Method and apparatus for operating a laser material processing machine

Abstract

To find a process parameter for securing high quality laser material processing performed by a laser material processing machine.SOLUTION: In a computer-implemented method for operating a laser material processing machine (1, 2), an estimated result (ysim) is ascertained, which characterizes how good an actual result (yexp) of laser material processing will be, depending on a predefined process parameter (x), and the process parameter is varied by means of Bayesian optimization using a data-based model, until the actual result (yexp) of the laser material processing is sufficient enough.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to a method, a test bench, a computer program and a machine-readable storage medium for operating a laser material processing machine. [Background technology]

[0002] Prior art Laser beam drilling is a manufacturing method for creating holes in a wide variety of materials. In this case, a focused, pulsed laser beam is applied to the workpiece. The absorbed laser energy, due to its very high intensity, leads to pulsed, very rapid heating of the workpiece material, which, on a short time scale, is spatially highly localized, leading to the formation of a melt and, partially, also evaporation.

[0003] The molten material is forced out of the hole by the vapor pressure explosively generated by the process and the associated equally large pressure gradient, or also by an externally supplied gas flow. When particularly high intensities are achieved, for example, by using a laser beam with ultrashort laser pulses, a higher rate of evaporation can be achieved, and more precise drilling can be achieved.

[0004] At relatively long pulse durations and relatively low intensities, the formation of the perforations is dominated by melt extrusion, resulting in reduced accuracy but significantly increased productivity. Often, multiple laser pulses are required per hole to create the desired perforations. To improve the accuracy of the perforations, the laser beam can typically be guided by a suitable device on a circular or helical trajectory at the perforation location.

[0005] In laser drilling manufacturing processes, process development is typically characterized experimentally because the numerous, highly dynamic, interacting physical processes cannot currently be modeled with sufficient accuracy. This is partly due to the fact that workpiece characterization data regarding relevant pressures and temperatures is often unknown. At best, only highly simplified models are available that can predict to some extent the desired perforation geometry at given process parameters and within specific parameter ranges. These models currently do not allow reliable prediction of quality characteristics such as, for example, solidified molten deposits in the form of burrs inside the hole or at the perforation entrance, damage to the perforation edge, or roundness of the perforation.

[0006] Laser welding is an established manufacturing method for creating joints between workpieces made of different materials. In this process, a focused laser beam is applied to the workpieces to be joined. The absorbed laser energy, due to its very high intensity, leads to very rapid, localized heating of the workpiece material, which leads to the formation of a common molten pool that is highly spatially localized on a short time scale. After solidification of the molten pool, a joint is formed between the workpieces in the form of a weld seam.

[0007] To meet the requirements for joint strength (and fatigue strength), it may be desirable for the weld seam geometry to not fall below the minimum allowable weld seam depth and minimum allowable weld seam width. To achieve the desired weld seam shape, process parameters can be selected so that rapid, localized heating of the material by the laser beam results in evaporation in the weld pool. The explosively generated vapor pressure and associated large pressure gradients, or even an externally supplied gas flow, force the molten material out of the weld pool. The resulting metal spatter (so-called weld spatter) can lead to reduced part quality and / or require production interruptions to clean the laser welding system, significantly increasing manufacturing costs.

[0008] As with laser drilling, process development for laser welding (process optimization with the goal of minimizing weld spatter) is heavily characterized experimentally because the many highly dynamic and interacting physical processes cannot be modeled with sufficient accuracy. Summary of the Invention [Problem to be solved by the invention]

[0009] One modeling challenge in this case is that workpiece characteristic data regarding the relevant pressures and temperatures are often unknown. Manufacturing tolerances and material variations of individual workpieces can also have a very strong influence on the formation of weld spatter. While highly simplified models are available that can predict to some extent the desired weld seam geometry for given process parameters and within certain parameter ranges, these models do not allow reliable prediction of quality characteristics such as, for example, solidified weld spatter.

[0010] Due to the large number of configurable process parameters (often time- and position-dependent), such as laser power, focal diameter, focal position, welding speed, laser beam inclination, circular orbit frequency, and process protective gas, the optimization of the process parameters is a time-consuming process requiring a large number of experiments. Since these experiments require a large number of workpieces or parts, and the evaluation (creation of cross sections to measure the weld seam geometry) is also complex, the number of required trials must be reduced to a minimum.

[0011] Thus, for example, some process parameters are set to empirical values, and only relatively few parameters are specifically varied, and the optimum values that are actually achievable are largely never found. [Means for solving the problem]

[0012] Advantages of the invention It has been found that the precision and productivity achievable in laser material processing is highly dependent on the process parameters set, the workpiece material being used, and, to some extent, the geometry of the workpiece material.

[0013] There are many quality criteria for the drilling process. For example, the size of the drilled holes (e.g. the progression of the diameter depending on the depth), the roundness of the holes, the shape of the walls of the drilled holes, any melt deposits, droplet release during the drilling process, and the rounding of the edges of the drilled holes are important. Productivity is usually defined by the number of holes that can be produced per hour. Furthermore, in practice, of course, the cost of the required production equipment is also crucial and usually increases with increasing flexibility in the variable parameters.

[0014] In many cases, there are a large number of configurable process parameters (e.g., pulse duration, focal position, focal size, pulse repetition frequency, circular orbit diameter, circular orbit frequency, angle of attack, drilling duration, pulse energy, wavelength, type and pressure of process gas) that can also be varied in a time-dependent manner, making the optimization of process parameters a time-consuming process requiring a large number of experiments. These experiments, on the one hand, require a large number of workpieces or parts, and, on the other hand, the evaluation (especially the evaluation of the internal drilling shape) is also complicated, so the number of required trials must be reduced to a minimum.

[0015] Therefore, some process parameters are set to empirical values, and only relatively few parameters are specifically varied. Therefore, it is often difficult to find the optimum value that is actually achievable. Trial design techniques include a series of expert-specified trials and / or statistical trial design techniques.

[0016] In the case of laser welding, there are also a large number of configurable process parameters (often time- and position-dependent), such as laser power, focal diameter, focal position, welding speed, laser beam inclination, circular orbit frequency, and process protective gas, making the optimization of the process parameters a time-consuming process requiring a large number of experiments. Since these experiments, on the one hand, require a large number of workpieces or parts, and on the other hand, the evaluation (production of cross sections to measure the geometry of the weld seam) is also complicated, it is desirable to reduce the number of required trials to a minimum.

[0017] In contrast to this, the subject matter with the features of independent claim 1 has the advantage that process parameters of a laser material processing machine that ensure high-quality laser material processing can be found with only a few experiments.

[0018] Further aspects of the invention are the subject of further independent claims. Advantageous developments are the subject of the dependent claims.

[0019] Disclosure of the Invention The present invention relates to a method by which an efficient and targeted optimization of process parameters can be performed. For this purpose, a Bayesian optimization method is used. This method can be used to find the optimum value of an unknown function. The optimum value is determined by one or more quality characteristics (features) q defined by the user. i The target value q i,Ziel To obtain a single function to be optimized, several quality characteristics can be calculated in a so-called cost function K, which also needs to be specified by the user. One example is the sum of scaled deviations from the respective target values:

number

[0020] Bayesian optimization methods are suitable for finding parameter sets that lead to optimal function values for functions that map a multidimensional parameter space to scalar values. Depending on the optimization goal, the optimal value is defined as the maximum possible value that the function value can take, or as the minimum possible achievable value. In the context of process optimization, for example, a parameter set is given by a set of specific process parameters, i.e., the corresponding function value can be determined by the cost function described above.

[0021] Since determining the function values of the cost function requires conducting and evaluating experiments, the function essentially provides only a value table with data that still contains experimental "noise." Because experiments are so complex, averaging the results over multiple iterations of the same parameter set typically does not suppress this noise. Therefore, it would be advantageous to perform optimization in a way that allows for global optimization with good results despite a small number of trial evaluations, and that does so without the need to calculate the gradient of the cost function. Bayesian optimization has been found to satisfy these properties.

[0022] Bayesian optimization consists of a Gaussian process mathematical method that provides the most probable prediction of a function value, including variance, based on a given value table for each parameter set, and algorithmically formulated rules for which parameter sets further function evaluations (i.e., experiments in this specification) should be performed based on the Gaussian process predictions.

[0023] Specifically, the parameter set x N+1 The prediction about the outcome of the function evaluation at is the most probable value of the Gaussian process (the "mean value")

number

number

[0024] To estimate the mean and variance in the above formula, we use the whole parameter set x n(n=1..N) and the parameter set x to be predicted N+1 With respect to [k] n =k(x n ,x n+1 ) is calculated. There are various approaches regarding the core function to be used in a specific case, one simple approach is to use the following quadratic exponential core function:

number

[0025] The selection of the next set of parameters on which trials should be performed is based on the prediction of the mean and variance calculated using the formula above. Various strategies are possible here, for example the strategy of "expected improvement".

[0026] In this case, for the next experiment, we use the maximum (or minimum, depending on the optimization goal) known function value from the previous N iterations.

number

number

[0027] The "+" operator here means that only positive values are used, and negative values are set to zero. In Bayesian optimization, the optimization continues until the Determining new trial points (i.e., parameter sets), Conducting trials; Updating the Gaussian process with new function values is carried out iteratively.

[0028] Optimization of the Gaussian process using new trial points and new function values is performed such that new pairs of trial points and function values are added to already recorded trial data consisting of pairs of trial points and function values, and the hyperparameters are adapted to maximize the probability (e.g., likelihood) of the trial data.

[0029] This process is illustrated in relation to FIG.

[0030] By an iterative procedure of the above steps (performing experiments, evaluating quality criteria and determining cost function values, updating the Gaussian process, and proposing the next parameter set), a process model (mapped by the Gaussian process) can be successively constructed. The best parameter set of all evaluated function evaluations or all trials is then used as the best optimization result.

[0031] The advantage of implementing optimization is gained by incorporating existing process knowledge. According to the procedure described below, one or more process models P are optimized by replacing real experiments under specific conditions with simulated experiments. 1...nknowledge in the form of σ can be incorporated into the optimization, in which case the uncertainty in the model's mapping of the process and the number of quality criteria it describes become trivial.

[0032] If a process model is used that perfectly maps a real experiment, it would be possible to replace every real experiment with a simulation experiment. If the evaluation time is shorter than the real implementation, time as well as effort would be saved. However, the predictive accuracy of a process model is generally limited. In many cases, the predictive accuracy of a process model is valid only in a part of the parameter space and / or describes only a subset of the process results and does not take into account all physical effects, thus producing results only within the uncertainty region. Therefore, in principle, a process model cannot be completely replaced by a physical experiment, but only partially.

[0033] In the sense of the invention described herein, each iterative optimization step begins with calling up a process simulation model that can predict a subset of relevant characteristics with known accuracy. If, based on the predicted process results, it can be ruled out with sufficient certainty that the process results will be close to the target value, even within the prediction accuracy, no real experiments are performed. Instead, the results calculated by the process model are used instead as experimental results, and the optimization process continues.

[0034] If multiple process simulation models with different predictive accuracy are available for different regions of the parameter space, the process simulation model with the best predictive accuracy can be used.

[0035] Accordingly, in a first aspect, the present invention relates to a computer-implemented method for operating a laser material processing machine, comprising: determining, depending on specified process parameters, by simulation without controlling the laser material processing machine, an estimated outcome of the laser material processing characterizing how good an actual outcome of the laser material processing will be for the process parameters; and varying the process parameters by Bayesian optimization using a data-based model until the actual outcome of the laser material processing is sufficiently good, the data-based model being configured to estimate the outcome of the laser material processing depending on the process parameters.

[0036] This can be done by determining a value of the cost function as a function of estimated variables that characterize the estimated results of laser material processing, or as a function of actual variables that characterize the actual results of laser material processing, and then determining whether this value of the cost function is below a specifiable threshold. The variables that characterize the estimated or actual results of laser material processing can characterize the product and / or the product process produced by laser material processing.

[0037] In this case, the value of the cost function can be determined depending on how much the estimated or actual variables deviate from the target variables that characterize the desired result of the laser material processing.

[0038] Bayesian optimization allows for rapid identification of an optimum within a specifiable parameter range without the need to identify a gradient, which may require multiple actual steps of laser material processing and may only be identified uncertainly by a quotient of differences due to unavoidable experimental noise. While multiple trials may be required to sufficiently reduce this noise, Bayesian optimization can reduce this. Furthermore, Bayesian optimization allows for the identification of a global optimum.

[0039] To reduce the number of laser material processing steps actually required as much as possible, the process parameters can first be varied until the estimated results are good enough, and only then can the actual results of laser material processing for those process parameters be detected. In other words, actual experiments to identify actual results are performed only when the simulation experiments suggest that good actual results, i.e., experimental results, can be expected.

[0040] In that case, a data-based model can be trained depending on the actual outcome, i.e., depending on the actual variables that characterize the actual outcome.

[0041] In particular, it is possible to train, ie, update, a data-based model depending on the estimated outcome, ie, depending on the estimated variables that characterize the estimated outcome.

[0042] Despite the inadequacies of the estimated results, it may be advantageous to use them to train a data-based model in order to reduce the number of laser material processing steps actually required.

[0043] To limit possible mistraining of the data-based model, the estimated results (y sim ) is sufficiently good, i.e., close enough to the optimization goal, then the data-based model can be trained solely on the actual outcomes, rather than on estimated outcomes.

[0044] As mentioned at the beginning, the data-based model can advantageously be a Gaussian process model, which allows for particularly targeted variations of the process parameters, since in addition to the estimated results, in particular the uncertainty of the estimated results due to noise and the uncertainty of the actual results can be identified and taken into account.

[0045] Alternatively or additionally, the estimated results can be determined by a physical model of laser material processing, and in cases where the evaluation of the physical model may be performed for parameters outside the specifiable range, the estimated results can be determined by a data-based model, which allows possible known inadequacies of the physical model to be compensated for particularly easily.

[0046] Of course, the estimated outcome may involve multiple variables, in which case the data-based model may be a multidimensional model, or may use multiple one-dimensional data-based models corresponding to the multiple variables, or may use a mixture of one-dimensional and multidimensional models.

[0047] Because the physical simulation model may only be able to predict a subset of the characteristics relevant to the optimization, the use of heuristics still allows for the specification of estimated outcome values. Thus, in a further aspect, estimated outcomes are specified by the physical model evaluated at specified process parameters and by actual outcomes specified at other process parameters.

[0048] In the following, embodiments of the invention will be described in more detail with reference to the accompanying drawings. [Brief explanation of the drawings]

[0049] [Figure 1] 1 is a schematic diagram of the structure of a laser drilling machine. [Figure 2] 1 is a schematic diagram of the structure of a laser welding machine. [Figure 3] FIG. 1 is a schematic diagram of the structure of a test stand. [Figure 4] 1 is a flow chart of one embodiment for operating a test stand. [Figure 5] 1 is a flow chart of one embodiment for operating a test stand. DETAILED DESCRIPTION OF THE INVENTION

[0050] Description of the Examples Figure 1 shows a schematic diagram of the structure of a laser drilling machine (1). A control signal (A) is provided by control logic (40) to control a laser (10a). The laser beam strikes a piece of material (12) and creates a perforation (11) therein.

[0051] 2 shows a schematic diagram of the structure of a laser welder (2). Again, a control signal (A) is provided by control logic (40) to control a laser (10b). The laser beam strikes two pieces of material (13, 14) and creates a weld seam (15) there.

[0052] A laser cutter (not shown) is similarly contemplated.

[0053] FIG. 3 shows a schematic diagram of the structure of a test bench (3) for identifying optimal process parameters (x). Current process parameters (x) are supplied from a parameter memory (P) via an output interface (4) of a laser material processing machine, such as a laser drilling machine (1) or a laser welding machine (2). The laser material processing machine performs laser material processing depending on these supplied process parameters (x). Sensors (30) detect sensor variables (S) that characterize the results of the laser material processing. These sensor variables (S) are then input via an input interface (50) to the test bench (3) to determine quality characteristics (y). exp ) to the machine learning block (60).

[0054] The machine learning block (60) includes a Gaussian process model in this embodiment, which is configured to calculate the quality characteristics (y exp ) and can be trained depending on the Gaussian process model to provide modified process parameters (x'), which are stored in a parameter memory (P).

[0055] Alternatively or in addition to being provided via the output interface (4), the process parameters (x) can also be provided to an estimation model (5), which estimates the actual quality characteristics (y exp ) instead of the estimated quality characteristic (y sim ) to the machine learning block (60).

[0056] The test stand, in this embodiment, includes a processor 45 configured to execute a computer program stored on a computer-readable storage medium 46. The computer program includes instructions that, when executed, cause the processor 45 to perform the method illustrated in Figures 4-5. The computer program can be implemented in software, in hardware, or in a combination of hardware and software.

[0057] Figure 4 shows a flow chart of an exemplary method for operating the test stand (3). The method begins with determining the initial process parameters (x init ) are provided as process parameters (x), and previously recorded trial data is initialized as an empty set (100). Optionally, the process parameters (x) are specified by a design of experiments, and as described in more detail below, the laser material processing machine (1, 2) is controlled by these process parameters (x), and the variables (y exp ) is identified, and a Gaussian process is trained using the trial data thus identified.

[0058] In the case of laser drilling, these process parameters (x) include, in one embodiment, the pulse duration, and / or the focal spot position resolved as a function of time via a characteristic map, and / or the focal spot size, and / or the pulse repetition frequency, and / or the circular orbit diameter resolved as a function of time via a characteristic map, and / or the circular orbit frequency, and / or the angle of attack resolved as a function of time via a characteristic map, and / or the drilling duration, and / or the pulse energy resolved as a function of time via a characteristic map, and / or the wavelength, and / or parameters characterizing the process protective gas, such as the type or pressure of the process gas. The aforementioned circular orbit is a known feature in many drilling methods, such as helical drilling or trepanning drilling.

[0059] In the case of laser welding, these process parameters (x) include the laser power, resolved in a time-dependent and / or position-dependent manner via characteristic maps, and / or the focal spot diameter, and / or the focal spot position, and / or the welding speed, and / or the inclination of the laser beam, and / or the circular orbit frequency of the laser wobble, and / or parameters characterizing the process protective gas.

[0060] The laser material processing machine (1, 2) is controlled (110) by the current process parameters (x) and the variables (y) that characterize the actual results of the laser material processing. exp ) is identified (120).

[0061] For laser drilling, these variables (y exp ) in one embodiment includes variables characterizing the size of the perforation (11), and / or the roundness of the perforation (11), and / or the shape of the walls of the perforation (11), and / or the presence or absence of molten deposits, and / or the amount of droplet ejection during the perforation process, and / or the roundness of the edges of the perforation (11), and / or productivity.

[0062] In the case of laser welding, these variables (y exp) in further embodiments includes variables characterizing the minimum weld seam depth and / or the minimum weld seam width and / or the productivity and / or the number of weld spatters and / or the number of holes and / or the weld distortion and / or the weld residual stress and / or the weld cracks along the weld seam (15).

[0063] Depending on these variables, a cost function K is evaluated (130), for example, as may be given by equation (1), where the features (q i ) as a variable (y exp ) and the corresponding target values of these variables (q i,Ziel ) will be supplied.

[0064] A cost function K may be envisaged that penalizes deviations of features from target values, especially if the deviations exceed a specifiable tolerance, and rewards high productivity. The "penalty" may be achieved, for example, by a high value of the cost function K, and the "reward" may be achieved correspondingly by a low value.

[0065] It is then determined whether the cost function K indicates that the current process parameters (x) are good enough, i.e., if the penalty implies a high value and the reward implies a low value, then the cost function K is checked to see if it is below a specifiable maximum cost value 140. If this is the case ("yes"), the method ends with the current process parameters (x) 150.

[0066] If this is not the case ("No"), then the process parameters (x) and the corresponding variables (y) that characterize the outcome are exp ) and the data points (x, y exp ) is added to the identified trial data (160), and the Gaussian process is retrained, i.e., the hyperparameters (Θ,Θ) of the Gaussian process are adjusted to maximize the probability that the trial data is obtained from the Gaussian process. d ) is adapted.

[0067] Then (170), the acquisition function, such as that illustrated in equation (7), is evaluated, thereby identifying the new process parameters (x'). A branch is then taken back to step (110).

[0068] Figure 5 shows a flow chart of a further exemplary method for operating the test stand 3. Steps 100 to 170 are similar to those shown in Figure 4 and will not be described separately.

[0069] However, after successfully identifying the new process parameters (x'), the actual variables (y exp ) instead of the estimated variables (y sim The simulation model is invoked with these new process parameters (x') to identify (180).

[0070] In the case of laser drilling, this can be done, for example, as follows: for the radius r of the drilling 11 along the depth coordinate z, r(z) is calculated by the following equation:

number

number

[0071] in this case:

number

number

[0072] This model cannot predict some characteristics, such as the presence or absence of molten deposits and / or the amount of droplets emitted during the drilling process. To identify these characteristics, an empirical model can be specified, or a result can be determined from the experimentally determined values up to this point, e.g., the average of all these values, or the experimentally determined actual values can be weighted depending on the distance between the process parameters when the actual experimentally determined values were determined and the current process parameters. In particular, a Gaussian process prediction trained on the actual variables can be used as the estimated values.

[0073] Alternatively or additionally, it may be that at least some features cannot be reliably calculated for all process parameters (x), and it is possible to check whether the current process parameter (x) is within a specifiable range, and if this is not the case, identify the features using one of the approaches listed above.

[0074] In the case of laser welding, the estimated variables (y sim ) can be identified, for example, as follows:

number

number

number

[0075] Subsequently, in (190), a cost function K is identified similarly to step (130), in this case using the experimentally identified variables (y exp ) instead of the variables estimated by simulation (y sim ) is used.

[0076] Then (200), using the cost function K as in step (140), it is checked whether the current process parameters (x) are good enough, in which case a second maximum cost value greater than the maximum specifiable cost value can be used instead of the maximum specifiable cost value.

[0077] If the check indicates that the current process parameters (x) are good enough, a branch is taken back to step (110), otherwise a branch is taken back to step (160).

Claims

1. A computer-implemented method for operating a laser material processing machine (1, 2), comprising: Depending on the specified process parameters (x), the actual results of laser material processing (y ｅｘｐ ) will be. ｓｉｍ ) and Varying the process parameter (x) by Bayesian optimization using a data-based model; the data-based model is configured to predict an outcome of the laser material processing depending on the process parameter (x); determining a cost function value depending on estimated variables characterizing the estimated result (y sim ) of the laser material processing or depending on actual variables characterizing the actual result (y exp ) of the laser material processing; The method of claim 1, wherein the process parameter (x) is varied by Bayesian optimization using the data-based model until the value of the cost function falls below a specified threshold.

2. The process parameter (x) is varied until the value of the cost function falls below the specified threshold, and only once the value of the cost function falls below the specified threshold is the actual result of the laser material processing (y ｅｘｐ ) to detect The method of claim 1.

3. The data-based model is then calculated based on the actual results (y ｅｘｐ ) to train, The method of claim 2.

4. The data-based model is then used to calculate the estimated results (y ｓｉｍ ) and training, The method of claim 3.

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

6. The estimated result (y ｓｉｍ ) is determined by the physical model of laser material processing.

6. The method according to any one of claims 1 to 5.

7. If the evaluation of the physical model may be performed for a parameter (x) outside the specifiable range, the estimated result (y ｓｉｍ ) is identified by a model based on the data; The method of claim 6.

8. The estimated result (y ｓｉｍ ) by the physical model evaluated at the specified process parameter (x) and the actual results (y) determined at other process parameters (x'). ｅｘｐ ) to identify, 8. The method according to claim 6 or 7.

9. The laser material processing machine is a laser drilling machine (1), 9. The method according to any one of claims 1 to 8.

10. The estimated result (y ｓｉｍ ) and / or to characterize the actual results (y ｅｘｐ ) using variables characterizing the geometry of the holes (11) drilled by the laser drilling machine (1), 10. The method of claim 9.

11. The laser material processing machine is a laser welding machine (2).

9. The method according to any one of claims 1 to 8.

12. The estimated result (y ｓｉｍ ) and / or to characterize the actual results (y ｅｘｐ ) using variables characterizing the geometry of the weld seam (15) welded by the laser welding machine (2) to characterize The method of claim 11.

13. Following the setting of the process parameter (x), the laser material processing machine (1, 2) is operated according to the process parameter (x) thus set.

13. The method according to any one of claims 1 to 12.

14. A test stand (3) configured to supply process parameters (x) to a laser material processing machine (1, 2) for carrying out a method according to any one of claims 1 to 12.

15. A computer program arranged to carry out the method according to any one of claims 1 to 12.

16. A machine-readable storage medium having stored thereon the computer program of claim 15.

Images