Method for monitoring a joining process

Abstract

The present invention relates to a method for monitoring a joining process, in which a first joining partner and a second joining partner are provided (1), a simulation model of the first joining partner and of the second joining partner and of a joining process to be carried out is provided (2), and an equivalent model is generated on the basis of the simulation model and is trained (3) by means of a machine learning algorithm. The first joining partner and the second joining partner are joined (4) in a real joining process by means of a joining device, at least one process variable of the real joining process being determined. The determined process variable is taken into account as a boundary condition in the equivalent model, and the equivalent model is simulated (5) at the same time as the real joining process is carried out, the real joining process being open-loop or closed-loop controlled on the basis of a result of the equivalent model and / or a result of the real joining process being monitored (6) on the basis of the result of the equivalent model.

Description

[0001] Method for monitoring a joining process

[0002] The present invention relates to a method for monitoring a joining process.

[0003] During joining, the connection between components is usually concealed, making it impossible to assess from the outside. Therefore, it can only be evaluated using destructive or non-destructive testing methods. These testing methods are performed outside the production line, and destructive testing renders the component being manufactured unusable. This is time-consuming and costly for manufacturing companies. Therefore, process monitoring is essential in joining technology. An example of a non-destructive testing method is described in Matiszik, Christian et al.: Magnetic Characterization of the Nugget Microstructure at Resistance Spot Welding, Crystals 2022, 12, 1512. Figure 1 of this publication also shows typical geometric parameters, such as the diameter of the joint, that need to be determined for the quality of the joining.

[0004] The goal of process monitoring is to assess the quality of the joint based on process data. Process data consists of measurable parameters during the joining process, such as the welding current and the required welding force. Currently, expert systems and, less frequently, machine learning (ML) methods are used for monitoring. Expert systems rely on the specialized knowledge of individual experts through the application of statistical methods, such as limit value analysis or defined validity ranges. Machine learning methods train an algorithm to recognize patterns in the process data that also indicate the quality of the joint. Both monitoring methods are black-box models capable of real-time operation, but they do not provide the user with explainability or traceability of the decisions.Due to the lack of physical traceability, higher safety factors are applied, as false positive welds are unacceptable.

[0005] Numerical process simulations can also be used to evaluate a hidden joint. These are based on physical laws and are suitable for calculating internal physical quantities over time. The finite element method (FEM) is frequently used for this purpose. Due to its explainability, this is a white-box model. Therefore, it is used in process development and to improve process understanding. However, due to its high complexity and the required computing power, it is unsuitable as a monitoring method.

[0006] Unless non-destructive testing is used in the process, the problem is currently solved by evaluating the process parameters of the joining system or device, depending on the joining method. However, the accuracy of this prediction varies considerably and depends on the material being joined, the geometry, the component treatment, and supply variations. Existing solutions are therefore disadvantageous due to the lack of in-process testing methods that directly assess the quality of the joint and the associated costs. Statistical methods have the disadvantage that they are not based on physical laws and only provide an estimate of the joint quality. This reduces acceptance, as making a quality decision is difficult in the borderline area between "acceptable" and "not acceptable."This places high safety factors on the limits of process monitoring, which increases the reject rate and thus negatively impacts costs and productivity.

[0007] The present invention therefore aims to propose a method that avoids the aforementioned disadvantages and enables real-time process monitoring of a joining process.

[0008] This problem is solved according to the invention by a method according to the main claim. Advantageous embodiments and further developments are contained in the dependent claims.

[0009] In a method for monitoring a joining process, a first joining partner and a second joining partner, as well as a simulation model of the first joining partner, the second joining partner, and the joining process to be carried out, are provided. A substitute model based on the simulation model is generated and trained using a machine learning algorithm. This substitute model typically comprises the simulation model and a machine learning model that is trained by the machine learning algorithm. The first joining partner and the second joining partner are joined in a real (i.e., not virtual or simulated) joining process using a joining device. At least one process parameter of the real joining process is determined, and this determined process parameter is considered as a boundary condition in the substitute model. The substitute model is applied simultaneously with the execution of the real joining process.In particular, input data is specified and a prediction is obtained as in the simulation model, whereby the real joining process is controlled or regulated based on a result of the substitute model and / or a result of the real joining process, i.e. a joined component, is controlled based on the result of the substitute model.

[0010] By using a surrogate model, the joining process can be monitored much more efficiently with regard to the quality of the joint, as a physical explanation is provided. While the simulation model is often high-resolution and computationally slow, the surrogate model is provided with a feature space derived from the simulation model. All events within this feature space produce reliable results, while results outside the feature space are subject to greater fluctuations. The simulation model thus creates a database on which the surrogate model, and in particular the machine learning model, is trained. The applicability of the simulation model is ensured by experimental validation before the described procedure is carried out.The simulation model depends, for example, on the mesh architecture, material properties, boundary conditions, contact definitions, and model assumptions. Experimental validation is performed by comparing the simulation result with the actual joint (which usually involves a geometric comparison). The simulation model is typically calibrated to the experiment by adjusting numerical parameters to measured process variables. Specifically, this might mean that for an electrical simulation, the current per unit time is used as input data and the electrical voltage as output. In this example, the current is usually specified experimentally, and the electrical voltage is measured. Both values ​​match at the end of the calibration, and the model architecture—in particular, the solver, mesh resolution, and time step—is optimized to ensure accurate validation and calibration.

[0011] The described method can be performed in real time, specifically within the cycle time of the joining process, allowing for decisions to be made between individual steps. For spot welding, the cycle time is typically 1 second. The method can accommodate various boundary conditions of joining processes or joining methods, such as machine parameters, component geometry, or the material used, and, based on the training data, provides the most probable result without further calculation. Thus, accurate simulation results can be delivered in real time even for varying boundary conditions.These can be used for real-time control or regulation, but it can also be intended to classify the result of the joining process without having to perform destructive or non-destructive testing, solely based on the result of the substitute model, i.e., essentially a "virtual representation" of the process. This enables fast and cost-effective quality inspection with high accuracy.

[0012] The joining process can in principle be arbitrary, but preferably the joining process is resistance spot welding or projection welding.

[0013] The at least one process parameter is typically selected from an electric current, preferably a welding current, an electric voltage, preferably a welding voltage, an electrode force, a contact condition, and / or an electrode movement. Alternatively or additionally, production data such as material properties, geometric dimensions such as a sheet thickness combination, and / or disturbance variables can also be considered. The process parameters can generally be adjusted or change during the joining process and are therefore well suited for monitoring.

[0014] The simulation model can be a finite element model, a finite volume model, or a computational fluid dynamics model, since these models can be reliably calculated even if the computation time is usually very long.

[0015] The machine learning algorithm is typically implemented as an artificial neural network, preferably a deep neural network, a support vector machine, a decision tree, a random forest, a regression method (especially linear regression), or a gradient descent algorithm such as XGboost or CatBoost. Typically, the algorithm is a supervised learning algorithm and is used for classification. As explained above, the joining process can be regulated or controlled in real time based on the simulation results, and / or the outcome can be monitored. This allows for the fastest possible response to deviations from the expected or desired behavior or process flow.

[0016] A device for monitoring a joining process comprises a joining unit configured to connect a first joining partner and a second joining partner, wherein at least one sensor is configured to detect at least one process parameter. The device also comprises an electronic computing unit configured to provide a simulation model of the first joining partner and the second joining partner, as well as of the joining process to be carried out, taking into account the at least one process parameter, and to generate a substitute model based on the simulation model and train it using a machine learning algorithm.The determined process parameter is taken into account as a boundary condition in the substitute model, the substitute model is applied simultaneously with the execution of the real joining process, whereby the electronic computing unit is also designed to control or regulate the real joining process based on a result of the substitute model and / or to check a result of the real joining process based on the result of the substitute model.

[0017] The described procedure is typically carried out using the described device, i.e., the described device is set up to carry out the described procedure.

[0018] A computer program product includes a computer program that provides software means for carrying out the described procedure or for controlling the described device when the computer program is executed in the electronic computing unit.

[0019] An embodiment of the invention is shown in Figure 1 and will be discussed below with reference to this figure.

[0020] Figure 1 shows a schematic view of the process flow of a method for monitoring a joining process. In a first step 1, a first joining partner and a second joining partner are provided. In a next step 2, a simulation model is provided as a multiphysics simulation model of the first joining partner and the second joining partner, as well as of the joining process to be carried out. The simulation model typically has an electrical-thermal and a mechanical-thermal sub-model, which are coupled to each other in order to consider all effects in the simulation, depending on the joining process. The simulation model 2 is a finite element model that has already been calculated and whose result has been calibrated and validated, i.e., checked against the actual manufactured component.In the following step 3, a replacement or surrogate model is generated based on the simulation model and (at least a section, usually referred to as the "region of interest") trained using a machine learning algorithm (in the illustrated example, using a deep neural network). This means that a database for all expected process parameters and boundary conditions is created using the simulation. The training data for the neural network is process-specific and typically consists of location-dependent temperature and displacement values ​​from the simulation of the simulation model for different time steps.

[0021] In step 4, a joining device is used to join the first and second joining partners in a real joining process. At least one process parameter of the real joining process is determined by a sensor, for example, an electrical welding current. This determined process parameter is considered as a boundary condition in the surrogate model (after the sensor has transmitted the process parameter to the processing unit, which acts as an evaluation unit, control unit, or monitoring unit). The surrogate model is then simulated in real time in step 5, simultaneously with the execution of the real joining process. Once the determined result is available on the evaluation unit, the real joining process is also controlled or regulated in real time in step 6, based on a result from the surrogate model. Alternatively, a result of the real joining process is checked based on the result of the surrogate model.The result can also be displayed on an output unit such as a screen or a data management system. Typically, depending on the joining process, the diameter of a formed weld bead is considered to classify the quality of the joint. Steps 5 and 6, as iterations, generally run concurrently with step 4, since everything can be done in real time; however, step 5 does not run concurrently with step 6. For series production, steps 4-6 are repeated continuously, while steps 1-3 are only performed once initially.

[0022] The described method can be used as additional software or a computer program on existing joining devices, meaning that machine modifications to the joining device itself are often unnecessary. For example, in resistance welding, a database can be generated from the simulation results using the electro-thermal-mechanical simulation model. This database can depend on different welding parameters. In this case, the welding current was varied between 5 kA and 11 kA, the electrode force between 3 kN and 7 kN, and the welding time between 200 ms and 400 ms. For each parameter setting, the simulation model is solved and calibrated and validated through experiments. The input data for the simulation are the process variables current and force from the experiments. The simulation results obtained are used to create a database.

[0023] For resistance welding, the temperature distribution is of particular interest for determining the weld nugget geometry. This can only be determined through destructive or non-destructive testing methods. The displacement is also important, for example, to determine the electrode indentation. The simulation time is approximately 100 hours per parameter setting. The database contains not only the results at the end of the welding process but also the time-dependent changes every 10 milliseconds. This is crucial because the machine learning model is trained on the simulation's behavior and must also train the temperature generation to accurately represent the physical behavior. A deep neural network with seven hidden layers is used for training. The input data consists of the temperature and displacement values ​​from the simulation. The simulation results are then reduced to a smaller region of interest to improve performance.Hyperparameter optimization determines the optimal model architecture. The trained model is then applied to generate new simulation results for previously unknown boundary conditions (current, force, time). The simulation's behavior can be accurately represented because the maximum temperature deviation is 7 Kelvin. The machine learning model is solved in 47 ms.

[0024] In practice, welding is usually performed with a fixed set of parameters. However, disturbances such as contamination, tolerances, or positioning inaccuracies during production can mean that the desired result cannot always be achieved, despite the parameters. To account for these disturbances in the simulation, additional process data is considered. For example, fluctuations in current and voltage occur when contamination is present. This can also be captured as a boundary condition in the simulation. This allows for the creation of a database that considers different fluctuations in process variables for a given welding parameter. The machine learning model used is then able to reliably predict a simulation result for new, unknown fluctuations in the time series.

Claims

Patent claims 1. Method for monitoring a joining process, wherein a first joining partner and a second joining partner are provided (1), a simulation model of the first joining partner and the second joining partner as well as of the joining process to be carried out is provided (2), a substitute model is generated based on the simulation model and is trained using a machine learning algorithm (3), the first joining partner and the second joining partner are joined in a real joining process using a joining device (4), wherein at least one process parameter of the real joining process is determined and the determined process parameter is taken into account as a boundary condition in the substitute model, and the substitute model is applied simultaneously with the execution of the real joining process (5),wherein the actual joining process is controlled or regulated based on a result of the substitute model and / or a result of the actual joining process is controlled based on the result of the substitute model (6).

2. Method according to claim 1, characterized in that the joining process is resistance spot welding or projection welding.

3. Method according to claim 1 or claim 2, characterized in that the at least one process parameter is selected from an electric current, preferably a welding current, an electric voltage, preferably a welding voltage, an electrode force, a contact condition and / or an electrode movement.

4. Method according to one of the preceding claims, characterized in that the simulation model is a finite element model, a finite volume model or a computational fluid dynamics model.

5. Method according to one of the preceding claims, characterized in that the machine learning algorithm is selected from an artificial neural network, preferably a deep neural network, a support vector machine, a decision tree method, a random forest method or a gradient method.

6. Method according to one of the preceding claims, characterized in that the joining process is regulated or controlled in real time depending on the simulation result and / or a result is checked.

7. Device for monitoring a joining process, comprising a joining unit configured to connect a first joining partner and a second joining partner, wherein at least one sensor is configured to detect at least one process parameter, an electronic computing unit is configured to provide a simulation model of the first joining partner and the second joining partner as well as of the joining process to be carried out, taking into account the at least one process parameter, and to generate a substitute model based on the simulation model and to use a machine to train a learning algorithm, whereby the determined process parameter is taken into account as a boundary condition in the substitute model, to apply the substitute model simultaneously with the execution of the real joining process, whereby the electronic computing unit is also trained to control or regulate the real joining process based on a result of the substitute model and / or to check a result of the real joining process based on the result of the substitute model.

8. Computer program product comprising a computer program comprising software means for carrying out a method according to any one of claims 1 to 6 or for controlling a device according to claim 7 when the computer program is executed in the electronic computing unit.

Images