A method for rapid prediction of residual stress in laser powder bed fusion by fusing experimental data, finite element simulation and neural network

Abstract

The application discloses a kind of fusion experimental data, laser powder bed melting residual stress fast prediction method of finite element simulation and neural network, belong to stress measurement field, this method is based on " finite element simulation→experimental data calibration→neural network training " three-stage modeling strategy, organically integrates the advantage of three kinds of information sources, while avoiding their respective disadvantages, break through the bottleneck problem of high-precision experimental detection high cost, precision and efficiency difficult trade-off in finite element simulation, weak generalization ability and insufficient physical consistency;The method combines high-fidelity finite element simulation results, spatially sparse experimental observations and physical law constraints in thermal-mechanical control equations, to build a multi-source information-driven neural network model with strong interpretability, high prediction efficiency and good adaptability, achieving fast prediction of laser powder bed melting residual stress.

Description

Technical Field

[0001] This invention belongs to the field of stress measurement, and more specifically, relates to a method for rapid prediction of residual stress in laser powder bed fusion that integrates experimental data, finite element simulation and neural networks. Background Technology

[0002] In high-end manufacturing fields such as aerospace, precision manufacturing, and high-performance equipment, laser powder bed fusion (LPBF) technology is widely used due to its ability to integrally form complex metal structures. However, the widespread adoption of this technology has been severely hampered by significant residual stress generated during the forming process. These residual stresses mainly originate from the extremely high temperature gradient and non-uniform phase transformation caused by the rapid melting and cooling of the laser, which can easily lead to warping, cracking, dimensional inaccuracies, and even fatigue performance degradation of components during manufacturing or service. This has become a major bottleneck restricting the application of this technology in the manufacture of critical load-bearing structural components.

[0003] Currently, the detection of residual stress mainly relies on two technical approaches: experimental measurement and numerical simulation.

[0004] In experimental measurement, mainstream methods include blind-hole method, X-ray diffraction, ultrasonic method, and neutron diffraction. Among them, neutron diffraction technology can measure the internal three-dimensional stress distribution of the sample without damaging it, with a penetration depth of up to the centimeter level, making it particularly suitable for non-destructive testing of internal stress in commonly used additive manufacturing materials such as titanium alloys. However, these experimental methods generally rely on large-scale equipment, resulting in high testing costs and long cycles. Furthermore, they typically only provide the final stress state after forming and cannot monitor and predict stress evolution during the manufacturing process in real time, thus making them difficult to apply to the rapid iteration and optimization design of process parameters.

[0005] In numerical simulation, the finite element method (FEM) based on thermo-mechanical coupling theory is currently the mainstream tool for studying the formation mechanism and evolution of residual stress. This method, by constructing a physical model that incorporates material nonlinearity, latent heat of phase transition, and powder-to-solid transformation, can simulate the transient behavior of temperature and stress fields with high fidelity. Existing technologies, such as the transient model based on the "birth and death element" strategy proposed in Chinese Patent Publication No. CN114386303A, or methods using the Johnson-Cook constitutive model to consider dynamic material responses, have shown good prediction accuracy under specific conditions. However, high-fidelity finite element simulations consume enormous computational resources, have extremely long solution times, and their accuracy heavily depends on the precise setting of initial and boundary conditions. Although simplified models such as the equivalent variation method have been developed, as provided by Chinese Patent Publication Nos. CN115455776A and CN119272549A, which significantly improve computational efficiency and enable their application to large-size components, their prediction accuracy often decreases accordingly. This contradiction between computational accuracy and efficiency greatly limits the application of traditional numerical simulation technology in rapid evaluation and large-scale parameter optimization in industrial settings.

[0006] In recent years, data-driven methods, represented by neural networks, have provided new insights for the rapid prediction of residual stress. These methods, by learning mapping relationships in historical data, can achieve near real-time stress field inference. For example, the paper "Residual Stress Prediction Based on Neural Networks in Selective Laser Melting" uses a BP neural network to model the ultrasonic residual stress measurement results of 316L stainless steel components, significantly improving prediction efficiency. The paper "Geometrically-informed predictive modeling of melt pool depth in laser powder bed fusion using deepMLP-CNN and metadata integration" proposes a deep MLP-CNN network model integrating geometric, image, and process parameters, effectively improving the accuracy of predicting complex morphologies. However, the performance of neural network models is highly dependent on the scale and quality of the training data. In the field of LPBF (Laser-Based Burnout), obtaining sufficient, high-precision experimental data is extremely costly, while relying solely on limited metadata will cause the model to inherit its errors and physical consistency is difficult to guarantee, resulting in insufficient generalization ability when facing new processes or structures.

[0007] Comprehensive analysis reveals that existing technologies generally face a "triangular dilemma" when addressing the problem of residual stress prediction in LPBF (Limited Particulate Airflow): the cost of high-precision experimental testing, the trade-off between accuracy and efficiency in finite element simulation, and the dual dependence of data-driven models on high-quality data and physical consistency—all three are difficult to balance. Currently, there is a lack of a rapid residual stress prediction solution that can achieve a balance between prediction accuracy, computational efficiency, and physical reliability at an industrially feasible cost. Summary of the Invention

[0008] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a rapid prediction method for residual stress in laser powder bed melting that integrates experimental data, finite element simulation, and neural networks. This method employs a phased, multi-source information fusion hybrid modeling framework, enabling residual stress prediction with high accuracy, high efficiency, and strong physical consistency.

[0009] To achieve the above objectives, according to a first aspect of the present invention, a method for rapid prediction of residual stress in laser powder bed fusion, integrating experimental data, finite element simulation, and neural networks, is provided, comprising: Training phase: S1. The three-dimensional residual stress of the component sample under multiple different laser powder bed melting process conditions is obtained through simulation software, and the simulation data of residual stress are obtained. ; n=1,2,…,N, where N is the total number of stress measurement points during simulation; For the nth stress simulation measurement point The coordinates; S2, prepare the solid component sample under the specified laser powder bed melting process conditions, measure the three-dimensional residual stress of the solid component sample, and obtain the actual measurement data of the residual stress. ; m = 1, 2, ..., M, where M is the number of stress measurement points during the actual measurement, and M <N, For the m-th actual stress measurement point The coordinates; where the set of actual stress measurement points is a subset of the set of stress simulation measurement points; S3, for and Perform linear fitting to obtain the fitting parameters. and ;Will Substitute into the formula The calibrated residual stress simulation data were calculated. Use it as the training dataset, The coordinates, the specified laser powder bed melting process conditions, and the geometric characteristic parameters of the component sample are used as inputs. To supervise the training of the physical information neural network; in, , These are the global scaling factors for the stress components in the x, y, and z directions, respectively. These are the compensation amounts for the stress components in the x, y, and z directions, respectively. Application phase: The coordinates of the target stress measurement point of the target component, the laser powder bed melting process conditions of the target component, and the geometric feature parameters of the target component are input into a trained physical information neural network to obtain the three-dimensional residual stress at the target stress measurement point.

[0010] According to a second aspect of the present invention, an electronic device is provided, comprising: a computer-readable storage medium and a processor; The computer-readable storage medium is used to store executable instructions; The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the method as described in the first aspect.

[0011] According to a third aspect of the invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to perform the method as described in the first aspect.

[0012] According to a fourth aspect of the invention, a computer program product is provided, comprising a computer program or instructions that, when executed by a processor, implement the method described in the first aspect.

[0013] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: The method provided by this invention is based on a three-stage modeling strategy of "finite element simulation → experimental data calibration → neural network training." It organically integrates the advantages of three types of information sources while avoiding their respective disadvantages, overcoming the bottlenecks of high cost in high-precision experimental testing, difficulty in balancing accuracy and efficiency in finite element simulation, weak generalization ability, and insufficient physical consistency. This method organically combines high-fidelity finite element simulation results, spatially sparse experimental observations, and physical constraints from thermo-mechanical control equations to construct a multi-source information-driven neural network model with strong interpretability, high prediction efficiency, and good adaptability, enabling rapid prediction of residual stress in laser powder bed melting. This method effectively combines the advantages of three types of data sources, overcoming the problems of traditional neural networks being difficult to train and converge, lacking simulation error correction, and having scarce experimental data that hinders generalization. It has the following advantages: 1) High prediction efficiency: The well-trained neural network model can complete the prediction in seconds or even milliseconds. The computational efficiency is more than two orders of magnitude higher than that of traditional finite element simulation, which can meet the needs of rapid iteration and real-time evaluation of process parameters. 2) Good consistency between prediction accuracy and physics: By embedding physical laws such as stress balance equations as constraints into the neural network and calibrating the system error in combination with experimental data, it is ensured that the prediction results not only match the experimental measurements (error controlled within ±20%), but also conform to the basic physical laws. 3) Strong generalization ability and low data cost: By augmenting data with limited metadata and combining it with transfer learning techniques, the dependence of pure data-driven methods on massive experimental data is effectively overcome, enabling the model to be economically and efficiently transferred to prediction tasks of new materials and new configurations. 4) Convenient engineering integration: As a software-level solution, this method does not rely on specific models of forming equipment. It can be seamlessly integrated into the existing digital design-manufacturing process using only conventional simulation and experimental testing tools. It has a low implementation threshold and strong practicality. In summary, the rapid prediction method for residual stress in laser powder bed melting, which integrates experimental data, finite element simulation, and neural networks, provided by this invention is suitable for predicting residual stress in laser powder bed melting, especially for large-sized complex components where experimental costs are high and simulation efficiency is low. It is particularly suitable for the rapid prediction and active control of residual stress in large-sized complex components in aerospace, precision manufacturing, and other fields where experimental costs are extremely high and traditional simulations are difficult to handle. Attached Figure Description

[0014] Figure 1 One of the flowcharts for a rapid prediction method of residual stress in laser powder bed fusion, which integrates experimental data, finite element simulation, and neural networks, is provided in an embodiment of the present invention. Figure 2 The second flowchart of the method for rapid prediction of residual stress in laser powder bed fusion, which integrates experimental data, finite element simulation and neural network, is provided for the embodiments of the present invention. Figure 3 A schematic diagram illustrating the calibration mapping from simulation results to experimental results provided in this embodiment of the invention; Figure 4 A diagram of a multi-branch neural network structure incorporating physical constraints provided in an embodiment of the present invention; Figure 5 This is a schematic diagram comparing the predicted results with the experimental residual stress provided in the embodiments of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0016] This invention provides a method for rapid prediction of residual stress in laser powder bed fusion, which integrates experimental data, finite element simulation, and neural networks. Figures 1-2 As shown, it includes: Training phase: S1. The three-dimensional residual stress of the component sample under multiple different laser powder bed melting process conditions is obtained through simulation software, and the simulation data of residual stress are obtained. ; n=1,2,…,N, where N is the total number of stress measurement points during simulation; For the nth stress simulation measurement point The coordinates.

[0017] Step S1 completes the construction of a high-fidelity simulation database: Under multiple different laser powder bed melting process conditions and typical component geometric combinations, a thermo-mechanical coupled finite element model is established. The model should consider material nonlinearity, latent heat of phase transition, and powder-to-solid material transformation, and use the "birth and death element method" to simulate the layer-by-layer deposition process of the material, thereby obtaining the three-dimensional residual stress distribution under different process conditions and component configurations, forming the original high-fidelity simulation database.

[0018] Preferably, the laser powder bed melting process conditions include process parameters and a scanning strategy; wherein, the process parameters include laser power, scanning speed, and powder bed thickness; the scanning strategy includes a scanning path and rotation angle, and the scanning path includes a scanning vector, thermal history, etc. S2, prepare the solid component sample under the specified laser powder bed melting process conditions, measure the three-dimensional residual stress of the solid component sample, and obtain the actual measurement data of the residual stress. ; m = 1, 2, ..., M, where M is the number of stress measurement points during the actual measurement, and M <N, For the m-th actual stress measurement point The coordinates; where the set of actual stress measurement points is a subset of the set of stress simulation measurement points.

[0019] Step S2 completes the multi-source experimental data acquisition and key area observation: For the typical structure and representative process parameter combinations selected in S1, a component entity is fabricated. Using experimental methods such as X-ray diffraction or neutron diffraction, observation points are systematically deployed in the key areas of the component to obtain experimental reference data on residual stress. The key areas should cover: ① regions with high stress gradient changes; ② regions with abrupt geometric changes (such as rounded corners and hole edges); ③ regions sensitive to overall stress balance.

[0020] S3, for and Perform linear fitting to obtain the fitting parameters. and The linear fitting formula is: ;Will Substitute into the formula The calibrated residual stress simulation data were calculated. Use it as the training dataset, The coordinates, the specified laser powder bed melting process conditions, and the geometric characteristic parameters of the component sample are used as inputs. To supervise the training of the physical information neural network; in, , These are the global scaling factors for the stress components in the x, y, and z directions, respectively. These are the compensation amounts for the stress components in the x, y, and z directions, respectively.

[0021] Step S3 includes: S31, Simulation-Experimental System Error Calibration A linear mapping model between simulation and experimental data is established to accurately correct systematic errors in finite element simulations using a combination of global and local methods. This step transforms high-cost, sparse experimental measurement data into a systematic correction of high-density simulation data, generating a high-quality dataset equivalent to the experimental domain.

[0022] The linear mapping model is represented as:

[0023] in: This is a global scaling factor for each stress component, used to correct the overall deviation of the simulation model.

[0024] S32, Multi-feature fusion and physical information neural network training Using S31-calibrated stress data as the supervision target, training samples for the neural network are constructed. Input features It is a multidimensional vector that integrates the coordinates of stress measurement points, process conditions, and component geometric field information (i.e., the geometric characteristic parameters of the component, including curvature, wall thickness, etc.).

[0025] A multi-branch physical information neural network structure is adopted, in which different branch sub-networks extract features from the heterogeneous input features, and then the features are fused in a high-level feature fusion module to finally output the predicted residual stress tensor. .

[0026] like Figure 4 As shown, the physical information neural network includes a spatial coordinate branch, a process parameter branch, a scanning strategy branch, a geometric feature parameter branch, and a feature fusion module; The spatial coordinate branch is used to input the coordinates of the stress measurement point; The process parameter branch, scanning strategy branch, and geometric feature parameter branch are all fully connected networks, which are used to extract features from process parameters, scanning strategies, and laser powder bed melting process conditions, respectively. The feature fusion module is used to fuse the features extracted from the process parameter branch, scanning strategy branch, and geometric feature parameter branch to obtain the predicted three-dimensional residual stress.

[0027] The network training employs the following composite loss function to simultaneously ensure data accuracy and physical consistency:

[0028] in, The mean square error between the predicted stress and the calibrated experimental stress; For the stress balance equation constraints applied based on automatic differentiation techniques, such as , For the Nabra operator, For residual stress tensor; Boundary condition constraints, including stress boundary constraints and displacement continuity constraints, are used to improve the predictability of boundary behavior; for example, the boundary stress constraint is: , Given surface forces, Here is the boundary normal vector; the displacement continuity constraint is: , The displacement is known.

[0029] All of these are adjustable hyperparameter weights, and it is recommended to use an adaptive adjustment strategy to optimize the training process.

[0030] Application phase: The coordinates of the target stress measurement point of the target component, the laser powder bed melting process conditions of the target component, and the geometric feature parameters of the target component are input into a trained physical information neural network to obtain the three-dimensional residual stress at the target stress measurement point.

[0031] Using a trained neural network model, the residual stress distribution of new target structures or process parameters can be predicted at the second or even millisecond level.

[0032] The following relative error indices can be used to evaluate prediction accuracy:

[0033] If the prediction error is below a preset threshold (e.g., 20%), the model is considered validated and can be directly used for rapid process evaluation, structural optimization, and failure risk assessment. If the prediction error exceeds the limit, an adaptive update mechanism is triggered: for typical structural features with large prediction errors, supplementary finite element simulations and experimental verifications are automatically initiated, and new data is added to the training set to fine-tune the original model until the prediction accuracy converges. This mechanism ensures the continuous evolution and robustness of the model during application.

[0034] Preferably, if the target component and the component sample have different shapes, the trained physical information neural network is used as a pre-trained network before the application stage is performed on the target component to retrain the trained physical information neural network.

[0035] Specifically, if the target component and the component sample have different shapes, the prediction accuracy of the physical information neural network can be further improved through transfer learning.

[0036] The method provided by this invention will be further illustrated below with two specific examples.

[0037] Example 1: Rapid prediction of residual stress in 316L stainless steel cubic specimens formed by laser powder bed fusion molding This example aims to illustrate in detail the basic process and verification procedure of the method of the present invention using a simple cubic sample.

[0038] (1) Finite element modeling and simulation measurement Using a 316L stainless steel cubic specimen (10 mm × 10 mm × 10 mm) as the model, the laser power was set to 200 W, the scanning speed to 800 mm / s, the powder layer thickness to 40 μm, and the scanning strategy to be a strip scan with 67° interlayer rotation. A thermo-mechanical coupling model was established in Abaqus finite element software. The material constitutive model considering strain rate effects was adopted, the latent heat of phase transformation was handled using the equivalent specific heat capacity method, and the "birth and death element" technique was used to simulate the layer-by-layer deposition of the material. After the simulation was completed, the three-dimensional residual stress tensor of the entire cubic domain was output. .

[0039] (2) Experimental measurement Under the same process conditions as described in step (1), the cubic specimen was formed. Using X-ray diffraction (XRD), nine measurement points were selected at 1 mm intervals along the scanning direction (X) and the vertical direction (Y) in the central region of the upper surface of the specimen to measure residual stress and obtain experimental stress data. Before measurement, the sample is electropolished to remove the surface stress layer.

[0040] (3) Simulation and experimental data calibration Extract the stress data obtained from the experiment in step (2). And compare it with the residual stress obtained from the simulation in step S1. Matching is performed using linear regression. For example, suppose the global scaling factor β = [1.12, 0.95, 1.08] is determined. =[5MPa, 6MPa, 17MPa]. This result indicates that the simulation model's prediction in the X-direction (scanning direction) is too low and needs to be corrected. The prediction result in the Y direction (scanning direction) is too large and needs to be corrected. However, the prediction in the Z direction (layer thickness direction) is too small and needs to be corrected. This calibration step effectively quantified and corrected the systematic biases of the simulation model.

[0041] (4) Training of physical information neural networks Based on the data calibrated in step (3), training samples are constructed, where the input features are... This includes spatial coordinates (x, y, z), process parameters (such as power, speed, layer thickness), geometric features of the component, and scanning strategies (such as scanning path, rotation angle). For example, a fully connected neural network with four hidden layers (128, 64, 32, and 16 neurons respectively) is used as the backbone of the Physical Information Neural Network (PINN), meaning each branch of the network is a fully connected neural network. In the loss function, in addition to the mean squared error of the data (MSE_data), a stress balance constraint is applied ( The weighting coefficient λ1 is set to 0.8 to prioritize the physical rationality of the prediction results.

[0042] (4) Model accuracy evaluation and results The neural network trained in step (4) improves computational efficiency by more than three orders of magnitude in predicting the stress field of the entire cube compared to the original finite element method. At nine verification points, the average relative error ε(x) between the predicted stress and experimental measurements was 11.5%, with a maximum error of 16.8%. The neural network model's single prediction time for the stress field of the entire cube is less than 0.1 seconds, representing an efficiency improvement of more than five orders of magnitude compared to traditional finite element simulation (approximately 8 hours). The maximum error meets engineering application requirements (<20%), validating the effectiveness and high efficiency of the basic process of this method.

[0043] Example 2: Rapid prediction of residual stress in TC4 titanium alloy triangular pyramidal structural components formed by laser powder bed fusion molding This example aims to illustrate the adaptability of the method of the present invention to structural components with complex geometric features and scanning paths, and to demonstrate its advantage in improving efficiency through transfer learning.

[0044] (1) Finite element modeling and simulation A TC4 titanium alloy triangular pyramid structure (15 mm height, 8 mm side width) was used as the model. The process parameters were: laser power 280 W, scanning speed 1000 mm / s, and layer thickness 30 μm. Due to its geometric complexity, a partitioned scanning strategy was adopted to control deformation. A thermo-mechanical coupling model was built in Simufact Additive software, with the mesh refined at edges and sharp corners to accurately capture the stress concentration effects in these areas. The simulation output showed the residual stress field of the structure. .

[0045] (2) Experimental measurement The same triangular pyramid specimen was prepared under the same process conditions as described in step (1). Using X-ray diffraction, 12 measurement points were set up, focusing on the edges and top regions with high stress concentration risk, to obtain the experimental stress. .

[0046] (3) Simulation and experimental data calibration Extract the stress data obtained from the experiment in step (2). And compare it with the residual stress obtained from the simulation in step (1). Perform a match.

[0047] (4) Network training based on transfer learning The neural network model trained on a 316L cube in Example 1 was used as a pre-trained model. The data from the TC4 triangular pyramid in this example (containing new geometric features) was used as a new training set to fine-tune the network again. This allows the model to quickly learn the effects of complex geometry and the new material (TC4) while maintaining its understanding of fundamental physical laws. Results show that, compared to training from scratch, transfer learning reduces the number of epochs required for training convergence by approximately 60%, significantly reducing data and computational costs.

[0048] (5) Model evaluation and adaptive iterative optimization The residual stress prediction neural network model trained in step (4) was used to predict the residual stress of the target complex structure. Preliminary predictions showed an error of 22% in the edge region, slightly exceeding the target of 20%. Based on this, an adaptive update mechanism was triggered: two sets of finite element simulation data of triangular pyramid structures with different sizes (scaling ratios of 0.8 and 1.2) were added to the training set. After a round of fine-tuning, the prediction error in the edge region decreased to 17.5%, and the overall model average error decreased to 13.1%, achieving the required accuracy. This demonstrates the rapid self-evolution capability of the method in this invention when facing new structures.

[0049] This invention provides an electronic device, including: a computer-readable storage medium and a processor; The computer-readable storage medium is used to store executable instructions; The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the method as described in any of the above embodiments.

[0050] This invention provides a computer-readable storage medium storing computer instructions that cause a processor to perform the method described in any of the above embodiments.

[0051] This invention provides a computer program product, including a computer program or instructions, which, when executed by a processor, implement the method described in any of the above embodiments.

[0052] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for rapid prediction of residual stress in laser powder bed fusion, integrating experimental data, finite element simulation, and neural networks, characterized in that, include: Training phase: S1. The three-dimensional residual stress of the component sample under multiple different laser powder bed melting process conditions is obtained through simulation software, and the simulation data of residual stress are obtained. ; n=1,2,…,N, where N is the total number of stress measurement points during simulation; For the nth stress simulation measurement point The coordinates; S2, prepare the solid component sample under the specified laser powder bed melting process conditions, measure the three-dimensional residual stress of the solid component sample, and obtain the actual measurement data of the residual stress. ; m = 1, 2, ..., M, where M is the number of stress measurement points during the actual measurement, and M <N, For the m-th actual stress measurement point The coordinates; where the set of actual stress measurement points is a subset of the set of stress simulation measurement points; S3, for and Perform linear fitting to obtain the fitting parameters. and ;Will Substitute into the formula The calibrated residual stress simulation data were calculated. Use it as the training dataset, The coordinates, the specified laser powder bed melting process conditions, and the geometric characteristic parameters of the component sample are used as inputs. To supervise the training of the physical information neural network; in, , These are the global scaling factors for the stress components in the x, y, and z directions, respectively. These are the compensation amounts for the stress components in the x, y, and z directions, respectively. Application phase: The coordinates of the target stress measurement point of the target component, the laser powder bed melting process conditions of the target component, and the geometric feature parameters of the target component are input into a trained physical information neural network to obtain the three-dimensional residual stress at the target stress measurement point.

2. The method as described in claim 1, characterized in that, The loss function of the physical information neural network is: in, All are weighting coefficients; The mean square error between the residual stress predicted by the network and the calibrated residual stress simulation data; Constrained by the stress balance equations, These are boundary condition constraints, including stress boundary constraints and displacement continuity constraints; For residual smoothing regularization term, This represents the difference between the network output and the monitoring target.

3. The method as described in claim 1 or 2, characterized in that, The laser powder bed melting process conditions include process parameters and scanning strategies; The physical information neural network includes a spatial coordinate branch, a process parameter branch, a scanning strategy branch, a geometric feature parameter branch, and a feature fusion module; The spatial coordinate branch is used to input the coordinates of the stress measurement point; The process parameter branch, scanning strategy branch, and geometric feature parameter branch are all fully connected networks, which are used to extract features from process parameters, scanning strategies, and laser powder bed melting process conditions, respectively. The feature fusion module is used to fuse the features extracted from the process parameter branch, scanning strategy branch, and geometric feature parameter branch to obtain the predicted three-dimensional residual stress.

4. The method as described in claim 3, characterized in that, The process parameters include laser power, scanning speed, and powder layer thickness; The scanning strategy includes the scanning path and rotation angle; The geometric feature parameters include curvature and wall thickness.

5. The method as described in claim 1, characterized in that, If the target component and the component sample have different shapes, the trained physical information neural network is used as a pre-trained network before the application stage is performed on the target component to retrain the trained physical information neural network.

6. An electronic device, characterized in that, include: Computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the method as described in any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a processor to perform the method as described in any one of claims 1-5.

8. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the method as described in any one of claims 1-5.

Images