Period, for example, held constant until a game change.

Decisive for the quality of the produced sliver FB _ab is in particular the uniformity of the length-specific mass. To check this uniformity, it is known from the prior art, by means of an evaluation unit 15 from the signals S of the measuring funnel 10 to calculate a coefficient of variation indicating the percent band nonuniformity for a given reference length of the sliver FB _ab . The reference length is also called cut length, the coefficient of variation also CV% value. In practice, the cutting lengths used are in the range between a few centimeters and a few meters.

To perform manually controlled test runs, which is used to determine the control point REP, the determined CV% value to the operator by means of a display 16 displayed. in the As part of such test runs, the operator varies the control point and the intensity of the rule until the CV% value reaches the optimum, namely the smallest, value. However, to carry out automatically controlled test runs, the CV% value could also be sent directly to the control device 13 be transmitted.

Such test runs, which can be done manually or automatically, result in a greater amount of waste produced. Furthermore, such test runs are associated with a high expenditure of time, so that the productivity of the track 1 is severely restricted.

The rule intensity RI is another input of the control device 13 , which can be specified by an operator. The rule intensity determines the value of the draft in the drafting system 5 is changed when the length-specific mass of slivers FB submitted _to changes. The rule intensity RI thus determines the strength of a control intervention taking place under certain conditions. Also, the rule intensity RI is determined according to the prior art in off-the-job test runs.

To determine the rule intensity RI, fiber slivers FB _of different length-specific total mass are successively introduced in the course of the test runs according to the prior art. The rule intensity RI is then set so that the average length-specific mass M of the generated sliver FB _ab , which from the signals S of the measuring funnel 10 is calculated and transmitted to the display, regardless of the mass of fiber slabs FB submitted _to . This will produce an extra scrap and the possible production time of the line 1 further reduced.

2 shows a route 1 , which is formed in accordance with the invention. It is characterized by a device 17 for carrying out the method according to the invention. The device 17 is used to determine the control point REP and the rule intensity RI as a function of several influencing variables EG, VE, VV, VVD, HVD, which the effect of the control point REP and the rule intensity RI on the CV% value of the means of the route 1 influenced sliver FB _from influence. These are in the illustrated embodiment, the pull-in voltage VE, the pre-delay VV, the Vorverzugsdistanz VVD and the main drafting distance HVD. These numerically known influencing variables are determined by the operator by means of the operating unit 14 to the device 17 transmitted. Likewise, the set position of Bandumlenkstäbe 4a . 4b . 4c . 4d to the device 17 be transmitted.

Another influencing factor is the intake speed EG of the drafting system 5 by means of the inlet sensor unit 3 determined and continuously to the device 17 automatically transmitted. This is advantageous because the feed rate during operation of the route 1 conditionally changes due to control interventions. By contrast, the parameters VE, VV, VVD, HVD specified by the operator remain in operation of the route 1 constant over a longer period of time. An adaptation of these influencing variables VE, VV, VVD, HVD normally occurs only during a batch change.

The device 17 is completely in the track 1 integrated. The influencing factors EG, VE, VV, VVD, HVD become integral components of the line 1 automatically to the device 17 transmitted. Also the output of the device 17 namely, the rule intensity RI and the rule use point REP, are internally and automatically sent to the controller 13 transmitted. During operation of the route, therefore, the rule intensity RI and the rule use point REP, which by means of the device 17 were determined automatically be used to control the textile machine. If, during operation of the line, the pull-in speed EG of the drafting system 5 changes, the rule intensity RI and the rule use point REP are automatically adjusted, so that in the result of the CV% value of the generated sliver FB _AB always assumes the smallest possible value.

3 shows the basic structure of a device according to the invention 17 , Core of the device 17 is a neural network 18 , The neural network 18 comprises a plurality of neurons N, which are arranged in an input layer ES, a hidden layer VS and an output layer AS. The indexing of the neurons N refers to the layer in which the respective neuron N is located and to the serial number of the neuron N in the respective layer. For example, the neuron N _{V3 is} in the hidden layer S at the third position.

Each neuron N _E1 , N _E2 , N _E3 , N _E4 , N _{E5 of} the input layer ES is connected to each neuron N _V1 , N _V2 , N _V3 , N _V4 , N _V5 , N _{V6 of} the hidden layer VS. Likewise, each neuron N _V1 , N _V2 , N _V3 , N _V4 , N _V5 , N _{V6 of} the hidden layer VS is connected to each neuron N _A1 , N _{A2 of} the output layer AS.

To supply the influencing variables EG, VE, VV, VVD, HVD is an input-side interface 19 intended. The input-side interface 19 is for normalization of the respective influencing variables EG, VE, VV, VVD, HVD formed and so with the neural network 18 connected to each of said factors EG, VE, VV, VVD, HVD a neuron N _E1 , N _E2 , N _E3 , N _E4 , N _{E5 of} the input layer ES can be supplied as neuron input quantities NE _E1 , NE _E2 , NE _E3 , NE _E4 , NE _E5 . For example, the pull-in voltage VE is supplied to the neuron N _E2 as a normalized neuron input NE _E2 .

Based on the neuron input NE _E1 , NE _E2 , NE _E3 , NE _E4 , NE _E5 supplied to it, each neuron N _E1 , N _E2 , N _E3 , N _E4 , N _{E5 of} the input layer ES generates an output which is sent to each neuron N _V1 , N _V2 , N _V3 , N _V4 , N _V5 , N _{V6 is} transmitted to the hidden layer. Each neuron N _V1 , N _V2 , N _V3 , N _V4 , N _V5 , N _{V6 of} the hidden layer VS therefore receives five input quantities. Each of the neurons N _V1 , N _V2 , N _V3 , N _V4 , N _V5 , N _V6 combines the input quantities supplied to them and converts them into an output signal. This output signal is in turn supplied to each neuron N _A1 , N _{A2 of} the output layer AS. Based on the respective six applied signals, the neurons N _A1 , N _A2 generate an output signal NA _A1 , NA _A2 , wherein the neuron output signal NA _{A1 contains} the determined control application point REP and the neuron output signal NA _{A2 contains} the rule intensity RI in normalized form. For denormalization and for the output of the control application point REP and the rule intensity RI is an output-side interface 20 intended.

The outlined determination of the control point REP and the rule intensity RI as forward propagation denoted as the information flow exclusively from the input layer ES takes place in the direction of the output layer AS.

Prerequisite for this is that the neural network 18 is trained accordingly. This can be effected in particular by supervised learning. This will be the neural network 18 Data sets D _i presented, which values EG _i , VE _i , VV _i , VVD _i , HVD _i the input variables of the neural network, ie the factors EC, VE, VV, VVD, HVD, included. Furthermore, the data records D _i contain empirically determined desired values REP _ie , RI _ie for the output variables of the neural network, ie values of the control application point REP and the rule intensity RI. Based on the input variables, output variables REP _ib , RI _{ib are} calculated, which are then compared with the empirical values REP _ie , RI _ie . Due to this comparison, the information processing operations in the neural network can 18 be adjusted. It is usually started at the output layer AS, then the hidden layer VS and finally the input layer ES adapted. This is also referred to as backward propagation of the errors. The data sets D _i required for training are of the device 17 via an interface 21 for receiving records D _{i can be} fed.

4 shows an example of the information processing in the context of the forward propagation through the Neuron N _V3 . The neuron N _V3 is a neuron of the hidden layer VS. It is therefore connected to each neuron N _E1 , N _E2 , N _E3 , N _E4 , N _{E6 of} the input layer ES and to each neuron N _A1 , N _{A2 of} the output layer AS via a direct link. Each link is assigned a variable weight W. The injection of the links W contains the index of the transmitting neuron and the index of the receiving neuron. The weight W _E2V3 is therefore the weight of the connection which connects the neuron N _E2 to the neuron N _V3 .

In the forward propagation, the neuron output _signals NA _E1 , NA _E2 , NA _E3 , NA _E4 , NA _{E5 of} the neurons N _E1 , N _E2 , N _E3 , N _E4 , N _{E5 of} the upstream input layer ES are multiplied by the weight of the respective connection, so that the result of the multiplication can be further processed by the receiving neuron N _V3 . For example, the neuron _output signal NA _{E2 of} the neuron N _{E2 is} multiplied by the weight W _E2V3 and further processed in this form by the neuron N _V3 . Overall, the neuron N _V3 receives five neuron input signals, which are processed by means of a so-called input function IF. Common input functions are the
Sum function y = Σ x or the
Product function y = Π x.

The result calculated by means of the input function is further processed by means of an activation function. The most common activation functions are the
Identity y = x, the
Sigmoid function y = 1/1 + e ^-gy or the
Hyperbolic tangent y = e ^gy - e ^-gy / e ^gy + e ^-gy

From the result of the activation function, the neuron output NA _{V3 of} the neuron N _{V3 is then} calculated by means of an output function. For this purpose, the
Identity y = x
Be used.

This neuron output quantity NA _V3 is in each case weighted, forwarded to the neurons N of the output layer AS. As part of the training of the neural network 18 the weights W of the neuron connections are changed until the output quantities REP _ib , RI _{ib of} the neural network 18 _ie with those found by empirical tests REP values, RI match _ie approximately. The basis for this are error minimization methods, such as, for example, gradient descent methods. The procedure is such that initially the weights W _VjAj between the output layer and the hidden layer VS are optimized and then the weights W _EkVk between the hidden layer and the input layer ES. This procedure is also called Reverse propagation called.

5 shows an exemplary overview of the procedure in the inventive determination of the Regeleinsatzpunktes REP and the rule intensity RI. The method according to the invention is an experimental phase VP, a learning phase LP, an operational phase BP and a training phase TP.

Within the scope of the test phase VP, for a given set of values of the influencing variables EG _i , VE _i , VV _i , VVD _i , HVD _i, the empirical values REP _ie , RI _ie are determined such that the CV% value of the generated fiber band FB _decreases is optimized. In this case, recourse may be had to the test methods known from the prior art. Now, a data set D _{i is} formed which comprises the values EG _i , VE _i , VV _i , VVD _i , HVD _{i of} the influencing variables and the empirical values REP _ie , RI _ie . As part of the experimental phase, the influencing variables are varied so that a large number of data sets D _i is available as a result.

In the learning phase, the influencing variables _i contained in a data set D EC _i, VE _i are calculated VV _i, VVD _i, HVD _i supplied to the neural network as an input variable and _ie by means of the neural network output values REP, RI _ie. Now, the difference ΔREP _i and the difference ΔRI _i between the empirically determined values REP _ie , RI _ie and the calculated values REP _b and RI _{b is} determined. The weights W _VjAj , namely the weights of the connections between the starting layer and the hidden layer, are then adjusted from the differences ΔREP _i ΔRI _i thus determined. From the adjusted weights W _VjAj between the output layer and the hidden layer, the weights W _EkVk between the hidden layer and the input layer are then determined. The weights W are set so that the error for each record D _{i is} minimized.

After completion of the training, the neural network in the operational phase BP for calculating the input variables of the control device 13 to be used. It is now possible to determine the best possible values for the control application point REP and the rule intensity RI for combinations of influencing variables EG, VE, VV, VVD, HVD.

In this case, it is possible to check the values of the control application point REP and the control intensity RI determined in the operational phase BP in a test phase TP by determining [REP - Δ, REP + Δ] and an interval [RI - Δ, RI + Δ] empirical values REP _e , RI _{e are} determined for which the CV% value is optimal. If the empirical values REP _e , RI _e thus determined correspond to the values REP, RI previously determined by means of the neural network, then the test phase can be ended. If this is not the case, new data records D _i can be generated on the basis of the new empirical values REP _e , RI _e . These new data sets D _i can in a further learning phase LP for training the neural network 17 can be used to increase the accuracy of the neural network 17 values determined in the operational phase to improve REP, RI.

alternative or additionally between the learning phase LP and the operational phase BP such Test phase TP be provided.

The Invention is not on the illustrated and described embodiment limited. through the method according to the invention can also other input variables control devices be determined by any textile machinery. It also can not mentioned Factors considered become. Likewise, the invention can be applied when a control device, which is part of a control loop, supplied with input variables shall be. Also can Method known from the prior art for automatic control point search combined with the use of a neural network. So For example, the control point determined with a neural network may be as starting value for a known method can be used.

Claims (42)

Method for controlling a textile machine ( 1 ), in particular a spinning preparation machine ( 1 ), such as a card or track ( 1 ), or a spinning machine, for example a rotor spinning machine, in which one or more input variables (REP, RI) of a control device ( 13 ) of the textile machine ( 1 ) are determined so that one or more parameters (CV%) of the textile machine ( 1 ) Textile material (FB generated are optimized _ab), characterized in that said input variables (REP, RI) as a function of one or more factors (EC, VE, VV, VVD, HVD) that the action of said input variables (REP, RI ) to said characteristics (CV%), by means of a neural network ( 18 ). Method according to Claim 1, characterized in that each said influencing quantity (EG, VE, VV, VVD, HVD) belongs to a respectively assigned neuron (N _E1 , N _E2 , N _E3 , N _E4 , N _E5 ) of an input layer (ES) of the neural Network ( 18 ) is supplied as a neuron input (NE _E1 , NE _E2 , NE _E3 , NE _E4 , NE _E5 ). A method according to claim 2, characterized in that at least one said influencing variable (EG, VE, VV, VVD, HVD), preferably each be said influencing variable (EG, VE, VV, VVD, HVD) is normalized before it is supplied to the respective associated neuron (N _E1 , N _E2 , N _E3 , N _E4 , N _E5 ) of the input layer (ES). Method according to one of claims 1 to 3, characterized in that at least one said input variable (REP, RI), preferably each said input variable (REP, RI), from a respective associated neuron (N _A1 , N _A2 ) of an output layer (AS) of the neural network as a neuron output (NA _A1 , NA _A2 ). Method according to one of claims 1 to 4, characterized in that at least one said input variable (REP, RI), preferably each said input (REP, RI), in standardized form. Method according to one of Claims 1 to 5, characterized in that a said neural network ( 18 ) having at least one hidden layer (VS) with hidden neurons (N _V1 , N _V2 , N _V3 , N _V4 , N _V5 , N _V6 ). Method according to one of Claims 1 to 6, characterized in that a said neural network ( 18 ), wherein each neuron (N _E1 , N _E2 , N _E3 , N _E4 , N _E5 , N _V1 , N _V2 , N _V3 , N _V4 , N _V5 , N _V6 , N _A1 , N _A2 ) of a layer (ES, VS, AS) with each neuron of an adjacent layer (ES, VS, AS), preferably with each neuron (N _E1 , N _E2 , N _E3 , N _E4 , N _E5 , N _V1 , N _V2 , N _V3 , is N _A1, N _A2) each adjacent layer (ES, VS, AS) connected N _{V4, V5} N, N _V6. Method according to one of Claims 1 to 7, characterized in that a said neural network ( 18 ), in which all connections between the neurons (N _E1 , N _E2 , N _E3 , N _E4 , N _E5 , N _V1 , N _V2 , N _V3 , N _V4 , N _V5 , N _V6 , N _A1 , N _A2 ) is assigned a variable weight (W _EkVk , W _VjAj ). Method according to one of claims 1 to 8, characterized that said input quantities (REP, RI) by forward propagation said influencing variables (EG, VE, VV, VVD, HVD) are determined. Method according to one of claims 1 to 9, characterized in that said input quantities (REP, RI) before an operational phase (BP) of the textile machine ( 1 ) be determined. Method according to one of claims 1 to 10, characterized in that said input variables (REP, RI) in an operational phase (BP) of the textile machine ( 1 ), preferably on occurrence of an event, for example, when changing a said influencing variable (EG, VE, VV, VVD, HVD), be determined periodically or continuously. Method according to one of claims 1 to 11, characterized in that said input variables (REP, RI) after their detection automatically to the control device ( 13 ) of the textile machine ( 1 ) be transmitted. Method according to one of Claims 1 to 12, characterized in that data records (D _i ) are generated in an off-field trial phase (VP), each data record (D _i ) having values (EG _i , VE _i , VV _i , VVD _i , HVD _i ) said influencing variables (EG, VE, VV, VVD, HVD) as well as empirically determined values (REP _ie , RI _ie ) of said input quantities (REP, RI) which at the respective values (EG _i , VE _i , VV _i , VVD _i , HVD _i ) of said influencing variables (EG, VE, VV, VVD, HVD) lead to an optimum value of the at least one parameter (CV%). Method according to one of claims 1 to 13, characterized in that the neural network ( 18 ) is trained in an off-field learning (LP), preferably by reverse propagation. A method according to claim 14, characterized in that in the learning phase (LP) the variable weights (W _EkVk , W _VjAj ) of the connections between the neurons (N _E1 , N _E2 , N _E3 , N _E4 , N _E5 , N _V1 , N _V2 , N _V3 , N _V4 , N _V5 , N _V6 , N _A1 , N _A2 ). Method according to claim 14 or 15, characterized in that the neural network ( 18 ) in the learning phase (LP) is trained by supervised learning. Method according to one of Claims 1 to 16, characterized in that, in an external test phase (TP), the tests performed by means of the neural network ( 18 ) and to control the textile machine ( 1 ) values of said input quantities (REP, RI) are checked by using a plurality of different test values for said input quantities (REP, RI) to produce at least one measurement result of the at least one characteristic variable (CV%) of the textile material produced (FB to get _down), preferably records (Di) for the training of the neural network ( 18 ) be generated. Method according to one of claims 1 to 17, characterized in that the neural network ( 18 ) for determining said input quantities (REP, RI), at least one said influencing variable (EG, VE, VV, VVD, HVD) is preset manually by an operator. Method according to one of claims 1 to 18, characterized in that the neural network ( 18 ) for determining said input quantities (REP, RI) at least one influencing variable (EG, VE, VV, VVD, HVD) by a sensor device ( 3 ) and / or a computing device of the textile machine ( 1 ) or a plant control is automatically specified. Method according to one of claims 1 to 19, characterized in that at least one said influencing variable (EG, VE, VV, VVD, HVD) is a material size (M), wherein the material size (M) one or more properties of the textile machine ( 1 ) submitted material (FB _to ), for example, the fiber type, the fiber length, a quality-indicative size and / or the type of pretreatment represented. Method according to one of Claims 1 to 20, characterized in that at least one said influencing variable (EG, VE, W, VVD, HVD) is an environmental variable, the environmental variable having one or more properties of the environment of the textile machine ( 1 ), for example the ambient temperature and / or the ambient air humidity. Method according to one of Claims 1 to 21, characterized in that at least one said influencing variable (EG, VE, VV, VVD, HVD) is an operating variable (EG, VE, VV, VVD, HVD), the operating variable having one or more properties the textile machine ( 1 ), for example, the intake speed (EG), the intake tension (VE), the pre-draft (VV), the Vorverzugsdistanz (VVD), the main drafting distance (HVD) of a drafting system ( 5 ) of the textile machine ( 1 ) and / or the position of one or more of the drafting system ( 5 ) associated Bandumlenkstäbe ( 4a . 4b . 4c . 4d ). Method according to one of Claims 1 to 22, characterized in that a said input variable (REP, RI) is the control application point (REP), the control intensity (RI) or the contact pressure of a top roller ( 5a ' . 5b ' . 5c ' ) of a roller pair ( 5a . 5a ; 5b . 5b '; 5c . 5c ' ) of a drafting system ( 5 ) of the textile machine ( 1 ). Method according to one of claims 1 to 23, characterized in that a said parameter (CV%) is a quality-indicative quantity, for example a CV% value or a variable derived from one or more CV% values, of the textile product produced by the textile machine ( FB _ab ), in particular one by means of a drafting system ( 5 ) of the textile machine ( 1 ) stretched sliver (FB _ab ) is. Method according to one of claims 1 to 24, characterized in that the at least one input variable (REP, RI) by means of a neural network ( 18 ), which is separate from the textile machine ( 1 ), wherein said input quantities (REP, RI) are fed manually or preferably automatically to the textile machine ( 1 ) be transmitted. Method according to one of claims 1 to 24, characterized in that the at least one input variable (REP, RI) by means of a neural network ( 18 ) is determined, which in the textile machine ( 1 ) is integrated. Method according to one of claims 1 to 26, characterized in that the at least one input variable (REP, RI) by means of a neural network ( 18 ) is determined, which is designed as a software-controlled calculation device or as part of such. Method according to one of claims 1 to 26, characterized in that the at least one input variable (REP, RI) by means of a neural network ( 18 ), which is designed as a hardware-controlled computing device or as part of such. Contraption ( 17 ), in particular for carrying out a method according to one of claims 1 to 28, having an output-side interface ( 20 ) for outputting one or more input variables (REP, RI) of a control of a textile machine ( 1 ), by which at least one parameter (CV%) of the with the textile machine ( 1 ) textile material (FB _ab ) can be optimized, characterized in that an input-side interface ( 19 ) for inputting one or more influencing variables (EG, VE, VV, VVD, HVD) which influence the effect of said input variables (REP, RI) on the at least one characteristic variable (CV%), the input-side interface ( 19 ) with a neural network ( 18 ), so that said influencing variables (EG, VE, W, VVD, HVD) of an input layer (ES) of the neural network ( 18 ) and wherein the output-side interface ( 20 ) with an output layer (AS) of the neural network ( 18 ), so that said input quantities (REP, RI) can be output as a function of said influencing variables (EG, VE, VV, VVD, HVD). Contraption ( 17 ) according to claim 29, characterized in that the input-side interface ( 19 ) is designed to normalize said influencing variables (EG, VE, VV, VVD, HVD). Contraption ( 17 ) according to claim 29 or 30, characterized in that the output side Interface ( 20 ) for denormalization of said input variables (REP, RI) is formed. Contraption ( 17 ) according to one of claims 29 to 31, characterized in that the neural network ( 18 ) has at least one hidden layer (V) with hidden neurons (N _V1 , N _V2 , N _V3 , N _V4 , N _V5 , N _V6 ). Contraption ( 17 ) according to any one of claims 29 to 32, characterized in that each neuron (N _E1 , N _E2 , N _E3 , N _E4 , N _E5 , N _V1 , N _V2 , N _V3 , N _V4 , N _V5 , N _V6 ; N _A1 , N _A2 ) of a layer (E, V, A) with each neuron of an adjacent layer (ES, VS, AS), preferably with each neuron (N _E1 , N _E2 , N _E3 , N _E4 , N _E5 , N _V1 , N _V2 , N _V3 , N _V4 , N _V5 , N _V6 , N _A1 , N _A2 ) of each adjacent layer (ES, VS, AS). Contraption ( 17 ) according to one of claims 29 to 33, characterized in that all connections between the neurons (N _E1 , N _E2 , N _E3 , N _E4 , N _E5 , N _V1 , N _V2 , N _V3 , N _V4 , N _V5 , N _V6 , N _A1 , N _A2 ) is each assigned a variable weight (W _EkVk , W _VjAj ). Contraption ( 17 ) according to one of claims 29 to 34, characterized in that the neural network ( 18 ) is designed so that said input quantities (REP, RI) can be determined by forward propagation of said influencing variables (EG, VE, VV, VVD, HVD). Contraption ( 17 ) according to one of claims 29 to 35, characterized in that the neural network ( 18 ) is designed so that it can be trained by backward propagation. Contraption ( 17 ) according to one of claims 29 to 36, characterized in that an interface ( 21 ) for receiving data sets (D _i ), which values of said influencing variables (EG _i , VE _i , VV _i , VVD _i , HVD _i ) as well as empirically determined values of said input variables (REP _ie , RI _ie ), which at the respective Values of said influencing variables (EG _i , VE _i , VV _i , VVD _i , HVD _i ) to an optimum value of the at least one characteristic variable (CV%) are provided, so that the neural network ( 18 ) is trainable on the basis of these data sets (D _i ). Contraption ( 17 ) according to one of claims 29 to 37, characterized in that it is separate from the textile machine ( 1 ), for example as part of a system control or as an independent, portable device, wherein preferably said input variables (REP, RI) are automatically connected to the textile machine by means of a data connection which can be produced at least temporarily. 1 ) are transferable. Contraption ( 17 ) according to one of claims 29 to 37, characterized in that it enters the textile machine ( 1 ) is integrated. Contraption ( 17 ) according to one of claims 29 to 39, characterized in that the neural network ( 18 ) is implemented as a software-controlled calculation device. Contraption ( 17 ) according to one of claims 29 to 39, characterized in that the neural network ( 18 ) is realized as a hardware-controlled calculation device. Textile machine ( 1 ), in particular a spinning preparation machine ( 1 ), such as a card or track ( 1 ), or a spinning machine, for example a rotor spinning machine, characterized in that a device ( 17 ) is provided according to one of claims 29 to 41.