DE19518804A1 - Process control - Google Patents

Abstract

To monitor a prodn. process, sensors generate signals at various prodn. stages for evaluation of the quality of the whole product. Process indices (Pk1-Pki) are developed from the signal runs during the process stages, to be correlated with given quality characteristics (Qk1-Qkm). The process indices are assembled in a neuron network (NN1) as input values. In an evaluation phase of the neuron netwrok (NN1) the quality characteristics (Qk1-Qkm) are arranged with the process indices to give output values. The neuron network (NN1) has a training phase to ensure with high reliability that quality characteristic classifications (Qk1-Qkm) are assigned for the process indices sets for each prodn. stage, which match the prodn. stage for the finished product. In a monitoring phase, the process indices are converted to quality characteristics or classes by the trained neuron network (NN1) to be displayed and/or printed and/or for sorting the finished product either in or out of order and/or to control the process.

Description

The invention relates to a method for monitoring a Production process in which a measurement of signal curves at several points in the process by sensors and one Evaluation of the products produced takes place.

Production monitoring with the help of statistical Process control (SPC) is essentially based on today random product testing of finished products. The Checking the products, be it constantly, or just as Sampling, however, is very expensive. A regulation of a Process by changing the setting values is because of complex relationships are often very difficult and happen  therefore often without firmly assignable rules, so that they are not suitable for automation.

The object of the invention is therefore a method of to create the kind mentioned that a surveillance and the associated documentation without checking the products themselves enables.

DE 42 09 746 A1 describes a method for optimization of a technical neuro-fuzzy system in which the data with which the neuro-fuzzy controller is trained, must be generated by simulation. This simulation is but a complex process and also requires knowledge of Causal relationship between process flow and training the data required by the neuro-fuzzy controller.

DE 44 16 317 A1 describes one method and one Control device for controlling a material processing Process became known, taking into account process relevant material properties certain process parameters be pre-calculated to preset the system to serve. This calculation also makes knowledge causal Relationships between the factors influencing the process and ahead of the process parameters. 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 solution to the problem must be the difficulty overcome that the causal relationship between parameters the process, the process parameters, the quality characteristics and / or the feature classes of the products produced is known and that those characteristic of a process Data result in a wealth of data that runs parallel to the process flow cannot be recorded and processed. For the sake of explanation added. A quality feature is e.g. B. the length of a Product. At least two classes of characteristics (OK or n.i.o.) provided. But it can also be more, e.g. B. 10-11 mm, 11-12 mm,. . . , 19-20 mm, etc.

According to the invention the stated object is achieved in that

• (a) from the waveforms during one for generation of a product relevant process section Process key figures (Pk1... Pkn) are determined that such are determined that changes in the process key figures with a change in certain quality characteristics (Qk1... Qkm) the products correlate;
• (b) the one that occurs in the production of products Process indicators as a neural network (NN1) whose input variables are assigned;
• (c) the neural network (NN1) in an evaluation phase output values assigned to the process indicators of quality characteristics (Qk1... Qkm) are assigned, which are found in the production of the products;  
• (d) the neural network (NN1) in a training phase is trained in such a way that it is safe for everyone in a process section generated set of Process key figures quality feature classes (Qk1... Qkm) assigns the same as that in the process section produced product, and
• (e) from the process indicators in a monitoring phase Help of the trained neural network (NN1) Quality characteristic values or classes are derived, which are displayed and / or documented and / or for Sort the generated products (OK or NOK) and / or used to control the process.

A special feature of the method according to the invention lies in that neural networks are used that are based on actual processes and the process interactively determined values of quality features trained so that afterwards - after the training phase - with Help these procedures without checking the products produced by the Process can be monitored and / or regulated so that a Statement about whether the products are OK or not in order (n.i.o) are possible. The input parameters of this neural network are process indicators that are made up of Waveforms measured by sensors in the process will be determined in such a way that their significance for the Quality of the products with simultaneous data reduction preserved.  

The same idea can then be improved that

• (f) the process indicators (Pk1 ... Pkn) further a second neural network (NN2) supplied as input variables become;
• (g) in an assessment phase of a change in Process key figures specific change values (ΔEi) for the Setting variables (Ei) of the process can be assigned to the a return of the process key figures to the Effect output values;
• (h) that the second neural network (NN2) in one Training phase is trained; and that
• (i) an activation logic a deviation of the quality values (Qk1... Qkm) from a desired set of Determines quality values and at such Determination by means of the second neural network (NN2) determined change values (ΔEi) of the setting variables defines new process parameters (egg, new).

In this development of the method according to the invention So the process is regulated so that both Monitoring and control run fully automatically and only few products that are not in order (not ok) are produced become.  

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 by sensors waveforms;

Fig. 4 different mold pressure profiles for IO parts;

Fig. 5 different mold pressure profiles for nio parts;

Fig. 6, the neural network NN1 for monitoring the process;

Fig. 7 is a further illustration of the method;

Fig. 8, the various processes for the evaluation phase, the training phase, monitoring phase and a regulation phase;

Figure 9 is a diagram for generalization ability of the neural networks.

Fig. 10, the neural network NN2 for controlling the process;

Fig. 11 shows the dependence of the overall quality of two process parameters for explaining an evolution-based optimization strategy.

Fig. 1 shows the sequence of monitoring and controlling a process, which is represented by the rectangle "process". Injection molding of plastic is considered as an example below. A quality monitoring should be derived from the course of this process, which is displayed in the rectangle "quality monitoring", and without examination or control of the product itself. From the course of a process itself, which is causal for whether a product "in Order "(= OK) or" not OK "(= NOK), it should be deduced whether the manufactured product is OK or NOK. Furthermore, a process control should also be used which, if the process does not deviate from a certain desired working point, enables traceability to this by deriving change variables ΔEi (i = 1,2,...), By means of which the setting variables of the process are such can be corrected that new setting variables Ei, new = Ei + ΔEi are generated, which set the process back to the optimal working point.

Because the causal relationship between the individual measurable Process variables - not to be confused with the setting variables Ei on the machine or machines - and the process result, So the quality of the products produced is so complex that a closed representation - approximately in the form of a Algorithm - is not possible without this knowledge A model of this relationship can be developed Monitoring the quality of the product produced and a process control geared to this.

When injection molding plastics, the quality of the product depends not only on the adjustable and set function of the machine and the tool, which can be controlled by individual sizes Ei, but also on the ambient conditions (temperature, air pressure, air humidity, etc.) and the material from. This is shown schematically in FIG. 2. The properties of the material can vary not only from batch to batch in the case of intermediate products manufactured industrially according to specific specifications; they are, for example in the increasing processing of recyclate, often very different even in the short term, without a conclusive connection between these properties and the quality of the products produced being able to be represented, even without the changes in the material properties even being known.

In a first step, the method according to the invention provides for the derivation of process parameters Pki which characterize the process sequence itself and which, on the one hand, are dependent on the influencing variables 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, such as the hydraulic pressure pMasch, the screw path sSchn, a certain distribution pressure pV and a pressure curve pF in a mold cavity. Of course, such quantities can be measured at several points, i.e. the internal tool pressure pF at several points, e.g. B. close to the gate and distant from the gate. 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. to measure nozzle pressure, nozzle temperature, or reduced.

In Figs. 4 and 5 show two angußnahe cavity pressure curves for io products are, again as an example, (Fig. 4) and nio products (FIG. 5), respectively. 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 cavity pressure curves pF1 (io), pF2 (io) in Fig. 4 and pF1 (nio), pF2 (nio) in Fig. 5 clearly shows that typical characteristics of these pressure curves with the quality of the parts (io or nio) are correlated. 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.

Taking into account such findings, which are only to be understood as examples in FIGS . 3 and 4 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 profiles, and these with the quality correlate the products created, in the simplest case with the determination io or nio.

The derivation of process parameters instead of a direct one Processing the waveforms is used per Process section and measuring point per signal waveform with a lot less data and a lot less processor and Storage capacity to get along as otherwise required would. Such a derivation of process parameters Pki allows  to reduce the data so that the subsequent processing for monitoring and control becomes manageable at all. Otherwise the waveforms would in digitized form depending on the sampling frequency and others Characteristics of an excessive number of to be processed Keep data.

To reduce the data, significant sizes of the Waveforms are determined and displayed, for. B. the Amplitude of the maximum, the time of occurrence of the Maximum, the steepness of the signal curve at certain Digits, integrals of signal segments, etc.

The derivation of process parameters is shown in the summary representation according to FIG. 8 in lines 1, 2, 5 and 6 in the second block from the left as a "process parameter calculation". For example, 20 process parameters can be provided for monitoring an injection molding process (i = 20).

A feature of the invention is to model the relationship between process parameters Pk1. . . Pki and the quality of the parts produced by quality features Qm and feature classes Qk1. . . . Qkm is to be marked to use a neural network NN1. Such a neural network NN1 is shown as an example in FIG. 6. In the simplest case (m = 2) Qk1 = 1.0. and Qk2 = nio for the monitored feature. An example of m = 3 is e.g. B. that a certain measure of the product is too long (Q1), too short (Q2) or satisfactory (Q3). By using several such neural networks, the neural network NN1 can be modularly expanded for any number of quality features.

The theory of neural networks is known per se. Such a neural network may consist of several layers, in the example according to FIG. 6 three layers, via which a weighted linkage 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. the 1st line in FIG. 8), the process parameters Pki derived from the signal profiles are assigned interactively certain quality features Qkj (Qk1... Qkm), which are assessed by visual assessment and / or measurement of the generated products can be determined. 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 either recorded in a "training data file" and then serves in a later - separate - process for "training" the neural network.

The training of a neural network takes place on the basis of the training data file or a partial data record thereof using one of the generally known learning algorithms such as e.g. B. the "back propagation" algorithm instead. FIG. 6 shows the change in the weights in the neural network NN1 on the basis of an error signal which represents deviations in the learning phase.

Through a "training" that - in the example of the Injection Molding - using 60 process stages (i.e. the Production of 60 products) can be effected, that a neural network provides classification security, d. H. correct allocation of process parameters and Quality features, on the order of more than 95% reached.

So in this way it is possible to go through

• - Deriving process parameters from the Waveforms during a current Process section and
• - Training of the neural network NN1 based on an assignment of quality parameters in the neural network NN1 to create a process model that simulates the causal relationship between the process parameters Pki and the quality parameters Qkj and after completion of the evaluation phase ( Fig. 8, line 1 ) and the training phase ( Fig. 8, line 3 ) based on the respective process parameters allows a statement as to whether the quality of a product corresponds to a certain set of quality classes (= OK) or not (= NOK), and that without checking the product.

For everything used to determine the quality Quality feature (i.e.: appearance of the surface, filling the shape, dimensions, etc.) you can also one separate neural network to use a modular Structure and hierarchical model structures possible.

In the training phase, e.g. B. by hierarchical Arrangement of networks first monitoring attributive Characteristics are generated and if these i.o. are a quantitative feature monitoring can be triggered. In the Training phase of the neural network thus takes place Attributive assignment of certain process parameters to one specific process operation 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 the Training data file saved.

In Fig. 7 the entire system is shown again in context. The sensor signals (= process data) are derived from the sensors in the machine. These then trigger the corresponding signal curves, which reach the central data processing unit and are further processed there in the manner described.

In Fig. 8 the process of interactive evaluation is shown in the first line. The process parameters are derived and then assigned to the quality characteristic. In the third line of FIG. 8, the training sequence is shown using the first neural network NN1.

The fifth line of FIG. 8 then shows the monitoring phase in which - after the training phase has been completed - sorting into OK parts and NOK parts is carried out directly from the signal profiles, with simultaneous display and documentation.

The neural network NN1 should be trained in this way - cf. Fig. 9 - that it is based on measured process key figure sets, symbolized by the points P1, P2,. . . P8, defines a space in which the quality features are fulfilled for both io and nio classes, and also for points Px that were not recorded during the training phase. One speaks here of a generalization or interpolation ability of the neural network.

The network should not only be able to recognize whether a certain trained set of process parameters for a specific Result (ok or not ok) leads. It should also be a sentence Recognize process parameters as such, for which there is not yet has been trained and that outside of a defined i.o./n.i.o. area, such as B. the cuboid P1. . . P8. A simple example should clarify this: If you have a neural network trains a cat from a dog too distinguish and then show him a giraffe, so it should don't try to decide whether the giraffe is a dog or is a cat, but it is supposed to recognize that it is the giraffe does not know. This is achieved e.g. B. with the help of a so-called. "RBF" (= Radial Basis Function) network. Not something Known or not realizable is considered to be faulty displayed. The display can possibly be used for a post-training session of neural network.

Another section of the process can, as already mentioned, consist in carrying out a regulation. These If the quality parameters Qk deviate from the desired values the process parameters  change that the desired values of the quality characteristics be reached again.

This takes place with the aid of a further neural network NN2, which is shown in FIG. 10. The process parameters Pk1. . . Pki entered. They become specific values for changing the setting values ΔEk1. . . Assigned to ΔEkn, initially arbitrarily in value and then again during a training phase in such a way that the error signal changes the internal network weighting of the individual influencing variables. At the end of the training phase, the neural network contains a model which, in the event of deviations (PA ′) in the working point of the process, defined by process parameters Pki, from a desired nominal working point PA changes the setting values ΔEk1. . . ΔEkn indicates that the process is brought back to the operating point PA. The path of the deviation is shown in FIG. 9 as path 1, the return path as path 2.

Again, this should be emphasized - after the training phase - without any specific examination of the products produced. This control phase (including the preliminary assessment and training) is shown in lines 2, 4 and 6 in FIG. 8. The labeling is self-explanatory based on the explanations given above.

In addition to FIG. 10, it must be added that before a change in the setting values is initiated, an activation logic f (Qk1... Qkm), which is provided in the second neural network NN2, must be activated from the first neural network NN1; the quality features derived from the process key figures Pki reach the activation logic from the neural network NN1. 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 be used until 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 case 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. 9). Area. So you can have the activation logic addressed even if the operating point PA has emigrated to a point PA ', which is still OK itself, but is close to an unsuccessful area.

Fig. 11 is the view of an evolution-based control. With this component it is provided that, when the detected process parameters Pki leave the detected area (the cuboid in FIG. 9), the setting values are changed step by step and then the resulting process parameters are determined and checked whether the quality features move in the direction of a Move or not improve or deteriorate the overall quality and so move to a "Max" point where the overall quality is a maximum.

The evolution course starting from the "start" in FIG. 11 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 tries a new attitude in another direction, in which an improvement in quality can be determined again on the basis of the process key figures.

Claims (4)

1. A method for monitoring a production process in which a measurement of signal curves at several points in the process by sensors and an evaluation of the products produced takes place, characterized in that

• (a) Process key figures (Pk1... Pkn) are determined from the signal profiles during a process section relevant for the production of a product, which are determined such that changes in the process key figures correlate with a change in certain quality characteristics (Qk1... Qkm) of the products ;
• (b) the process indicators occurring during the generation of products are assigned to a neural network (NN1) as its input variables;
• (c) values of quality characteristics (Qk1... Qkm) are assigned to the neural network (NN1) in an evaluation phase as output variables assigned to the process indicators, which are determined during the production of the products;
• (d) in a training phase the neural network (NN1) is trained in such a way that it assigns with a high degree of certainty quality feature classes (Qk1... Qkm) to each set of process indicators generated in a process section, which are equal to those of the product generated in the process section, and
• (e) in a monitoring phase, quality characteristic values or classes are derived from the process key figures 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 for regulating the Process.

2. The method according to claim 1, characterized in that (f) the process indicators (Pk1... Pkn) are also fed to a second neural network (NN2) as input variables;

• (g) in an evaluation phase of a change in the process key figures, specific change values (ΔEi) for the setting variables (Ei) of the process are assigned, which bring the process key figures back to the initial values;
• (h) that the second neural network (NN2) is trained in a training phase; and that
• (i) an activation logic determines a deviation of the quality values (Qk1... Qkm) from a desired set of quality values and, in the event of such a determination, using the change values (ΔEi) of the setting variables determined by the second neural network (NN2), new setting variables (Ei, new) ) of the process.

3. The method according to claim 2, characterized in that for carrying out an evolution-based optimization of the control

• (j) the process indicators (Pki) are changed such that they leave a trained area (P1... P8),
• (k) the setting values (Ei) are changed and the resulting process indicators are determined,
• (l) A check is made as to whether the values of the quality characteristics (Qkm) are moving in the direction of an improvement in the overall quality and, if this is the case, the process is continued with the new process key figures as the operating point (PA).

4. The method according to claim 3, characterized in that

• (m) if the test carried out after step (1) results in a deterioration in the overall quality, the system goes back to a previous operating point and looks for the further changes in a different direction.