DE102018220064A1 - Determination of values of production parameters - Google Patents

Abstract

The invention relates to a method for determining values of production parameters (PI1-PI2; PI1-PI4; PI7-PI9; PI1-PI4, PI7-PI9) of a production process, a product being produced by the execution of the production process, the method following Steps (S2a, S4) comprises: - receiving a model (1; 2; 3; 4; 5) of the production process that is a result of machine learning, wherein at least one production parameter (PI1-PI2; PI1-PI4; PI7-PI9 ; PI1-PI4, PI7-PI9) input variable (11-12; 11-14; I7-I9; 11-14, I7-I9) of the model (1; 2; 3; 4; 5) and at least one product property ( PO1, PO2; PO3) is the output variable (O1, O2; O; O3) of the model; - Using at least one value of at least one production parameter (PI1-PI2; PI1-PI4; PI7-PI9; PI1-PI4, PI7-PI9) a reversal model (U1, U2) of the production process, the reversal model (U1, U2) being generated by reversing the model (1; 2; 3; 4; 5) of the production process, so that the at least est a product property (PO1, PO2; PO3) is the input variable and the at least one production parameter (PI1-PI2; PI1-PI4; PI7-PI9; PI1-PI4, PI7-PI9) is the output variable.

Description

The invention relates to a method and an arrangement for determining values of production parameters of a production process, during the execution of which the production process produces a product.

Various production parameters can be decisive for a production process. The production parameters include, for example, ambient conditions of the production environment, for example the temperature or the humidity, properties of materials and / or products used in the production process, for example a density or a composition of the material, and / or properties of machines used in the production process, for example For example a pressing force or a rotational speed.

Production parameters of the production process can influence the product that is generated by the execution of the production process. Frequently, at least part of the product properties of the products produced should be constant and the values of the product properties should be within predetermined tolerance ranges, for example. This can also apply to all product properties, especially in series production. As an alternative or in addition, it may be desirable to achieve different product properties in each case in accordance with a specification, the values of the product properties then being, for example, within tolerance ranges which correspond to the respective specification.

In particular, values of production parameters that are unsuitable for the production process, for example an incorrectly selected and / or incorrectly set and / or fluctuating value of a production parameter, can lead to undesirable product properties or negatively influence product properties. Such undesirable product properties can result in reduced quality of the product and / or uselessness of the product.

Especially when a large number of production parameters occur in the production process, relationships between values of product properties and values of production parameters can only be determined with difficulty.

In particular, combinations of production parameters in which a value of a production parameter is unfavorable, at least in combination with a value of a further production parameter, can have a negative effect on product properties. Such influences in particular are difficult to identify and / or quantify. Such a combination can be, for example, the combination of a forming temperature, a forming force, a forming speed and a sheet thickness of a semi-finished product to be formed.

Values of production parameters that lead to desired values of product properties can be determined empirically. As a rule, however, this is associated with great effort, in particular if a large number of production parameters are to influence the production process and / or a large number of product properties are to be determined.

Values of production parameters can also be determined on the basis of empirical values. However, the usability of the values depends on whether the experience covers all relevant production circumstances.

It is also possible to simulate production processes, whereby values of production parameters can be specified. In the case of extensive and / or complex production processes, complete or almost complete physical modeling and / or simulation of the influences of production parameters, in particular the influences of interactions between production parameters, is not possible or only with a high level of computation effort. In addition, it can be difficult and time-consuming to validate such simulation models.

The use of machine learning methods (e.g. regression, support vector machine, neural network) can help. In this way, values of product properties can be forecast on the basis of values of known production parameters.

In connection with models that are obtained through machine learning, the so-called back propagation, also called "error feedback", can be used, especially of neural networks in one Learning a model can be used to reduce errors. Realistic values of both the production parameters and the product properties are required for both teaching and error feedback.

WO 2015/184729 A1 describes a method for hyper parameter optimization and feature tuning of machine learning algorithms. The method includes: A: randomly generating an initial collection pool that has a plurality of parameter groups; B: performing an iterative optimization based on EnKF (Ensemble Kalman filter) on the large number of parameter groups in the initial acquisition pool in order to obtain a large number of first optimized parameter groups in each case; C: evaluating the plurality of first optimized parameter groups in order to obtain a first acquisition pool and a first supplementary parameter acquisition pool; D: Update the initial collection pool with the supplemental parameter collection pool; Repeating steps B, C and D a first predetermined number of times to obtain a second acquisition pool and a set of optimal parameters. The method increases the calculation efficiency and accuracy of the hyper parameter optimization and the function optimization and has a high degree of generality. In addition, a system for hyper-parameter optimization and feature matching of machine learning algorithms is also described.

It is an object of the invention to provide a method for determining values of production parameters of a production process, during the execution of which a product is produced. In particular, the determination of values of the production parameters should be facilitated and / or their use should reliably lead to desired product properties. In particular, the method should enable the determination of sets from values of a plurality of production parameters for complex production processes and in particular for those production processes that have a large number of production parameters. It is a further object of the invention to provide an arrangement for carrying out the method.

It is proposed that a model of the production process be used to determine values of production parameters of the production process, the model being a result of machine learning. Machine learning in particular offers the possibility of determining and / or describing relationships between a large number of production parameters and a large number of product properties. In particular, these can be relationships that are not or difficult to describe using simulations of physical processes and / or analytical methods.

• - Empfangen eines Modells des Produktionsprozesses, das ein Ergebnis von maschinellem Lernen ist, wobei zumindest ein Produktionsparameter Eingangsgröße des Modells ist und zumindest eine Produkteigenschaft Ausgangsgröße des Modells ist;
• - Ermitteln zumindest eines Wertes zumindest eines Produktionsparameters mithilfe eines Umkehrmodells des Produktionsprozesses, wobei das Umkehrmodell eine Umkehrung des Modells des Produktionsprozesses ist und zumindest ein Wert der zumindest einen Produkteigenschaft in das Umkehrmodell eingeht (d.h. bei dessen Ausführung verwendet wird) und der zumindest eine Wert des zumindest einen Produktionsparameters aus dem Umkehrmodell erhalten wird.

In particular, the following is proposed: A method for determining values of production parameters of a production process, during the execution of which a product is generated, the method comprising the following steps:

• Receiving a model of the production process that is a result of machine learning, at least one production parameter being the input variable of the model and at least one product property being the output variable of the model;
• - Determining at least one value of at least one production parameter with the aid of a reversal model of the production process, the reversal model being a reversal of the model of the production process and at least one value of the at least one product property being included in the reversal model (ie being used in its execution) and the at least one value of the at least one production parameter is obtained from the reverse model.

Production parameters can in particular be any quantifiable quantities that occur during the production process. Such a production parameter can in particular be a physical quantity, for example a process parameter, a machine parameter, a material parameter, an environmental parameter and / or a property of a product that is used in the production process. Such a product can be, for example, a product produced in a previous production process.

Production parameters can be adjustable. An adjustable production parameter is, for example, a machine parameter or a property of a product that is used in the production process. Other production parameters cannot be adjustable. A non-adjustable production parameter can be an environmental parameter, for example.

The production process can in particular have or consist of at least one production and / or processing and / or processing. Manufacturing and / or processing and / or processing can relate in particular to a raw material, a semi-finished product and / or purchase an existing product. The existing product can in particular be a piece goods, a bulk goods or a flow goods.

The model of the production process can in particular be received via an interface. In particular, a computer that has the interface or is connected to it can receive the model. "Receiving" can in particular mean receiving data - for example from another computer and / or from a network and / or from a storage device. “Receiving” can also be understood to mean an input, for example with the aid of input means. As an alternative or in addition, “receiving” can include reading in the model or parts thereof. In particular, receiving the model can enable the model to be executed and / or started and / or adapted.

The model of the production process is designed in such a way that at least one production parameter is the input variable of the model. This means, in particular, that when the model is used (ie modeling the production process using the model), at least one value of the at least one production parameter is included in the model and, for example, at least one value and in particular one value of several production parameters is included in the model . The model can be configured in such a way that a plurality of values are included in the model with respect to at least one production parameter, for example if the production parameter is a process variable that changes in the course of the execution of the production process.

The model of the production process is in particular designed such that at least one product property is the starting variable of the model. In particular, this means that when the model is used, at least one value of the at least one product property is output from the model, and for example at least one value and in particular in each case exactly one value of a plurality of product properties is output from the model. The model can be designed such that a plurality of values are output from the model with respect to at least one product property, e.g. B. if several products are generated in succession in the production process and the product property changes in the course of execution.

A “plurality of values” can in particular denote a plurality of discrete values and / or one or more ranges of values, for example in the form of at least one interval. The entry of a plurality of values of a production parameter into the model can therefore be realized by entering a range of values. It is determined for the model that the production parameter is within the limits of the value range. The same can apply to the reversal model if a value range of a product property is received.

The model of the production process can be available in particular as a mathematical description or as a plurality of mathematical descriptions. In particular, the model of the production process can have an analytical model. Alternatively or additionally, the model of the production process can have a numerical model. The at least one production parameter (input variable) and / or the at least one product property (output variable) can be represented in the model of the production process in particular by variables.

Machine learning can be carried out, for example, using a neural network and / or using regression and / or an SVM (Support Vector Machine).

The inversion model, which is an inversion of the model of the production process, can in particular have an inversion of at least part of the mathematical description or at least part of one of the plurality of mathematical descriptions. In particular, the reversal model can have an inversion of at least a part of the analytical model. Alternatively or additionally, the reversal model can have an inversion of at least a part of the numerical model. The at least one production parameter (output variable of the reversal model) and / or the at least one product property (input variable of the reversal model) can in particular be represented by variables within the reversal model.

At least one product property is the input variable of the reversal model. This means that at least one value of the at least one product property is included in the reversal model. The value or one of the values can be entered for the purpose of a single application of the reversal model. Alternatively, the value or one of the values can be specified and accessed when the reversal model is used.

Furthermore, at least one production parameter is the output variable of the reversal model. This means that at least one value of the at least one production parameter is obtained from the reversal model, in particular is output when it is used to model a reversal of the production process.

The model of the production process can in particular be designed to predict that at least one concrete value of the at least one production parameter (input variable) in the production process leads to a concrete value of the at least one product property (output variable). The reversal model can in particular be designed to predict that at least one concrete value of the at least one product property (input variable) can be achieved if at least one concrete value of the at least one production parameter (output variable) is maintained in the production process.

The reversal model can be created immediately before and / or when determining the at least one value of the at least one production parameter from the model of the production process, in particular on the basis of the model of the production process. In a special case, the reversal model can arise solely from a modified application of the model of the production process. This is the case, for example, if a mathematical description (for example using a system of equations) describes the model of the production process and the mathematical description contains the input and output variables of the model. To use the reversal model, only at least one of the output variables of the model or the only output variable of the model is then used as the input variable of the reversal model. While at least one value of at least one production parameter is included in the mathematical description for an application of the model of the production process, at least one value of at least one product property is included in the mathematical description for an application of the reversal model.

Alternatively, the reversal model can already exist, in particular have been created and / or received beforehand, if the value or values of the product property (s) which serve as input variable (s) for the application of the reversal model are not yet available. The step of determining the at least one value of the at least one production parameter can in particular be carried out once, for example before the execution of the production process or before a repeated execution of the production process, or the step can be carried out repeatedly, for example before each execution of the production process.

In particular, the reversal model can be applied, adapted and / or newly created several times during production, for example repeatedly at intervals and / or for example on the basis of changed production parameters. In particular, these can be production parameters that cannot be set and / or are not taken into account in the reversal model.

An application of the reversal model can in particular be part of an industrial production system, the method being carried out with the aid of a computer server, which is also part of the industrial production system. Alternatively or additionally, the reversal model can be used in a cloud-based manner and / or using a virtual machine. Determined values of production parameters can be set manually and / or automatically within the industrial production system in order to carry out the production process accordingly.

A possible sequence of the two steps of the method is that the step of receiving the model of the production process is carried out before determining the at least one value of the at least one production parameter. Both steps can also be carried out several times in succession, for example in order to optimize the model of the production process and / or the reversal model.

The proposed method facilitates the determination of values of production parameters of a production process. With the aid of the reversal model, at least one value of at least one production parameter can be determined, in particular, according to the specification of at least one (desired) value of at least one product property. The step of receiving such a specification can be regarded as an explicit step of the method. If the reversal model exists, it can be used several times, for example during production, and it can be achieved that the value of the at least one product property remains the same, although the at least one value of the at least one production parameter changes or changes during production. Alternatively, the value of the at least one product property can be changed in the desired manner.

For example, at least one changed value of a production parameter can be reacted to by another change in at least one value of at least one further production parameter, which the reversal model determines, so that the at least one value of the at least one product property is, for example, constant or adjusted as desired. Corresponding combinations of values of different production parameters can be obtained by using the reversal model.

• - Durchführen von maschinellem Lernen, um das Modell des Produktionsprozesses anhand von Daten anzulernen, die zumindest einen Wert des zumindest einen Produktionsparameters, der Eingangsgröße des Modells ist, und zumindest einen Wert der zumindest einen Produkteigenschaft, die Ausgangsgröße des Modells ist, enthalten.

According to one embodiment of the method, the method has the following additional step, which is carried out before the step of receiving the model of the production process:

• Performing machine learning in order to teach the model of the production process on the basis of data which contain at least one value of the at least one production parameter, which is the input variable of the model, and at least one value of the at least one product property, which is the output variable of the model.

Machine learning can take place, for example, using known data, for example data from current and / or already completed production processes of copies of the product, or using empirical values and / or using empirically determined data, for example relating to one and / or some and / or all production parameters that are input variables of the model and / or one and / or some and / or all product properties that are output variables of the model.

As an alternative or in addition, teaching can take place, for example, using data from simulations, for example relating to one and / or some and / or all production parameters that are input variables of the model and / or one and / or some and / or all product properties, the output variables of the model.

Values of production parameters used for teaching can relate both to product parameters that can be changed and to those that cannot be changed.

Machine learning can take place in particular as part of an industrial production system, with machine learning being carried out using at least one computer of the industrial production system. Alternatively or additionally, machine learning can be cloud-based and / or using a virtual machine.

Learned knowledge can, for example, be stored in the form of mathematical quantities within the model of the production process. If machine learning is carried out using a neural network, for example, knowledge can be stored in the neural network, for example in particular in the form of at least one weight and / or at least one bias. Weights and bias are fundamentally essential components of neural networks, represent acquired knowledge and generally occur in multiplicity in neural networks.

Machine learning generally requires data to be able to implement learning and to generate a model as a result. Teaching the model of the production process creates a basis for creating the reversal model.

• - Ermitteln zumindest eines Werts zumindest einer Produkteigenschaft mithilfe des Modells des Produktionsprozesses, um das Modell des Produktionsprozesses zu validieren, wobei als Eingangsgröße zumindest ein Wert des zumindest einen Produktionsparameters vorgegeben wird und zumindest ein Wert der zumindest einen Produkteigenschaft als Ausgangsgröße ausgegeben wird.

According to one embodiment of the method, the method has the following additional step, which is carried out before the step of determining the at least one value of the at least one production parameter:

• - Determining at least one value of at least one product property using the model of the production process in order to validate the model of the production process, at least one value of the at least one production parameter being specified as the input variable and at least one value of the at least one product property being output as the output variable.

With this additional step, the model of the production process can be validated. For example, the at least one value of the at least one production parameter can be defined as an input value in the model of the production process. The at least one value of the at least one product property determined using the model can then be checked, for example on the basis of a predetermined plausibility check. In particular, the at least one determined value of the at least one product property can be compared using at least one known value of the at least one product property and the comparison can be evaluated. The at least one known value of the at least one product property can, for example, be based on at least one copy of the product are determined, wherein the at least one copy of the product was produced using the at least one value of the at least one production parameter.

Alternatively or additionally, for example, the at least one value of the at least one production parameter can be changed in a production, so that several copies of the product are produced with different values of the at least one product property. These models can then be used to validate the model of the production process.

As an alternative or in addition, a validation can be carried out, for example, on the basis of known data, for example data from current and / or already completed production processes of copies of the product, or using empirical values and / or using empirically determined data, for example relating to one and / or some and / or all production parameters that are input variables of the model and / or one and / or some and / or all product properties that are output variables of the model.

As an alternative or in addition, validation can be carried out, for example, using data from simulations, for example relating to one and / or some and / or all production parameters which are input variables of the model and / or one and / or some and / or all product properties, the output variables of the model.

The reversal model, which is generated by reversing the model of the production process, can also be at least partially or completely validated by validating the model of the production process.

• - Vorgeben zumindest eines Vorgabewertes zumindest eines Produktionsparameters, der Eingangsgröße des Modells des Produktionsprozesses ist, als Randbedingung im Umkehrmodell, sodass der zumindest eine Wert des zumindest einen Produktionsparameters, der aus dem Umkehrmodells erhalten wird, in Abhängigkeit des zumindest einen Vorgabewertes ermittelt wird.

According to one embodiment of the method, the method has the following additional step:

• - Specifying at least one default value of at least one production parameter, which is the input variable of the model of the production process, as a boundary condition in the reversal model, so that the at least one value of the at least one production parameter, which is obtained from the reversal model, is determined as a function of the at least one default value.

The at least one default value can in particular be a value that is difficult or impossible to set for the production process. As an alternative or in addition, the at least one default value can be a value that can be set for the production process, but should nevertheless be predetermined for other reasons. For example, depending on the default value, it may be of interest to obtain suitable sets of the values of production parameters from the reversal model.

The at least one production parameter, the value of which is the at least one default value, is an input variable of the model of the production process. This enables the reversal model to take the at least one default value into account. In addition, the model of the production process can be taught using suitable values of all production parameters. This ensures that the reversal model can be used reliably taking into account the at least one default value and that at least one value of at least one production parameter can be obtained from the reversal model.

With regard to a specific production parameter, the boundary condition can have a single default value and / or several default values, for example a default value range and / or several discrete default values. The same can apply to a plurality of specific production parameters.

The boundary condition can be formulated mathematically, even if the at least one default value is a value of a physical quantity.

In particular, the reversal model can be used in relation to the boundary condition in such a way that the at least one determined value of the at least one production parameter, which is determined with the aid of the reversal model, is determined as a function of the boundary condition. This means that the value is determined for different default values of the same production parameter or for different default values of the same production parameter.

There are various options for this, which can also be combined with one another. On the one hand, the reversal model can be used repeatedly for the various default values of the same or the same production parameter. On the other hand, the boundary condition can be applied in such a way that values of those production parameters or those production parameters from the reversal model are also obtained for which the default value or the default values are defined as a boundary condition. In this case, a plurality of sets or combinations of values of production parameters are obtained from the reversal model. The set that meets the boundary condition, that is to say contains the default value or the default values, can then be selected from these records. The last-mentioned embodiment makes it possible, in particular, to apply the reversal model before a longer period in which the production process is carried out several times under different boundary conditions. The associated set of production parameters can then be selected for the applicable boundary conditions.

In particular, the different sets or combinations of values of production parameters that are obtained from the reversal model lead to the same combination of values of the product properties. Therefore, if a specific production parameter varies, the other production parameters also change and this leads to compensation for the change in the specific production parameter.

According to one embodiment of the method, the at least one production parameter for which the at least one default value is specified is a material parameter and / or an environmental parameter and / or a machine parameter and / or a process parameter and / or a property of a product that is used in the production process .

For example, this production parameter can be an environmental parameter, for example an ambient temperature or a humidity. Alternatively, it can be, for example, a machine property and / or a process parameter, for example a degree of wear of a machine, or a holding time in a heat treatment cycle, the change of which would have undesirable effects. Alternatively, it can be, for example, a material property, for example a degree of hardness or a carbon content of a steel used, a viscosity of a paste, a particle size of a substance or proportions of different substances in a substance mixture that has already been produced. It can optionally also be a material property of an intermediate product that arises only in the production process, for example a thickness of a coating or a particle size of a mixture of substances.

According to one embodiment of the method, the method for a production process in which at least one intermediate product is formed between the start and the end of the production process is carried out in such a way that properties of the intermediate product are not used as input variables of the reversal model. In other words, properties of the intermediate product are not included in the reverse model in the form of at least one input variable or in the form of at least one boundary condition. Properties of the intermediate product therefore do not need to be measured and / or determined by simulation and / or determined in any other way in order to use the reversal model. In contrast, such measurements and / or simulations can be useful for the purpose of teaching and / or validating the model of the production process. The output variable of the reversal model can therefore in particular be at least one production parameter, which is for example a material parameter and / or a property of a product that is used in the production process, but is not itself produced within the production process. In this embodiment of the method, therefore, using the reversal model, values of production parameters for the entire production process can be obtained before the start of the production process, without the product properties of the intermediate product having to be determined in the meantime or in any case having to be established.

According to another embodiment of the method, a plurality of models of the production process are received in relation to a production process in which at least one intermediate product is produced, the various models corresponding in each case to temporal sections of the production process which end with or under the production of the intermediate product Use of the intermediate product. In particular, a first model for the time period up to the production of the intermediate product and a second model for the time period thereafter are therefore received. For each of these plurality of models, an associated reversal model is used to obtain at least one value of at least one production parameter. Specific embodiments of this embodiment of the method will still be discussed.

In accordance with one embodiment of the method, the at least one determined value of the at least one production parameter is checked to determine whether it fulfills an especially predetermined criterion, and / or is evaluated using an especially predetermined evaluation scheme.

The criterion can in particular be used to check the suitability for production of the at least one determined value of the at least one production parameter. If several determined values of the at least one production parameter have been determined by the reversal model, these values can be checked separately using the criterion. If a value from different production parameters has been determined in each case, that is to say a combination of values from different production parameters, in particular a respective test can be carried out on the basis of different criteria and / or a common criterion. The common criterion can in particular show dependencies between different production parameters.

If several values from different production parameters have been determined in each case, that is to say several combinations of values from different production parameters, in particular a respective test can be carried out on the basis of different criteria and / or a common criterion. The common criterion can, in particular, have dependencies between different production parameters.

With the aid of a criterion, the suitability of the at least one determined value of the at least one production parameter for a production can be checked. As an alternative or in addition, the plausibility and / or the admissibility can be checked in particular. Alternatively or additionally, it can be checked whether the at least one determined value of the at least one production parameter corresponds to a production strategy, for example with regard to the lowest possible costs, the fastest possible production and / or the least possible effects on the environment.

As an alternative or in addition, the at least one determined value of the at least one production parameter can be evaluated using a predetermined evaluation scheme.

If several determined values of the at least one production parameter have been determined by the reversal model, these values can be evaluated together using the evaluation scheme. In particular, a ranking of the determined values of the at least one production parameter can then be created.

If a value from different production parameters has been determined in each case, that is to say a combination of values from different production parameters, in particular a respective evaluation can be carried out using different evaluation schemes and / or a common evaluation scheme. The common evaluation scheme can in particular have dependencies between different production parameters.

The suitability of the at least one determined value of the at least one production parameter for a production can in particular be assessed with the aid of an evaluation scheme. Alternatively or additionally, the plausibility and / or the admissibility can be assessed. Alternatively or additionally, it can be assessed to what extent the at least one determined value of the at least one production parameter corresponds to a production strategy, for example with regard to the lowest possible costs, the fastest possible production and / or the lowest possible effects on the environment.

According to one embodiment of the method, the model of the production process is a trained artificial neural network.

The number of input variables of the neural network can in particular be greater than the number of output variables. This results in a non-symmetrical neural network. Alternatively, the number of input variables can match the number of output variables. This can result in a neural network that is symmetrical in structure. Alternatively, the number of input variables can be smaller than the number of output variables. This results in a non-symmetrical neural network.

The structure of the neural network can in particular be predetermined. In particular, the number of inputs and / or outputs can be specified. A number of neurons and / or a number of levels of neurons and / or a number of neurons lying on one level, which are located between the inputs and the outputs, can also be predetermined.

As already explained, the neural network can be learned using machine learning. In particular, machine learning can be used to determine values of weights and / or biases that are located within the neural network.

The model of the production process, which is a result of machine learning, can in particular be in the form of a mathematical rule or several mathematical rules that describe the neural network at least partially mathematically. Such a mathematical regulation and / or several such mathematical regulations can in particular have at least one input variable (at least one production parameter), at least one output variable (at least one product characteristic), at least one weight and / or at least one bias.

In particular, such a mathematical rule can have the following form in the case of an asymmetrical neural network with one output variable and several input variables: $$\text{O\_1}=\text{f}\left(\text{l\_1}...\text{l\_m}\text{, b\_1}...\text{b}\_\text{n}\text{, w\_1}...\text{w\_p}\right).$$

Here is O_1 an output quantity, I_1 to In the are m input variables, b_1 to b_n are n bias, w_1 to w_p are p weights, where m, n and p are natural numbers.

Are for the bias b_1 to b_n and weights w_1 to w_p Known values, the form of such a mathematical rule is simplified as follows: $$\text{O\_1}=\text{f}\left(\text{l\_1}...\text{l\_m}\right).$$

The initial size O_1 is from the input quantities I_1 to In the dependent.

$$\text{l\_2}=\text{f}\left(\text{O\_1}\text{, l\_1}\text{, l\_3}\dots \text{l\_m}\right)$$

$$\text{l\_m}-\text{1}=\text{f}\left(\text{O\_1}\text{, l\_1}\text{,}\dots \text{ l\_m}-\text{2}\text{, l\_m}\right)$$

$$\text{l\_m}=\text{f}\left(\text{O\_1}\text{, l\_1}\text{,}\dots \text{l\_m}-1\right).$$

The reversal model, which is generated by reversing the model of the production process, can in particular be a mathematical reversal of such a mathematical rule, which allows an output variable O_1 to specify and solutions for the input variables I_1 to In the to determine. The initial size O_1 through the inversion it becomes an input quantity of the inversion model, the input quantities I_1 to In the become output variables of the reversal model. Such an inversion can be represented mathematically, for example, as follows: $$\text{l\_1}=\text{f}\left(\text{O\_1}\text{, l\_2}...\text{l\_m}\right)$$

$$\text{l\_2}=\text{f}\left(\text{O\_1}\text{, l\_1}\text{, l\_3}...\text{l\_m}\right)$$

$$\text{l\_m}-\text{1}=\text{f}\left(\text{O\_1}\text{, l\_1}\text{,}...\text{l\_m}-\text{2nd}\text{, l\_m}\right)$$

$$\text{l\_m}=\text{f}\left(\text{O\_1}\text{, l\_1}\text{,}...\text{l\_m}-1\right).$$

Such a system of equations cannot be uniquely solved because it is only an input quantity O_1 , but several output variables I_1 to In the having. So it has a bevy of solutions. One or more solutions can, in particular, be selected and / or evaluated using a criterion and / or using an evaluation scheme.

wobei diese Gleichung unter Variation der Eingangsgrößen I_1 bis I_m für die vorgegebene Ausgangsgröße O_1 minimiert wird, zum Beispiel mithilfe eines numerischen Lösungsverfahrens.As an alternative or in addition, such an inversion can also be represented mathematically, for example, as follows: $$0=\text{f}\left(\text{l\_1}...\text{l\_m}\right)-\text{O\_1}\text{,}$$

this equation varying the input variables I_1 to In the for the given output variable O_1 is minimized, for example using a numerical solution method.

The cases described above and below are only exemplary embodiments. In more general terms, a system of equations that describes a neural network can always be solved unambiguously and therefore analytically, for example, if the described production process has more input variables O as output variables I. has (in this case the solution of the system of equations would even be overdetermined if the inverse model was used) or if the described production process had just as many input variables O like output quantities I. having. In all of these cases it is assumed that all other variables and constants of the system of equations are known, in particular due to the learning of the neural network.

Not only in the two mathematical representations of neural networks described above, it is also possible to record at least one value of an output variable. Such a value can be, in particular, a value of the at least one further production parameter that cannot be changed or can only be changed with difficulty in the production process. Alternatively or additionally, the at least one value of the at least one further production parameter can be a value that can be changed in the production process, but is nevertheless intended to be predetermined.

Once the reversal model has been generated, the reversal model can be used, for example, during production. A (desired) input variable (e.g. O_1 , see above) of the reversal model can be specified, at least one (for example desired and / or not or only difficult to change) output variable can also be specified. The result is then a value of the at least one (unspecified) output variable.

According to one embodiment of the method, the product that is to be generated in the production process is an energy store, in particular an accumulator, or a component of an energy store, in particular an accumulator, and becomes at least one value of a production parameter of the energy store or of the component of the energy store with the aid of the Reversal model determined.

The energy store can in particular be a store for thermal energy and / or a store for chemical energy and / or a store for mechanical energy and / or a store for electrical energy.

In particular, the energy store can be a memory which is designed to be used in a vehicle, in particular for providing drive power and / or for providing power to other energy consumers of the vehicle (for example equipment details and / or a control unit) and / or Energy consumers connected to the vehicle. Such an energy consumer can be, for example, a device carried by a passenger while driving, which is connected to the energy store via an interface.

In particular, the energy store can be an accumulator that is rechargeable and stores electrical energy on an electrochemical basis. For example, it can be a lithium-ion battery and / or a lithium-polymer battery and / or a lithium-iron-phosphate battery and / or a lithium-titanate battery and / or a lithium-sulfur battery and / or a lithium-air battery and / or a nickel-metal hydride battery and / or a silver-zinc battery and / or a tin-sulfur-lithium battery and / or an aluminum-ion battery and / or a nickel Zinc accumulator and / or a nickel-hydrogen accumulator and / or a nickel-iron accumulator and / or a nickel-metal hybrid accumulator and / or a sodium-nickel chloride accumulator and / or a sodium-sulfur accumulator and / or a lead accumulator and / or a solid-state battery and / or an accumulator in which silicon and / or magnesium and / or sodium and / or titanium and / or a ceramic material, for example titanium nitride, and / or carbon and / or a hydrocarbon is used.

There are many production parameters in the production of accumulators that influence the properties of the product. In particular, complex systems with a large number of adjustable production parameters are used in the production of accumulators. Influences of production parameters on product properties are often not or not directly measurable. They sometimes only have an effect later in the production process and / or in the product. The proposed method offers the possibility of taking such influences into account in the production of accumulators within the model of the production process and within the reversal model.

• - eine Schnittstelle, ausgestaltet zum Empfangen des Modells des Produktionsprozesses, das ein Ergebnis von maschinellem Lernen ist, wobei zumindest ein Produktionsparameter Eingangsgröße des Modells ist und zumindest eine Produkteigenschaft Ausgangsgröße des Modells ist;
• - eine Ermittlungseinrichtung, ausgestaltet zum Ermitteln zumindest eines Wertes zumindest eines Produktionsparameters mithilfe eines Umkehrmodells des Produktionsprozesses, wobei das Umkehrmodell eine Umkehrung des Modells des Produktionsprozesses ist und wobei die Ermittlungseinrichtung ausgestaltet ist, das Umkehrmodell derart anzuwenden, dass der zumindest eine Wert der zumindest einen Produkteigenschaft in das Umkehrmodell eingeht und der zumindest eine Wert des zumindest einen Produktionsparameters aus dem Umkehrmodell erhalten wird.

Furthermore, an arrangement is proposed which is designed to carry out a method for determining values of production parameters of a production process, during the execution of which a product is produced, the arrangement comprising:

• an interface configured to receive the model of the production process, which is a result of machine learning, at least one production parameter being the input variable of the model and at least one product property being the output variable of the model;
• a determination device configured to determine at least one value of at least one production parameter using a reversal model of the production process, the reversal model being a reversal of the model of the production process and the determination device being designed to use the reversal model such that the at least one value of the at least one product property enters the reversal model and the at least one value of the at least one production parameter is obtained from the reversal model.

The interface can in particular be a data interface that is designed to receive digital data. For example, it can be a network interface and / or a USB interface and / or another interface that is used in the field of computers. As an alternative or in addition, the interface can be an input device configured in such a way that the model of the production process can be input. The interface can also be arranged within a computer, for example the interface of the main memory into which the model of the production process can be loaded.

The determination device can in particular be a computer or a computer network that is designed to use the reversal model. The reversal model can be described, for example, by at least one mathematical rule. This mathematical regulation can in particular be based on the at least one mathematical regulation (e.g. be the same as or be derived from this regulation), which describes the model of the production process.

The method for determining values of production parameters can in particular be the method described above.

Refinements of the arrangement result from refinements of the method which are described in this description.

• 1a ein Beispiel für ein Modells eines Produktionsprozesses als symmetrisches neuronales Netz,
• 1b ein Umkehrmodell des in 1a dargestellten Modells,
• 2a ein weiteres Beispiel für ein Modell eines Produktionsprozesses als unsymmetrisches neuronales Netz,
• 2b ein Umkehrmodell des in 2a dargestellten Modells,
• 3a ein Flussdiagramm gemäß einem Ausführungsbeispiel des erfindungsgemäßen Verfahrens,
• 3b ein Beispiel eines Modells des Produktionsprozesses, den das in 3a gezeigte Verfahren betrifft,
• 4 ein weiteres Beispiel eines Modells eines Produktionsprozesses,
• 5 noch ein weiteres Beispiel eines Modells eines Produktionsprozesses,
• 6 ein Ausführungsbeispiel einer Anordnung, die zur Durchführung des Verfahrens zur Ermittlung von Werten von Produktionsparametern ausgestaltet ist.

Embodiments of the invention will now be described with reference to the accompanying drawings. The individual figures of the drawing show:

• 1a an example of a model of a production process as a symmetrical neural network,
• 1b a reverse model of the in 1a model shown,
• 2a another example of a model of a production process as an asymmetrical neural network,
• 2 B a reverse model of the in 2a model shown,
• 3a 2 shows a flow chart according to an embodiment of the method according to the invention,
• 3b an example of a model of the production process that the in 3a shown method concerns
• 4th another example of a model of a production process,
• 5 yet another example of a model of a production process,
• 6 an embodiment of an arrangement which is designed to carry out the method for determining values of production parameters.

In various exemplary embodiments, the same reference numerals are used for devices and elements with the same function. However, this does not mean that the exemplary embodiments are the same exemplary embodiment or parts of the same exemplary embodiment.

1a shows a model of a production process in the form of a symmetrical neural network 1 . The neural network 1 has two neurons N11 and N12 on an entrance level (left in 1a shown). The neuron N11 receives the input variable I1 . The neuron N12 receives the input variable I2 . The neural network 1 also shows two neurons N21 and N22 in an intermediate level (middle in 1a shown). The neuron N21 has an intermediate output size L1 , the neuron N22 has an intermediate output size L2 .

In addition, the neural network points 1 two neurons N31 and N32 in an exit level (right in 1a shown). The neuron N31 has the initial size O1 , the neuron N32 has the initial size O2 . The neural network 1 has already been learned, it shows bias values for the bias b1 , b2 on and weight values for the weights w111 , w112 , w121 , w122 , w211 , w212 , w221 , w222 on. The bias values for the bias b1 , b2 and weight values for the weights w111 , w112 , w121 , w122 , w211 , w212 , w221 , w222 are determined with the help of machine learning (for example by means of optimization, gradient descent, back propagation).

$$\text{O2}=\text{L1*w212}+\text{L2*w222}$$

mit: $$\text{L1}=\text{a}\left(\text{I1*w111}+\text{I2*w121}+\text{b1}\right)$$

$$\text{L2}=\text{a}\left(\text{I1*w112}+\text{I2*w122}+\text{b2}\right).$$

This in 1a shown neural network 1 can be described mathematically as follows: $$\text{O1}=\text{L1 * w211}+\text{L2 * w221}$$

$$\text{O2}=\text{L1 * w212}+\text{L2 * w222}$$

With: $$\text{L1}=\text{a}\left(\text{I1 * w111}+\text{I2 * w121}+\text{b1}\right)$$

$$\text{L2}=\text{a}\left(\text{I1 * w112}+\text{I2 * w122}+\text{b2}\right).$$

The activation function a is used here. The activation function a is the sigmoid function in this example. Alternatively or additionally, it can also be another function, for example the identity function or the ReLu function.

$$\text{O2}=\text{a}\left(\text{I1}+\text{w111}+\text{I2}+\text{w121}+\text{b1}\right)*\text{w212}+\text{a}\left(\text{I1}+\text{w112}+\text{I2}+\text{w122}+\text{b2}\right)*\text{w222}\text{.}$$

If equations 1-4 are combined, the result is: $$\text{O1}=\text{a}\left(\text{I1}+\text{w111}+\text{I2}+\text{w121}+\text{b1}\right)*\text{w211}+\text{a}\left(\text{I1}+\text{w112}+\text{I2}+\text{w122}+\text{b2}\right)*\text{w221}$$

$$\text{O2}=\text{a}\left(\text{I1}+\text{w111}+\text{I2}+\text{w121}+\text{b1}\right)*\text{w212}+\text{a}\left(\text{I1}+\text{w112}+\text{I2}+\text{w122}+\text{b2}\right)*\text{w222}\text{.}$$

$$\text{O2}=\text{f}\left(\text{I1}\text{, I2}\right).$$

In short form: $$\text{O1}=\text{f}\left(\text{I1}\text{, I2}\right)$$

$$\text{O2}=\text{f}\left(\text{I1}\text{, I2}\right).$$

$$\text{I2}=\text{f}\left(\text{O1}\text{, O2}\right).$$

The system of equations consisting of equations (8) and (9) can be clearly solved. In addition, it can be reversed, so that the reversal model U1 results in 1b is shown. The reverse model U1 can be described mathematically as follows: $$\text{I1}=\text{f}\left(\text{O1}\text{, O2}\right)$$

$$\text{I2}=\text{f}\left(\text{O1}\text{, O2}\right).$$

The input variables I1 , I2 are in the reverse model U1 Output variables. The output variables O1 , O2 are in the reverse model U1 Input variables.

The input variables I1 , I2 correspond to production parameters PI1 , PI2 . The initial size O1 , O2 correspond to product properties PO1 , PO2 . Values for the output variables O1 , O2 are considered incoming values of the reverse model U1 fixed. They correspond to values of product properties that are desired or required, for example.

Values for the input variables I1 , I2 are called outgoing values of the reverse model U1 determined. They correspond to values of production parameters for which the reverse model U1 predicts that they will lead to the values of the product properties.

The production process can be a paste manufacturing process, for example. For example, it can be a paste that is required in the manufacture of an accumulator. The production parameters can, for example, be a quantity of a material (according to the input size I1 ) and a duration of a mixing process (according to the input variable I2 ) be. Alternatively, the production parameters can describe other sizes. The product properties can, for example, the viscosity (according to the initial size O1 ) and the homogeneity (according to the initial size O2 ) of a paste produced. Alternatively, the product properties can be other sizes.

2a shows a model of a production process in the form of an asymmetrical neural network 2nd . The neural network 2nd has two neurons N11 and N12 on an entrance level (left in 2a shown). The neuron N11 receives the input variable I1 . The neuron N12 receives the input variable I2 . The neural network 2nd also shows two neurons N21 and N22 in an intermediate level (middle in 2a shown). The neuron N21 has an intermediate output size L1 , the neuron N22 has an intermediate output size L2 .

In addition, the neural network points 2nd a neuron N3 in an exit level (right in 2a shown). The neuron N3 has the initial size O . The neural network 2nd has already been learned, it shows bias values for the bias b1 , b2 on and weight values for the weights w111 , w112 , w121 , w122 , w21_ , w22_ on. The bias values for the bias b1 , b2 and the weight values for the weights w111 , w112 , w121 , w122 , w21_ , w22_ are determined using machine learning.

mit: $$\text{L1}=\text{a}\left(\text{I1*w111}+\text{I2*w121}+\text{b1}\right)$$

$$\text{L2}=\text{a}\left(\text{I1*w112}+\text{I2*w122}+\text{b2}\right).$$

This in 2a shown neural network 2nd can be described mathematically as follows: $$\text{O1}=\text{L1 * w21}\_+\text{L2 * w22}\_$$

With: $$\text{L1}=\text{a}\left(\text{I1 * w111}+\text{I2 * w121}+\text{b1}\right)$$

$$\text{L2}=\text{a}\left(\text{I1 * w112}+\text{I2 * w122}+\text{b2}\right).$$

The activation function a is used here. The activation function a is the sigmoid function in this example. Alternatively or additionally, it can also be another function, for example the identity function or the ReLu function.

If equations 12-14 are combined, the result is: $$\text{O}=\text{a}\left(\text{I1 * w111}+\text{I2 * w121}+\text{b1}\right)*\text{w21\_}+\text{a}\left(\text{I1 * w112}+\text{I2 * w122}+\text{b2}\right)*\text{w22\_}$$

In short form: $$\text{O1}=\text{f}\left(\text{I1},\text{I2}\right).$$

wobei Gleichung 17 unter Variation der Eingangsgrößen I1 und I2 für die Ausgangsgröße O gelöst wird, zum Beispiel mithilfe eines numerischen Lösungsverfahrens. Die Eingangsgrößen I1, I2 sind um Umkehrmodell U2 Ausgangsgrößen. Die Ausgangsgröße O ist im Umkehrmodell U2 eine Eingangsgröße.Equation (16) cannot be clearly solved. It can be shaped so that the reverse model U2 results in 2 B is shown. The reverse model U2 can be described mathematically as follows: $$0=\text{f}\left(\text{l1}\text{, l2}\right)-\text{O}\text{.}$$

where equation 17 varies the input variables I1 and I2 for the output size O is solved, for example using a numerical solution method. The input variables I1 , I2 are about reverse model U2 Output variables. The initial size O is in the reverse model U2 an input variable.

The input variables I1 , I2 correspond to production parameters ( PI1 , PI2 ). The initial size O corresponds to a product property ( PO ). A value for the output quantity O is used as the incoming value of the reverse model U2 fixed. It corresponds to a value of a product property that is desired or required, for example.

Values for the input variables I1 , I2 are called outgoing values of the reverse model U2 determined. They correspond to values of production parameters for which the reverse model U2 predicts that it will lead to the value of the product property.

• PI1: Menge Material 1,
• PI2: Menge Material 2,
• PI3: Dauer des Mischvorgangs,
• PI4: Leistungsaufnahme Mischer,
• PI5: Partikelgröße Material 1,
• PI6: Partikelgröße Material 2.

In 3a a flowchart of an exemplary embodiment of the method according to the invention is shown. In the first step S1 A production process that is to be described is carried out several times. The production process is a manufacturing process of a paste, which is a sub-step in the production of an accumulator. Six production parameters PI1-PI6 of the production process are considered. Examples are:

• PI1: amount of material 1 ,
• PI2: amount of material 2nd ,
• PI3: duration of the mixing process,
• PI4: power consumption mixer,
• PI5: particle size material 1 ,
• PI6: Particle size material 2nd .

• PO1: Viskosität der Paste,
• PO2: Homogenität der Paste.

Two product features PO1 , PO2 are considered. Examples are:

• PO1: viscosity of the paste,
• PO2: homogeneity of the paste.

The production process is the first step S1 carried out several times and here the value of each production parameter PI1-PI6 varies. The values of the production parameters PI1-PI6 can be varied, for example, within an interval of a relevant range of values. The production process can then be used in particular for points of support for the values of the production parameters PI1-PI6 , which are within the intervals.

On the basis of the copies of the product obtained, values of the product properties are shown PO1 , PO2 determined, for example by measurement and / or simulation. There is one record for each copy of the product D , the values of the production parameters PI1-PI6 and values of product properties PO1 , PO2 having.

The records D be in the second step S2 used for machine learning based on a given neural network 3rd that simplifies in 3b is shown to perform. That means in particular that values of weights and or or bias of the neural network 3rd be determined and / or adjusted. Because six production parameters PI1-PI6 and two product features PO1 , PO2 the neural network possesses 3rd six input variables I1-I6 (for values of the production parameters PI1-PI6 ) and two output variables O1 , O2 (for values of the product properties PO1 , PO2 ).

Once the neural network 3rd learned with the help of machine learning, it is a model of the production process, the relationship between the six production parameters PI1-PI6 and the two product features P01-P06 contains. In particular, the model of the production process for a value of at least one production parameter PI1-PI6 , which lies between two support points, in a determination of the values of the product properties PO1 , PO2 interpolate. The model of the production process is in step S3 validated by first different combinations of values of production parameters PI1-PI6 be specified, the model is executed in each case and the values of product properties PO1 , PO2 be determined. The determined values of product properties PO1 , PO2 can then, for example, with measured product properties of appropriately produced copies of the product and / or are compared with the results of corresponding simulations of the production process and / or checked for plausibility.

In one step S2a after the second step S2 and before the fourth step S4 the model of the production process is received, for example using a computer. This can in particular be carried out in such a way that the computer receives the model via an interface. This can be an internal interface - the model of the production process can, for example, be loaded into a working memory for use for processing - and / or an external interface, for example a network connection or a reading device, for example a reading device of a data carrier. As an alternative or in addition, “receiving” can include reading in the model or parts thereof. In particular, receiving the model enables the model to be executed and / or started.

$$0=\text{f}\left(\text{l1}\text{, l2}\text{, l3}\text{, l4}\text{, l5}\text{, l6}\right)-\text{O2}\text{.}$$

In the fourth step S4 the reversal model is created. Since that in 3b shown neural network 3rd is asymmetrical, the creation of the reversal model can be done, for example, in an analogous manner to that in 2a shown neural network 2nd is described. For example, there may be a mathematical description of the model that can be represented in simplified form as: $$0=\text{f}\left(\text{l1}\text{, l2}\text{, l3}\text{, l4}\text{, l5}\text{, l6}\right)-\text{O1}$$

$$0=\text{f}\left(\text{l1}\text{, l2}\text{, l3}\text{, l4}\text{, l5}\text{, l6}\right)-\text{O2}\text{.}$$

The equations can then be varied by varying values of the input variables I1-I6 (i.e. values of the production parameters PI1-PI6 ) for specified values of output variables O1 , O2 (i.e. predetermined values of the product properties PO1 , PO2 ) can be solved, for example using a numerical solution method. The result is a bevy of solutions of values of the production parameters PI1-PI6 .

• Entsprechende Werte der Produktionsparameter PI5 und PI6 können in den Gleichungen (19) und (20) für die Eingangsgrößen I5 und I6 eingesetzt werden. Die Gleichungen können dann unter Variation von Werten der Eingangsgrößen I1-I4 (also Werten der Produktionsparameter PI1-PI4) für vorgegebene Werte von Ausgangsgrößen O1, O2 (also vorgegebene Werte der Produkteigenschaften PO1, PO2) gelöst werden, zum Beispiel mithilfe eines numerischen Lösungsverfahrens. Resultat ist eine Schar von Lösungen von Werten der Produktionsparameter PI1-PI4.

In the fifth step S5 become two boundary conditions R1 , R2 set up: values for the production parameters PI5 and PI6 should be given, because the particle size is material 1 and the particle size material 2nd production parameters that cannot be influenced. The two boundary conditions R1 , R2 can be implemented in two ways:

• Corresponding values of the production parameters PI5 and PI6 can be found in equations (19) and (20) for the input variables I5 and I6 be used. The equations can then be varied by varying values of the input variables I1-I4 (i.e. values of the production parameters PI1-PI4 ) for specified values of output variables O1 , O2 (i.e. predetermined values of the product properties PO1 , PO2 ) can be solved, for example using a numerical solution method. The result is a bevy of solutions of values of the production parameters PI1-PI4 .

Alternatively or additionally, from the host of solutions of values of the production parameters PI1-PI6 (see above) the solutions can be selected without recalculating the boundary conditions in relation to the production parameters PI5 and PI6 fulfill.

In step S6 For example, a set of values of the production parameters is based on a criterion and / or on the basis of a predefined evaluation scheme PI1-PI4 selected to be used in a paste production.

The product property obtained when the paste was produced PO1 , the viscosity of the paste, can be used as a boundary condition in a subsequent production step. The subsequent production step is a step of coating a metal foil using the paste. The paste is filled in a first storage container, from which it is pumped into the coater and then applied to the metal foil by an application device, which moves along under the application device.

The subsequent production step is again based on a second model of the production process, the neural network 4th is designed and simplified in 4th is described. Machine learning has already been done, the neural network 4th has already been trained.

• PI7: Auftragungsdicke,
• PI8: Druck der Paste an der Auftragvorrichtung,
• PI9: Bahngeschwindigkeit,
• PI10: Viskosität der Paste.

Four production parameters PI7-PI10 of the production process are considered. Examples are:

• PI7: application thickness,
• PI8: printing the paste on the applicator,
• PI9: web speed,
• PI10: viscosity of the paste.

• PO3: Beschichtungsdicke.

A product feature PO3 is considered. For example, this is:

• PO3: coating thickness.

Because four production parameters PI7-PI10 and a product feature PO3 the neural network possesses 4th four input variables I7-I10 (for the production parameters PI7-PI10 ) and an output variable O3 (for the product property PO3 ). A reversal model can correspond to the neural network 3rd (please refer 3b) be generated.

If the paste in the storage container is now used up in production and switched to another storage container which contains paste of a different viscosity, the viscosity of the paste on the application device changes over a period of time while coating processes are being carried out. Using a parallel and repeated application of the second model of the production process, the coating thickness (product property PO3 ) are specified and kept approximately constant. For this, the currently measured value of the viscosity of the paste (production parameters PI10 ) as a boundary condition for the input variable I10 considered.

• PI1: Menge Material 1,
• PI2: Menge Material 2,
• PI3: Dauer des Mischvorgangs,
• PI4: Leistungsaufnahme Mischer,
• PI5: Partikelgröße Material 1,
• PI6: Partikelgröße Material 2,
• PI7: Auftragungsdicke,
• PI8: Druck der Paste an der Auftragvorrichtung,
• PI9: Bahngeschwindigkeit.

In 5 becomes another neural network 5 shown that the functions of the neural network 3rd ( 3b) and the neural network 4th ( 4th ) partially combined. It has nine input variables I1-I9 (for values of the production parameters PI1-PI9 ) and an output variable O3 (for product property values PO3 ) on. The nine production parameters PI1-PI9 are:

• PI1: amount of material 1 ,
• PI2: amount of material 2nd ,
• PI3: duration of the mixing process,
• PI4: power consumption mixer,
• PI5: particle size material 1 ,
• PI6: Particle size material 2nd ,
• PI7: application thickness,
• PI8: printing the paste on the applicator,
• PI9: path speed.

• PO3: Beschichtungsdicke.

The initial size PO3 is:

• PO3: coating thickness.

Regarding the neural network 5 It should be noted that product properties of the paste, which is an intermediate product, are not directly as product properties in the neural network 5 are taken into account: The viscosity of the paste occurs in the neural network 5 not on. The neural network 5 In other words, it is a common model of two successive production processes.

The viscosity of the paste occurs in the neural network 5 not on and also not in its reverse model. In the reverse model, only the coating thickness (output size PO3 ) specified. Values of the production parameters become the boundary conditions PI5 and PI6 specified (particle size material 1 and particle size material 2nd ). The initial parameter is the coating thickness (product property PO3 ).

6 schematically shows an embodiment of an arrangement 61 to carry out the procedure. The order 61 can be realized in particular by a computer or a computer network.

By arrows in the left 6 It is shown that information, in particular data, that describe the model of the production process via an interface 71 the arrangement 61 are fed can. Alternatively or additionally, at least some of the information that describes the model can be input means 67 which have, for example, at least one keyboard and / or a pointing device (e.g. computer mouse or trackball) in the arrangement 61 , especially the computer. The input means 67 are like arrows in below and to the right 6 indicated with the arrangement 61 connected.

In the bottom right 6 is also a sensor 69 schematically represented, which generates measured values during its operation and the arrangement 61 feeds.

In the specifically illustrated embodiment of the arrangement 61 this has a data memory 65 and a computing device 63 on. The computing device 63 can have one or more computing units such as CPU (central data processor unit) and / or GPU (graphics processor unit). Shown in 6 is that the computing device 63 includes a reverse model, for example the reverse model U2 . The data store 65 is on the input side with the interface 71 connected and on the output side to the computing device 63 , the computing device 63 also data in the data storage 65 can drop. Right in the 6 An arrow is shown, which indicates that the arrangement 61 and in particular the computing device 63 a result of applying the reverse model U2 can spend.

In particular, in 6 overall arrangement shown 61 during their operation via the interface 71 Receive information that clearly and completely defines a model of a production process. This information can be stored as data in the data store 65 be filed. The model is, for example, a taught-in neural network in which, for example, weights and / or other variables and / or constants are already known from the teaching.

In a next step, the computing device 63 access this data and from the model the reverse model U2 produce. This step can be done during a step of applying the reverse model U2 and / or executed before. The reversal model can optionally be in the form of suitable data in the data memory 65 get saved. This enables the computing device 63 later, for another application of the reverse model U2 to access it again.

Alternatively, this can be done via the interface 71 received model has not yet been trained or needs to be further approximated or validated. This function can be the arrangement 61 under the control of the computing device 63 take over. Information, in particular data, for teaching and / or validation can be sent via the interface 71 and / or the input means 67 and / or the at least one sensor 69 the arrangement 61 are fed.

When using the reverse model U2 Optionally, information about at least one boundary condition can be taken into account, this information for example about the input means 67 can be entered. Alternatively or additionally, the at least one sensor delivers 69 for example at least one value of a production parameter, such as an ambient temperature, as a boundary condition. By using the reverse model U2 at least one value of at least one production parameter of the production process is determined and output.

Reference symbol list

1

neural network

2nd

neural network

3rd

neural network

4th

neural network

5

neural network

61

arrangement

63

Investigative facility

65

Data storage

67

Input device

69

sensor

71

Input interface

b1

Bias

b2

Bias

D

record

I1-I10

Input variable

L1

Intermediate output size

L2

Intermediate output size

N11

Neuron

N12

Neuron

N21

Neuron

N22

Neuron

N3

Neuron

N31

Neuron

N32

Neuron

O

Output size

O1-O3

Output size

PI1-PI6

Production parameters

PO

Product property

PO1, PO2

Product property

R1, R2

boundary condition

S1-S6

steps

S2a

step

U1

Reverse model

U2

Reverse model

w111

Weight

w112

Weight

w121

Weight

w122

Weight

w211

Weight

w212

Weight

w221

Weight

w222

Weight

QUOTES INCLUDE IN THE DESCRIPTION

This list of documents listed by the applicant has been generated automatically and is only included for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Patent literature cited

• WO 2015/184729 A1 [0012]

Claims (10)

Method for determining values of production parameters (PI1-PI2; PI1-PI4; PI7-PI9; PI1-PI4, PI7-PI9) of a production process, during the execution of which a product is generated, the method having the following steps (S2a, S4) : - Receiving a model (1; 2; 3; 4; 5) of the production process that is a result of machine learning, wherein at least one production parameter (PI1-PI2; PI1-PI4; PI7-PI9; PI1-PI4, PI7-PI9 ) Input variable (I1-I2; I1-I4; I7-I9; I1-I4, I7-I9) of the model (1; 2; 3; 4; 5) and at least one product characteristic (PO1, PO2; PO3) output variable ( O1, O2; O; O3) of the model; - Determining at least one value of at least one production parameter (PI1-PI2; PI1-PI4; PI7-PI9; PI1-PI4, PI7-PI9) using a reversal model (U1, U2) of the production process, the reversal model (U1, U2) being a reversal of the model (1; 2; 3; 4; 5) of the production process and at least one value of the at least one product property (PO1, PO2; PO3) is included in the reversal model (U1, U2) and the at least one value of the at least one production parameter ( PI1-PI2; PI1-PI4; PI7-PI9; PI1-PI4, PI7-PI9) is obtained from the reverse model (U1, U2). Procedure according to Claim 1 , characterized in that the method comprises the following additional step (S2) which is carried out before the step (S2a) of receiving the model (1; 2; 3; 4; 5) of the production process: - performing machine learning in order to to learn the model of the production process on the basis of data (D), the at least one value of the at least one production parameter (PI1-PI2; PI1-PI4; PI7-PI9; PI1-PI4, PI7-PI9), the input variable (11-12; 11 -14; I7-I9; 11-14, I7-I9) of the model (1; 2; 3; 4; 5), and at least one value of the at least one product property (PO1, PO2; PO3), the output variable (O1 , O2; O; O3) of the model (1; 2; 3; 4; 5) is included. Procedure according to one of the Claims 1 - 2nd , characterized in that the method has the following additional step (S3) which, prior to step (S4), of determining the at least one value of the at least one production parameter (PI1-PI2; PI1-PI4; PI7-PI9; PI1-PI4, PI7-PI9) is carried out: - Determining at least one value of at least one product property (PO1, PO2; PO3) with the aid of the model (1; 2; 3; 4; 5) of the production process in order to obtain the model (1; 2; 3; 4 ; 5) validate the production process, with at least one value of the at least one production parameter (PI1-PI2; PI1-PI4; PI7-) as input variable (11-12; 11-14; I7-I9; 11-14, I7-I9). PI9; PI1-PI4, PI7-PI9) is specified and at least one value of the at least one product property (PO1, PO2; PO3) is output as the output variable (O1, O2; O; O3). Procedure according to one of the Claims 1 - 3rd , characterized in that the method has the following additional step (S5): - Specifying at least one default value of at least one production parameter (PI5, PI6, PI10), the input variable (I5, I6, I10) of the model (1; 2; 3; 4; 5) of the production process, as a boundary condition (R1, R2) in the reversal model (U1, U2), so that the at least one value of the at least one production parameter (PI1-PI2; PI1-PI4; PI7-PI9; PI1-PI4, PI7 -PI9), which is obtained from the reversal model (U1, U2), is determined as a function of the at least one default value. Procedure according to Claim 4 , characterized in that the at least one further production parameter (PI5, PI6, PI10) is a material parameter and / or an environmental parameter and / or a machine parameter and / or a process parameter and / or a property of a product that is used in the production process. Procedure according to one of the Claims 1 - 5 , characterized in that the method for a production process in which at least one intermediate product arises between the beginning and the end of the production process is carried out in such a way that properties (PO1) of the intermediate product are not used as input variables of the reversal model (U1, U2). Procedure according to one of the Claims 1 - 6 , characterized in that the at least one determined value of the at least one production parameter (PI1-PI2; PI1-PI4; PI7-PI9; PI1-PI4, PI7-PI9) is checked to determine whether it fulfills a criterion and / or using one predefined evaluation scheme is evaluated. Procedure according to one of the Claims 1 - 7 , characterized in that the model (1; 2; 3; 4; 5) of the production process is set up as an artificial neural network (1; 2; 3; 4; 5). Procedure according to one of the Claims 1 - 8th , characterized in that the product is an energy store, in particular an accumulator, or a component of an energy store, in particular an accumulator, and that at least one value of a production parameter (PI1-PI2; PI1-PI4; PI7-PI9; PI1-PI4, PI7-PI9) of the energy store or of the component of the energy store is determined using the reversal model (U1, U2). Arrangement (61) designed to carry out a method for determining values of production parameters (PI1-PI2; PI1-PI4; PI7-PI9; PI1-PI4, PI7-PI9) of a production process, during the execution of which a product is produced, the Arrangement has: - an interface (71) designed to receive the model of the production process, which is a result of machine learning, at least one production parameter (PI1-PI2; PI1-PI4; PI7-PI9; PI1-PI4, PI7-PI9) input variable ( 11-12; 11-14; I7-I9; 11-14, I7-I9) of the model (1; 2; 3; 4; 5) and at least one product property (PO1, PO2; PO3) output variable (O1, O2 ; O; O3) of the model; - A determination device (63) configured to determine at least one value of at least one production parameter (PI1-PI2; PI1-PI4; PI7-PI9; PI1-PI4, PI7-PI9) with the aid of a reversal model (U1, U2) of the production process, the Reversal model (U1, U2) is a reversal (1; 2; 3; 4; 5) of the model of the production process and the determining device (63) is designed to use the reversal model (U1, U2) in such a way that the at least one value of at least one product property (PO1, PO2; PO3) enters the reversal model (U1, U2) and the at least one value of the at least one production parameter (PI1-PI2; PI1-PI4; PI7-PI9; PI1-PI4, PI7-PI9) the reverse model (U1, U2) is obtained.

Images