WO2023083698A1

WO2023083698A1 - Method and device for operating an injection molding machine by means of machine learning

Info

Publication number: WO2023083698A1
Application number: PCT/EP2022/080727
Authority: WO
Inventors: Alexander Kroschel; Marcus Hlavac; Andreas Michalowski; Stephanie Karg; Adina Kerstin Dais; Matthias Musialek; Attila Reiss; Paul Sebastian Baireuther; Patrick Ganter
Original assignee: Robert Bosch Gmbh
Priority date: 2021-11-11
Filing date: 2022-11-03
Publication date: 2023-05-19
Also published as: DE102021212730A1

Abstract

The invention relates to a computer-implemented method for operating an injection molding machine (1), wherein an estimated result (y sim ) is determined as a function of predefined process parameters (x) to characterize how good an actual result (y exp ) of the injection molding will be, and wherein the process parameters (x) are varied using a Bayesian optimization process with the aid of a data-based model until an actual result (y exp ) of the injection molding is sufficiently good.

Description

title

METHOD AND APPARATUS FOR OPERATING AN INJECTION MOLDING MACHINE USING MACHINE LEARNING

The invention relates to a method for operating an injection molding machine, a test stand, a computer program and a machine-readable storage medium.

State of the art

Injection molding is an established and effective manufacturing process for manufacturing components of varying complexity from polymer materials (e.g. duroplastics, thermoplastics and elastomers). Due to the wide range of uses for components made of polymer materials (which are also referred to as "plastics") and the possibility of also producing components from different polymer materials (which is known, for example, as 2-component materials), the injection molding manufacturing process is widely used Industrial sectors strongly represented.

Process development is currently still strongly characterized by experimental tests on the machine, although many recursions in tool manufacture and component design could be avoided through the use of simulation techniques. This requires appropriate simulation techniques, which is challenging. The reason for this are the numerous dynamic and mutually interacting physical and chemical effects, which still cannot be modeled with sufficient accuracy. Furthermore, the material data and conditions of the workpiece and influences on the part of the injection molding system (e.g. wear on the screw) are often not known with sufficient accuracy. Although simplified models are available, with which Given material data and process parameters as well as in certain parameter ranges, a certain prediction of the injection molding result achieved (e.g. a degree of filling of the injection molding tool or a degree of shrinkage of the component) is possible. Reliable predictions of quality properties such as dimensional accuracy of the component are currently often not possible despite the simulation-supported tool design.

It turns out that the achievable injection molding result and the productivity of the injection molding process depend very much on the set process parameters and the workpiece material used as well as the ambient temperatures. The criteria used to measure the quality of an injection molded plastic component are numerous. The component dimensions achieved and complete material filling at all points of the component are often important. Productivity can typically be defined by the required process time, also known as the cycle time, and the resources consumed (mold tool wear, percentage of rejects and energy) per component.

Due to the many adjustable process parameters (e.g. temperatures of a mold and/or a melt, injection speed and holding pressure), which can often also be varied over time, optimizing the process settings is a time-consuming process that requires a lot of experiments. Because, on the one hand, many workpieces or components are required for these experiments and, on the other hand, the evaluation (in particular the dimensional and shape measurements (e.g. using tactile 3D measuring methods) as well as the mechanical testing of the produced components is complex, it is desirable to limit the number of required Try to keep it to a minimum.

It is possible to set some process parameters to values based on experience and vary only some of the parameters in experiments. Test series specified by experts and/or methods of statistical test planning can be used as a planning method for the tests. In this case, the actually achievable optimum is generally not found.

Advantages of the Invention The object with the features of independent claim 1 has the advantage that process parameters of injection molding machines can be found with just a few experiments, by means of which a high quality of injection molding can be achieved.

Further aspects of the invention are the subject matter of the independent claims. Advantageous developments are the subject of the dependent claims.

Disclosure of Invention

The invention relates to the way in which an efficient and targeted optimization of the process parameters can be carried out. The method of Bayesian optimization is used for this. This method can be used to find optima in unknown functions. In the context of this invention, the functions are so-called cost functions. A cost function K characterizes the quality of the process and/or the process result. The cost function depends on one or more quality properties (features) q _; and is set up in such a way that it assumes an optimum (in this context the optimum is usually a minimum) if the quality properties reach certain target values (also called target values) or target intervals (also called target intervals) q _i target specified by a user become. Several quality properties can be calculated in a cost function to get a single function to be optimized. This cost function must also be specified by the user. An example is the sum of scaled deviations from the respective target value:

The parameters Sj are specifiable scaling parameters. In order to find the optimum of the cost function, process parameter sets (hereinafter also referred to as parameter sets for short) can be proposed for the next experiment by applying Bayesian optimization. After After carrying out the experiment, the resulting values of the quality criteria and thus the current cost function value can be determined and made available to the optimization process as a data point together with the set parameter set.

The Bayesian optimization method is suitable for finding that parameter set for a function that maps a one-dimensional or multi-dimensional parameter space to scalar values, which leads to an optimal function value. Depending on the optimization goal, the optimum is defined as the largest possible or alternatively also the smallest possible achievable value that the function values can assume. In terms of process optimization, for example, the set of parameters is given by a specific set of process parameters; the associated function value can be determined using the cost function described above. In the case of the cost function, the lowest possible value is usually desirable.

Because experiments have to be carried out and evaluated to determine the function values of the cost function, there is basically only a value table with data available for the function, which also shows an experimental "noise". Because the experiments are very complex, this noise cannot normally be suppressed by numerous repetitions with the same parameter set and subsequent averaging of the results. It is therefore advantageous to carry out the optimization using a method which, despite fewer test evaluations, enables global optimization with good results and does not require a calculation of gradients of the cost function. It has been found that Bayesian optimization satisfies these properties.

Bayesian optimization consists of determining a mathematical model which, based on a given table of values, results in a prediction of the function value in the form of a probability distribution for each parameter set, in particular a prediction of the expected value and the variance of the function value, and an algorithmically formulated rule , for which parameter set a further function evaluation (here an Ex- periment) is to be carried out, which is based on the predictions of the mathematical model. The mathematical models are advantageously Gaussian processes. These are advantageous because they can include existing (expert) knowledge and because they are suitable for describing noisy data and because test or simulation data that are already available can be fed in directly. The mathematical basics of Bayesian optimization with Gaussian processes are presented below (however, other models, such as Bayesian linear regression, are of course also possible).

In concrete terms, the prediction for the result of the function evaluation with a parameter set x _N+1 is given by the expectation value, which in the Gaussian process is given by the mean value.

and by the variance

Here, C _N means the covariance matrix, which is given by

[Ostlnm ⁼ (x _n , x _m ) + ß 8 _nm , with n, m = 1. . N, (4) where x _n or x _m are parameter sets for which a function evaluation has already taken place. The size ß ^-1 represents the variance of the normal distribution, which describes the scatter of function evaluations with the same set of parameters, 8 _nm is the Kronecker symbol. The scalar c is conventionally given by c = /c(x _w+1 , x _w+1 ) + β ^-1 . The vector t contains the respective results for the individual parameter sets x _n (n=1../V) for which a function evaluation has taken place. The core function k(x _n , x _m ) describes the extent to which the result of the function evaluation with a parameter set x _n still has an influence on the result of the function evaluation with a parameter set x _m . Large values stand for a high influence, if the value is zero, there is no longer any influence.

For the prediction of the mean and the variance in the above formula, the vector k, with [k] _n = kx _n ,x _N+1 ), with respect to all parameter sets x _n (n = 1.. /V) and the parameter set x _w+1 calculated. For the concrete There are different approaches for the kernel function to be used, a simple approach is the following quadratic exponential kernel function:

with the selectable hyperparameters 0 _O and 0 _d (d = 1.. O), where D is the dimensionality of the parameter space. In this core function, 0 _O describes the scale on which the function values vary and 0 _d describes the influence of the "distance" in the parameter space on the correlation of two function values for the parameter sets x _n and x _m . Other core functions are possible.

The selection of the next set of parameters to test on is based on the mean and variance predictions calculated using the formulas above. Different strategies are possible here; for example that of “expected improvement”.

The parameter set for the next experiment is selected for which the expected value for finding a function value is greater (or smaller, depending on the optimization goal) than the largest (or smallest, depending on the optimization goal) known function value from the previous N iterations

so

becomes maximum. The possible function values f(x) at the point x are normally distributed with mean according to formula (2) and variance according to formula (3), each with x _w+1 = x. Such a function to be optimized is also referred to as an acquisition function. Other acquisition functions are possible, for example a knowledge gradient or an entropy search.

The "+" operator here means that only positive values are used and negative values are set to zero. Bayesian optimization is now iterative - a new test point (i.e. parameter set) is determined,

- performed an experiment

- evaluated the test result and calculated the (cost) function value,

- the Gaussian process is updated with the new value pair from parameter set and function value until the optimization is aborted.

The optimization of the Gaussian process with a new test point and the associated new function value is done, for example, in such a way that the new pair of test point and function value is added to the test data already recorded, consisting of pairs of test points and function values, and the hyperparameters are adjusted in such a way that a likelihood of the test data is maximized.

This process is illustrated in connection with FIG.

Through the iterative procedure of the steps described above (performing an experiment, evaluating the quality criteria and determining the cost function value, updating the Gaussian process and suggesting the next set of parameters), a process model (represented by the Gaussian process) can be built up successively. The best parameter set of all evaluated function evaluations or tests is then used as the best optimization result.

You gain advantages when carrying out the optimization by including existing process knowledge. With the procedure described below, knowledge in the form of one or more process models Pi _n can be included in the optimization by replacing real experiments with simulation experiments under certain conditions. It is conceptually irrelevant with which uncertainty the models depict the process and how many of the quality criteria they describe.

With a process model that would perfectly represent the real experiment, every real experiment could be replaced by a simulation experiment. If the evaluation duration were shorter than the actual implementation, time saved in addition to effort. In general, however, the prediction accuracy of the process models is limited. They are often only valid in a part of the parameter space and/or describe only a subset of the process results, and do not take into account all physical effects and therefore produce results only within an uncertainty band. As a rule, process models can therefore not completely replace physical experiments, but only partially.

In terms of the invention described here, the process simulation models, which can predict at least a subset of the relevant features with a known or estimated accuracy, are first called up for each iterative optimization step. If, based on the predicted process result, it can be ruled out with sufficient certainty within the scope of the prediction accuracy that the process result will be close to the target values, no real experiment is carried out. Rather, the results calculated with the process models are used as an alternative as an experimental result and the optimization process is continued.

If several process simulation models with different prediction accuracy for different areas in the parameter space are available, the one with the best prediction accuracy can be used.

In a first aspect, the invention therefore relates to a computer-implemented method for operating an injection molding machine, wherein, depending on predetermined process parameters, in particular without controlling the injection molding machine, an estimated result of the injection molding is determined simulatively, which characterizes how well an actual result of the injection molding is these process parameters will be, and wherein the process parameters are varied by means of Bayesian optimization using a data-based model that is set up to estimate the result of the injection molding depending on the process parameters, until an actual result of the injection molding is sufficiently good. This can be done by determining a value of a cost function as a function of estimated variables or as a function of actual variables, the estimated variables characterizing the estimated result of the injection molding and the actual variables characterizing the actual result of the injection molding, and then determining whether this value of the cost function falls below a predeterminable threshold value. Variables that characterize an estimated or actual result of the injection molding can characterize the product produced with the injection molding and/or the production process.

In this case, the value of the cost function can be determined as a function of how much the estimated or actual values deviate from target values that characterize a target result of the injection molding.

Bayesian optimization allows an optimum to be determined quickly in a specifiable parameter range without having to determine gradients. Finding gradients of the cost function, even without experimental noise, would require numerous additional actual injection molding steps. A further complication would be that because of the unavoidable experimental noise, the gradients determined, for example, by difference quotients would also be subject to noise. In order to get this noise sufficiently small, many tests would be necessary, which can be saved by using Bayesian optimization. In addition, Bayesian optimization makes it possible to determine the global optimum.

In order to reduce the number of injection molding steps actually required as best as possible, the process parameters can first be varied until the estimated result is sufficiently good, and only then is the actual injection molding result recorded for these process parameters. In other words, an actual experiment to determine the actual result is only carried out if the simulation experiment suggests that a good actual, i.e. experimental, result can be expected. The data-based model can then be trained depending on the actual result, ie depending on actual variables that characterize the actual result.

In particular, it is possible for the data-based model to be trained, i.e. updated, depending on the estimated result, i.e. depending on estimated variables that characterize the estimated result.

Despite the shortcomings of the estimated result, it can be advantageous to use it to train the data-based model in order to achieve a reduction in the injection molding steps that are actually required.

In order to suppress any incorrect training of the data-based model, it can be provided that the data-based model is not dependent on the estimated result, _but only depending on the actual result.

As described above, the data-based model can advantageously be a Gaussian process model. This permits a particularly targeted variation of the process parameters since, in addition to the estimated result, an uncertainty in the estimated result and an uncertainty in the actual result, in particular due to noise, can also be determined and taken into account.

Alternatively or additionally, it can be provided that the estimated result is determined using a physical model of the injection molding, it being possible for the estimated result to be determined using the data-based model if the physical model were to be evaluated with parameters outside a specifiable range is determined.

Any known inadequacies in the physical model can be compensated for in a particularly simple manner. It is understood that the estimated result can include a plurality of quantities. In this case, it can be provided that the data-based model is a multi-dimensional model, or that a plurality of one-dimensional data-based models corresponding to the plurality of variables is used, or that a mixture of one-dimensional and multi-dimensional models is used.

Since physical simulation models can sometimes only predict a subset of the features relevant for the optimization, values for the estimated result can still be determined using heuristics. In a further aspect, it is therefore provided that the estimated result is determined using a physical model evaluated for the specified process parameters and using actual results determined for other process parameters.

Embodiments of the invention are explained in more detail below with reference to the accompanying drawings. In the drawings show:

FIG. 1 shows schematically a structure of an injection molding machine;

FIG. 2 shows schematically a structure of a test stand;

FIG. 3 shows an embodiment for operating the test bench in a flowchart;

FIG. 4 shows an embodiment for operating the test bench in a flowchart.

Description of the exemplary embodiments

FIG. 1 schematically shows a structure of an injection molding machine (1). A control signal (A) is provided by a control logic (40) in order to control a control of heating elements (5) and/or a rotational movement of a worm (7). In a funnel (2) filled granules (3) in a Cylinder (4) introduced, inside which the screw (7) transports the granules (3) by the rotation of the screw (7) in the direction of a tip of the screw (7). The heat from the heating elements (5) melts the granulate (3) and plasticizes it and feeds it to an attached tool (6), in which the cooling melt then forms a component to be manufactured.

FIG. 2 shows a diagram of the structure of a test stand (3) for determining optimal process parameters (x). Current process parameters (x) are provided by a parameter memory (P) via an output interface (40) of the injection molding machine (1). This carries out the injection molding depending on the provided process parameters (x). Sensors (30) determine sensor variables (S) that characterize the result of the injection molding. Quality properties (y _exp ) determined from these sensor variables (S) are made available to a machine learning block (60) via an input interface (50).

In the exemplary embodiment, the machine learning block (60) comprises a Gaussian process model which, as illustrated in FIG. 3 or FIG. 4, is trained as a function of the provided quality properties (y _exp ). Depending on the Gaussian process model, varied process parameters (x') can be provided, which are stored in the parameter memory (P).

As an alternative or in addition to being provided via the output interface (40), the process parameters (x) can also be provided to an estimation model (15), which provides the machine learning block (60) with estimated quality properties (y _sim ) instead of the actual quality properties (y _exp ).

In the exemplary embodiment, the test stand comprises a processor (45) which is set up to play back a computer program which is stored on a computer-readable storage medium (46). This computer program includes instructions that cause the processor (45) to carry out the method illustrated in Figures 3 or 4 when the computer program is played. This computer program can be implemented in software, or in hardware, or in a hybrid of hardware and software. FIG. 3 shows an exemplary method for operating the test bench (3) in a flowchart. The method begins (100) in that initial process parameters (x _init ) are provided as process parameters (x) and test data recorded so far are initialized as an empty set. Optionally, process parameters (x) are specified with a design-of-experiment method and, as explained in more detail below, the injection molding machine (1) is controlled with these process parameters (x), variables (y _exp ) are determined and the Gaussian process is trained with the experimental data determined in this way . These process parameters (x) can include one or more parameters that characterize a time profile of the rotational movement of the screw (7) and/or a time profile of a heating current of the heating elements (5).

The injection molding machine (1) is controlled (110) with the current process parameters (x) and variables (y _exp ) are determined (120), which characterize the actual result of the injection molding.

These variables (y _exp ) can include variables that include the geometric dimensions of the manufactured component (including distortion and/or shrinkage) and/or a weight of the manufactured component and/or optical properties (such as reflectivity) of the manufactured component and/or or characterize a surface condition (such as a roughness) of the manufactured component and/or a degree of filling and/or a filling error of the injection molding compound in the tool (6).

A cost function K is evaluated (130) as a function of ^these variables, as can be given by equation (1), for example, the variables (yexp) being provided ^as features (q;) and corresponding target values of these variables (q^ztei). become.

A cost function K is also conceivable, which penalizes deviations of the features from the target values, in particular if they exceed a specifiable tolerance distance, and rewards high productivity. The "punish" can For example, be realized by a high value of the cost function K, the "reward" accordingly by a low value.

Then it is determined whether the cost function K indicates that the current process parameters (x) are sufficiently good; in the event that a punishment is expressed by a high value and a reward by a low value, by checking whether the cost function K falls below a predeterminable maximum cost value (140). If this is the case (“Yes”), the method ends (150) with the current process parameters (x).

If this is not the case ("No"), the data point (x,y _exp ) determined in this way from process parameters (x) and associated variables characterizing the result is added to the determined test data (160) and the Gaussian process is retrained, i.e. the hyperparameters ( 0 _O , 0 _d ) of the Gaussian process is adjusted in such a way that a probability that the experimental data result from the Gaussian process is maximized.

Then (170) an acquisition function is evaluated, as illustrated by way of example in formula (7), and new process parameters (x′) are thereby determined. A branch is then made back to step (110), where the new process parameters (x') replace the current process parameters (x).

FIG. 4 shows a further exemplary method for operating the test stand (3) in a flowchart. Steps (100) to (170) are the same as illustrated in FIG. 3, and a separate description is therefore omitted.

However, after new process parameters (x') have been determined that replace the current process parameters (x), a simulation model _is called up (180) with these new process parameters (x') in order to use estimated sizes (y _sim ) to determine. This simulation can, for example, use known simulation methods to determine component dimensions and/or distortion and/or shrinkage of the injection molding compound.

In the injection molding of thermoplastics, the cooling process or solidification process initiated by initial contact of the melt with a wall of the mold (6) is started after the thermoplastic melt has been injected into the mold (6). The variables mentioned can be determined here, for example, with the help of the Tait equation, which describes a change in density (1/v) of the cooling material as a function of pressure (p) and temperature (T). This equation is given by

Here, C is a constant that can be determined for the material to be used via appropriate experiments. v ₀ (T), v _t ,B(T) can be given by suitable parameterized functions, such as those from PG Tait, Physics and Chemistry of the Voyage of HMS Challenger. Vol. 2, Part 4 (HMSO, London, 1888) or F. Baumgärtner, C. Bonten, Approach for the Description of PvT Behavior of Thermoplastics at High Cooling Rates, AIP Conference Proceedings 2289, 2020 are known.

In the injection molding of elastomers or duroplastics, on the other hand, the material is not cooled when it comes into contact with a wall of the mold (6), but rather heated (mold temperature > 150 °C) in order to activate the cross-linking reaction. The change in density, which is directly related to the degree of crosslinking, can be described with the help of so-called crosslinking reactions, such as those described in H. Ou, M. Sahli, T. Barriere, J. Gelin, Experimental characterization and modeling of rheokinetic -properties of different silicone elastomers, Int J Adv Manuf Technol, 92, 4199 - 4211, 2017 or L. Granado, R. Tavernier, G. Foyer, G. Davida, S. Caillol, Comparative curing kinetics study of high char yield formaldehyde- and terephthalaldehyde-phenolic thermosets, Thermochimica Acta, 667, 42-49, 2018 are known. The determination of some of the estimated variables (y _sim ) such as a degree of filling or a filling error is not possible in one embodiment with this model. These variables can be determined, for example, using characteristic diagrams. Typical domain knowledge can be stored in these characteristic diagrams. Alternatively or additionally, these variables can be estimated based on the experimentally determined values up to this point in time, for example by an average of all these values, or these experimentally determined actual values can be weighted depending on a distance between the current process parameters and those process parameters to which the respective experimentally determined actual values have been determined. In particular, it is possible that predictions of Gaussian processes that were trained using actual variables are used as estimated values. Provision can also be made for these estimated values to be updated accordingly as soon as future actual values become known.

Then (190), analogously to step (130), the cost function K is determined, with the quantities (y _sim ) estimated by simulation being used instead of the quantities (y _Prn ) determined experimentally.

Then (200), analogously to step (140), the cost function K is used to check whether the current process parameters (x) are sufficiently good or not, whereby instead of the specifiable maximum cost value, a specifiable second maximum cost value can be used that is greater than predeterminable maximum cost.

If the check has shown that the current process parameters (x) are sufficiently good, the process branches back to step (110). Otherwise, the process branches back to step (160).

Claims

Expectations

1. Computer-implemented method for operating an injection molding machine (1), an estimated result (y _sim ) being determined as a function of predetermined process parameters (x), which characterizes how good an actual result (y _exp ) of the injection molding will be, and wherein the process parameters (x) are varied by means of Bayesian optimization using a data-based model until an actual result (y _exp ) of the injection molding is sufficiently good.

2. The method according to claim 1, wherein the process parameters (x) are varied until the estimated result (y _sim ) is sufficiently good, and only then is the actual result (y _exp ) of the injection molding machine recorded.

3. The method according to claim 2, wherein the data-based model is trained depending on actual results (y _exp ).

4. The method according to claim 3, wherein the data-based model is also trained as a function of the estimated result (y _sim ).

5. The method according to any one of claims 1 to 4, wherein the data-based model is a Gaussian process model.

6. The method according to any one of claims 1 to 5, wherein the estimated result (y _sim ) is determined by means of a physical model of the injection molding.

7. The method according to claim 6, wherein if the evaluation of the physical model would take place with parameters (x) outside a predeterminable range, the estimated result (y _sim ) using the data-based model is determined. Method according to one of claims 6 to 7, wherein the estimated result (y _sim ) is determined using a physical model evaluated for the specified process parameters (x) and using actual results (y _exp ) determined for other process parameters (x'). Method according to one of the preceding claims, wherein to characterize the estimated result (y _sim ) and / or to characterize the actual result (y _exp ) sizes are used, the geometric dimensions of the manufactured component, and / or a distortion of the manufactured component, and /or shrinkage of an injection molding compound in a tool (6) and/or a weight of the manufactured component and/or optical properties, in particular reflectivity, of the manufactured component and/or a surface quality, in particular roughness, of the manufactured component and/or a Characterize the degree of filling and/or a filling error of the injection molding compound in the tool (6). Method according to one of Claims 1 to 9, the injection molding machine (1) being operated with the process parameters (x) set in this way following the setting of the process parameters (x). Test stand (3) for an injection molding machine (1), which is set up to carry out the method according to one of Claims 1 to 10. Computer program set up to carry out the method according to one of Claims 1 to 10. Machine-readable storage medium on which the computer program according to claim 12 is stored.