A shot peening residual stress prediction method based on a DEM-FEM-LSTM-attention three-dimensional process model

Abstract

The application discloses a shot peening residual stress prediction method based on a DEM-FEM-LSTM-Attention three-dimensional process model, relates to the technical field of metal material surface strengthening, and comprises the following steps: a DEM-FEM coupling shot peening three-dimensional simulation model is constructed and verified; after original data set of shot peening residual stress is generated by using the verified three-dimensional simulation model and normalized processing is performed, the effective sample set is divided into a training set, a test set and a verification set; the training set is input into the constructed LSTM-Attention prediction model for training, and the verification set is used for hyperparameter optimization, so that an optimal prediction model is obtained; the test set is used for verifying the optimal prediction model, and after verification, the shot peening process parameters and target residual stress layer depth of a to-be-tested working condition are input into the optimal prediction model, and corresponding shot peening residual stress prediction values are output, so that efficient and accurate prediction of residual stress is realized.

Description

Technical Field

[0001] This invention relates to the field of surface strengthening technology for metallic materials, and specifically to a method for predicting residual stress in shot peening based on a three-dimensional process model of DEM-FEM-LSTM-Attention. Background Technology

[0002] Shot peening is a key process that improves the fatigue properties of materials by creating residual compressive stress on the surface of a workpiece through high-speed impact of a large number of shot particles. It is widely used in the manufacture of aerospace components such as TC4 titanium alloy. The magnitude and distribution of residual stress from shot peening directly determine the strengthening effect; therefore, accurate prediction of residual stress is crucial for process optimization.

[0003] Existing methods for predicting residual stress mainly fall into two categories: experimental testing and numerical simulation. Experimental testing often involves point-by-point measurement through corrosion peeling, which is cumbersome, time-consuming, and labor-intensive, making it difficult to meet the rapid prediction requirements of multiple process parameter combinations. In terms of numerical simulation, although the discrete element-finite element (DEM-FEM) coupled method can accurately reproduce the shot peening process, it requires multiple explicit dynamic calculations, resulting in high computational costs and low efficiency, and cannot meet the real-time prediction needs of engineering applications.

[0004] In recent years, machine learning models, such as artificial neural networks (ANN), support vector machines (SVM), and long short-term memory networks (LSTM), have been increasingly applied to residual stress prediction. However, traditional machine learning models have significant limitations. ANN and SVM struggle to capture the long-term dependencies between process parameters and layer depth gradients, resulting in limited prediction accuracy. A single LSTM model fails to consider the weight differences in the influence of different features on residual stress, making it unable to effectively handle complex relationships involving multiple coupled factors. Existing research has not yet introduced attention mechanisms to address these issues, leading to insufficient model generalization ability and prediction accuracy, making it difficult to meet the high-precision requirements for digital prediction of shot peening surface integrity. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, this application proposes the following technical solution: This application provides a method for predicting residual stress in shot peening based on a three-dimensional process model using DEM-FEM-LSTM-Attention, including: A DEM-FEM coupled shot peening three-dimensional simulation model was constructed, and the effectiveness of the three-dimensional simulation model was verified based on the simulation results and experimental test results to obtain a verified three-dimensional simulation model. The original dataset of residual stress from shot peening was generated using the verified 3D simulation model. The original dataset of residual stress from shot peening is normalized to obtain an effective sample set, which is then divided into a training set, a test set, and a validation set. The training set is input into a pre-built LSTM-Attention prediction model for training, and the hyperparameters are optimized using the validation set to obtain the optimal LSTM-Attention prediction model. The optimal LSTM-Attention prediction model is validated using the test set. After successful validation, the shot peening process parameters and target residual stress layer depth of the test condition are input into the optimal LSTM-Attention prediction model. After forward computation of the model, the corresponding shot peening residual stress prediction value is output.

[0006] In one possible implementation, a DEM-FEM coupled shot peening three-dimensional simulation model is constructed, and the effectiveness of the three-dimensional simulation model is verified based on simulation results and experimental test results to obtain a verified three-dimensional simulation model, including: Using TC4 titanium alloy as the workpiece target material and cast steel as the pellet material, a three-dimensional geometric model including a pellet model and a workpiece target material model was constructed using ABAQUS software. The Johnson-Cook constitutive model is applied to the workpiece target to describe its dynamic mechanical properties. The nozzle is meshed using SFM3D4R elements, and the workpiece target is meshed using C3D8R elements. The mesh size of the area where the target surface contacts the pellet is set to one-tenth of the pellet diameter. The continuous shot blasting process is simulated using the Particle generator module, and the initial velocity of the shot is calculated based on the shot blasting pressure and shot flow rate. Using the initial velocity of the pellets as a boundary condition, DEM-FEM coupled simulation calculations were performed on the three-dimensional simulation model to obtain the simulated value of residual stress. TC4 titanium alloy specimens were prepared, and physical tests were conducted using a pneumatic shot peening machine under preset shot peening coverage conditions. The residual stress of the specimens was measured using an X-ray residual stress detection device to obtain the test results. The simulated residual stress value is compared and verified with the experimental residual stress value obtained from physical experiments. When the error between the two meets the preset threshold, the current DEM-FEM coupled shot peening three-dimensional simulation model is confirmed to be effective, and a verified three-dimensional simulation model is obtained.

[0007] In one possible implementation, the continuous shot blasting process is simulated using a Particle generator module. The formula for calculating the initial velocity of the shot based on the shot blasting pressure and shot flow rate is as follows: in, V The initial velocity of the pellet. For pellet flow rate, Where is the shot peening pressure, and D is the shot diameter.

[0008] In one possible implementation, the original dataset of shot peening residual stress is generated using the verified three-dimensional simulation model, including: Shot diameter, shot peening velocity and shot peening coverage were selected as experimental factors. Multiple levels were set for each factor, and multiple sets of orthogonal simulation experiments were designed. The process parameters of each group of orthogonal simulation experiments are input into the verified three-dimensional simulation model for coupled simulation calculation; After each simulation is completed, multiple measuring points are set up at preset intervals along the depth direction of the workpiece target surface. The residual stress value in the preset direction at each measuring point is extracted, and the residual stress layer depth of the corresponding measuring point is recorded. This results in the construction of an original dataset containing shot diameter, shot peening speed, shot peening coverage, residual stress layer depth and corresponding residual stress value.

[0009] In one possible implementation, the original dataset of shot peening residual stress is normalized to obtain an effective sample set, and the effective sample set is divided into a training set, a test set, and a validation set, including: The original dataset of residual stress from shot peening was normalized using the Min-Max normalization method to obtain an effective sample set. The calculation formula is as follows: in, This is the normalized residual stress output value. These are the original feature values ​​for a single sample. It is the minimum value within the entire set of feature data. It represents the maximum value within the entire set of feature data; The effective sample set is then divided into a training set, a test set, and a validation set according to a preset ratio.

[0010] In one possible implementation, the LSTM-Attention prediction model includes an input layer, the input of which is used to receive normalized features of shot diameter, shot peening velocity, shot peening coverage, and residual stress layer depth. The output of the input layer is connected to the input of an LSTM layer, the output of the LSTM layer is connected to the input of an attention mechanism layer, the output of the attention mechanism layer is connected to the input of a fully connected layer, and the output of the fully connected layer is connected to an output layer.

[0011] In one possible implementation, the loss function used when inputting the training set into a pre-built LSTM-Attention prediction model for training is: in, Let n be the loss function, and n be the total number of samples. k For sample index, For the first k The true value of each sample For the first k The predicted value for each sample.

[0012] In one possible implementation, the optimal LSTM-Attention prediction model is validated using the test set. After successful validation, the shot peening process parameters and target residual stress layer depth of the working condition to be predicted are input into the optimal LSTM-Attention prediction model. After forward computation, the corresponding predicted shot peening residual stress value is output, including: Each sample in the test set is input into the optimal LSTM-Attention prediction model to obtain the predicted residual stress value corresponding to each sample. The predicted residual stress value of each sample is compared with the corresponding real residual stress value. The model performance evaluation index is calculated. When each evaluation index meets the preset requirements, the optimal LSTM-Attention prediction model is confirmed to have passed the verification. Obtain the shot diameter, shot peening velocity, shot peening coverage, and residual stress layer depth under the predicted working condition; The shot diameter, shot peening velocity, shot peening coverage and residual stress layer depth of the working condition to be predicted are normalized, and the preprocessed feature data are converted into a three-dimensional temporal vector adapted to the temporal network. The three-dimensional temporal vector is input into the validated optimal LSTM-Attention prediction model, and the standardized temporal feature sequence is output through the input layer; The standardized time-series feature sequence is input into the LSTM layer. Through the collaborative work of the internal forget gate, input gate and output gate, the coupling and correlation information of multiple process parameters in the time-series feature sequence is filtered, and multiple sets of hidden state feature vectors corresponding to different time-series nodes are output. Then, all the hidden state feature vectors output by the LSTM layer are input into the attention mechanism layer. The weight coefficients corresponding to each set of hidden state feature vectors are calculated, and differentiated weights are assigned according to the contribution of each process feature to the prediction target. The weight coefficients are summed with the corresponding hidden state feature vectors to output a weighted feature vector that integrates global key features. The weighted feature vector is input into the fully connected layer, and the dimensionality of the weighted feature vector is compressed and the feature is fused through the fully connected network. The normalized residual stress prediction is output by the single-output regression neuron. Finally, Min-Max inverse normalization is used to restore the predicted residual stress at the normalized scale to obtain the predicted residual stress at the real physical scale.

[0013] Compared with the prior art, the beneficial effects of this application are as follows: This application obtains high-quality datasets through the DEM-FEM coupling model and combines the feature capture advantages of the LSTM-Attention model to achieve efficient and accurate prediction of residual stress.

[0014] This application obtains a high-precision dataset through a DEM-FEM coupled model, and optimizes feature weight allocation by combining an attention mechanism, thus improving the model's coefficient of determination. R 2 It exceeds 0.98, which is 4%, 20%, and 32% higher than BP, SVR, and single LSTM models, respectively. The error indicators such as MAE, RMSE, and MAPE are reduced by 47%-84% compared with ANN, SVM, and single LSTM models, and the overall performance is improved by nearly 60%.

[0015] This application avoids the cumbersome explicit dynamic calculations and experimental testing required by traditional numerical simulations. The trained model can output prediction results in real time, significantly reducing time costs and improving prediction efficiency. Simultaneously, the orthogonal experimental design covers multiple combinations of process parameters, and the LSTM-Attention model effectively captures the coupling relationships of multiple factors, making it suitable for residual stress prediction under different shot peening process conditions with strong generalization ability. It provides a digital tool for optimizing shot peening strengthening processes and can be directly applied to the production and manufacturing of titanium alloy components in aerospace and other fields, improving product fatigue performance and reliability. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a shot peening residual stress prediction method based on a DEM-FEM-LSTM-Attention three-dimensional process model provided in this application embodiment; Figure 2 This is a schematic diagram of the LSTM-Attention residual stress time-series prediction model structure provided in the embodiments of this application; Figure 3 A comparison chart of predicted and actual residual stress values ​​obtained using the BP model in the embodiments of this application is provided. Figure 4 A comparison chart of predicted and actual values ​​of shot peening residual stress obtained using the SVR model is provided for embodiments of this application. Figure 5 A comparison chart of predicted and actual shot peening residual stress values ​​obtained using a single LSTM model, provided for embodiments of this application; Figure 6A comparison chart of the predicted and actual values ​​of shot peening residual stress obtained by the LSTM-Attention model in an embodiment of this application. Detailed Implementation

[0017] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0018] Figure 1 A flowchart illustrating a shot peening residual stress prediction method based on a DEM-FEM-LSTM-Attention three-dimensional process model, provided in this application embodiment, is shown below. Figure 1 This embodiment presents a method for predicting residual stress in shot peening based on a three-dimensional process model using DEM-FEM-LSTM-Attention, comprising: S101. Construct a DEM-FEM coupled shot peening three-dimensional simulation model. Verify the effectiveness of the three-dimensional simulation model based on the simulation results and experimental test results to obtain a verified three-dimensional simulation model.

[0019] In this embodiment, TC4 titanium alloy is used as the workpiece target material and cast steel is used as the pellet material. A three-dimensional geometric model including the pellet model and the workpiece target material model is constructed using ABAQUS software.

[0020] The pellets adopt three spherical structures: S70, S110, and S170, with corresponding diameters of 0.18mm, 0.28mm, and 0.38mm, respectively. By editing keywords in the generated INP file, discrete element method (DEM) definitions are performed on the pellets, and particle properties, particle generator, particle velocity, and flow rate are set. After editing keywords, a new INP file is submitted in ABAQUS for calculation. The target model size is designed as 4D×4D×5D, where D is the pellet diameter.

[0021] The dynamic mechanical properties of the workpiece target are described using the Johnson-Cook constitutive model, with parameters including yield stress A = 1098 MPa, power factor B = 1094 MPa, hardening factor n = 0.93, strain rate sensitivity coefficient C = 0.014, and temperature sensitivity coefficient m = 1.1. The basic mechanical parameter is density. The elastic modulus is 114 GPa and Poisson's ratio is 0.342; the shot is made of standard cast steel shot, defined as a rigid body, and its basic mechanical parameter is density. The elastic modulus is 210 GPa, and Poisson's ratio is 0.3. Regarding contact properties, the Coulomb friction model and penalty function method are used between the pellets and the workpiece target, with a friction coefficient of 0.3 and a contact stiffness scaling factor of 0.1. The Hertzian contact model is used between the pellets, based on the equivalent contact radius. R Equivalent Young's modulus E * overlap with normal δEstablish normal contact force F The quantitative relationship between the contact stiffness and the penetration displacement is considered, while also taking into account the normal contact stiffness. K n Tangential contact stiffness and damping were determined by substituting the corresponding geometric and material parameters into the two sets of contact parameters, namely pellet-to-pellet and pellet-to-workpiece target, to simulate impact and interaction behavior.

[0022] in, This refers to the Hertzian normal contact force between two contacting particles. For equivalent Young's modulus ，R For the equivalent contact radius, This represents the normal overlap when the two particles are in contact. , and , Let the effective Young's modulus and Poisson's ratio be the values ​​of the two contacting particles. and These are the radii of the two contacting particles, respectively. This represents the normal contact stiffness.

[0023] The nozzle was meshed using SFM3D4R elements, and the workpiece target was meshed using C3D8R elements. The mesh size of the area where the target surface contacts the shot was set to one-tenth of the shot diameter to ensure simulation accuracy. The continuous shot blasting process was simulated using the Particle generator module. The initial velocity of the shot was calculated based on the shot blasting pressure and shot flow rate, using the following formula: in, V The initial velocity of the pellet. For pellet flow rate, Where is the shot peening pressure, and D is the shot diameter.

[0024] Using the initial shot velocity as a boundary condition, a DEM-FEM coupled simulation was performed on the 3D simulation model to obtain the simulated residual stress values. For the boundary conditions, full constraints were applied to the bottom surface of the workpiece target. The initial shot velocities covered three levels: 60 m / s, 80 m / s, and 100 m / s. The shot peening power source was clean compressed air with a pressure of 0.6 MPa ± 15%. The temperature boundary was set to room temperature (25℃). The analysis step was set to Dynamic Explicit type with a time increment step of 1e-7 s. The field output function was enabled, outputting key parameters such as residual stress in the S11 direction, PEEQ equivalent plastic strain, and Z-direction displacement every 10 increment steps.

[0025] Subsequently, TC4 titanium alloy specimens conforming to GB / T3075-2008 standard were prepared. Nine verification conditions were established using three shot peening intensities (0.1 mmA, 0.2 mmA, 0.3 mmA) and three coverage rates (100%, 200%, 300%). Physical tests were conducted using a pneumatic shot peening machine with cast steel shot as the peening medium and a shot peening pressure of 0.6 MPa ± 15%. For physical testing, a PROTOX residual stress detector was used with a tube voltage of 30 kV, a tube current of 5 mA, and a diffraction angle of 156°. Three test points were selected on the specimen to measure the principal stresses in the X and Y directions. A TIME3220 handheld roughness meter was used to measure surface roughness with Rz as the evaluation parameter. Simultaneously, DEM-FEM coupled simulations were performed for the corresponding conditions. By calculating the initial shot velocity and using it as a boundary condition, the residual stress and surface roughness values ​​were extracted from the simulation results. The comparison results show that the simulated value of residual stress is consistent with the experimental value, with the largest error of 1.66% at a coverage of 300%. The variation of surface roughness with shot peening intensity and coverage is consistent with the experiment, with a maximum error of only 6.78%, confirming the effectiveness of the model.

[0026] S102 uses a validated 3D simulation model to generate the original dataset of shot peening residual stress.

[0027] In this embodiment, based on the significant influence of shot diameter, shot peening velocity, and shot peening coverage on the residual stress of shot peening, the following parameters are selected: Nine sets of orthogonal array simulation experiments were designed to ensure comprehensive coverage and balanced combination of experimental factors and levels. The experimental factor levels were set as follows: shot diameters of 0.18mm, 0.28mm, and 0.38mm cast steel shot; shot peening velocities of 60m / s, 80m / s, and 100m / s; and shot peening coverage of 100%, 200%, and 300%. The specific process parameter combinations and corresponding surface residual stress reference values ​​for the nine sets of orthogonal conditions are defined as follows: SP-1: Shot diameter 0.18mm, shot peening speed 60m / s, shot peening coverage 100%, surface residual stress -761.61MPa; SP-2: Shot diameter 0.18mm, shot peening speed 80m / s, shot peening coverage 300%, surface residual stress -1161.33MPa; SP-3: Shot diameter 0.18mm, shot peening speed 100m / s, shot peening coverage 200%, surface residual stress -1157.82MPa; SP-4: Shot diameter 0.28mm, shot peening speed 60m / s, shot peening coverage 300%, surface residual stress -839.46MPa; SP-5: Shot diameter 0.28mm, shot peening speed 80m / s, shot peening coverage 200%, surface residual stress -992.71MPa; SP-6: Shot diameter 0.28mm, shot peening speed 100m / s, shot peening coverage 100%, surface residual stress -955.06MPa; SP-7: Shot diameter 0.38mm, shot peening speed 60m / s, shot peening coverage 200%, surface residual stress -1059.87MPa; SP-8: Shot diameter 0.38mm, shot peening speed 80m / s, shot peening coverage 100%, surface residual stress -1002.54MPa; SP-9: Shot diameter 0.38mm, shot peening speed 100m / s, shot peening coverage 300%, surface residual stress -1182.78MPa.

[0028] During simulation, all operating conditions followed the DEM-FEM coupled model parameters determined in the preceding steps, including an analysis step time of 0.5 ms, a time increment step of 1e-7 s, a contact friction coefficient of 0.3, and mesh generation standards. Only the experimental factor parameters were changed. The data acquisition process employed a multi-path stress averaging method. Using Python to further develop ABAQUS, multiple extraction paths were created in areas with a coverage rate greater than or equal to 98%. Sixteen measuring points were placed at 20 μm intervals along the depth direction of the workpiece target surface. The residual stress value in the S11 direction was extracted from each measuring point. Simultaneously, the shot diameter, shot peening velocity, shot peening coverage, and measuring point depth information for the corresponding operating condition were recorded. Averaging of multi-path data reduced outlier interference. The final raw dataset was stored in CSV format, with data fields including operating condition number, shot diameter, shot peening velocity, shot peening coverage, residual stress layer depth, and residual stress value. The total sample size was 144.

[0029] S103, normalize the original dataset of shot peening residual stress to obtain an effective sample set, and divide the effective sample set into a training set, a test set and a validation set.

[0030] In this embodiment, the Min-Max normalization method is used to uniformly transform the input parameters of effective samples—pellet diameter, peening velocity, peening coverage, residual stress layer depth, and residual stress value output parameters—to the [0,1] interval. This eliminates dimensional and numerical differences between different parameters, avoiding convergence bias caused by numerical imbalance during model training. The normalization calculation formula is as follows: in, This is the normalized residual stress output value. These are the original feature values ​​for a single sample. It is the minimum value within the entire set of feature data. It represents the maximum value within the entire set of feature data.

[0031] The effective sample set was randomly divided into training, testing and validation sets in a 6:2:2 ratio, with a random seed of 42 to ensure the repeatability of the partitioning results.

[0032] S104. Input the training set into the pre-built LSTM-Attention prediction model for training, and use the validation set to optimize the hyperparameters to obtain the optimal LSTM-Attention prediction model.

[0033] See Figure 2 In this embodiment, the LSTM-Attention prediction model includes an input layer. The input of the input layer receives normalized features such as shot diameter, peening velocity, peening coverage, and residual stress depth. The output of the input layer is connected to the input of the LSTM layer, the output of the LSTM layer is connected to the input of the attention mechanism layer, the output of the attention mechanism layer is connected to the input of the fully connected layer, and the output of the fully connected layer is connected to the output layer. The LSTM layer controls the flow of information through input gates, forget gates, and output gates, capturing long-term dependencies between parameters. The attention mechanism layer calculates the weight coefficients of each feature, assigning higher weights to important features to strengthen the influence of key information. The fully connected layer maps the processed features to predicted residual stress values.

[0034] The loss function used when inputting the training set into the pre-built LSTM-Attention prediction model for training is: A four-factor, four-level orthogonal experiment was used to optimize the core hyperparameters. The factors included optimizer type (Adam, SGD, Nadam, RMSprop), learning rate (0.1, 0.01, 0.001, 0.0001), batch size (5, 10, 20, 30), and number of hidden layer neurons (50, 75, 100, 125). The orthogonal experiment results showed that the optimal hyperparameter combination was: Adam optimizer, learning rate 0.01, batch size 10, and number of hidden layer neurons 75.

[0035] Training process monitoring: During training, the loss values ​​and MAPE of the training and validation sets are recorded every 10 iterations; the RMSE, MAPE, MAE, and R² metrics of the three datasets are calculated every 50 iterations; loss curves and performance metric curves are plotted to monitor the model training status in real time. If an overfitting trend occurs, indicating a continuous decrease in training set loss while an increase in validation set loss, the dropout rate is dynamically adjusted (range 0.1-0.3) to ensure the model's generalization ability. The calculation formulas are as follows: Where n is the total sample size, and n > 0; k For sample index, k =1,2,...; For the first k The true value of each sample For the first k The predicted value for each sample, This is the average of all samples.

[0036] S105. The optimal LSTM-Attention prediction model is validated using the test set. After the validation is successful, the shot peening process parameters and the target residual stress layer depth of the test condition are input into the optimal LSTM-Attention prediction model. After the model performs forward computation, the corresponding shot peening residual stress prediction value is output.

[0037] In this embodiment, each sample in the test set is input into the optimal LSTM-Attention prediction model to obtain the predicted residual stress value corresponding to each sample. The predicted residual stress value of each sample is compared with the corresponding true residual stress value, and the model performance evaluation index is calculated. When each evaluation index meets the preset requirements, the optimal LSTM-Attention prediction model is confirmed to have passed the verification.

[0038] After training, the model performance was evaluated on the test set and compared with feedforward neural network (BP), support vector regression (SVR), and single LSTM model. The performance metrics are shown in Table 1 below.

[0039] Table 1 Comparison of Evaluation Indicators for Different Models The results show that the LSTM-Attention prediction model proposed in this application significantly outperforms the comparative models in all performance metrics, with a determination coefficient of [missing information]. R 2 The model achieved an accuracy of 0.98, which is 4%, 20%, and 32% higher than BP, SVR, and single LSTM models, respectively. The error metrics such as RMSE, MAE, and MAPE are reduced by 47%-84% compared to ANN, SVM, and single LSTM models. The overall performance is improved by nearly 60%, which verifies the superiority and innovation of the model.

[0040] When predicting residual stress, the shot diameter, shot peening velocity, shot peening coverage, and residual stress layer depth under the predicted working condition are obtained. The shot diameter, shot peening velocity, shot peening coverage, and residual stress layer depth under the predicted working condition are normalized, and the preprocessed feature data is converted into a three-dimensional temporal vector adapted to the temporal network. The three-dimensional temporal vector is input into the validated optimal LSTM-Attention prediction model. The input layer outputs a standardized temporal feature sequence, which is then input into the LSTM layer. Through the collaborative work of the internal forget gate, input gate, and output gate, the coupling and correlation information of multiple process parameters in the temporal feature sequence is filtered, and multiple sets of hidden state feature vectors corresponding to different temporal nodes are output. Then, all hidden state feature vectors output by the LSTM layer are input into the attention mechanism layer. The weight coefficients corresponding to each set of hidden state feature vectors are calculated, and differentiated weights are assigned according to the contribution of each process feature to the prediction target. The weight coefficients are weighted and summed with the corresponding hidden state feature vectors to output a weighted feature vector that fuses global key features. The weighted feature vector is input into the fully connected layer, and the weighted feature vector is compressed in dimension and fused through the fully connected network. The normalized residual stress prediction is output by a single-output regression neuron. Finally, the residual stress prediction at the normalized scale is restored by Min-Max inverse normalization to obtain the shot peening residual stress prediction value at the real physical scale.

[0041] To validate the shot peening residual stress prediction method based on DEM-FEM-LSTM-Attention proposed in the above embodiments, this embodiment selects the BP model, SVR model, and single LSTM model as comparison models. Test set data is input into each trained model in sample order to obtain the corresponding predicted shot peening residual stress values, which are then compared with the simulated actual residual stress values ​​to form the actual value curve and predicted value curve for each model. The results are as follows: Figures 3-6 As shown. Figure 3 , Figure 4 and Figure 5 The diagrams show the residual stress prediction results of the BP model, SVR model, and single LSTM model, respectively. The predicted curves of the BP model and SVR model deviate significantly from the actual curves, and there is obvious distortion at the stress extremum locations. Although the single LSTM model can roughly follow the trend of residual stress changes, there are still significant prediction deviations at the stress peaks and compressive stress troughs. Figure 6 This diagram illustrates the residual stress prediction results of the LSTM-Attention model used in this application. The predicted curve is basically consistent with the actual residual stress curve, indicating that the method proposed in this application can effectively fit the distribution variation law of shot peening residual stress under the coupling of multiple process parameters. The prediction accuracy is significantly better than BP, SVR, and the single LSTM comparison model, demonstrating that introducing an attention mechanism to differentiate and weight process features can effectively improve the accuracy of shot peening residual stress prediction. The process parameters of this method are not only applicable to the specific values ​​mentioned above, but also have good predictive applicability across continuous ranges: the applicable shot diameter range is 0.05mm to 2mm, the applicable shot peening velocity range is 30m / s to 180m / s, the applicable shot peening coverage range is 50% to 500%, and the applicable residual stress layer depth range is 0μm to 1000μm. Furthermore, this method is not limited to cast steel shot and TC4 titanium alloy samples; the shot material can also be applied to commonly used shot peening media such as stainless steel shot, glass shot, ceramic shot, and chromium-molybdenum alloy steel shot; the sample material can also be applied to metal materials such as TA15 titanium alloy, 7075 aluminum alloy, 4340 high-strength steel, and nickel-based high-temperature alloys. Under the above wide parameter range and multi-material systems, this method maintains high predictive accuracy and generalization ability, without the need for remodeling or extensive supplementary simulations, and can be directly used for rapid prediction of residual stress under multiple working conditions and multi-material systems in engineering sites.

[0042] In this embodiment, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0043] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application.

Claims

1. A method for predicting residual stress in shot peening based on a three-dimensional process model of DEM-FEM-LSTM-Attention, characterized in that, include: A DEM-FEM coupled shot peening three-dimensional simulation model was constructed, and the effectiveness of the three-dimensional simulation model was verified based on the simulation results and experimental test results to obtain a verified three-dimensional simulation model. The original dataset of residual stress from shot peening was generated using the verified 3D simulation model. The original dataset of residual stress from shot peening is normalized to obtain an effective sample set, which is then divided into a training set, a test set, and a validation set. The training set is input into a pre-built LSTM-Attention prediction model for training, and the hyperparameters are optimized using the validation set to obtain the optimal LSTM-Attention prediction model. The optimal LSTM-Attention prediction model is validated using the test set. After successful validation, the shot peening process parameters and target residual stress layer depth of the test condition are input into the optimal LSTM-Attention prediction model. After forward computation of the model, the corresponding shot peening residual stress prediction value is output.

2. The method for predicting residual stress in shot peening based on a three-dimensional process model of DEM-FEM-LSTM-Attention as described in claim 1, characterized in that, A DEM-FEM coupled shot peening three-dimensional simulation model was constructed. The effectiveness of the three-dimensional simulation model was verified based on the simulation results and experimental test results, resulting in a verified three-dimensional simulation model, including: Using TC4 titanium alloy as the workpiece target material and cast steel as the pellet material, a three-dimensional geometric model including a pellet model and a workpiece target material model was constructed using ABAQUS software. The Johnson-Cook constitutive model is applied to the workpiece target to describe its dynamic mechanical properties. The nozzle is meshed using SFM3D4R elements, and the workpiece target is meshed using C3D8R elements. The mesh size of the area where the target surface contacts the pellet is set to one-tenth of the pellet diameter. The continuous shot blasting process is simulated using the Particle generator module, and the initial velocity of the shot is calculated based on the shot blasting pressure and shot flow rate. Using the initial velocity of the pellets as a boundary condition, DEM-FEM coupled simulation calculations were performed on the three-dimensional simulation model to obtain the simulated value of residual stress. TC4 titanium alloy specimens were prepared, and physical tests were conducted using a pneumatic shot peening machine under preset shot peening coverage conditions. The residual stress of the specimens was measured using an X-ray residual stress detection device to obtain the test results. The simulated residual stress value is compared and verified with the experimental residual stress value obtained from physical experiments. When the error between the two meets the preset threshold, the current DEM-FEM coupled shot peening three-dimensional simulation model is confirmed to be effective, and a verified three-dimensional simulation model is obtained.

3. The method for predicting residual stress in shot peening based on a three-dimensional process model of DEM-FEM-LSTM-Attention as described in claim 2, characterized in that, The continuous shot blasting process is simulated using the Particle generator module. The formula for calculating the initial velocity of the shot based on the shot blasting pressure and shot flow rate is as follows: in, V The initial velocity of the pellet. For pellet flow rate, Where is the shot peening pressure, and D is the shot diameter.

4. The method for predicting residual stress in shot peening based on a three-dimensional process model of DEM-FEM-LSTM-Attention as described in claim 1, characterized in that, The validated 3D simulation model is used to generate a raw dataset of shot peening residual stress, including: Shot diameter, shot peening velocity and shot peening coverage were selected as experimental factors. Multiple levels were set for each factor, and multiple sets of orthogonal simulation experiments were designed. The process parameters of each group of orthogonal simulation experiments are input into the verified three-dimensional simulation model for coupled simulation calculation; After each simulation is completed, multiple measuring points are set up at preset intervals along the depth direction of the workpiece target surface. The residual stress value in the preset direction at each measuring point is extracted, and the residual stress layer depth of the corresponding measuring point is recorded. This results in the construction of an original dataset containing shot diameter, shot peening speed, shot peening coverage, residual stress layer depth and corresponding residual stress value.

5. The method for predicting residual stress in shot peening based on a three-dimensional process model of DEM-FEM-LSTM-Attention as described in claim 1, characterized in that, The original dataset of residual stress from shot peening is normalized to obtain a valid sample set, which is then divided into a training set, a test set, and a validation set, including: The original dataset of residual stress from shot peening was normalized using the Min-Max normalization method to obtain an effective sample set. The calculation formula is as follows: in, This is the normalized residual stress output value. These are the original feature values ​​for a single sample. The minimum value within the entire set of feature data. It represents the maximum value within the entire set of feature data; The effective sample set is then divided into a training set, a test set, and a validation set according to a preset ratio.

6. The method for predicting residual stress in shot peening based on a three-dimensional process model of DEM-FEM-LSTM-Attention as described in claim 1, characterized in that, The LSTM-Attention prediction model includes an input layer, the input of which is used to receive normalized features of shot diameter, shot peening velocity, shot peening coverage and residual stress layer depth. The output of the input layer is connected to the input of the LSTM layer, the output of the LSTM layer is connected to the input of the attention mechanism layer, the output of the attention mechanism layer is connected to the input of the fully connected layer, and the output of the fully connected layer is connected to the output layer.

7. The method for predicting residual stress in shot peening based on a three-dimensional process model of DEM-FEM-LSTM-Attention as described in claim 1, characterized in that, The loss function used when inputting the training set into the pre-built LSTM-Attention prediction model for training is: in, Let n be the loss function, and n be the total number of samples. k For sample index, For the first k The true value of each sample For the first k The predicted value for each sample.

8. The method for predicting residual stress in shot peening based on a three-dimensional process model of DEM-FEM-LSTM-Attention as described in claim 1, characterized in that, The optimal LSTM-Attention prediction model is validated using the test set. After successful validation, the shot peening process parameters and target residual stress layer depth of the working condition to be predicted are input into the optimal LSTM-Attention prediction model. After forward computation, the corresponding predicted values ​​of shot peening residual stress are output, including: Each sample in the test set is input into the optimal LSTM-Attention prediction model to obtain the predicted residual stress value corresponding to each sample. The predicted residual stress value of each sample is compared with the corresponding real residual stress value. The model performance evaluation index is calculated. When each evaluation index meets the preset requirements, the optimal LSTM-Attention prediction model is confirmed to have passed the verification. Obtain the shot diameter, shot peening velocity, shot peening coverage, and residual stress layer depth under the predicted working condition; The shot diameter, shot peening velocity, shot peening coverage and residual stress layer depth of the working condition to be predicted are normalized, and the preprocessed feature data are converted into a three-dimensional temporal vector adapted to the temporal network. The three-dimensional temporal vector is input into the validated optimal LSTM-Attention prediction model, and the standardized temporal feature sequence is output through the input layer; The standardized time-series feature sequence is input into the LSTM layer. Through the collaborative work of the internal forget gate, input gate and output gate, the coupling and correlation information of multiple process parameters in the time-series feature sequence is filtered, and multiple sets of hidden state feature vectors corresponding to different time-series nodes are output. Then, all the hidden state feature vectors output by the LSTM layer are input into the attention mechanism layer. The weight coefficients corresponding to each set of hidden state feature vectors are calculated, and differentiated weights are assigned according to the contribution of each process feature to the prediction target. The weight coefficients are summed with the corresponding hidden state feature vectors to output a weighted feature vector that integrates global key features. The weighted feature vector is input into the fully connected layer, and the dimensionality of the weighted feature vector is compressed and the feature is fused through the fully connected network. The normalized residual stress prediction is output by the single-output regression neuron. Finally, Min-Max inverse normalization is used to restore the predicted residual stress at the normalized scale to obtain the predicted residual stress at the real physical scale.

Images