Method for predicting residual stress of laser directional energy deposition thin-wall component driven by mechanism and data
By constructing a method combining the mechanism model based on thermal-elasticity theory and the LSTM-Attention algorithm, the efficiency and accuracy of residual stress prediction in laser directional energy deposition are solved, and a more efficient prediction effect is achieved.
Patent Information
- Application Number
- CN202510442416.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-10
- Publication Date
- 2025-07-25
AI Technical Summary
In the prior art, residual stress prediction during laser directional energy deposition has problems such as high computational cost, complex parameters and difficult model verification. The data-driven method cannot explain the internal operating principle of the algorithm and lacks generalization capabilities.
Using a method combining mechanism and data driving, a residual stress formation mechanism model based on thermal-elasticity theory was constructed, and modeled and experimented through ABAQUS software, combined with LSTM-Attention algorithm for residual stress prediction, and fused mechanism model and experimental data.
It significantly improves the efficiency and accuracy of residual stress prediction, solves the limitations of traditional methods, and provides a more efficient prediction solution.
Smart Images
Figure CN120372850A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for predicting residual stress of thin-walled components in laser directed energy deposition driven by mechanism and data, belonging to the technical fields of laser manufacturing and intelligent manufacturing. Background Technique
[0002] Laser directed energy deposition is a cutting-edge technology in the field of metal additive manufacturing, which can achieve the efficient manufacturing and repair of metal parts. However, excessive residual stress generated during the deposition process will cause failure problems such as product deformation, cracking and delamination, seriously affecting the quality and performance of products. Therefore, establishing a reliable residual stress prediction method has important engineering value for realizing the quality control and performance optimization of the laser directed energy deposition process.
[0003] Traditional mechanism-driven residual stress prediction methods have limitations such as high computational cost, complex parameters and difficult model verification; data-driven residual stress prediction methods have problems such as inability to explain the internal operation principle of the algorithm and insufficient generalization ability. Therefore, combining the advantages of mechanism and data-driven, and integrating the physical characteristics generated by the mechanism model with the experimental data of residual stress measurement, has important engineering significance for improving the efficiency and accuracy of laser directed energy deposition residual stress prediction. Summary of the Invention
[0004] In view of the above problems, the present invention designs a method for predicting residual stress of thin-walled components in laser directed energy deposition driven by mechanism and data. Taking the thin-walled components in laser directed energy deposition as the object, the aim is to accurately predict the residual stress formed during the deposition process through a prediction method that combines mechanism and data, improve the efficiency and accuracy of prediction, and provide support for optimizing the laser directed energy deposition process and improving product quality and performance.
[0005] The present invention relates to a method for predicting residual stress of thin-walled components in laser directed energy deposition driven by mechanism and data, which is characterized by including the following steps:
[0006] S1: Construct a residual stress formation mechanism model based on thermo-elastoplastic theory;
[0007] S2: Design and implement an in-situ test experiment for residual stress of thin-walled components in laser directed energy deposition;
[0008] S3: Propose a method for predicting residual stress of thin-walled components in laser directed energy deposition that combines mechanism and data.
[0009] Preferably, the step S1 includes the following sub-steps:
[0010] S11: Construct a model for the formation mechanism of residual stress in laser direct energy deposition based on thermo-elastoplastic theory: Analyze the heat conduction during the laser direct energy deposition process, take the temperature load of any node obtained as the body load to apply the temperature load increment, and finally obtain the stress and strain increment equations for each element as: dσ = [D]dε - [C]dT, where: dσ is the stress increment vector, [D] is the elastic matrix or elastoplastic matrix of the material, dε is the strain increment vector, [C] is the temperature-stress matrix, and dT is the temperature increment vector. In thermo-elastoplastic calculations, when the node temperature increment is known, the internal stress increment of the structure can be solved through the equilibrium equation, and finally the stress field distribution of the model can be obtained. The equilibrium equation is: {dF} e +{dR} e =[K] e {dδ} e , where: {dF} e represents the load increment on the node, {dR} e represents the force increment on the node caused by temperature, [K] e represents the element stiffness matrix, and {dδ} e represents the displacement increment on the element.
[0011] S12: Laser direct energy deposition residual stress modeling based on ABAQUS (a finite element software for engineering simulation): Establish a thin-walled component model with dimensions of 50mm × 50mm × 5mm and a layer height of 0.32mm. Use a double ellipsoidal heat source as the moving heat source model for simulation, and its formula is:
[0012]
[0013] where, q r and q f respectively represent the volume heat source densities of the front hemisphere and the rear hemisphere, f r and f f are the heat distribution functions of the front and rear ellipsoids, x, y, z are the local coordinate systems of the double ellipsoidal heat source model, Q represents the total input heat power, a, b, and c respectively represent the melt length, melt width, and melt depth, a r and a f respectively represent the front semi-axis length and the rear semi-axis length of the actual molten pool.
[0014] S13: Parametric modeling of laser direct energy deposition residual stress based on secondary development of ABAQUS: With the help of the GUI and RSG functions of ABAQUS, realize the parametric automatic modeling process.
[0015] S14: Analysis of the numerical simulation results of the temperature field and stress field during the laser directed energy deposition process: Four points with different spacings are randomly selected along the horizontal direction of the thin-walled component model. Based on the center line passing through the center point, four points with the same interval are selected on both sides of the upper line of the center point to analyze the temperature field and stress field during the deposition process.
[0016] S15: Comparative analysis of the calculation results of the residual stress formation mechanism model and the numerical simulation results: After numerical comparison and verification, the calculation results are highly consistent with the simulation results in the distribution trend of residual stress, and the error is within 20%.
[0017] Preferably, the step S2 includes the following sub-steps:
[0018] S21: Design and implement a single-pass orthogonal experiment for laser directed energy deposition: Select to deposit In718 nickel-based superalloy powder on a 45# steel experimental substrate. Based on the orthogonal experiment results, the optimal process parameter combination is: laser power 1100W, powder feeding rate 1.4g / min, and scanning speed 400mm / s.
[0019] S22: Design and implement an experiment on laser directed energy deposition of thin-walled components: Based on the optimal process parameters obtained from the single-pass orthogonal experiment, 40 layers of deposition are carried out in a serpentine scanning manner.
[0020] S23: Design and implement in-situ testing of residual stress in laser directed energy deposition thin-walled components: Four detection points are selected, and the residual stress of the specimen is detected by X-ray diffraction method.
[0021] S24: Comparison of the calculation, numerical simulation and experimental test results of the residual stress formation mechanism model of laser directed energy deposition thin-walled components: Based on the comparison results, the numerical simulation results are consistent with the detection results by the X-ray diffraction method in the overall residual stress distribution, and can more accurately reflect the residual stress state, and the numerical simulation accuracy reaches more than 80%.
[0022] Preferably, the step S3 includes the following sub-steps:
[0023] S31: Construct a dataset of residual stress in laser directed energy deposition thin-walled components: 5004 surface nodes are selected from the input dataset, and about 33000 temperature data of all time analysis steps are obtained for a single node temperature field. Four types of stress data are extracted from the output dataset, with each type of stress corresponding to 5004 nodes, totaling about 20000 stress results. The input and output are divided into an 80% training set and a 20% test set.
[0024] S32: Construct a residual stress prediction model for laser directed energy deposition thin-walled components that integrates mechanism and data: Three data-driven algorithms, namely alternative artificial neural network (ANN), long short-term memory network (LSTM), and extreme gradient boosting (XGBoost), are selected for residual stress prediction.
[0025] S33: Comparison of residual stress prediction results: Based on the comparison results, ANN performs mediocrely for multiple outputs, is prone to overfitting problems due to interference between outputs, and has a strong dependence on network depth. The synergistic relationship between the input and output of XGBoost is easily overlooked, is greatly affected by hyperparameter settings, and has a weak ability to capture global dependencies. LSTM has obvious performance advantages in capturing the dependency relationships of data in multi-output tasks, and its learning tends to be stable, making it suitable for subsequent optimization and expansion. Considering comprehensively, LSTM is selected as the final model.
[0026] S34: Introduce the self-attention mechanism to optimize the LSTM algorithm: To solve the problem of the decreasing processing ability of the LSTM algorithm for long-term dependencies, the present invention introduces the self-attention mechanism. The calculation formula of the self-attention mechanism is:
[0027]
[0028] S ki = V·tanh(Wh k + Uh i + b)
[0029]
[0030] In the formula, α ki represents the normalized attention weight, S ki represents the attention score, T x represents the length of the input sequence, V represents the learnable weight vector, tanh represents the activation function, W and U represent the learnable weight matrices, h k and h i respectively represent the hidden states at the k-th position and the i-th position in the input sequence, b represents the bias term, and C represents the weighted sum of the hidden states at all positions.
[0031] S35: Analysis of residual stress prediction results: The improved LSTM-Attention (LSTM-attention) algorithm uses the fitting explanatory index R 2 value to evaluate the accuracy of the model's fitting explanation. The results show that the improvements in the fitting accuracy of each stress are 3.1%, 7.7%, 13.9%, and 1% respectively, and the comprehensive performance improvement reaches 6.4%.
[0032] Compared with the prior art, the beneficial effects of the present invention are as follows: By integrating mechanism and data, the present invention breaks through the limitations of traditional single prediction methods, effectively avoiding various problems existing in pure mechanism-driven and pure data-driven methods, significantly improving the efficiency and accuracy of residual stress prediction, and providing important theoretical value and practical contributions to the development and application of laser directed energy deposition technology. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 It is a numerical simulation model diagram of a thin-walled component based on ABAQUS of the present invention.
[0034] Figure 2 It is a schematic diagram of a double-ellipsoid heat source of the present invention.
[0035] Figure 3 It is a demonstration diagram of the parametric modeling result based on ABAQUS of the present invention.
[0036] Figure 4 It is a schematic diagram of the selected points for numerical simulation based on ABAQUS of the present invention.
[0037] Figure 5 It is a comparison and verification diagram of the calculation result and simulation result of the mechanism model of the present invention.
[0038] Figure 6 It is a physical diagram of the laser directed energy deposition equipment of the present invention.
[0039] Figure 7 It is an experimental result diagram of a single pass of laser directed energy deposition of the present invention.
[0040] Figure 8 It is a result diagram of a thin-walled component of laser directed energy deposition of the present invention.
[0041] Figure 9 It is a schematic diagram of the selection of residual stress detection points for the thin-walled component of the present invention.
[0042] Figure 10 It is a comparison and verification diagram of the calculation result, test result and simulation result of the present invention.
[0043] Figure 11 It is a schematic diagram of the internal structure of the LSTM-Attention algorithm of the present invention.
[0044] Figure 12(1)-12(4) It is a prediction result diagram of the LSTM-Attention algorithm of the present invention. DETAILED DESCRIPTION OF THE INVENTION
[0045] The present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the embodiments are used to explain the present invention and are not used to limit the present invention.
[0046] The present invention discloses a method for predicting residual stress of thin-walled components by laser directed energy deposition based on mechanism and data. Figure 1 It is a numerical simulation model diagram of thin-walled components based on ABAQUS of the present invention. Figure 2 It is a schematic diagram of a double-ellipsoid heat source of the present invention. Figure 3 It is a demonstration diagram of the parametric modeling result based on ABAQUS of the present invention.
[0047] Figure 4 It is a schematic diagram of point selection for numerical simulation based on ABAQUS of the present invention. Figure 5 It is a comparison and verification diagram of the calculation result and simulation result of the mechanism model of the present invention. Figure 6 It is a physical diagram of the laser directed energy deposition equipment of the present invention. Figure 7 It is a diagram of the experimental results of single-pass laser directed energy deposition of the present invention. Figure 8 It is a diagram of the results of laser directed energy deposition of thin-walled components of the present invention. Figure 9 It is a schematic diagram of the selection of residual stress detection points for thin-walled components of the present invention. Figure 10 It is a comparison and verification diagram of the calculation result, test result and simulation result of the present invention. Figure 11 It is a schematic diagram of the internal structure of the LSTM-Attention algorithm of the present invention. Figure 12(1)-12(4) It is a prediction result diagram of the LSTM-Attention algorithm of the present invention.
[0048] The overall technical solution of the present invention is a method for predicting residual stress of thin-walled components by laser directed energy deposition based on mechanism and data, including the following steps:
[0049] S1: Construct a residual stress formation mechanism model based on thermo-elastoplastic theory;
[0050] S2: Design and implement an in-situ test experiment on residual stress of thin-walled components by laser directed energy deposition;
[0051] S3: Propose a method for predicting residual stress of thin-walled components by laser directed energy deposition by integrating mechanism and data.
[0052] The step S1 includes the following sub-steps:
[0053] S11: Construct a model for the formation mechanism of residual stress in laser directed energy deposition based on thermo-elastoplastic theory: Analyze the heat conduction during the laser directed energy deposition process. Take the temperature load of any node obtained as the body load and apply the temperature load increment. Finally, the stress and strain increment equations for each element are obtained as: dσ = [D]dε - [C]dT, where: dσ is the stress increment vector, [D] is the elastic matrix or elastoplastic matrix of the material, dε is the strain increment vector, [C] is the temperature-stress matrix, and dT is the temperature increment vector. In thermo-elastoplastic calculations, when the node temperature increment is known, the internal stress increment of the structure can be solved through the equilibrium equation, and finally the stress field distribution of the model can be obtained. The equilibrium equation is: {dF} e +{dR} e =[K] e {dδ} e ,where: {dF} e represents the load increment on the node, {dR} e represents the force increment on the node caused by temperature, [K] e represents the element stiffness matrix, and {dδ} e represents the displacement increment on the element.
[0054] S12: Modeling of residual stress in laser directed energy deposition based on ABAQUS (a finite element software for engineering simulation): Establish a thin-walled component model with dimensions of 50mm × 50mm × 5mm and a layer height of 0.32mm. Use a double-ellipsoid heat source as the moving heat source model for simulation, and its formula is:
[0055]
[0056] where, q r and q f respectively represent the volume heat source densities of the front hemisphere and the rear hemisphere, f r and f f are the heat distribution functions of the front and rear ellipsoids, x, y, z are the local coordinate systems of the double-ellipsoid heat source model, Q represents the total input heat power, a, b, and c respectively represent the melt length, melt width, and melt depth, and a r and a f respectively represent the front semi-axis length and the rear semi-axis length of the actual molten pool.
[0057] S13: Parametric modeling of residual stress in laser directed energy deposition based on secondary development of ABAQUS: With the help of the GUI and RSG functions of ABAQUS, realize the parametric automatic modeling process.
[0058] S14: Analysis of the numerical simulation results of the temperature field and stress field during the laser directed energy deposition process: Four points with different spacings are randomly selected along the horizontal direction of the thin-walled component model. Based on the center line passing through the center point, four points with the same interval are selected on both sides of the upper line of the center point to analyze the temperature field and stress field during the deposition process.
[0059] S15: Comparative analysis of the calculation results of the residual stress formation mechanism model and the numerical simulation results: After numerical comparison and verification, the calculation results are highly consistent with the simulation results in the distribution trend of residual stress, and the error is within 20%.
[0060] The step S2 includes the following sub-steps:
[0061] S21: Design and implement a single-pass orthogonal experiment for laser directed energy deposition: Select to deposit In718 nickel-based superalloy powder on a 45# steel experimental substrate. The orthogonal experimental data is shown in Table 1. Based on the orthogonal experimental results, the optimal process parameter combination is: laser power 1100W, powder feeding rate 1.4g / min, and scanning speed 400mm / s.
[0062] Table 1 Orthogonal experimental data
[0063]
[0064] S22: Design and implement an experiment on laser directed energy deposition of thin-walled components: Based on the optimal process parameters obtained from the single-pass orthogonal experiment, 40 layers of deposition are carried out in a serpentine scanning manner.
[0065] S23: Design and implement in-situ testing of residual stress in laser directed energy deposition thin-walled components: Four detection points are selected, and the residual stress of the specimen is detected by X-ray diffraction method.
[0066] S24: Comparison of the calculation, numerical simulation and experimental test results of the residual stress formation mechanism model of laser directed energy deposition thin-walled components: Based on the comparison results, the numerical simulation results are consistent with the detection results by the X-ray diffraction method in the overall residual stress distribution, and can accurately reflect the residual stress state. The numerical simulation accuracy reaches more than 80%.
[0067] The step S3 includes the following sub-steps:
[0068] S31: Construct a dataset of residual stress in laser directed energy deposition thin-walled components: 5004 surface nodes are selected from the input dataset, and about 33000 temperature data of all time analysis steps are obtained for a single node temperature field. Four types of stress data are extracted from the output dataset, and each type of stress corresponds to 5004 nodes, with a total of about 20000 stress results. The input and output are divided into an 80% training set and a 20% test set.
[0069] S32: Build a residual stress prediction model for laser directed energy deposition thin-walled components that integrates mechanism and data: Three data-driven algorithms, namely artificial neural network (ANN), long short-term memory network (LSTM), and extreme gradient boosting (XGBoost), are selected for residual stress prediction.
[0070] S33: Comparison of residual stress prediction results: The comparison results of each stress evaluation index predicted by the basic algorithms are shown in Table 2. Based on the comparison results, it can be seen that ANN performs mediocrely for multiple outputs, is prone to overfitting problems due to interference between outputs, and has a strong dependence on network depth. The synergy relationship between the input and output of XGBoost is easily overlooked, is greatly affected by the setting of hyperparameters, and has a weak ability to capture global dependencies. LSTM has obvious performance advantages in capturing the dependence relationship of data in multi-output tasks, and at the same time, the learning tends to be stable, making it suitable for subsequent optimization and expansion. Considering comprehensively, LSTM is selected as the final model.
[0071] Table 2 Comparison results of each stress evaluation index predicted by the basic algorithms
[0072]
[0073] S34: Introduce the self-attention mechanism to optimize the LSTM algorithm: To solve the problem of the decline in the processing ability of the LSTM algorithm for long-term dependencies, the self-attention mechanism is introduced in the present invention. The calculation formula of the self-attention mechanism is:
[0074]
[0075] S ki =V·tanh(Wh k +Uh i +b)
[0076]
[0077] In the formula, α ki represents the normalized attention weight, S ki represents the attention score, T x represents the length of the input sequence, V represents the learnable weight vector, tanh represents the activation function, W and U represent the learnable weight matrices, h k and h i represent the hidden states at the k-th position and the i-th position in the input sequence respectively, b represents the bias term, and C represents the weighted sum of the hidden states of all positions.
[0078] S35: Analysis of residual stress prediction results: The stress evaluation results predicted by the improved LSTM-Attention (LSTM-attention) algorithm are shown in Table 3. The fitting interpretability index R 2Values are used to evaluate the accuracy of the fitting interpretation of the model. The results show that the improvements in the fitting accuracy of each stress are 3.1%, 7.7%, 13.9% and 1% respectively, and the comprehensive performance improvement reaches 6.4%.
[0079] Table 3 Results of the optimized model predicting each stress evaluation index
[0080]
[0081] The foregoing are the preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art of the present invention can still modify the above technical solutions or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.
Claims
1. A mechanism and data-driven prediction method for residual stress of thin-walled components by laser directed energy deposition, characterized in that The steps include: S1: Construct a residual stress formation mechanism model based on thermo-elastic-plastic theory; S2: Design and implement in-situ residual stress test experiments of thin-walled components deposited by laser directed energy deposition; S3: A residual stress prediction method for thin-walled components using laser directed energy deposition is proposed that integrates mechanism and data.
2. The residual stress prediction method for thin-walled components by laser direct energy deposition based on mechanism and data driving according to claim 1, characterized in that Step S1 includes the following sub-steps: S11: Construct a model for the formation mechanism of residual stress in laser directed energy deposition based on thermo-elastoplastic theory: Analyze the heat conduction during the laser directed energy deposition process, and use the temperature load of any node obtained as a body load to apply the temperature load increment. Finally, the stress and strain increment equations for each element are obtained as: dσ = [D]dε - [C]dT, where: dσ is the stress increment vector, [D] is the elastic matrix or elastoplastic matrix of the material, dε is the strain increment vector, [C] is the temperature-stress matrix, dT is the temperature increment vector. In thermo-elastoplastic calculations, when the node temperature increment is known, the internal stress increment of the structure can be solved through the equilibrium equation, and finally the stress field distribution of the model is obtained. The equilibrium equation is: {dF} e +{dR} e =[K] e {dδ} e , where: {dF} e represents the load increment on the node, {dR} e represents the force increment on the node caused by temperature, [K] e represents the element stiffness matrix, {dδ} e represents the displacement increment on the element; S12: Modeling of residual stress of laser directed energy deposition based on ABAQUS: A thin-walled component model with a size of 50mm×50mm×5mm and a layer height of 0.32mm was established. A double ellipsoid heat source was used as the mobile heat source model for simulation. The formula is: where q r and q f represent the volumetric heat source density of the front hemisphere and the rear hemisphere respectively, f r and f f are the heat distribution functions of the front and rear ellipsoids, x, y, z are the local coordinate system of the double ellipsoid heat source model, Q represents the total input heat power, a, b, and c represent the melting length, melting width, and melting depth respectively, a r and a f represent the front semi-axis length and the rear semi-axis length of the actual molten pool respectively; S13: Parametric modeling of residual stress in laser directed energy deposition based on secondary development of ABAQUS: With the help of ABAQUS GUI and RSG functions, the parametric automatic modeling process is realized; S14: Analysis of numerical simulation results of temperature field and stress field during laser directed energy deposition: Four points with different spacings were randomly selected along the horizontal direction of the thin-walled component model, and four points with the same spacings were selected on both sides of the upper line of the center point based on the midline passing through the center point to analyze the temperature field and stress field during the deposition process; S15: Comparative analysis of the calculation results of the residual stress formation mechanism model and the numerical simulation results: After numerical comparison and verification, the calculation results and the simulation results are highly consistent in the residual stress distribution trend, and the errors are within 20%.
3. The mechanism and data-driven method for predicting residual stress of thin-walled components by laser directed energy deposition according to claim 1 or 2, characterized in that, Step S2 includes the following sub-steps: S21: Design and implement a single-pass orthogonal experiment of laser directed energy deposition: In718 nickel-based high-temperature alloy powder was deposited on a 45 steel experimental substrate. Based on the results of the orthogonal experiment, the optimal process parameter combination is: laser power 1100W, powder feeding rate 1.4g / min, scanning speed 400mm / s; S22: Design and implement laser directed energy deposition thin-walled component experiments: Based on the optimal process parameters obtained from single-pass orthogonal experiments, 40 layers of deposition were performed using a serpentine scanning method; S23: Design and implement in-situ residual stress testing of thin-walled components by laser directed energy deposition: select 4 testing points and perform residual stress testing on the samples by X-ray diffraction method; S24: Comparison of residual stress formation mechanism model calculation, numerical simulation and experimental test results of laser directed energy deposition thin-walled components: Based on the comparison results, it can be seen that the numerical simulation results are consistent with the X-ray diffraction method detection results in the overall residual stress distribution, which can more accurately reflect the residual stress state, and the numerical simulation accuracy reaches more than 80%.
4. The mechanism and data-driven method for predicting residual stress of thin-walled components by laser directed energy deposition according to claim 3, characterized in that, Step S3 includes the following sub-steps: S31: Construct residual stress data set of laser directed energy deposition thin-walled components: 5004 surface nodes are selected from the input data set, and the single-node temperature field obtains about 33,000 temperature data of all time analysis steps; four types of stress data are extracted from the output data set, each stress corresponds to 5004 nodes, a total of about 20,000 stress results, and the input and output are divided into 80% training set and 20% test set; S32: Construct a residual stress prediction model for laser directed energy deposition thin-walled components that integrates mechanism and data: Select three data-driven algorithms, namely artificial neural network (ANN), long short-term memory network (LSTM), and extreme gradient boosting (XGBoost), for residual stress prediction; S33: Comparison of residual stress prediction results: Based on the comparison results, it can be seen that ANN performs mediocrely for multiple outputs, is prone to overfitting problems due to interference between outputs, has a strong dependence on network depth, the synergistic relationship between the input and output of XGBoost is easily ignored, it is greatly affected by hyperparameter settings, and its ability to capture global dependencies is weak. LSTM has obvious performance advantages in capturing the dependence relationship of data in multi-output tasks, and at the same time, the learning tends to be stable, making it suitable for subsequent optimization and expansion. Considering comprehensively, LSTM is selected as the final model; S34: Introduce self-attention mechanism to optimize the LSTM algorithm: To solve the problem of the decline in the processing ability of the LSTM algorithm for long-term dependencies, the present invention introduces a self-attention mechanism, and the calculation formula of the self-attention mechanism is: S ki = V·tanh(Wh k + Uh i + b) Where, α ki represents the normalized attention weight, S ki represents the attention score, T x represents the length of the input sequence, V represents the learnable weight vector, tanh represents the activation function, W and U represent the learnable weight matrices, h k and h i represent the hidden states at the k-th and i-th positions in the input sequence respectively, b represents the bias term, and C represents the weighted sum of the hidden states at all positions; S35: Analysis of Residual Stress Prediction Results: For the improved LSTM-Attention algorithm, the fitting explanatory index R 2 value is used to evaluate the accuracy of the fitting explanation of the model. The results show that the improvements in the fitting accuracy of each stress are 3.1%, 7.7%, 13.9% and 1% respectively, and the comprehensive performance improvement reaches 6.4%.