Method for operating production plant, computer program and data carrier

Abstract

The invention relates to a method for operating a production plant which produces a component from a semi-finished product in a process, in which a parameter set is determined by means of a model and the stability of the model is determined by means of the following steps: detecting at least one error value of a semi-finished product property and / or a state variable and / or a process parameter; (S1)-forming a data set for the at least one semi-finished product, which data set comprises, for a semi-finished product property, a semi-finished product value having an associated error value, the state variable having an associated error value, and the process parameter having an associated error value; (S2) selecting at least one control value (10) in an interval (12) formed by the respective error value; (S3) forming a control set comprising the corresponding data by means of the at least one selected control value (10) by means of the model; (S4)-forming a difference between the data set and the control set; (S5) and-deriving a threshold value from the difference. And (S6). The invention further relates to a computer program and to a data carrier.

Description

Technical Field

[0001] The invention relates to a method for operating a production system that produces components from a semi-finished product having at least one semi-finished product characteristic in a process characterized by a parameter set including at least one process parameter. The invention also relates to a computer program according to claim 10 and an electronically readable data carrier according to claim 11. Background Art

[0002] Production plants are frequently used in motor vehicle manufacturing, such as passenger car manufacturing. Such production plants can, for example, be part of a stamping press for producing body components. The production of body components in a stamping press is divided into several process steps. First, blanks are cut from a coil on a coil production line. This allows for the formation of a traceable stack of blanks, which are typically stored temporarily before processing in a press line. In the press line, for example, deep drawing of flat sheet metal, formed from the cut blanks, can occur. Further process steps, such as trimming and / or shaping, may follow the deep drawing.

[0003] Semi-finished products or thin sheets processed in a press typically exhibit fluctuations in their properties, particularly those of the semi-finished product. For example, sheet thickness, lubricant content, roughness, and elastic-plastic material properties may fluctuate. Depending on the nature of these fluctuations, process parameters of the production process may need to be adapted in order to achieve the desired quality of the produced components, i.e., component quality. In most presses, the adaptation of process parameters associated with the downtime of the system is usually carried out by the system operator based on experience. Depending on the duration of determining a suitable combination of process parameters, especially due to downtimes in the production process, significant costs may be incurred.

[0004] In order to avoid or reduce the aforementioned downtimes while searching for process parameters for problem solving, some approaches have introduced process control using algorithms. These algorithms can take into account the semi-finished product characteristics of the blank being processed, the state of the blank, such as its temperature, the state of the mold, such as its surface temperature, and the state of the press, such as the temperature of the hydraulic fluid in the drawing mat, and generate recommendations for selecting a suitable combination of process parameters.

[0005] Problems that may arise here are that, for example, measurement errors when determining the semi-finished product properties and / or states can change the predictions of the algorithm and the algorithm can provide erroneous results or recommendations due to instabilities, especially when there are nonlinear relationships between the individual states or semi-finished product properties. Summary of the Invention

[0006] The object of the present invention is therefore to provide a method, a computer program and a data carrier by means of which a production plant can be operated particularly advantageously in such a way that instabilities in the determination of parameters used for the operation of the process of the production plant can be advantageously detected.

[0007] This object is achieved according to the invention by the subject matter of the independent claims. Advantageous embodiments and further developments of the invention are given in the dependent claims as well as in the description and the drawings.

[0008] A first aspect of the present invention relates to a method for operating a production plant that produces components from a semi-finished product having at least one semi-finished product property in a process characterized by a parameter set including at least one process parameter, wherein the set of process parameters is predetermined and / or the component quality is determined by means of a trained model using the at least one semi-finished product property, and the stability of the trained model is determined using the following steps:

[0009] In a first step, at least one error value for at least one semi-finished product property and / or at least one state variable and / or at least one process parameter and / or the component quality or mass is detected. In a second step, a corresponding data set is generated for the at least one semi-finished product, which data set includes, for the at least one semi-finished product property, a semi-finished product value and an associated error value, and / or the state variable and an associated error value, and / or the at least one process parameter with an associated error value, and / or the component quality with an associated error value. In a third step, at least one control value is selected within an interval formed by the corresponding error value around the semi-finished product value and / or the state variable and / or the process parameter and / or the component quality. In a fourth step, a control set containing corresponding data is generated using the trained model with the aid of the at least one selected control value. In a fifth step, a difference between the corresponding data set and the control set is generated. Finally, in a sixth step, a threshold value is derived from the generated difference.

[0010] The production facility may, for example, be a press line for automobile manufacturing. The semi-finished product is, for example, a blank, which may have, as the at least one semi-finished product property, sheet thickness, lubricant content, or roughness. The semi-finished product value specifies the magnitude of the semi-finished product property and its actual value, such as sheet thickness in millimeters. The at least one state variable may, for example, describe the current temperature of the semi-finished product, the temperature of the production facility's mold, the air humidity, or the like. If the production facility includes a press, the process parameters may, for example, describe press parameters that characterize, for example, the drawing pad force, the kinematics of the impact motion, the position of the sheet support, and / or the application of additional lubricant to the semi-finished product. Other process parameters may be mold parameters, such as the determination of the stretching aid, the gas filling of the gas spring, and / or the position of the guide. The process parameters can thus characterize the adjustment options of the mold. Furthermore, the process parameters can also be used to adjust the straightening device on the coil production line.

[0011] The associated data for the control set include the values ​​of the at least one semi-finished product value, the at least one process parameter, the at least one state variable, and / or the component quality determined by the method for the selected control value(s). The selection can be performed, for example, using an optimization method and / or by random selection, particularly with a uniform distribution across the control value space. By prescribing a desired number of control values, for example by a user, the time required to create the control set can be influenced.

[0012] The component is, for example, a deep-drawn sheet metal component, which is formed from a semi-finished product by extrusion or deep drawing using a production facility, in particular, including a press. The trained model, in particular, includes a self-learning algorithm and / or a neural network. "Trained" means that the model is trained using training data to predetermine process parameters that achieve a desired component quality when the component is produced using the production facility. Additionally or alternatively, the model can be designed such that, given input of semi-finished product properties and a parameter set representing a set of process parameters, the component quality can be estimated using the model.

[0013] The method according to the present invention now offers the possibility of analyzing the stability of predictions of trained or self-learning models and, therefore, machine learning models. The method is suitable for evaluating the at least one process parameter and, in addition or alternatively, predicting component quality. Thus, in a first case, the control value can, for example, depend on the at least one semi-finished product value and the at least one state variable. In a second case, the control value can depend on the at least one semi-finished product value, the at least one state variable, and the at least one process parameter.

[0014] Furthermore, the method according to the present invention is suitable for analyzing the stability of the trained or self-learning model both before startup and during ongoing operation. Both use cases require or involve the determination or evaluation of measurement errors and, therefore, error values ​​for all measured variables used as input data for the model. Thus, the detection in the first step of the method involves, in particular, the measurement and / or storage and / or determination or retrieval of corresponding error values, which can describe, in particular, the measurement error or uncertainty of the at least one semi-finished product characteristic, the state variable, and / or the at least one process parameter of the parameter set, or the component quality. Typically, the process parameters are setpoints, which are set or predetermined at the production facility. The actual values ​​of the process parameters may deviate from these setpoints. These values ​​may be referred to as measured process parameters. Uncertainties are determined only for the measured process parameters, and these uncertainties can form the basis for the associated error values. For non-measured process parameters, uncertainty cannot be determined in terms of measurement accuracy. However, an evaluation can be performed for the detection of the associated error values, for example, based on assumed tolerances in the production facility. At least one of the corresponding error values ​​can also be predetermined by component quality requirements, or the quality of the prediction can be derived from quality requirements. Furthermore, it is conceivable to evaluate the corresponding error values ​​with the aid of a standard.

[0015] The present invention is based on the recognition that when suggesting process parameters by means of a self-learning model, measurement errors or error values, for example of sensors used for determining properties and / or states of semi-finished products, should be taken into account.

[0016] For example, it is possible to assume that a first blank or a first semi-finished product has semi-finished product properties. A second blank has semi-finished product properties that differ from the first blank, wherein these semi-finished product properties may differ only slightly within measurement tolerances or measurement accuracy and thus lie within error values.

[0017] As already mentioned, the trained model or self-learning algorithm can be used, on the one hand, to output recommendations for process parameters or to output a prediction of the quality of the produced component. For the first case, it is now desirable for the algorithm or model to propose a parameter set that includes the at least one process parameter for both blanks or semi-finished products. In this case, assuming identical values ​​for the state variables, the proposed process parameters should also be identical or should differ only slightly from the corresponding other recommendations for the respective semi-finished products or blanks, since these process parameters correspond to identical values ​​for the state variables or semi-finished product properties within the limits of measurement accuracy. If the predictions or recommendations for the process parameter set by the model now show significant differences in the values ​​of the process parameters, this may indicate that the model is not sufficiently stable to propose reliable values ​​for the at least one process parameter.

[0018] In particular, in machine learning models or trained models that take into account nonlinear relationships between semi-finished product characteristic state variables and process parameters, instabilities cannot be ruled out. When instabilities occur, the model's predictions become unreliable. Possible causes of such instabilities include insufficient measurement accuracy of the input variables, local discontinuities in the model, and / or so-called overfitting.

[0019] Crucially, the instability can be assessed for points in the space of input data that are not part of the data used for training, validating, and testing the model. The method can also be used for the training phase as well as for the data used to train the model.

[0020] Furthermore, the method can also be used in models based on linear regression models, for example, because a large number of parameters, such as those based on semi-finished product properties and conditions, can lead to stability problems due to error propagation. For example, the linear expansion of a material depends on the ambient temperature, which can constitute a state variable. If the ambient temperature is not sufficiently determined, the measured thickness can be incorrectly determined based on the temperature-dependent linear expansion, for example. Therefore, the method can provide an indication of potentially insufficient measurement accuracy in many variables used in the model, such as semi-finished product properties, state variables, and process parameters.

[0021] Instabilities can arise in machine learning models when predicting the quality of produced components. If, in the example described above, the quality prediction for a first blank based on values ​​characterizing the semi-finished product differs significantly from the prediction for a second blank in the aforementioned example, this can also indicate model instability. However, in this case, the cause can lie in both the model used and insufficient measurement accuracy of the semi-finished product's properties or conditions.

[0022] It is crucial to identify such instabilities in order to ensure satisfactory operation of the production machine or the model used therein for recommending process parameters and / or component quality. The advantage of the method according to the invention described is that such instabilities can be discovered in an advantageous manner.

[0023] In an advantageous embodiment of the present invention, if the difference is less than a threshold value, the model is used for operating the production plant and / or the at least one component is produced using a set of process parameters determined by the model. In other words, if it is concluded that no instability is expected based on the model, the model used in the previous method is continued to be used. Additionally or alternatively, process parameters or parameter sets generated by the model for operating the plant can be used to produce components using the production plant. This results in the advantage that components produced using the method have particularly high component quality.

[0024] In another advantageous embodiment of the present invention, the data set is formed as a hypercube, the center point of which is formed by the at least one semi-finished product value and / or the at least one state variable and / or the at least one process parameter and / or the component mass, and the corresponding, in particular corresponding or associated, side lengths of the hypercube corresponding to the respective values ​​or parameters are predetermined by associated error values, which are determined, in particular, in a first step and each form an interval (error interval predetermined side length). The at least one semi-finished product value, the at least one state variable, the at least one process parameter, and the component mass can be considered in combination as input variables for the model. In other words, the hypercube forms a space of input variables defined by measurement errors, wherein the center point of the hypercube is a point that describes all measured values ​​of the semi-finished product properties and states and, in addition, possibly, the process parameter or component mass. In other words, the center point represents the actually measured or determined value. The input variables can thus also be combined into an input space. The data set thus provides the coordinates for the center point, wherein the dimension of the hypercube corresponds to the number of entries in the data set. The number of entries is equal to the sum of the semi-finished product values, state variables, process parameters, and / or component masses contained in the data set. Depending on the type of error value, the space formed by the hypercube may contain values ​​for semi-finished product properties or state variables that cannot physically occur. For example, when measuring the lubricant quantity, a value of zero may be obtained. If a measurement accuracy of ±0.1 g / m*m is used, the hypercube can include negative values ​​for the lubricant quantity. Therefore, the hypercube should be defined in such a way that no physically impossible values ​​for semi-finished product properties and / or state variables are generated. In this way, the extent of the hypercube can be correspondingly reduced in terms of the corresponding dimensions, so that the hypercube does not contain physically impossible semi-finished product properties or states. In principle, this measure only reduces the size of the hypercube. The advantage of the hypercube is that the selection in the third step of the method can be carried out particularly advantageously.

[0025] In another advantageous embodiment of the present invention, the at least one control value is selected using a sampling method and / or an optimization method. In other words, the control value is determined from the space of input variables, or in particular, for example, from a hypercube, using a sampling method, for example, based on random sampling, and / or using an optimization method, for example, by searching for a maximum value around the center of the hypercube. Using an optimization method can yield the advantage of finding the control value that corresponds to the most adverse effect of instability. Using a sampling method can also yield the advantage of, for example, providing a uniform distribution of multiple control values, thereby advantageously detecting instabilities. This can be particularly the case when the optimization method is terminated prematurely due to a discontinuity in the objective function.

[0026] In another advantageous embodiment of the present invention, the sampling method comprises a Latin hypercube method. In other words, points are generated in the hypercube that serve as control values ​​selected using the Latin hypercube. This results in the advantage that the at least one control value corresponds to a sample that reflects a uniform distribution of the at least one selected control value, and in particular a plurality of selected control values, over the corresponding interval of error values ​​and can therefore be advantageous compared to random samples.

[0027] In another advantageous embodiment of the present invention, when at least two control values ​​are available and the difference can therefore include at least two differential values, an aggregated difference is formed. In other words, for the corresponding data set (if multiple control values ​​are determined for the data set), all differences are determined and aggregated into a single difference. Thus, for example, it is conceivable to use the largest difference as the value for further observation, wherein this value can form the aggregated difference. This results in the advantage that the threshold value can be formed in a particularly advantageous manner. Furthermore, the threshold value, in particular the largest threshold value, forms a very advantageous value for assessing instability.

[0028] In another advantageous embodiment of the present invention, the threshold value is derived through a statistical evaluation of the aggregated variances and / or through target setting, particularly for component quality. For example, a limit value that can be used as a threshold value can be derived from the distribution of the calculated variances. Additionally or alternatively, the threshold value can be generated based on the quality requirements set for the process. For the statistical evaluation, standardized statistical methods and values ​​such as mean, variance, standard deviation, and others can be used. This results in the advantage that instabilities can be particularly advantageously inferred based on the determination of the threshold value.

[0029] In another advantageous embodiment of the present invention, a warning signal is output when the difference exceeds the threshold value. Additionally or alternatively, a different trained model and / or updated training data set can be suggested for the currently used model. In other words, when instability is detected, the user of the production system can be warned and / or alternative methods for determining process parameters or parameter sets can be suggested. This results in particularly advantageous operation of the production system. This allows for stability checks, particularly during the use of the model.

[0030] In another advantageous embodiment of the present invention, the at least one data set can use the at least one semi-finished product value and / or the at least one state variable and / or the at least one process parameter and / or the component quality from a training data set and / or a validation data set and / or an operational data set and / or an artificial data set, the operational data set comprising data generated during model operation and the artificial data set comprising artificially generated data. In other words, the data set is formed from the amount of training data and / or data used to verify the predicted quality. Additionally or alternatively, the data set is formed from the at least one semi-finished product characteristic and / or the at least one state variable and / or the at least one process parameter and / or the component quality currently used in operation. Additionally or alternatively, the data for the at least one semi-finished product characteristic and / or the at least one state variable and / or the at least one process parameter and / or the component quality can be artificially generated, for example, by interpolation. This allows for evaluations to be performed for points in the input data space that are not part of the data used for model training, validation, and testing. This results in the advantage of enabling particularly advantageous verification of the model and, therefore, determination of its stability. Furthermore, there is the advantage that the model can be checked for instabilities already during the training phase of the model.

[0031] A second aspect of the present invention provides a computer program that can be loaded into a memory of a computer device of a production system and includes program components for executing the steps of the method when the computer program is executed in the computer device or control unit.

[0032] Advantages and advantageous embodiments of the first aspect of the present invention are hereby regarded as advantages and advantageous embodiments of the second aspect of the present invention, and vice versa.

[0033] A third aspect of the present invention relates to an electronically readable data carrier having stored thereon electronically readable control information, which has at least one computer program as just described and is designed such that, when the data carrier is used in an electronic computing device, the method described herein according to the first aspect of the present invention can be executed.

[0034] Advantages and advantageous embodiments of the third aspect of the present invention are hereby regarded as advantages and advantageous embodiments of the second aspect and the first aspect of the present invention, and vice versa.

[0035] Further features of the invention are apparent from the claims, the drawings, and the description of the drawings. The features and feature combinations described above in the description and the features and feature combinations described subsequently in the description of the drawings and / or shown individually in the drawings can be used not only in the respectively indicated combination but also in other combinations or alone. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The present invention will now be explained in more detail with the aid of preferred exemplary embodiments and with reference to the accompanying drawings.

[0037] Figure 1 a schematic flow chart showing a method for operating a production plant; and

[0038] Figure 2 A schematic representation of a data set constructed as a two-dimensional hypercube for the method is shown. DETAILED DESCRIPTION

[0039] For example, a production facility can be used to manufacture vehicle body components. The production facility can include a press or a press line and produces components in the form of vehicle body components from semi-finished products, such as blanks cut from coils. The production facility is operated with process parameters that are favorable for the process of producing the components and that enable a certain component quality. These process parameters are adapted to the respective semi-finished product properties for the shaping process, for example. The process parameters can be suggested by a self-learning algorithm or as a self-learning or trained model. Such a model can be unstable.

[0040] The method presented here makes it possible to analyze the stability of the model and thus to operate the production plant in an advantageous manner. Figure 1 A schematic diagram shows a method for operating a production plant, which produces a component from at least one semi-finished product having semi-finished product properties in a process characterized by a parameter set including at least one process parameter, wherein the parameter set is predetermined and / or the component quality is determined by a trained model using the at least one semi-finished product property. Figure 1 Steps S1 to S6 shown in FIG. 5 determine the stability of the trained model:

[0041] In a first step S1, at least one error value for at least one semi-finished product property and / or state variable and / or at least one process parameter and / or the component quality is detected. In a second step S2, a data set is generated for the at least one semi-finished product. The data set includes, for the at least one semi-finished product property, a semi-finished product value that characterizes or describes the semi-finished product property, along with an associated error value, which can be understood, in particular, as a measurement error, and / or includes the state variable with an associated error value and / or the at least one process parameter with an associated error value and / or the component quality with an associated error value. The respective error values ​​for the at least one semi-finished product value, the at least one state variable, and the at least one process parameter indicate, for example, uncertainty, in particular measurement uncertainty, and thus the deviation between a target value and an actual value. In the case of component quality, the error value can be understood as a predetermined quality requirement, and in particular, a value of zero can be acceptable. In a third step S3, at least one control value 10 is selected within a range 12 formed by the respective error values ​​and surrounding the semi-finished product value and / or the state variable and / or the process parameter and / or the component quality. In a fourth step S4, a control set including the corresponding data is formed by the trained model using the at least one selected control value 10. In a fifth step S5, a difference between the data set and the control set is formed. Finally, in a sixth step S6, a threshold value is derived from the formed difference, which describes the degree of instability for the model.

[0042] The trained model can be designed, in particular, as a machine learning model and thus, for example, implemented as a self-learning algorithm and / or a neural network. Additionally or alternatively, the trained model can describe a simulation model and / or a mathematical function. To facilitate the operation of the production plant, the model used in the method is used to operate the production plant when the difference is less than a threshold value, i.e., when no instabilities are present, so that the at least one component can be produced using a parameter set determined by the model. The method allows the stability of the model to be analyzed both before startup and during ongoing operation, wherein error values ​​and thus corresponding magnitudes, i.e., corresponding measurement errors of state variables of semi-finished product values, process parameters, or component quality, are determined and / or evaluated or at least one detection is performed for both use cases, and a data set is then formed. The data of this data set can, in particular, be used, at least partially, as input data for the trained model.

[0043] In this way, the observation of the start-up or training of the model can first be carried out in one phase. After the training of the model or the corresponding model, each of the data sets can be formed from a certain amount of training data and from the data used for verifying the prediction quality or component quality.

[0044] It is also conceivable to artificially generate combinations of semi-finished product properties and state variables during the training phase and to test these combinations as well. This allows the model's stability to be analyzed even during the training phase when extrapolating or previously unknown points in the space of semi-finished product properties and state variables. For example, it may be particularly necessary to check the edges of the semi-finished product's specification limits, as the training data typically does not cover the entire permissible range of the semi-finished product's properties. Outside the specification limits, the model does not need to be checked for instabilities, as such materials cannot be processed. Therefore, it is conceivable to form a second hypercube that simulates the specification limits and to perform a targeted search of the entire space for instabilities through sampling. New points are selected in this space, thereby forming a data set, and then at least one control point is generated for each data set. A similar approach can be used for the state variables. However, no specifications are given for these values. Therefore, for such a test, the possible limits of the state variables must be evaluated. Therefore, the model can be particularly well tested by repeating the method on different data sets.

[0045] The data set contains data or quantities resulting from or related to the production or creation of the individual components. In the method, the environment is observed within a space of input variables, including semi-finished product properties and state variables. This space is defined by the measurement accuracy or corresponding error values.

[0046] When considering error values, it is important to note that process parameters are typically set values ​​and therefore do not have measurement accuracy. However, the actual value may deviate from the set value, so the associated error value can be formed by the difference between the theoretical value and the actual value. Similarly, target settings can be used as error values ​​when predicting component quality.

[0047] Each state variable, process parameter and / or semi-finished product value forms a separate dimension, so that the boundaries formed by the error values ​​can be the surfaces of a hypercube 16. As the center point or center of gravity of the hypercube 16, a point 14 is used, which describes all measured values ​​of the semi-finished product properties and state variables. Additionally or alternatively, this point can also describe the component quality or parameter set. All of these values ​​can be combined into a vector. This is advantageous when the data set is formed as described and in Figure 2 In the case of the hypercube 16 shown in FIG, the respective side lengths of the hypercube are formed by the respective intervals 12. If the respective intervals 12 have an extent in at least one direction which, combined with the position of the associated values ​​around which the intervals 12 are formed, results in parts of the intervals accommodating physically unsuitable and, in particular, impossible values, the intervals 12 can be tailored accordingly. This allows the point 14 to be offset from the center or midpoint of the hypercube.

[0048] so Figure 2 A two-dimensional hypercube is shown, wherein, in the height direction, for example, mold properties such as sheet metal thickness are plotted, and the intervals approximately represent the intervals of the error values. In the longitudinal or transverse or horizontal direction, for example, state variables such as the temperature of the semi-finished product can be simulated, including the corresponding measurement errors, as further intervals 10.

[0049] Figure 2 Furthermore, the control values ​​10 shown are shown to be selected using a special sampling method, the so-called Latin hypercube method, and this is done by a special uniform distribution over the space of input variables. This sampling method or Latin hypercube method constitutes sampling in order to obtain a sample that describes the control value 10. This makes it possible to generate n points (in the space of measurement accuracies or in the space of error values) within the hypercube 16. Figure 2 , where n is equal to four), wherein for each generated point a prediction of the machine learning model or the trained model is implemented for at least one process parameter of the parameter set and / or the component quality. Subsequently, by forming a difference, a comparison or formation can be made between the prediction at the point 14, i.e. the center of the hypercube, and the prediction implemented for analyzing the stability of the learning model, i.e. the prediction calculated at the corresponding control value 10. The calculation of this difference is carried out in particular for each available data set. For further observation, it may be suitable to aggregate the differences calculated for each data set into a single difference. That is, if at least two control values ​​10 are used, the difference may include at least two difference values, which can be combined into an aggregated difference value.

[0050] Alternatively, instead of a sampling method, an optimization method can be used, for example, to determine or select the at least one control value 10. This optimization method can be used to find the largest possible predicted difference between point 14 and the other points determined by the control value 10. Thus, for example, the optimization method can be used to easily find the optimal value and thus the maximum value of the aggregated difference. This makes it possible to particularly advantageously assess instabilities.

[0051] Advantageously, thresholds or limit values ​​characterizing stability can be derived through a statistical evaluation of the calculated aggregated differences and / or based on target settings. The thresholds can also be derived for points used for training. At these points, there is a particularly low risk of instability. For example, the thresholds or limit values ​​can be derived from the distribution of the calculated differences. All differences exceeding the threshold value indicate instability. If instability is detected, the user can be alerted. The user can then change the selection or method used for training the model again, so that a model with less instability and / or advantageously no confirmed instability is used for operating the production system. This can also be suggested by the method itself.

[0052] Alternatively or additionally, if a model is used to predict component quality, the threshold value can be derived, for example, from a measurement error in the component quality of the produced or generated component. This method essentially provides two embodiments: first, the model is used to determine the parameter set (the stability of the model is to be tested), and alternatively, the model or another model is used to determine the component quality, resulting in corresponding combinations according to the exemplary embodiments or the first claim. If the model suggests a parameter set, the threshold value can also be determined by experimentally and / or empirically determining the magnitude of the changes in the process parameters that result in a measurable change in the component quality.

[0053] Additionally, the stability can also be assessed during the operation of the model. The stability is determined only for the current data set for which a prediction should be made, wherein the limit value or threshold value can be obtained, for example, from a previous process of the method.

[0054] The described method can also be used for other types of predicted sizes or for other production systems as press lines.

[0055] In addition to the method, a computer program and a data carrier for evaluating the stability of predictions of a machine learning model for a production process should be introduced, wherein the computer program is designed to carry out the method when executed on an electronic computing device and the data carrier can include the corresponding program steps of the computer program.

[0056] An advantage of the method, computer program, and data carrier is, for example, the ability to predetermine the number of control values ​​to be observed. This makes it possible to evaluate the computational requirements of the method and determine a suitable selection of the number of control values ​​based on the task. This ensures that the predetermined duration for executing and controlling production is not exceeded. This applies both to sampling and to the optimization method. In principle, the predictions can also be calculated in parallel on multiple computing cores of the electronic computing device, thereby enabling scalability of the method.

[0057] In order to use the method, no knowledge of the distribution of the measured values ​​with respect to semi-finished product properties, state variables or measured process parameters is required.

[0058] Since the measurement accuracy of sensors is usually determined in the laboratory, an increase in uncertainty can be expected when used in an industrial environment. This can be taken into account by hypercubic scaling, where, for example, a separate scaling can be performed for each dimension of the input space.

[0059] Reference Signs List

[0060] S1 First step

[0061] S2 second step

[0062] S3 Step 3

[0063] S4 Step 4

[0064] S5 Step 5

[0065] S6 Step 6

[0066] 10 control value

[0067] 12 intervals

[0068] 2 p.m.

[0069] 16 Hypercube

Claims

1. A method for operating a production plant for producing components from a semi-finished product having at least one semi-finished product property in a process characterized by a parameter set including at least one process parameter, wherein: The parameter set and / or the component quality are predetermined by means of the at least one semi-finished product property using a trained model, and the stability of the trained model is determined using the following steps: - detecting at least one error value of the at least one semi-finished product property and / or at least one state variable and / or at least one process parameter and / or the component quality; (S1) - forming a data set for the at least one semi-finished product, said data set comprising, for the at least one semi-finished product property, a semi-finished product value with an associated error value and / or the state variable with an associated error value and / or the at least one process parameter with an associated error value and / or the component mass with an associated error value; （S2） - selecting at least one control value (10) in an interval (12) formed by corresponding error values ​​around the semi-finished product value and / or the state variable and / or the process parameter and / or the component quality; (S3) - forming a control set comprising corresponding data by means of the trained model with the aid of the at least one selected control value (10); (S4) - forming a difference between the data set and the control set; (S5) and - deriving a threshold value from said difference (S6).

2. The method according to claim 1, characterized in that When the difference is smaller than the threshold, the model is used to operate the production equipment and generate a parameter set.

3. The method according to claim 1 or 2, characterized in that The data set is formed as a hypercube (16), the center point of which is formed by the at least one semi-finished product value and / or the at least one state variable and / or the at least one process parameter and / or the component quality, and the respective side lengths of the hypercube are predetermined by the associated error values.

4. The method according to any one of the preceding claims, characterized in that The at least one control value (10) is selected by a sampling method and / or by an optimization method.

5. The method according to claim 4, characterized in that The sampling method includes the Latin Hypercube method.

6. The method according to any one of the preceding claims, characterized in that When at least two control values ​​(10) are used and the difference thus comprises at least two difference values, an aggregate difference can be formed.

7. The method according to claim 6, characterized in that The threshold value is derived by statistical evaluation of the aggregate differences and / or by target setting.

8. The method according to any one of the preceding claims, characterized in that When the difference is greater than the threshold, a warning signal is output and / or an additional trained model and / or an updated training data set is suggested.

9. The method according to any one of the preceding claims, characterized in that For the at least one data set, the at least one semifinished product value and / or the at least one state variable and / or the at least one process parameter and / or the component quality of a training data set and / or a validation data set and / or an operating data set and / or an artificial data set is used.

10. A computer program that can be directly loaded into a memory of an electronic computer of a production system, the computer program comprising program components for executing the steps of the method according to claim 1 when the computer program is executed.

11. An electronically readable data carrier having electronically readable control information stored thereon, the control information comprising at least one computer program according to claim 9 and being designed such that, when the data carrier is used in an electronic computing device of a production plant, the control information executes a method according to one of claims 1 to 8.

Images