DE102019212171A1 - Method and device for optimizing an inductive heat treatment process using a Bayesian optimization process - Google Patents

Abstract

The invention relates to a method for determining process parameters for a heat treatment process of a component, in particular for hardening a workpiece (2), the process parameters being determined using a Bayesian optimization method with the following steps: Provision of a trainable quality function that includes a Gaussian process model and Map process parameters to quality measures that each indicate a quality of a component with regard to one or more component properties, - during the implementation of the Bayesian optimization process, measurement (S2) of components that are subjected to a heat treatment process according to test process parameters, with regard to one or more several component properties in a measurement method, - training (S4) or updating the quality function based on the result of the measurement method for determining the one or more component parameters, - optimizing (S5) one Acquisition function that depends on at least one parameter of the quality function in order to obtain a next test process parameter set for measuring the one or more component properties of a correspondingly manufactured component, and selecting (S7) the process parameters by determining an optimized quality measure by optimizing the quality function .

Description

Technical area

The invention relates to inductive heat treatment processes, and in particular to methods for optimizing process parameters when using a heat treatment process, in particular for hardening a metallic workpiece. In particular, the present invention relates to measures for the selection of process parameters for the heat treatment process, which leads to desired material properties, in particular to a desired degree of hardness and a desired material composition.

Technical background

Heat treatment processes are used to harden metallic materials. For example, in inductive heat treatment, an alternating magnetic field is generated with the aid of an inductor. The alternating magnetic field induces eddy currents in the workpiece to be hardened, which locally lead to an increase in the temperature in the workpiece. The induction hardening process enables component properties to be adapted, such as the hardening depth, SHD and the like.

On the one hand, variations in the alloy composition and different structural characteristics of the starting material of workpieces due to the resulting distribution of the local electrical resistance lead to fluctuations and variations in the process temperatures and thus to sometimes considerable variations in the final component properties. On the other hand, if the material, the component geometry and / or the processing tools are changed, extensive re-parameterization of the heat treatment process is necessary.

To parameterize a heat treatment process, extensive input tests are carried out for each component geometry and material in order to determine a stable process window for the process parameters of the heat treatment process. Experiences from previous measurements of other workpiece geometries and materials can only be partially taken into account. This increases the time required to parameterize the heat treatment process for a component type, since a single measurement of a workpiece is very time-consuming.

Disclosure of the invention

According to the invention, a method for optimizing and performing a heat treatment process by determining process parameters with the aid of a Bayesian optimization method according to claim 1 and a corresponding device and a heat treatment system according to the independent claims are provided.

Further refinements are given in the dependent claims.

• - Bereitstellen einer trainierbaren Qualitätsfunktion, die ein Gaußprozessmodell umfasst und Prozessparameter auf Qualitätsmaße abbilden, die jeweils eine Qualität eines Bauteils hinsichtlich einer oder mehrerer Bauteileigenschaften und/oder Prozesseigenschaften angeben,
• - während des Bayes'schen Optimierungsverfahrens Vermessen von Bauteilen, die entsprechend Test-Prozessparameter einem Wärmebehandlungsprozess unterzogen werden, hinsichtlich der einen oder den mehreren Bauteileigenschaften in einem Vermessungsverfahren,
• - Trainieren oder Aktualisieren der Qualitätsfunktion basierend auf dem Ergebnis des Vermessungsverfahrens zum Bestimmen des einen oder der mehreren Bauteilparameter,
• - Optimieren einer Akquisitionsfunktion, die von mindestens einem Parameter der Qualitätsfunktion abhängt, um jeweils einen nächsten Test-Prozessparametersatz zur Vermessung der einen oder der mehreren Bauteileigenschaften eines entsprechend hergestellten Bauteils zu erhalten, und
• - Auswählen der Prozessparameter durch Bestimmen eines optimierten Qualitätsmaßes durch Optimieren der Qualitätsfunktion.

According to a first aspect, a method for determining process parameters for a heat treatment process of a component is provided, in particular for hardening the component, the process parameters being determined using a Bayesian optimization method with the following steps:

• - Provision of a trainable quality function that includes a Gaussian process model and that map process parameters to quality measures that each indicate a quality of a component with regard to one or more component properties and / or process properties,
• - during the Bayesian optimization process, measurement of components that are subjected to a heat treatment process in accordance with test process parameters, with regard to the one or more component properties in a measurement process,
• - Training or updating the quality function based on the result of the measurement method for determining the one or more component parameters,
• - Optimizing an acquisition function that depends on at least one parameter of the quality function in order to obtain a next test process parameter set for measuring the one or more component properties of a correspondingly manufactured component, and
• - Selecting the process parameters by determining an optimized quality measure by optimizing the quality function.

A heat treatment process usually provides for a heat treatment of a workpiece over a total period of time, the temperature profile following a predetermined trajectory during the entire period of time. The temperature trajectory is chosen so that certain component properties are to be achieved. The process parameters for setting the temperature trajectory are parameterized in accordance with a result of a Bayesian optimization method based on a trainable quality function in the form of a Gaussian process. The quality function forms the process parameters of the heat treatment process to a relevant quality measure, the component parameters such as the structure composition and / or process properties (such as the process duration) of the manufacturing process. In addition, requirements for the process result, such as a minimum hardness of the component to be hardened, can be taken into account as secondary conditions. These secondary conditions are taken into account in the optimization process with further Gaussian process models.

Since a larger number of process parameters of the heat treatment process can be optimized with the aid of the Bayesian optimization method than can be done manually with justifiable effort, the material properties of hardened workpieces can be significantly improved through an improved heat treatment process. In addition, the unpredictable influences of the manual setting of process parameters on the material properties, and in particular their variability, can be reduced.

Since the results of test procedures through which the material properties are determined experimentally or physically after a heat treatment process in a measurement procedure, the process parameters can be automatically optimized with the aid of reinforcement learning, which is carried out by the Bayesian optimization procedure. This enables higher quality workpieces with regard to their material properties.

• - eine Zeitdauer für eine Bearbeitungsphase;
• - einen Zeitpunkt des Wechselns zwischen Bearbeitungsphasen;
• - eine Bearbeitungstemperatur in einer Bearbeitungsphase; und
• - eine Leistungszufuhr zum Erwärmen des Bauteils in einer Bearbeitungsphase.

Furthermore, the heat treatment process can be divided into several successive processing phases, the process parameters specifying one or more of the following parameters:

• - a period of time for a processing phase;
• - a point in time of switching between processing phases;
• a machining temperature in a machining phase; and
• - A power supply for heating the component in a processing phase.

According to one embodiment, the value of the quality measure specified by the quality function can be higher with a higher quality of the component than with a lower quality, the test process parameters being determined by maximizing the acquisition function, or
the value of the quality measure specified by the quality function can be lower in the case of a higher quality of the component than in the case of a lower quality, the test process parameters being determined by minimizing the acquisition function.

Furthermore, the quality measure can be determined from an evaluation measure of a structural composition and / or an overall process duration.

The acquisition function can take into account at least one constraint, a trainable constraint function, which in particular contains a further Gaussian process model, is measured with respect to a component property relating to the constraint in the measurement method and then trained, the acquisition function being determined depending on a probability that the constraint is determined by the respective process parameter is met, is determined.

In particular, the component property of the secondary condition can correspond to a degree of hardness or a hardness depth.

It can be provided that the quality function takes into account a predefined priority function, which is dependent on process parameters, from a previous optimization process.

According to one embodiment, several measurement methods, in particular a destructive measurement and a non-destructive measurement, can be provided, the measurement method being selected based on the acquisition function, which is selected in particular as an information-theoretical acquisition function, and depending on the costs associated with the measurement method

It can be provided that the quality function depends on a basic function, which indicates a known relationship between process parameters and the quality measure, and the Gaussian process model.

Figure list

• 1 eine schematische Darstellung eines Wärmebehandlungssystems;
• 2 eine beispielhafte Darstellung für eine Temperaturtrajektorie, die durch Prozessparameter definiert wird;
• 3 ein Flussdiagramm zur Veranschaulichung eines Verfahrens zur Bestimmung von Prozessparametern für den Wärmebehandlungsprozess; und

Embodiments are explained in more detail below with reference to the accompanying drawings. Show it:

• 1 a schematic representation of a heat treatment system;
• 2 an exemplary representation for a temperature trajectory that is defined by process parameters;
• 3 a flowchart to illustrate a method for determining process parameters for the heat treatment process; and

Description of embodiments

1 shows a schematic representation of a heat treatment system 1 for performing heat treatment on a workpiece 2 . The heat treatment system 1 has a heat treatment device 3 on that a workpiece holder 4th to hold the workpiece 2 (Component). The heat treatment facility 3 is with a control unit 5 connected, which controls a heat treatment process.

The control unit 5 carries out the method described below and controls the heat treatment device for this purpose 3 in order to carry out a heat treatment process with a specified workpiece or component when the process parameters are specified.

With the help of a test facility 6th component properties such as the hardness depth, SHD and the like can be determined. The resulting test results are sent to the control unit as corresponding measured variables 5 transmitted so that it can optimize the process parameters that determine the course of the heat treatment process.

The aim of the method described below is to optimize the process parameters for performing a heat treatment process. The process parameters should be selected so that the components treated with the heat treatment process determined thereby have one or more specified component properties.

The process parameters can define a heat treatment trajectory which specify a profile of a temperature at a specific component position or a profile of an energy supply by the heat treatment device. For example, in 2 a temperature profile with different temperature plateaus T1 , T2 , T3 shown, which are each retained for time periods predetermined by the times t1, t2, t3. The values of the different temperatures T1 , T2 , T3 and the periods of time or times t1, t2, t3 correspond to variable process parameters that are to be optimized by the method described below.

In general, Bayesian optimization is used when an unknown system behavior f, e.g. according to a so-called "black box" function, is to be optimized. This system behavior f can only be evaluated and observed (possibly affected by noise) for a value x. The observed value y results as y = f (x) + e, where e denotes the noise. In addition, it is assumed that every evaluation or measurement of the system behavior f is expensive, ie it causes costs, in the sense that the implementation of a test method for measuring the system behavior causes a lot of effort, as is the case, for example, when executing a test method in the test facility of Case is. Due to the "expensive" measurement of the system behavior, it is desirable that only a few measurements have to be carried out for the optimization or that the costs for the measurements (especially determined by their time and material expenditure) are as low as possible.

Under certain assumptions, such as the continuity of the unknown function, the quality function can be approximated with a Gaussian process regression in a Gaussian process model or a function model. A Gaussian process is a universal function approximator that is used as a surrogate function for the unknown quality function. Usually numerical optimization algorithms are based on many, very cheap evaluations / measurements of the optimization functions. However, if the function evaluations are complex, such as B. performing one of the above test procedures, one can no longer use the conventional optimization algorithms. Instead, the Bayesian optimization method is used, which includes a modeling of the quality function in the form of a surrogate function. The quality function describes the behavior of the system and gives a quality value depending on the parameters with which the system is operated.

For this purpose, after measuring the quality function at several evaluation points, i.e. test process parameters and observing the corresponding function values, i.e. the respective quality measures, a model of the quality function can be set up using the Gaussian process. A property of the Gaussian process is that in areas in the vicinity of the measured test process parameters the model prediction is very good and the quality function is approximated well. This is reflected in a low level of uncertainty in the functional model. Far away from evaluation points, the model predictions via the quality function can become poor. In any case, the uncertainty increases as the distance to the measured test process parameters increases, since there are no observations for the areas of the measured test process parameters.

A possible strategy to optimize the quality function is to evaluate the quality function at many different places (e.g. on a regular grid) and to find the lowest / highest observed function value ( Quality measure that indicates the highest possible quality with regard to the assessed property (s) to be assumed as the result of the optimization. However, this approach is inefficient and many measurement processes with the test procedures are necessary with a correspondingly high level of effort in order to find the optimum. If several quality functions are relevant for the given optimization task, such as thermal conductivity and processability, several Gaussian process models can be used.

Instead of the above approach, a Gaussian process model is used as a quality function within the scope of a Bayesian optimization method in order to select new test process parameters for measuring the heat treatment process. A new test process parameter set for measuring the heat treatment process is selected in such a way that, on the one hand, this improves the Gaussian process model of the quality function, so that the overall estimation uncertainty of the Gaussian process is reduced. For this purpose, the test process parameters are usually selected in areas in which the quality function has not yet been evaluated (exploration). On the other hand, the new test process parameters for measuring the quality function are selected in such a way that the goal of optimizing the quality function, i.e. maximizing it, is achieved as quickly as possible or with the lowest possible number of measurements with the test process parameters. For this purpose, a test process parameter set is preferred, which promises high quality measures based on the quality function (exploitation). These two contradicting criteria are weighed up by a so-called predetermined acquisition function.

In the Bayesian optimization method, the acquisition function is used to optimize the measurement of the system behavior to determine the quality function in such a way that the quality function does not necessarily converge quickly to the lowest uncertainty, but rather that the highest possible informational value about the position of the optimum, i.e. H. have the material parameter set at which the highest possible quality value can be achieved.

The acquisition function generally uses parameters of the quality function that is described by a Gaussian process model, such as the Gaussian process mean value µ (x) and the Gaussian process standard deviation σ (x). Examples are the so-called Lower Confidence Bound (LCB) acquisition function or the Upper Confidence Bound (UCB) acquisition function, which are described as follows: LCB (x) = µ (x) - kσ (x) or UCB (x) ) = µ (x) + kσ (x). In practice, the exploration factor k is often set constant, for example at a certain value, such as k = 2. This new criterion can be efficiently minimized or maximized with common gradient-based methods and the minimum of LCB (x) or the maximum from UCB (x) then forms the new test process parameters for the quality function. It should be noted here that an optimization domain must be defined for the optimization of the acquisition function in which the next test process parameters are searched for. This domain is typically chosen on the basis of experience and / or expert knowledge and / or technical / physical specifications. For example, technical limits can specify a maximum power / temperature of the induction coil or a maximum process duration.

In addition to the above acquisition functions, other acquisition functions are known, such as Expected Improvement (EI), Probability of Improvement (PI) or so-called entropy search methods that are based on information-theoretical considerations.

The method for optimizing the material parameters is described below using the flow chart of FIG 3 explained in more detail. The method uses a Bayesian optimization method.

In step S1 a set of process parameters is initially provided with which a first heat treatment process is to be carried out. The initial process parameters can be specified arbitrarily or in accordance with previously known expert knowledge.

In step S2 a heat treatment process is carried out on the component to be treated and one or more component properties are determined. The one or more component properties can include, for example, a structural composition, a process duration or the like.

In step S3 a quality measure is determined depending on the measured component property. This can be done in particular by forming the difference to a target value for a desired component property. This setpoint value can correspond, for example, to a specification of an optimal property of the component. Several component properties can also be taken into account when determining the quality measure. The quality measure should indicate a variable value for specifying a quality of the product obtained through the heat treatment process, e.g. in a value range between 0 and 1, where 0 can indicate a low quality and 1 an optimal quality of the component with regard to one or more component properties.

A data record for training a quality function is determined from the process parameters and the assigned quality measure.

In step S4 the quality function, which is specified as a Gaussian process model, is trained or updated based on the data sets previously determined by the heat treatment processes.

In step S5 an acquisition function is created which depends on the quality function learned as a Gaussian process, the acquisition function being determined in particular on the basis of the Gaussian process mean value μ (x) and the Gaussian process standard deviation σ (x). The acquisition function is optimized so that the process parameters x, which correspond to an optimum (for example a maximum or minimum) of the functional value of the acquisition function, provide test process parameters as the next measurement point.

In the present case, one or more secondary conditions can be taken into account. A secondary condition can be specified, for example, by a minimum degree of hardness or a minimum hardness depth. For example, one or more component properties that relate to a secondary condition can be modeled according to further Gaussian process models. This additional Gaussian process model can be trained after the component properties have been measured, depending on the underlying process parameters. The further Gaussian process model can be evaluated according to a secondary condition criterion. The evaluation can take place by determining the respective probability that the secondary condition criterion is met with the relevant process parameters. The further Gaussian process model can also be updated with the newly generated data set.

The acquisition function can then take into account the probabilities of one or more such secondary conditions as a factor, so that the fulfillment of the secondary condition criterion weights the acquisition function.

For example, the acquisition function AF used can be based on a UCB acquisition function and multiplied by the probability of fulfilling the secondary condition NB as follows: AF (x) = (u (x) + kσ (x)) _* P (NB (x) = true), where x corresponds to the process parameter set. The acquisition function AF is maximized in order to determine the next test process parameter set.

In step S6 a termination condition is checked. If the abort condition is met, the optimization of the process parameters is ended (alternative: yes) and the procedure with step S7 continued. Otherwise (alternative: no) becomes step S2 bounced back. The termination condition can relate to the period of time that is to be expended for the optimization of the quality function, or to the number of iterations or to a suitable convergence criterion.

In step S7 the heat treatment process is carried out with optimized process parameters. In particular, the optimized set of process parameters can be selected by maximizing the quality measure. It can be provided that, after the evaluation of the last experiment with process parameters that are optimal according to the model prediction, that process parameter set is used as the result of the optimization process that gave the best value of the quality function from the experiments carried out.

According to a further embodiment, a prior function from a previous optimization process can be used as mean prior mp (x) and taken into account in the quality function QF. The Gaussian process model of the quality function is then not trained on the quality measure, but on the difference in the quality measure for predicting the prior function mp (x). This implicitly assumes that the new heat treatment process (apart from affine transformations) behaves like the prior function. However, this assumption is locally canceled by deviating current measurements. This approach is particularly well suited for knowledge transfer when the quality function deviates locally from the prior function, but shares the same characteristics, e.g. the change in the rate of structural transformation with a changed initial structure or an effect that only occurs at certain temperatures when the initial structure changes. In such cases, the number of experiments required is significantly reduced by using such a prior.

Furthermore, for evaluating an experiment, the cycle time for determining the optimized set of process parameters can be reduced by providing various measurement methods. In particular, the measurement times for a non-destructive test are usually considerably shorter than for a destructive test. the cost of determining the optimized parameter set. Alternatively, the use of materials for evaluating an experiment can also be taken into account. The quality function is trained in such a way that the information content of the results of the various tests (destructive and non-destructive measurement) is taken into account.

In this case, the acquisition function is an information-theoretical acquisition function suitable, e.g. probabilistic Entropy Search or Max Value Entropy Search. In order to decide whether a destructive or non-destructive test should be carried out for the next survey, the expected change in entropy per unit of cost is maximized. For this purpose, “costs” of the destructive and non-destructive test are defined at the beginning of the optimization. These can include, for example, required working time or material costs. The arg max found defines the process parameters for the next measurement (including the selection of the measurement method). In addition, the value of the acquisition function α can be normalized with the costs in order to maximize the information gain per unit cost, ie the combined acquisition function becomes a _{d, nd} (9) = α (θ) / c (θ _e ), where c (θ _e ) are the costs of the destructive or non-destructive testing. The acquisition function is now maximized, ie θ * = arg max α _{d, nd} (θ). The arg max θ * found again defines the next experiment (including the selection of the measurement method, which is included as the last entry in θ *).

As a supplement to the existing approach of realizing the modeling of the process parameters using a Gaussian process as a quality function, it is intended to use a hybrid model with explicit basic functions. For some of the process parameters there is a first approximation known, such as a linear relationship between process parameters and the quality measure. For this purpose, the model of the quality function can be divided into a generalized linear regression model h (x) = w * phi (x), where w are the weights and phi (x) are so-called feature functions (e.g. polynomial functions) and a Gaussian process model g (x) can be divided so that a total quality function QF (x) = h (x) + g (x) results. The inference in this overall model takes place according to §2.7 in CE Rasmussen & CKI Williams, "Gaussian Processes for Machine Learning," the MIT Press, 2006, ISBN 026218253X, 2006 Massachusetts Institute of Technology .

In simplified terms, one can also first perform a linear regression independently of one another and then use this as the mean prior for the Gaussian process.

QUOTES INCLUDED IN THE DESCRIPTION

This list of the documents listed by the applicant was generated automatically and is included solely 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.

Non-patent literature cited

• C. E. Rasmussen & C. K. I. Williams, "Gaussian Processes for Machine Learning", the MIT Press, 2006, ISBN 026218253X, 2006 Massachusetts Institute of Technology [0048]

Claims (14)

Method for determining process parameters for a heat treatment process of a component, in particular for hardening a workpiece (2), the process parameters being determined using a Bayesian optimization method with the following steps: - Provision of a trainable quality function that includes a Gaussian process model and that map process parameters to quality measures, each of which indicates a quality of a component in terms of one or more component properties, - during the implementation of the Bayesian optimization process, measurement (S2) of components that are subjected to a heat treatment process according to test process parameters, with regard to the one or more component properties in a measurement process, - Training (S4) or updating the quality function based on the result of the measurement method for determining the one or more component parameters, - Optimizing (S5) an acquisition function which depends on at least one parameter of the quality function in order to obtain a next test process parameter set for measuring the one or more component properties of a correspondingly manufactured component, and - Selecting (S7) the process parameters by determining an optimized quality measure by optimizing the quality function. Procedure according to Claim 1 , wherein the heat treatment process is divided into several temporally successive processing phases, the process parameters specifying one or more of the following parameters: a time duration for a processing phase; - a point in time of the change between machining phases - a machining temperature in a machining phase; and - a power supply in a processing phase. Procedure according to Claim 1 or 2 , wherein the value of the quality measure specified by the quality function is higher for a higher quality of the component than for a lower quality, the test process parameters being determined by maximizing the acquisition function, or wherein the value of the quality measure specified by the quality function is lower for a higher quality of the component than with lower quality, the test process parameters being determined by minimizing the acquisition function. Method according to one of the Claims 1 to 3 , the quality measure being determined by an evaluation measure of a structural composition and / or an overall process duration. Method according to one of the Claims 1 to 4th , wherein the acquisition function takes into account at least one secondary condition, wherein a trainable secondary condition function, which in particular contains a further Gaussian process model, is measured with respect to a component property relating to the secondary condition in the measurement method and then trained, the acquisition function being determined depending on a probability that the secondary condition is met by the respective process parameters are met, is determined. Procedure according to Claim 5 , the component property of the secondary condition corresponding to a degree of hardness or a hardness depth. Method according to one of the Claims 1 to 6th , the quality function taking into account a predefined priority function, which is dependent on process parameters, from a previous optimization process. Method according to one of the Claims 1 to 7th , several measurement methods, in particular a destructive measurement and a non-destructive measurement, being provided, the measurement method being selected based on the acquisition function, which is selected in particular as an information-theoretical acquisition function, and depending on the costs associated with the measurement method Method according to one of the Claims 1 to 8th , whereby the quality function depends on a basic function, which indicates a known relationship between process parameters and the quality measure, and the Gaussian process model. Method for performing a heat treatment process of a component, wherein the heat treatment process is carried out based on process parameters that correspond to the method according to one of the Claims 1 to 9 are preserved. Device, in particular a control unit (5), for determining process parameters for a heat treatment process of a component, in particular for hardening a workpiece (2), the device being designed to determine the process parameters using a Bayesian optimization method with the following steps: Providing a trainable quality function, which includes a Gaussian process model and map process parameters to quality measures, each one Specify the quality of a component with regard to one or more component properties, - during the implementation of the Bayesian optimization process, measuring components that are subjected to a heat treatment process according to test process parameters, with regard to the one or more component properties in a measurement process, - training or updating the Quality function based on the result of the measurement method for determining the one or more component parameters, - Optimizing an acquisition function that depends on at least one parameter of the quality function in order to assign a next test process parameter set for measuring the one or more component properties of a correspondingly manufactured component and - selecting the process parameters by determining an optimized quality measure by optimizing the quality function. Heat treatment system (1) with a control unit according to Claim 11 . Computer program with program code means which is set up to implement a method according to one of the Claims 1 to 9 execute when the computer program is executed on a computing unit. Machine-readable storage medium with a computer program stored thereon Claim 13 .

Images