DE19743600A1 - Monitoring cyclic production process - Google Patents

Abstract

The method involves first coordinating the factors characteristic for the production of the product with the relevant quality features based on checking signal forms against envelope curves. These are used to train a neural network so that it can, with a high degree of reliability in subsequent production stages, identify that the characteristic factors recorded match the above quality features.

Description

The invention relates to a method for monitoring a Production process in which signal courses on several Positions of the production process are included and at least one quality statement about the products produced is derived. In particular, the invention relates to a Process for monitoring a cyclical production process in which the acquisition of signal curves several state variables of the process by sensors, one Check these signals for admissibility and, based on the Waveforms of generated process indicators, a forecast the quality of the products produced takes place at unauthorized expression via a readjustment of the  Setting variables of the process again in the allowed Area is returned.

A production monitoring z. B. with the help of statistical process control (SPC) is based today in essentially on a random product test of the finished products. The review of the products, be it constantly, even if only in the form of random samples, is very complex. In addition, the test results are mostly only with a great time delay to manufacture. A quality-related regulation of a process by change the setting values is only with immediate knowledge of the Quality characteristics and the resulting change in values possible. Because of the complex relationships between Quality deviations and the required Setting changes this is very difficult and takes place therefore often intuitive or heuristic, i.e. without a fixed assignable rules, so that this knowledge or this Do not regulate or only insufficiently for automation own.

DE 42 09 746 A1 describes a method for optimization of a technical neuro-fuzzy system, at the data with which the neuro-fuzzy controller trains will be generated by simulation. This simulation is but a complex process and also requires knowledge  the causal relationship between the process flow and the Training of the neuro-fuzzy controller required data ahead.

DE 44 16 317 A1 describes one method and one Control device for controlling a material processing Process became known, taking into account certain process-relevant material properties Process parameters are calculated in advance, which are used for Presetting the system. This calculation too establishes the knowledge of causal relationships between the influencing factors influencing the process and Process parameters ahead. But this is the case with many processes not possible.

The same applies to the procedure and the Control device according to DE 44 16 364 A1.

The object of the invention is a method of the above Way of creating an automatic manufacturing consistently high quality of the products produced enables. In particular, automatic monitoring is intended of the production process, the associated documentation and a quality-based regulation without checking the Products may be possible even after their manufacture.  

To complete this complex process with the required To design operational security, there is an upstream Check the underlying signal data for Admissibility in its expression.

The inventive solution to the problem overcomes the difficulty that the causal link between process parameters, the Process parameters, the quality characteristics and / or the Characteristic classes of the products produced are not known, and that the data characteristic of a process is a Data abundance that previously exist parallel to the process flow could not be recorded and processed.

The following should be added to explain: A variable A quality feature is the length of a product, for example Example 10.4 mm +/- 0.05 mm to capture the value forecast or judge. Attributive One speaks quality characteristics if that too Classifying characteristic of one of several characteristic classes is assigned. Are attributive to one Parts assessment at least two feature classes (e.g. OK or n.i.o.) required. But there can also be several be attributive classes, e.g. B. Class 1 = 10-11 mm, Class 2 = 11-12 mm,. . ., Class x = 19-20 mm, or, as named classes, too little, lower limit, good, upper  Limit, too much, etc. These quality features are from the information contained in the signals about process and thus also predicting quality fluctuations.

From U.S. Patent 5,282,261 and International Patent application WO 93/25943 are methods for controlling Known processes that build on the application trainable neural networks for the prediction of Product properties based on direct entry and input of process status measurement data or using combined input of process status measurement data and current values of Process control variables. The input quantities of the neural network each refer to one in the cited writings determined, equidistant monitoring time of a continuous process. This approach of Use of measurement data records as input of the Prediction network is usually based on cyclical processes not transferable, because then the neural network is one of the Number of measurement data of the total of the observed Correspondingly, extraordinarily large signal data curves Number (typically several thousand) of input neurons and a correspondingly large number of through training determining weighted connections between the neurons should have.

Since the necessary for training such networks,  correspondingly large number of sample or Experimental data sets not due to real conditions of the immense effort that can be generated is one process-adapted extraction of process parameters and a indirect process modeling based on this a minimized data set of process parameters for cyclical processes, with the usually required short system response time, a mandatory requirement.

Methods are disclosed in the published patent application DE 195 18 804 A1 for discrete key figure formation described how Maximum determination, integrals, gradients etc. as they do also limited use in the present invention are.

Similar approaches are also included in the forecast models neural networks detectable. It is crucial that at none of the known writings Plausibility check of the signal sources involved is listed. The key figures are also generated mainly through user-related, mathematical Algorithms, of course, only a limited part the information on which the signal curves are based represent the process fluctuations.

With the solution presented here, on the one hand  illegal waveforms, e.g. B. sensor defects or Different process limits are exceeded Forms of waste curve technology recognized. In addition, with the underlying signal coding by the combination of discrete mathematical parameter determination and / or PCA coding determines these parameters (Principal Component Analysis) only a powerful one and meaningful formation of parameters possible. This one shown plausibility check of the involved Signal sources and the type of key figure determination are essential properties that only work safely of forecasting and control models, especially with enables cyclical production processes. this is also valid for the modeling with neuronal shown here Nets.

According to the invention, the stated object is achieved by the Method according to claim 1 solved.

In particular, the object is achieved in that Waveforms from several state variables of the process during one for the creation of a product relevant process section by means of adaptive envelopes be examined for permissible expression. This Envelopes are made from a waveform Test series with different process settings and  conditions, e.g. B. changed machine setting based of a test plan generated self-adapting. Out permissible signal curves of the measurement data time series of several sensors that detect the process state during one for the creation of a product relevant process section of a cyclical Production process process key figures determined.

These process key figures are determined by a combined coding using discrete calculation and / or by means of process-specific PCA coding (Principal Component Analysis). This means discreet here Calculate individual signal values, such as B. the maximum of steepest rise, their occurrence etc. is the signal expression at both discrete points of the Signal course, but mainly the whole Signal curve, coded by the PCA method in the form of Process key figures described. Process key figures thus obtained different signal sources can be repeated PCA coding (hierarchical PCA coding) without significant loss of information significantly further compress. From these key figures generated with the PCA the original waveform can be almost identical be reconstructed. So that's one too data-efficient documentation of the extensive Signal curves over a longer manufacturing period  realizable.

The envelopes and the PCA coding are with signal data an experimental phase adapted to the respective process, thereby dividing into allowed and inadmissible Waveforms is possible. Then in one Evaluation phase the quality characteristics of the associated Values of process indicators of the products created assigned. The process indicators that appear in this way are neural networks or other statistical forecasting Assigned to models as its input variables. Then be the neural networks trained so that they are due to of unskilled process indicators the quality characteristics for can forecast these products with sufficient accuracy the shape so that it can be combined with a high degree of certainty Production cycle generated set of process indicators Assign quality characteristics or classes, if possible exactly with those of the generated in this process cycle Product match.

In a classification phase, the generated Process key figures using the trained neural Networks quality characteristic values or classes predicts, displays and / or documents and / or for sorting the generated products (ok or n.i.o.) and / or to stop the machine if there are too many  Serve bad parts.

If the permissible quality limits are exceeded additional neural networks to regulate the process are used that are comparable to the quality networks are designed, but their input / output relation by the process indicators from the waveforms and the Setting values of the machine from the test data sets be formed.

The training of these quality control networks using the Process parameters and setting parameters of the process takes place comparable to the quality networks during the experimental phase described above.

The inertia behavior of the setting parameters of the Processes and their delayed impact on the Product quality taken into account.

The entire system structure and software components are modular design. With a flexible, task-oriented, efficient and process-specific Design of the system layout a simple system adaptation to different process configurations in a simple way respectively.  

By means of a system part for optimization in the case of impermissible Product quality becomes a new, better process attitude either automatically in the system known Parameter space or found interactively outside of it.

A special feature of the method according to the invention lies in that in addition to the adaptive envelopes for testing the Different forms of signal coding and modular neural networks for the Quality forecast and quality-based regulation used be based on actually running process operations and the values of Quality characteristics are trained so that - after the training phase - with the help of these procedures without Checking the products produced the production process like this that a statement can be monitored and regulated about whether the products are in order (OK) or not Order (n.i.o) are possible. The input parameters of the neural network are process indicators that are made up of Waveforms measured by sensors in the process will be determined in different ways. The sentence of process indicators is chosen such that it Contains information about product quality.

Another essential advantage of the invention The procedure is that in addition to the classic  Determination of parameters a signal coding (process-specific PCA) is used, on the one hand allows a high degree of data reduction, however a maximum informational content to occur Process fluctuations in the signal curves to the neural Transmits networks.

A subsequent regulation is based on the fact that a Activation logic a deviation of the quality values from a desired set of quality values, and in such a determination by means of one further neural network determined change values of the Setting values new setting values of the process are determined become. Another logic level becomes active, the checks whether slowly changing setting parameters already exist have reached the steady state. Only after is used in the event of a predicted, illegal Quality deviation caused another rule intervention.

In this development of the method according to the invention So the process is only regulated if the Product quality leaves the permitted, selectable range. Both the monitoring of the signal curves, the Coding, monitoring as well as regulation fully automatic and there are only a few products that are not in order (n.i.o). Occur disturbed  Signals, their interpretation is very simple as Sensor defect or unauthorized process states interpret.

The invention will now be described with reference to the accompanying Drawings explained in more detail. Show it:

Figure 1 is a schematic representation of the method.

FIG. 2 shows the relationships between the quality of the products produced with the various influencing variables;

FIG. 3 shows various measured with sensors and the system signal waveforms;

FIG. 4a exemplary, various pressure curves of a mold cavity for Kunststoffspritzgießprozesses io parts;

FIG. 4b exemplary, various pressure curves of a mold cavity for Kunststoffspritzgießprozesses nio parts;

5a is an upper and a lower envelope and a permissible waveform for a production cycle of the mold cavity pressure of a Kunststoffspritzgießprozesses.

FIG. 5b is an upper and a lower envelope and an unacceptable waveform for a production cycle of the mold cavity pressure of a Kunststoffspritzgießprozesses;

Figure 5c the basic approach to signal related to the center of gravity Hüllkurvenbestimmung and monitoring.

Fig. 6 is an example of output signals, data reduction and signal reconstruction using PCA encoding;

Fig. 7, the basic principles for formation parameters by means of PCA;

Fig. 8 is a neural network with a known in- / output relation for process monitoring, which is automatically configured depending on the Input / Output elements of the network;

Figure 9 is a modular, feature related network arrangement for a plurality of different quality characteristics that are predominantly realized here usually due to the modular system structure and applied.

10a is a representation of the functional Gesamtprozeß- and system relationships and the actions / reactions to be carried by the system.

10b shows the principle of a possible design of the neural networks for quality-based control.

Figure 11 shows the principle, simplified representation of the system efficiency, in a parameter space based on process settings.

FIG. 12 is the functional phase to the system adaptation to a production process;

FIG. 13 is a schematic for process optimization with the underlying system development, is performed with the inside of the parameter space of FIG 11, an automatic, outside of which an interactive optimization of the process setting. and

Fig. 14 the basic, simplified illustration of a quality function in a parameter space based on settings of process parameters.

Fig. 1 shows the sequence of monitoring and control of a cyclic process, which is represented by the rectangle "process". In the following, the injection molding of plastic products is considered as an exemplary application. However, the system development described below is generally applicable to the monitoring and control of cyclically organized processes. Automated processes such as die casting processes, powder injection molding processes, forming processes, welding processes and, in particular, electrical welding processes, such as different forms of resistance welding, spot welding, submerged arc welding, etc., cinch processes, riveting processes, painting processes, galvanic processes, for example for coating, screen printing processes, are to be mentioned here as examples. Stamping processes, cutting processes and more.

Quality monitoring should be derived from the dynamic processes of these processes, which are explained below using the example of plastic injection molding ( FIG. 1), which is displayed in the lower rectangle, without any examination or control of the product itself. From the process A process itself, which is causally responsible for whether a product is "ok" (OK) or "not OK" (NOK), is to be derived whether the manufactured product is OK or NOK. Furthermore, in the case of no ok or well before the limit between io and no ok is reached, a process control should be used which, if the process does not deviate from a certain target operating point, is to be traced back to it by deriving change values ΔEi (i = 1, 2 ..) enables the setting variables of the process to be corrected so that new setting variables Ei, new = Ei + ΔEi with ΔEi = f (Qki, Pki) are generated, which set the process back to an optimal operating point.

Exemplary signal profiles of the mold cavity pressure for several production cycles during injection molding for the two quality classes are shown in FIGS. 4a and 4b. Two angußnahe cavity pressure curves for io products (Fig. 4a) and nio for products (Fig. 4b) in FIGS. 4a and 4b are again as an example, is shown. The same representation is found in the upper right quadrant, however, only for the initial phase of a few seconds, i.e. extended in time. A comparison of the same mold pressure profiles pF (io) in FIG. 4a and pF (nio) in FIG. 4b clearly shows that typical characteristics of these pressure profiles can be correlated with the quality of the parts (io or nio). So you can see that the maximum of the pressure curves is reached after approx. 2.5 to 3.5 seconds with the OK parts, while it occurs at 2.5 to 5.0 seconds with the OK parts. It can also be seen from the time-stretched images in the upper right quadrant of the two images that the pressure increase (gradient) fluctuates much more steeply in the nio parts than in the io parts.

Taking into account such findings, which are only to be understood as examples in FIGS . 4a and 4b and which are different for each process and each process parameter and therefore have to be analyzed process-specifically, process parameters Pki can be derived from the signal curves, and these with the quality correlate the products produced, in the simplest case with the determination of OK or NOK for the parts.

Since the causal relationship between the individual measurable process signals (see Fig. 3) - not to be confused with the setting variables Ei on the machine or the machine and the process result, i.e. the quality of the products produced - is so complex that a closed representation - for example in the form of an analytical expression - without this knowledge, a model of this connection must be developed without this knowledge, which enables the quality of the product produced to be monitored and the process to be regulated accordingly.

In many processes and also in the injection molding of plastics, the quality of the product depends not only on the adjustable and set functions of the machine and the tool, which can be controlled by individual sizes Ei, but also on the ambient conditions and also on the mostly unknown current properties the material from which the product is to be made. These relationships should provide FIG. 2.

In a first step, the method according to the invention provides for the monitoring of the signal profiles by means of adaptive envelopes for each signal profile used. Fig. 5a shows a Hüllkurvenbeispiel with permissible waveform, Fig. 5b shows a Hüllkurvenbeispiel of a production cycle with invalid waveform. Two different tolerance functions allow the upper and lower envelope curves to be slightly exceeded or fallen short of, in order to also enable the processing of signals that are in principle impermissible in the vicinity of the known work area of the process. Adaptive here means that the envelopes are generated automatically from the upper and lower envelopes of the maximum and minimum signal values of each permitted signal course of the underlying experiments. An error in the generation of the signals (e.g. sensor, measuring amplifier, cable defect, etc.) or exceeding a learned working area of the process is thus reliably detected. In addition, an inadmissible coding of unknown and thus unauthorized system states with unsecured quality forecast and incorrect control intervention is prevented.

An even more powerful method of checking signals for permissible values is shown in FIG. 5c. While the tolerance intervals for each measuring point of a signal curve are determined independently of one another in the above-mentioned method, this second method ( FIG. 5c) takes into account correlations between the measuring points. If one imagines each permitted signal curve as a point in a vector space, an envelope ellipsoid is constructed based on the covariance matrix of permitted signal curves, which contains all the permitted process states. This construction makes it much easier to separate permitted process states from unauthorized states.

The next step is the derivation of process parameters Pki, which characterize the process sequence and its fluctuations themselves, which thus depend on the one hand on the influencing parameters shown in FIG. 2, but which can also change if the setting values remain the same.

Fig. 3 shows for the example of plastic injection molding some essential signal curves of some essential process variables that occur on an injection molding machine during the process to be monitored, two trigger signals, the hydraulic pressure pHyd, the nozzle pressure pD, the screw path sSch, the pressure curve pFn and pFf close to the gate and remote in a mold cavity, the nozzle temperature TDIR and the tool wall temperature TW. These signal profiles characterize the real course of a process section while a product, for example an injection molded part, is being manufactured. They are determined by suitable sensors. The sensors can of course be extended depending on the process, e.g. B. by the screw speed, or reduced. Any number of discrete process parameters such as maximum, gradient, integral, etc. can be calculated from these signal profiles using known mathematical methods for each signal profile.

Significantly more efficient in terms of parameter formation however, is the method of PCA coding for the Waveforms. The PCA (Principal Component Analysis) is a methodology of multivariate statistics to analyze the statistical properties of a multidimensional Random vector. Starting from a representative Random sample, i.e. a set of signal profiles If the PCA tries, the subspace can be found  minimal dimension of the sample space such that the Projections of the patterns onto the subspace a given Do not exceed the barrier of the reconstruction error. If a permissible error of, for example, 1% is specified it is typically possible in applications to use one To achieve data compression of 1: 100 or more.

It is also possible with the implemented approach that Reduction factor by specifying the dimension of the Define subspace and the occurring To have coding errors determined.

Methodologically, the PCA requires an estimate of the Sample focus of a sample and the solution of the the covariance matrix of the random vector Eigenvalue problems, the covariance matrix also is estimated using the sample. The algorithmic Implementation of the PCA is under recourse Standard software packages for linear problems possible. The The present inventive method uses the PCA Deriving process key figures from the signal profiles of the cyclical measurement data time series from the process state sensing sensors.

In FIG. 6, 7 output waveforms whose PCA figures and the reconstructed from the key figures also 7 waveforms are shown an inner mold cavity pressure course exemplary. Fig. 7 summarizes the essential function blocks and calculation variants of the PCA.

The derivation of process parameters instead of a direct one Processing the waveforms is used per Process section and measuring point per signal curve with very much less data and much less processor and Storage capacity get along like this otherwise would be required. Such a derivative of Process parameters Pki mainly using the PCA method allows the data to be reduced so that the subsequent processing for monitoring and control becomes manageable at all, and this with the PCA method minimal loss of information. Otherwise they would Signal curves in digitized form depending on Sampling frequency and other characteristics to one lead to a high number of data to be processed.

For data reduction, on the one hand, significant quantities of the signal curves can be determined and displayed, e.g. the amplitude of the maximum, the time at which the maximum occurs, the steepness of the signal curve at certain points, integrals of signal segments, etc. The derivation of process parameters is in the an aggregated view of FIG. 12 are shown in the second block from the left in the second and third rows on the evaluation and in the third block of the monitoring and control as "characteristic-quantity calculation". For example, between 20 and 80 process parameters can be obtained for the monitoring of an injection molding process, depending on the type of process.

A feature of the invention is that the modeling of the relationship between process parameters Pk1. . . Pki and the quality of the parts produced, given by quality features Qm and feature classes Qk1. . . Qkm, characterized in that a neural network NN1 is used. Such a neural network NN1 is shown as an example in FIG. 8. In the simplest case, Qk1 = io and Qk2 = nio for the monitored feature. An example can be designed in such a way that a certain measure of the product is monitored too long (Q1), too short (Q2) or satisfactorily (Q3) on the basis of three feature classes. In another example, the current value of a feature is required, for this the network is trained on specific values of the feature. By using several such neural networks, the neural network NN1 can be expanded in a modular manner for any number of quality features ( FIG. 9).

The theory of neural networks is known per se. Such a neural network may consist of several layers, in the example according to FIG. 8 three layers, via which a weighted connection of the process parameters Pk1. . . Pki with quality feature classes Qk1. . . Qkm as they are characteristic of the quality characteristics and thus the quality of a product produced.

In a first “evaluation” phase (cf. FIG. 12), the process parameters Pk1 derived from the signal profiles are interactively assigned certain quality features Qkj (Qk1... Qkm) which are determined by visual inspection and / or measurement of the products produced. In the example of injection molding, the z. B. mean that the thickness, the surface quality at various points, geometric dimensions and their deviations from a nominal size, the strength, etc. are determined for a product, which are assigned to quality parameters Qkj. This assignment is now recorded in a training data file and is then used in a later, separate process for "training" the neural network (s).

The training of a neural network takes place on the basis of the training data file, or a partial data set thereof, using one of the generally known learning algorithms such as B. instead of the back propagation algorithm. Fig. 8 shows a learning step for the change of the weights in the neural network NN1 in consideration of a learning pattern (error signal is calculated by differences in the learning phase, and is due to the difference between the network output and learning default).

Through training, for example in the example of injection molding based on 60 process sections (i.e. the Production of 60 products), it is possible that a neural network has a classification and Forecast certainty, d. H. correct allocation of Process parameters and quality features in one Achieved magnitude, sometimes in the accuracy range the measuring equipment usually used for checking of these forecast results.

In this way it is possible, by deriving process parameters from the signal curves of measurement data time series and training the neural network NN1 on the basis of an assignment of quality parameters in the neural network NN1, to create a process model that shows the causal relationship between the process parameters Pki and the quality parameters Qkj simulates and after completion of the evaluation phase ( Fig. 12) and the training phase ( Fig. 12) based on the respective process parameters, a statement can be made as to whether the quality of a product corresponds to a certain set of quality forecast values (e.g. OK or 100 mm) or not (e.g. nio or 100.9 mm) without testing the product.

For each used to determine the quality Quality feature (i.e. the appearance of the surface, Filling the shape, dimensions, etc.) can also be done in each case use a separate neural network to one modular structure and hierarchical model structures possible close.

It is essential that the key figure determination and the Forecast of the networks only with the envelope Method allowed signal courses takes place, and thus a unauthorized system result with impermissible signal curve is prevented.

In the training phase, e.g. B. by hierarchical Arrangement of networks first monitoring attributive and then creating variable features. In the Training phase of the neural network thus takes place attributive or quantitative assignment of certain Quality parameters for process parameters of a certain one  Process cycle or section.

Before training the neural network, one can Database an assignment of the process parameters to the Quality parameters of a product take place, i.e. the Generation of a training vector. Multiple training vectors result in a training set. The training sets are in Training data file saved.

The entire system is shown again in context in FIG. 10a. The sensor signals are derived from the sensors in the machine. These then generate the corresponding signal profiles, which reach the central data processing unit and are further processed there in the manner described.

In Fig. 12, the operation of Hüllkurvenbestimmung is shown in the first line of the evaluation phase. The interactive assessment in the next two lines creates the training data file for the quality and control networks. The process parameters are derived and then assigned to the quality characteristics or the setting values. In the fourth and fifth lines of FIG. 12, the training sequence is shown using the neural networks NN1 and NN2.

In the sixth and seventh line of FIG. 12, the monitoring and control phase is then shown, in which, after the training phase has ended, the quality is predicted directly from the signal profiles and sorted into OK parts and NOK parts with simultaneous display and documentation becomes. The network NN2 only becomes active if an unauthorized quality status is reported by the NN1 ( FIG. 1).

The neural network NN1 ( FIG. 8, FIG. 9) should be trained in such a way that it is based on sets of process codes, symbolized by the points P1, P2,. . . Pj, works in a room, cf. 11 , in which both the quality characteristics for IO and NOK classes, as well as a representative distribution of variable characteristic values, occur, and in which a good forecast value is achieved for key figure values that were not generated during the training phase. One speaks here of a generalization capability of the neural network if a satisfactory quality forecast is achieved for sets of key figures that have not been trained.

So the network should not only be able to z from a certain trained set of process parameters. B. to arrive at a certain result (OK or NOK). It should also consist of a set of process parameters for which it has not yet been trained and which is within a defined io / n io range, such as, for. B. the cuboid P1. . . P8 in Fig. 11, is to provide a reliable result.

Another section of the process is at predicted no-go parts to a control function activate. In the event of deviation from the Quality parameters Qk from the permitted values Change the process parameters so that the desired values of the quality characteristics again achieved become.

This is done by means of a further neural network NN2, which is shown in Fig. 1 and Fig. 12. The process parameters Pk1. . . Pki entered. You will find certain setting values Ek1. . . Ekn assigned, in a training phase such that the error signal changes the internal network weighting of the individual influencing variables in a manner comparable to the training of the quality networks. At the end of the training phase, the neural network contains a model which, in the event of deviations from the working point of the process ( FIG. 11), defined by process parameters Pki, changes the setting values ΔEk1 from a desired target working point PA. . . ΔEkn indicates that the process is "brought back" to the currently best working point PA. The path of the deviation is shown in FIG. 11 as path 1 , the return path as path 2 .

This is also done, it should be emphasized again, with the trained network as described above without any specific testing of the products produced. This control phase and its specification is shown in lines 12, 3, 5 and 7 in FIG . The labeling is self-explanatory based on the explanations given above.

To FIGS. 1, 10a and 12, it should be added that, before a change of settings is initiated, an activation logic f (QK1... Qkm) which is provided in the second neural network NN2, from the first neural network NN1 forth must be activated. The activation logic can e.g. B. - in the simpler case - be designed so that it initiates a change in the setting values when a product is not ok.

On the other hand, the regulation should not only start when a product is not ok. One would like to use the control to change the process to an optimal operating point within the io in the event of changes that are still within an io range (i.e. in a smaller area around the point PA enclosed in the cuboid according to FIG. 11) Area. The activation logic can thus be activated even if the operating point PA has emigrated to a point PA 'which, although it itself allows the production of OK parts, is already close to a NOK area.

Fig. 13 is a representation of a basic solution of an evolution-based control and optimization. This component provides that if the control as described with NN2 is unsuccessful, or the detected process parameters Pki the learned area of the envelopes (FIGS . 5a, 5b, 5c), ie the detected area (the cuboid in FIG. 11) leave, the setting values are changed step by step ( FIG. 14), and then the resulting quality parameters are automatically determined or checked in the interactive case, i.e. it is examined whether the quality features move in the direction of an improvement or a deterioration in the overall quality or not, and thus move to a point "maximum" at which the overall quality has a maximum, ie is optimal.

The evolution course starting from the "start" in FIG. 14 can also show that a change in the setting parameters does not improve the overall quality, but rather worsens it. An evolution strategy then provides that, starting from a previous point, the system looks for a new attitude in a different direction, in which an improvement in quality can be determined again on the basis of the process key figures.

Claims (8)

1. A method for monitoring a production process, in which signal profiles are recorded at several points in the production process and at least one quality statement is derived for the products produced, characterized in that

• a) that the signal profiles of a process section relevant for the production of the products are examined for their admissibility with the aid of envelope curves,
• b) that process indicators (Pk1... Pkn) are determined from the permissible signal curves of the process section relevant for the production of the products in such a way that changes in the process indicators (Pk1... Pkn) with changes in quality characteristics (Qk1... Qkm) of the correlate manufactured products,
• c) that in an evaluation phase the process indicators (Pk1... Pkn) occurring during the production of the products are assigned to the associated quality characteristics (Qk1... Qkm) of the products produced,
• d) that the process key figures (Pk1... kn) obtained in this way are fed to a neural network (NN1) as its input variables,
• e) that the neural network (NN1) is trained in a training phase in such a way that it assigns with high certainty the quality indicators (Pk1... Pkn) generated in a later process section, which are equal to those of the in are the product of the process section, and
• f) that, in a monitoring phase, quality characteristic values and classes are derived from the process indicators (Pk1... Pkn) with the help of the trained neural network (NN1), which are displayed and / or documented and / or for sorting the products produced (OK or NOK) ) and / or used to regulate the production process.

2. The method according to claim 1, characterized in that the process key figures (Pk1 ... Pkn) either by a discrete calculation or by PCA coding (PCA = Principal Component Analysis) or by a Combination of which can be determined. 3. The method according to claim 1 or claim 2, characterized characterized in that the envelopes and / or the PCA Encodings with the help of waveforms Trial phase adapted to the production process  in the test phase, the settings of the production process can be changed in a targeted manner. 4. The method according to any one of the preceding claims, characterized in

• g) that the process indicators (Pk1... Pkn) are fed to a further neural network (NN2) as input variables,
• h) that in an evaluation phase the associated process parameters (Pk1... Pkn) are assigned the associated, varied setting parameters (Ei) of the production process,
• i) that the further network (NN2) is trained in a training phase,
• j) that an activation logic detects a deviation of the quality characteristics (Qk1... Qkm) from the desired quality characteristics, and
• k) that new setting variables (Ei, new) of the production processes are then generated by means of the change values (ΔEi) of the setting variables (Ei) determined by the further network (NN2).

5. The method according to claim 4, characterized in

• l) that the setting variables (Ei) are changed and the resulting process indicators (Pk1... Pkn) are determined,
• (m) it is checked whether the quality characteristics (Qk1... Qkm) are moving in the direction of an improvement in quality, and
• n) that, if this is the case, the production process is continued with the new process key figures (Pk1... PKn) as the working point (PA).

6. The method according to claim 5, characterized in

• o) that, if this is not the case, a previous operating point is selected and the setting variables (Ei) are changed in another direction by means of an optimization strategy.

7. The method according to any one of the preceding claims, characterized in that the inertial behavior of the Setting parameters (egg) of the production process in the Terms of its impact on the quality of the generated products is taken into account. 8. The method according to any one of the preceding claims, characterized in that the individual steps of Process as module-like software components be formed.