Method for resistance spot welding of aluminium sheets

Abstract

The invention relates to a method for resistance spot welding of aluminium sheets. According to the invention, it is provided that - a model (2) comprising a neural network (9) with experimental signal profiles from several experimental test welds, material parameters (5) and physical and engineering effects defined in the form of differential equations (10) which regularize the neural network (9) in terms of physical consistency, is trained, and - for the execution of a welding process (3) a welding control (4) before the welding process (3) welding parameters and / or reference curves (11) generated by the neural network (9) for various physical quantities are provided as setpoint (12) and / or controlled variable.

Description

[0001] The invention relates to a method for resistance spot welding of aluminium sheets.

[0002] As described in EP 3 895 834 B1, a method and a control unit for resistance welding are known from the prior art. In this method, welding processes are carried out in which welding electrodes are pressed against a weld point on the workpieces and energized with a welding current according to predefined welding parameters. During these welding processes, welding data describing the welding process are determined. An evaluation of the welding data is performed using a regression model, whereby the probability of weld spatter occurring in the next welding process, and the time at which the weld spatter is expected to occur in the next welding process, are determined.When the specified probability reaches a threshold, one of the predefined welding parameters is adjusted to bring the probability below the threshold. During this adjustment, a regression model is used to determine the relationship between the probability and the spatter timing (as the dependent variable) and the welding parameter (as the independent variable). Based on this relationship, the specified welding parameter is selected and adjusted.

[0003] US Patent 2002 / 0008086 A1 describes an assembly system that uses a neural network to control the joining process. A dynamic analog model is used for the neural elements that configure the network. The system includes a detector, a controller, and a neural network. The detector detects the connection state of a joining section when workpieces are joined. The controller regulates the performance of the assembly system. In response to the detector's output signals, the neural network transmits signals to the controller.

[0004] The invention is based on the objective of providing a method for resistance spot welding of aluminium sheets that is improved compared to the prior art.

[0005] The problem is solved according to the invention by a method for resistance spot welding of aluminium sheets with the features of claim 1.

[0006] Advantageous embodiments of the invention are the subject of the dependent claims.

[0007] In a method for resistance spot welding of aluminum sheets, a model comprising a neural network is trained with experimental signal profiles from several experimental test welds (i.e., test resistance spot welds), material parameters, and physical and engineering effects defined as differential equations that regularize the neural network for physical consistency. To execute a welding process, i.e., a resistance spot welding process, welding parameters generated by the neural network, such as current and / or electrode force, and / or reference curves for various physical quantities, such as current, resistance, electrode force, and / or electrode position, are provided to a welding controller as setpoints and / or controlled variables prior to the welding process.

[0008] The procedure described here can therefore comprise two combined procedure variants or one of the two procedure variants.

[0009] In one method variant, reference curves for various physical quantities of the welding control system are provided as setpoints and / or controlled variables before the welding process. These reference curves can be generated for parameters such as current, resistance, electrode force, and electrode position. This ensures that the welding control system knows the setpoint at every point during the welding process, allowing it to monitor or intervene accordingly, and in particular, to control the welding process.

[0010] The reference curves are generated by a physics-regulated neural network. This model is trained with material parameters, experimental signal waveforms, and known physical and engineering effects. The effects are defined as differential equations and regularize the neural network in terms of physical consistency. A key feature is the dynamic updating of the material parameters during the reference curve generation. The material parameters change depending on the temperature. These parameters were measured and are incorporated into the reference curve. This means that a thermophysical material model is integrated into the neural network. Furthermore, additional equations can be passed to the network.

[0011] The control can be achieved using pre-generated, physically consistent reference curves. This is a structured approach to the complex physical effects of resistance spot welding. The respective reference curve is parameter-specific and is determined for each weld spot using material parameters, experimental investigations, and differential equations.

[0012] In comparison to the state of the art, where the execution of welding processes is based on individual experimental curves or physically inconsistent formulas, In the described solution, reference curves are determined / learned by a physically consistent neural network.

[0013] The described solution addresses the problem of ensuring that, in the context of aluminum resistance spot welding, weld points are joined using a controlled process that meets quality criteria and prevents production stoppages. These quality criteria include a large weld point diameter and the prevention of weld spatter. Particularly in automotive body construction, a large weld point diameter is advantageous because it can withstand greater loads. Therefore, a larger diameter, or one calculated to meet the required diameter, is beneficial, for example, for crash testing.

[0014] For this purpose, a control system is used that monitors, evaluates, and intervenes as needed during the welding process. The control system must be comprehensible and, since it is a physical process, physically consistent. This technical challenge of the control system is particularly crucial for automotive body-in-white production, especially to ensure safety in the event of a crash.

[0015] The integrated approach described here, which incorporates experimental data (e.g., curves and / or measured point diameters and / or measured electrode indentation areas in the aluminum sheet), material parameters, and differential equations, generates accurate, realistic, and physically consistent reference curves. Experimental data and physical laws are integrated into the model. The neural network can balance these two elements and quantify the model's performance. This results in a physically consistent and verifiable reference curve.

[0016] The respective reference curve is used during the welding process to control and evaluate it. A better reference curve leads to improved control and a more consistent result / quality standard. The welding process itself is highly unstable. For example, the aluminum sheets may have scratches, oil, residue, gaps, and / or varying welding angles. The improved control enabled by the described solution can compensate for these uncontrollable factors.

[0017] Especially when resistance spot welding aluminum, the susceptibility to abnormal conditions is higher compared to steel. Abnormal conditions include, for example, electrode misalignment or the angle of attack.

[0018] In the other method variant, the welding parameters required to achieve a predefined quality criterion, particularly a large weld spot diameter, are provided in advance. The main welding parameters are current and electrode force. These parameters are generated by a physically regularized neural network. This model is trained using material parameters, (successful) experimental signal profiles, and known physical and engineering effects. The model training is identical to the first method variant. The effects are defined as differential equations and regularize the neural network in terms of physical consistency. A key advantage is the dynamic updating of the material parameters, especially during the reference curve. This means that a thermophysical material model is incorporated into the neural network.Furthermore, additional equations can be passed.

[0019] Automated parameter determination using physically consistent machine learning is possible. This is a structured approach to the complex physical effects of resistance spot welding.

[0020] In this method variant, the specified quality criterion, i.e., a formulation of the quality requirement, can be derived from design requirements. The described solution specifically addresses the problem of performing the resistance spot welding process in accordance with the quality requirements. This means, for example, that the weld spot diameter is achieved. In the prior art, the appropriate welding parameters for the welding process to achieve the spot diameter are unknown beforehand and must be determined experimentally in a time-consuming process. This problem is solved by the integrated approach of experimental curves, material parameters, and differential equations, which generates the relevant welding parameters according to the quality requirement. This allows access to a broader database / knowledge base. Experimental data and physical laws are integrated into the model.The neural network can balance between the two and its model performance can be quantified. This results in a physically consistent and comprehensible control system.

[0021] The main objective of this process variant is to find welding parameters that yield the desired quality indicator, particularly a good weld spot diameter. There is a wide search range available for finding these parameters. For example, the electric current can be varied between approximately 26 and 50 kiloamperes, and the electrode force between 4 kN and 8 kN, especially with common welding hardware manufacturers. Currently, this is done purely experimentally. With new materials, corresponding new experiments must be conducted, and all combinations must be retested. For example, in vehicle body construction, specific design requirements must be met, such as crash test performance and low weight. Therefore, it may be necessary to redesign an entire body during the design phase, using new materials, different material thicknesses, and a different construction method.This means that the joining techniques, in this case resistance spot welding, must be reviewed to meet the new requirements. New welding parameters must therefore be determined to ensure the joint's durability. As mentioned above, this currently involves conducting new experiments. The described solution reverses this process and incorporates the quality metric as a parameter. The current, force, and welding plan are generated from the quality indicator.

[0022] In summary, the second method variant generates the necessary welding parameters to achieve good weld points, while the first method variant enables good control of the welding process to obtain consistently good weld points.

[0023] Exemplary embodiments of the invention are explained in more detail below with reference to drawings.

[0024] This shows: Fig. 1. Schematic representation of an abstracted depiction of a first variant of a process for resistance spot welding of aluminium sheets, Fig. 2 schematically a detailed representation of the first procedure variant, Fig. 3. Schematic representation of an abstracted model training of the first procedure variant, Fig. 4 schematically a detailed representation of the model training of the first procedure variant, Fig. 5 schematically exemplary reference curves of the first method variant, Fig. 6 schematically an abstracted representation of a second process variant of the process for resistance spot welding of aluminium sheets, Fig. 7 schematically a detailed representation of the second procedure variant, Fig. 8 schematically an abstracted representation of a model training of the second procedure variant, Fig. 9 schematically a detailed representation of the model training of the second procedure variant, and Fig. 10 schematically shows an example of parameter finding for the second method variant.

[0025] Corresponding parts are marked with the same reference symbols in all figures.

[0026] Based on the Fig. Sections 1 to 10 below describe two exemplary process variants of a resistance spot welding process for aluminum sheets. These relate to... Fig. 1 to 5 the first procedure variant and the Fig. Points 6 to 10 relate to the second procedure variant. The two procedure variants can also be combined.

[0027] In summary, the method provides in particular that a model 2, comprising a neural network 9, is trained with experimental signal profiles from several experimental test welds, material parameters 5, and physical and engineering effects defined in the form of differential equations 10, which regularize the neural network 9 in terms of physical consistency. To carry out a welding process 3, welding parameters generated by the neural network 9, such as a current I and / or an electrode force F, and / or reference curves 11 for various physical quantities, such as current I(t), resistance Ω(t), electrode force F(t), and / or electrode position u(t), are provided to a welding control system 4 as setpoints 12 and / or controlled variables prior to the welding process 3.

[0028] This summarized procedure description therefore includes both procedure variants, which can be carried out individually or in combination.

[0029] The following will first be based on the Fig. Sections 1 to 5 describe the first procedure variant.

[0030] This first method variant includes in particular the determination of the reference curves 11 for resistance spot welding of aluminium sheets.

[0031] Resistance spot welding is used by the automotive industry to join sheet metal parts in vehicle body construction. Depending on the manufacturer and vehicle model, up to 60% of the joints are made using resistance spot welding.

[0032] Aluminum alloys are increasingly replacing steel alloys in vehicle body construction. Generally, resistance spot welding of aluminum alloys is associated with more problems compared to steel alloys. For this reason, a robust and traceable control system is required. There are generally various ways to control a welding process. These include envelope curves, individual experimental example curves, setpoints, or characteristic points.

[0033] In this first variant of the process, reference curves 11 for aluminum resistance spot welding are determined by pre-trained, physically regularized neural networks 9. Material parameters 5, process parameters, and differential equations 10 are used for this purpose. The neural network 9 can be directly regularized with the differential equations 10 via a loss function, thus forcing it to conform to physical laws. Physical equations with known factors can be used, for example, thermodynamic laws, the heat equation, and control functions. A special feature is that the material parameters 5 are dynamically updated during model training.

[0034] The reference curves 11 describe the electromagnetic and thermomechanical fields using curves for the current I(t), the electrical resistance Ω(t), the electrode force F(t), and the electrode position u(t). This requires the presence of sensors capable of measuring the actual values ​​14 of these quantities. The welding control system 4 acquires these reference curves 11 and uses them either for monitoring or to regulate the welding process 3 in such a way that the reference curves 11 are adhered to.

[0035] In the context of aluminum resistance spot welding, for example in vehicle body-in-white production, the technical challenge lies in creating weld points using a control system (13) that meets quality criteria and prevents production stoppages. These quality criteria include a large weld point and the prevention of weld spatter. A control system (13) is used to monitor and evaluate the welding process (3) and intervene if necessary. This control system (13) must be traceable and, since it involves a physical process, physically consistent. This technical challenge of the control system (13) is essential for automotive body-in-white production.

[0036] The technical solution consists of providing reference curves 11 for various physical quantities of the welding control system 4 as setpoints 12 and / or controlled variables before the welding process 3. The reference curves 11 can be generated for current, resistance, electrode force, and electrode position, among other parameters. This ensures that the welding control system 4 knows the setpoint 12 at all times during the welding process 3. It can then monitor or intervene accordingly.

[0037] The reference curves 11 are generated by a physics-regulated neural network 9. This model 2 is trained with material parameters 5, experimental signal waveforms, and known physical and engineering effects. The effects are defined in the form of differential equations 10 and regularize the neural network 9 in terms of physical consistency. A special feature is the dynamic updating of the material parameters 5 during the reference curve 11. This means that a thermophysical material model 19 is incorporated into the neural network 9. Furthermore, additional equations can be passed to it.

[0038] For example, a large number, say tens of thousands, of weld points with different process and material parameters 5 are joined and measured, and different thermophysical material parameters 5 are determined for each material batch. The model 2 is trained on this experimental data basis. The reference curves 11 are generated for each weld point.

[0039] The control 13 can be carried out using pre-generated, physically consistent reference curves 11. This is a structured approach for the complex physical effects of resistance spot welding. The reference curve 11 is parameter-specific and is determined for each weld spot using material parameters 5, experimental investigations, and differential equations 10.

[0040] The state of the art is based on individual experimental curves or physically inconsistent formulas. In contrast, the first method variant defines several reference curves 11 by a physically consistent model 2. Fig. Figure 1 shows an abstract representation of the first process variant. The process parameters, in particular material parameters 5 and current and force profiles 6, are entered via a user interface 1 of a resistance spot welding system and fed to the model 2, in particular a machine learning model 2 (machine learning). Furthermore, the welding process 3 is physically carried out. The welding control 4 regulates the welding process 3 according to expectations and specifications. It receives information from the model 2 and the welding process 3 for this purpose. Fig. Figure 2 shows a detailed representation of an example of the first method variant.

[0041] The material parameters 5 are determined by a material test 7. This involves an experimental or computer-aided investigation to determine the material parameters 5, for example thermal conductivity and / or electrical resistance.

[0042] The current and force profiles 6 are defined at the user interface 1.

[0043] The material parameters 5 and current and force profiles 6 are incorporated into the model 2, in particular into its neural network 9. In addition, the model 2 includes the differential equations 10, which are also incorporated into the neural network 9.

[0044] Model 2, in particular its neural network 9, provides the reference curves 11 and / or setpoints 12, which are used in the welding control 4, in particular in its regulation 13.

[0045] Information from the welding process 3, in particular actual values ​​14 of the welding process 3, which are determined by means of at least one measuring instrument 15, flows into the welding control 4, in particular into its control unit 13. The measuring instrument 15 has a sensor. This sensor is designed, for example, to detect the current, a voltage drop and / or the electrode force F.

[0046] Regulation 13 is in particular a regulation 13 of an error between the actual value 14 and the setpoint 12, for example by a proportional-integral-differential controller.

[0047] The regulatory behavior of rule 13 is, in particular, as follows: Control 13 can consist of reference curves 11 and / or master curves, envelopes, characteristic points, or setpoint exceedances. In this first method variant, the generated reference curve 11 serves as the basis for control 13. Depending on the implementation, the welding process 3 is thus monitored or controlled in such a way that the reference curve 11 is reached.

[0048] Differential equations 10 describe the physical fields. Physical quantities, such as the current E, can be determined from them. el and the electrode position v.

[0049] For example, the following differential equations 10 are used: - Law of conservation of momentum: ∇⋅S+ρmab=ρmau¨

[0050] This includes: S the stress tensor b the body force, in German volume force, v the acceleration, this is replaced by the second time derivative of the position “u”, p_ma is the density, ∇ is an operation called Divergence. - Maxwell's formulation of Gauss's law: ∇⋅Eel=ρε0

[0051] This includes: E el The electric field. The current intensity is a scalar that determines the field. Electromagnetism uses the electric field in its calculations. ρ is the charge carrier density, ε0 is the electric constant. - Heat conduction equation: ∂∂x(k∂T∂x)+∂∂y(k∂T∂y)+∂∂z(k∂T∂z)−ρmacP∂T∂t=−qρmacP

[0052] Formally (3) can be simplified to: ρmacpT˙=k∇⋅∇T+qs

[0053] Otherwise, in the next iteration, the equations can be expressed in detail as a function of x, y, z, t.

[0054] This includes: p_ma the density c_p is the specific heat capacity T the temperature k the thermal conductivity, q_s is the heat generation per m 3 .

[0055] Other equations include, for example, thermal expansion, models for the distribution of current density, the contact resistance over time, or hydrostatic pressure.

[0056] Fig. Figure 3 shows an abstract representation of the model training of the first procedure variant. It includes an input 16, a transformation 17, the model 2, and an output 18.

[0057] The input 16 are input variables for the model 2. These input variables are material parameters 5 and current and force profiles 6.

[0058] In transformation 17, the input variables are converted into a suitable format for model 2. This step is particularly important for physically regularized models 2.

[0059] Model 2 consists of the neural network 9, which is regularized by the differential equations 10. Regularization is a method by which Model 2 is penalized for physics inconsistencies, defined by the differential equations 10. This is done by the previously mentioned loss function.

[0060] Output 18 refers to the output variables of model 2. These are the reference curves 11 for various physical fields.

[0061] Fig. Figure 4 shows a detailed representation of the model training of the first procedure variant.

[0062] Input 16 includes the material parameters 5 and the current and / or force profiles 6.

[0063] The transformation 17 includes a material model 19 and a network generation 20.

[0064] Material model 19 is a thermophysical material model 19 and is generated based on material parameters 5. Material model 19 consists of material curves and is a function of material parameters 5 over temperature.

[0065] The input variables, designated as Input 16, are processed by the mesh generation 20. This generates the domains for the differential equations 10 (x, y, z coordinates and time t).

[0066] The differential equations 10 describe the physical laws of the welding process 3.

[0067] The material parameters 5 are incorporated into the material model 19. The material parameters 5 and the current and / or force profiles 6 are incorporated into the network generation 20. The material model 19 and the network generation 20, or their respective results, are incorporated into the neural network 9. Furthermore, the differential equations 10 are incorporated into the neural network 9. The neural network 9 provides at least one reference curve 11 as output 18.

[0068] Fig. Figure 5 shows exemplary reference curves 11 of the first process variant for aluminium sheets made of EN AW-6014 with a material thickness of 1.1 mm. Resistance spot welding is performed with a current I of 39 kA and an electrode force F of 5 kN.

[0069] The diagrams shown are: top left, a diagram of the current I over time t; bottom left, a diagram of an electrical resistance Ω over time t; top right, a diagram of the electrode force F over time t; and bottom right, a diagram of the electrode position u over time t.

[0070] The diagrams show a reference curve 11 and an actual value curve 21.

[0071] The graphics of Fig. Figure 5 shows the pre-generated reference curves 11 of the current I, the electrode force F, the resistance Ω and the electrode position u, and the actual values ​​14 of the measuring instruments 15. The quantification of the mean accuracy is given by the coefficient of determination: R2 = 0.94.

[0072] The integrated approach of experimental curves, material parameters 5, and differential equations 10 generates accurate, realistic, and physically consistent reference curves 11. Experimental data and physical laws are integrated into a model 2. The neural network 9 can balance between the two and its model performance can be quantified. This results in a physically consistent and comprehensible control system 13.

[0073] For a company producing aluminum sheets using resistance spot welding, especially in the automotive industry, the advantage is that a reliable, physically consistent, competition- and supplier-independent control 13 is enabled, supplier-independent because material parameters can be determined experimentally and then incorporated into the model.

[0074] The following will be based on the Fig. Sections 6 to 10 describe the second procedure variant.

[0075] This second method variant includes, in particular, the determination of the welding parameters for resistance spot welding of aluminium sheets.

[0076] Resistance spot welding, along with other joining technologies, is used by the automotive industry for the construction of vehicle body shells. Depending on the manufacturer and vehicle model, up to 60% of the joining points are made using resistance spot welding.

[0077] Aluminum alloys are increasingly replacing steel alloys in vehicle body construction. Generally, resistance spot welding of aluminum alloys is associated with more problems compared to steel alloys.

[0078] A major task is defining suitable parameters for the new welding points to meet the design requirements. In this second method variant, this parameter determination is achieved using machine learning. Quality requirements 22 can be derived from the design requirements. The welding parameters for the quality requirements 22 are determined using a neural network 9. Material parameters 5, process parameters, and differential equations 10 are used for this purpose. The neural network 9 can be directly regularized with the differential equations 10 via the loss function, thus ensuring consistency with physical laws. Physical equations with known factors can be used, for example, the heat conduction equation, thermodynamic laws, and control functions.

[0079] One special feature is that the material parameters 5 are dynamically updated during model training.

[0080] The welding control unit 4 is given the appropriate setpoint values ​​12 for monitoring and control 13. This requires the presence of sensors capable of measuring the actual values ​​14 of these quantities.

[0081] The formulation of quality requirement 22 can be derived from the design requirements. The technical problem lies in carrying out the resistance spot welding joining process in accordance with quality requirement 22. This means, for example, achieving the required spot diameter. The appropriate parameters for welding process 3, to achieve the required spot diameter, are unknown beforehand and must currently be determined experimentally through a complex process.

[0082] In the automotive industry, this problem is particularly relevant in the process development of joining technologies. Here, the starting parameters / parameter space for specific material combinations are defined for the entire vehicle body-in-white. Quality requirement 22 originates from vehicle development.

[0083] The technical solution consists of pre-defining the welding parameters required to meet the quality criteria. The main welding parameters are maximum current I and maximum electrode force F.

[0084] The welding parameters are generated by a physically regularized neural network 9. This model 2 is trained with material parameters 5, experimental signal profiles, and known physical and engineering effects. The effects are defined in the form of differential equations 10 and regularize the neural network 9 in terms of physical consistency. A special feature is the dynamic updating of the material parameters 5 during the reference curve 11. This means that a thermophysical material model 19 is incorporated into the neural network 9. Furthermore, additional equations can be passed to it, as described in the first variant.

[0085] For example, a large number, say tens of thousands, of weld points with different process and material parameters 5 are joined and measured, and various thermophysical material parameters 5 are determined for each material batch. Model 2 is trained on this (experimental) data basis. Model 2, in particular together with the relevant differential equations 10, is used to determine the required welding parameters based on the defined quality criteria.

[0086] This second method variant enables automated parameter determination using physically consistent machine learning. This is a structured approach to the complex physical effects of resistance spot welding.

[0087] Fig. Figure 6 shows an abstract representation of the second process variant. It includes a quality requirement 22, the user interface 1, the model 2, the welding process 3, and the welding control 4.

[0088] Quality requirement 22 is a formulation of quality criteria based on a target variable 23, in particular a specific spot diameter and the avoidance of weld spatter. The process parameters, in particular material parameters 5 and current and / or force profiles 6, are entered via the user interface 1 of the resistance spot welding system and fed to model 2, in particular a machine learning model 2 (machine learning). Furthermore, the welding process 3 is physically carried out. The welding control system 4 regulates the welding process 3 according to expectations and specifications. It receives information from model 2 and the welding process 3 for this purpose.

[0089] Fig. Figure 7 shows a detailed representation of an example of the second method variant.

[0090] The material parameters 5 are determined by material testing 7. This involves an experimental or computer-aided investigation to determine the material parameters 5, for example, thermal conductivity and / or electrical resistance.

[0091] The target variable 23 is determined from the quality requirement 22.

[0092] The current and force profiles 6 are determined by experiments 8.

[0093] The material parameters 5, the target quantity 23, and the current and force profiles 6 are incorporated into the model 2, in particular into its neural network 9. In addition, the model 2 includes the differential equations 10, which are also incorporated into the neural network 9.

[0094] Model 2, in particular its neural network 9, provides the setpoints 12, which are used in the welding control 4, in particular in its regulation 13.

[0095] Information from the welding process 3, in particular actual values ​​14 of the welding process 3, which are determined by means of at least one measuring instrument 15, flows into the welding control 4, in particular into its control unit 13. The measuring instrument 15 has a sensor. This sensor is designed, for example, to detect the current, the voltage drop and / or the electrode force F.

[0096] Regulation 13 is in particular a regulation 13 of an error between the actual value 14 and the setpoint 12, for example by a proportional-integral-differential controller.

[0097] The regulatory behavior of rule 13 is, in particular, as follows: Control 13 can consist of reference curves 11 and / or master curves, envelopes, characteristic points, or setpoint exceedances. In this second method variant, the generated reference curve 11 serves as the basis for control 13. Depending on the implementation, the welding process 3 is thus monitored or controlled in such a way that the reference curve 11 is reached.

[0098] The differential equations 10 describe the physical fields. From them, the physical quantities can be determined, for example the current I and the electrode position v.

[0099] For this second method variant, the above-mentioned differential equations 10 (1) to (4) are also used, for example.

[0100] Fig. Figure 8 shows an abstract representation of the model training of the second procedure variant. It includes the input 16, the transformation 17, the model 2, and the output 18.

[0101] The input 16 are input variables for the model 2. These input variables are material parameters 5 and current and force profiles 6.

[0102] In transformation 17, the input variables are converted into the appropriate format for model 2. This step is particularly important for physically regularized models 2.

[0103] Model 2 consists of the neural network 9, which is regularized by the differential equations 10. Regularization is a method by which Model 2 is penalized for physics inconsistencies defined by the differential equations 10.

[0104] The output variables of model 2 are designated as output 18. These are the reference curves 11 for various physical fields.

[0105] Fig. Figure 9 shows a detailed representation of the model training of the second procedure variant.

[0106] Input 16 includes the material parameters 5 and the current and / or force profiles 6.

[0107] The transformation 17 includes a material model 19 and a network generation 20.

[0108] Material model 19 is a thermophysical material model 19 and is generated based on material parameters 5. Material model 19 consists of material curves and is a function of material parameters 5 over temperature.

[0109] The input variables, designated as Input 16, are processed by the mesh generation 20. This generates the domains for the differential equations 10 (x, y, z coordinates and time t).

[0110] The differential equations 10 describe the physical laws of the welding process 3.

[0111] The material parameters 5 are incorporated into the material model 19. The material parameters 5 and the current and / or force profiles 6 are incorporated into the network generation 20. The material model 19 and the network generation 20, or their respective results, are incorporated into the neural network 9. Furthermore, the differential equations 10 are incorporated into the neural network 9. The neural network 9 provides at least one reference curve 11 as output 18.

[0112] Fig. Figure 10 shows an example of parameter finding for the second process variant for resistance spot welding of aluminium sheets made of EN AW-5812 with a material thickness of 1.5 mm.

[0113] The diagram shows current I and electrode force F, containing a spatter line 24 and a spot diameter line 25. This represents an automatically generated parameter space in which the conflicting objectives of weld spatter and minimum quality are met. Above spatter line 24, model 2 expects weld spatter, and below spot diameter line 25, model 2 expects the quality requirement 22 not to be met. The suitable parameter space lies between spatter line 24 and spot diameter line 25.

[0114] The integrated approach of experimental curves, material parameters 5, and differential equations 10 generates accurate welding parameters according to the quality requirement 22. This allows access to a broader data / knowledge base. Experimental data and physical laws are integrated into a model 2. The neural network 9 can balance between the two and its model performance can be quantified. This results in a physically consistent and traceable control system 13.

[0115] In this second method variant, the welding parameters are determined based on experimental data and physical laws. This allows for structured trials using machine knowledge, rather than relying on intuition or trial and error. Reference symbol list 1 User interface 2 Model 3 Welding process 4 Welding control 5 Material parameters 6 Current and / or force profile 7 Materials testing 8 Experiment 9 neural network 10 Differential equation 11 Reference curve 12 Target value 13 Regulation 14 Actual value 15 Measuring instrument 16 Input 17 Transformation 18 Output 19 Material model 20 Network generation 21 Actual Value Curve 22 Quality requirements 23 Target variable 24 spray line 25 Point diameter line F Electrode force I Current intensity t time Electrode position Ω resistance QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] EP 3 895 834 B1

[0002] US 2002 / 0008086 A1

[0003]

Claims

[1] Method for resistance spot welding of aluminium sheets, characterized by , that - a model (2) comprising a neural network (9) with experimental signal profiles from several experimental test welds, material parameters (5) and physical and engineering effects defined in the form of differential equations (10) which regularize the neural network (9) in terms of physical consistency, is trained, and - for the execution of a welding process (3) a welding control (4) before the welding process (3) welding parameters and / or reference curves (11) generated by the neural network (9) for various physical quantities are provided as setpoint (12) and / or controlled variable. [2] Method according to claim 1, characterized by , that reference curves (11) for current, resistance, electrode force (F) and / or electrode position (v) are generated. [3] Method according to claim 1 or 2, characterized by that a current (E) is used as a welding parameter el ) and / or an electrode force (F) is generated.

Images