A method for predicting mechanical properties of mechanical multi-joints based on deep learning

Abstract

The present invention discloses a method for predicting the mechanical properties of mechanical multiple joints based on deep learning, including: performing fine model simulation of multiple joints of SPR, FDS, and solder joints, creating several fine models of mechanical connection joints, and realizing data construction; collecting layout feature information of CAE fine models or experimental sheets; performing feature selection on the joint database based on the random forest method; constructing a strength feature model based on a convolutional neural network, using the feature information data selected by the random forest method as a training set, and using the corresponding joint strength feature information as a label to train the strength feature model of multiple joints; integrating multiple convolutional neural network models, comprehensively acquiring the predicted strength feature information of multiple joints, and evaluating and applying the model. The invention uses historical data and simulation results as training data, and the deep learning model can efficiently predict the mechanical properties when the number and position of joints change.

Description

Technical Field

[0001] The present invention belongs to the field of new energy vehicles, and specifically is a method for predicting the mechanical properties of mechanical multiple joints based on deep learning. Background Art

[0002] With the widespread popularity of new energy vehicles, automobile lightweighting, as an important way to achieve environmental protection and greenness in the vehicle production process, has been studied as a practical technology by major automobile companies. The large-scale application of lightweight materials has further promoted the process of automobile lightweighting. The most representative of these is the use of aluminum alloys. However, because traditional welding cannot achieve efficient connection between steel and aluminum, the connection between aluminum and high-strength steel requires innovative methods.

[0003] At present, technologies such as self-piercing riveting (SPR), hot melt self-tapping (FDS), Clinch and bonding are gradually being used on car bodies. These technologies are not only efficient and cost-effective, but also more environmentally friendly, and have already had a profound impact on the future of the automotive industry.

[0004] However, most of the current research only focuses on the mechanical properties of a single mechanical connection joint, but there is little research on the performance of multiple joints, especially the research on the working conditions of multiple mechanical connections of large parts. A model that can directly predict the performance of joints based on the layout has a more far-reaching significance for reducing the design redundancy of the automobile body and accelerating the optimization of structural optimization and material selection. Predicting the mechanical properties of multiple joints in an automated way can significantly improve production efficiency, reduce dependence on manual testing, and thus reduce production costs and time. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings existing in the prior art, and a method for predicting the mechanical properties of mechanical multiple joints based on deep learning is proposed, which can improve the efficiency of the arrangement of joints in the same part and reduce the design cost. On the basis of rapid prediction of the mechanical properties of multiple joints, appropriate strength redundancy design can be carried out or further lightweighting can be carried out while ensuring strength, that is, the same or even higher performance can be achieved with fewer joints.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: a method for predicting mechanical properties of mechanical multiple joints based on deep learning, characterized in that it includes:

[0007] S1: CAE simulation mother model construction, detailed model simulation of multiple joints of SPR, FDS, and solder joints, and benchmarking of simulation and experiment;

[0008] S2: Based on the step of S1, CAE simulation is used to create several fine models of mechanical connection joints to achieve data construction;

[0009] S3: Collect the layout feature information of the CAE fine model or the experimental sheet in steps S1 and S2: each group of models corresponds to a group of layout feature information, and the feature information includes the number of joints, the type of joints, and the arrangement of the joints; and collect the strength feature information: load peak value, the maximum force point in the stress-displacement curve;

[0010] S4: Preprocessing the acquired joint layout feature information and the corresponding strength feature information, integrating the preprocessed layout feature information and strength feature information of the joint samples to form a joint database, and performing feature selection on the joint database based on the random forest method;

[0011] S5: Building a strength feature model based on a convolutional neural network, using the feature information data selected by the random forest method as a training set, and using the corresponding joint strength feature information as a label to train the strength feature model of multiple joints;

[0012] S6: Integrate multiple convolutional neural network models to comprehensively obtain the predicted multi-joint strength feature information;

[0013] S7: Evaluation of the model. The mean absolute error and mean square error are selected as evaluation indicators. The model is evaluated using the test set data samples. Different evaluation indicators are obtained by adjusting the number and size of convolutional layers, the number of bidirectional long short-term memory network layers, the number of neurons in each layer, the time step, and the attention score function. The evaluation indicators are compared to select the parameter combination with a closer predicted value to the actual value and a more stable prediction performance.

[0014] S8: Model application: Apply the trained model to the actual test project, compare the multi-joint strength values ​​predicted by the model with the multi-joint strength values ​​obtained from the actual test, evaluate the application prediction effect of the model, retrain the model with new experimental data, and update the model parameters.

[0015] As a further description of the above technical solution: in step S1, the CAE simulation mother model is constructed to maintain the experimental loading conditions, so that the peak simulation accuracy under the shear load reaches more than 90% to be considered valid.

[0016] As a further description of the above technical solution: in step S2, the number of joints, the type of joints, and the arrangement of joints are arbitrarily changed in the overlap area, and finally a number of CAE overlap specimens are obtained, and quasi-static tensile simulation is performed on them to obtain force-displacement curves.

[0017] As a further description of the above technical solution: the types of connectors and the arrangement of connectors include but are not limited to the physical distance between connectors and the distance between connectors and the edge.

[0018] As a further description of the above technical solution: In step S2, the force-displacement curve obtained by CAE simulation or experiment is used to obtain the peak force through the function: Peak load = Max F (x), x∈ (0, x max ), x max= max{x|f(x)=0}; Peak load is the peak force, x is the displacement, and F(x) is the function expression of the force-displacement curve.

[0019] As a further description of the above technical solution: In step S4, the specific steps of feature selection for the layout feature information and the strength feature information of the multi-joint sample are as follows:

[0020] (1) Collect and preprocess layout feature information and intensity feature information data to ensure data quality.

[0021] (2) Integrate the data into a multi-joint sample database,

[0022] (3) Calibrate the layout feature information as feature variables.

[0023] (4) Calibrate the intensity feature information as the target variable;

[0024] (5) Use the characteristic variables and target variables in the rock sample characteristic database to train the random forest;

[0025] (6) After the random forest model training is completed, the permutation importance of each feature variable is calculated, and the weight index of each feature is calculated.

[0026] The calculation formula is: Among them, P i is the permutation importance index, pdx is the permutation importance of the x-th feature in the random forest, and the J features with the highest weights are selected based on the permutation importance of the features.

[0027] As a further description of the above technical solution: in step S5, after each feature selection, the top J feature data are integrated into a feature training set, a multi-joint strength prediction model based on a convolutional neural network is constructed, the layout feature data in the feature training set is input, and the corresponding multi-joint strength feature is used as the output value.

[0028] As a further description of the above technical solution: In step S6, the integration method of integrating multiple convolutional neural network models is: after feature selection by the random forest method, the feature training set is input into the multi-joint strength feature prediction model, after m times of feature selection by the random forest method, m multi-joint strength feature prediction models are generated, and the m prediction models are integrated by the model fusion method, specifically: Among them, M(x) is the output after model integration, is the prediction output of the nth model, is the weight index of the nth model.

[0029] Due to the adoption of the above technical solution, the beneficial effects achieved by the present invention are:

[0030] (1) By using historical data and simulation results as training data, the deep learning model can efficiently predict the mechanical properties when the number and position of joints change;

[0031] (2) The present invention significantly improves the accuracy and efficiency of mechanical performance prediction of multiple joints. Through deep learning technology, it can more accurately simulate and predict the mechanical behavior of joints under different working conditions, thereby providing more reliable data support for mechanical design and manufacturing;

[0032] (3) The present invention can reduce the number of experiments, reduce R&D costs, and shorten the time from product design to market;

[0033] (4) In practical applications, the present invention can provide engineers with more intuitive multi-joint strength prediction results, help optimize the design, and improve the overall performance and safety of the product. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is an implementation flow chart of a method for predicting mechanical properties of mechanical multiple joints based on deep learning proposed by the present invention;

[0035] Figure 2 This is an example diagram of the arrangement of multiple joints in a method for predicting mechanical properties of multiple joints based on deep learning proposed by the present invention;

[0036] Figure 3 This is an example diagram of the characteristics of multi-joint layout position information of a mechanical multi-joint mechanical performance prediction method based on deep learning proposed by the present invention;

[0037] Figure 4 This is a convolutional neural network model diagram for constructing a mechanical multi-joint mechanical property prediction method based on deep learning proposed by the present invention. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] Reference Figure 1-4 The present invention provides the following embodiments to describe the present invention in detail: Example 1

[0040] The present invention provides a method for predicting the mechanical properties of mechanical multiple joints based on deep learning, specifically a prediction model based on convolutional neural networks, which is used to evaluate the mechanical properties of multiple joints in mechanical connections. By using historical data and simulation results as training data, the deep learning model can efficiently predict the mechanical properties when the number and position of joints change. In this article, the CAE simulation mother model is constructed as follows: first, a fine model simulation of multiple joints of SPR, FDS, and welds is performed, and the simulation and experiment are compared. The experimental loading conditions are kept the same, and the peak simulation accuracy under shear load reaches 90% to be considered valid. Because the subsequent construction data are all constructed using this model, that is, only the position and type of the joints need to be changed in the fine model, and other settings of the simulation model are not allowed to be adjusted, it is called the mother model; SPR (self piercing rivet) is self-piercing rivet; FDS (flow drill screw) is hot-melt self-tapping. The specific steps are as follows:

[0041] 1. Construct the CAE simulation mother model. First, perform fine model simulation of multiple joints of SPR, FDS, and solder joints, and compare the simulation with the experiment. The experimental loading conditions are kept the same. The peak simulation accuracy under shear load reaches 90%, which can be considered valid.

[0042] 2. Data construction: CAE simulation is used to create several fine models of mechanical connection joints. As shown in the figure, the number of joints, the type of joints, and the arrangement of joints (including but not limited to the physical distance between joints and the distance from the edge of the joint) can be changed at will in the overlap area. Finally, several CAE overlap specimens are obtained, and quasi-static tensile simulation is performed on them to obtain the force-displacement curve.

[0043] The force displacement curve obtained by CAE simulation or experiment can be used to obtain the peak force through the following function: Peak load = Max F (x), x∈ (0, x max ), x max=max{x|f(x)=0}; Peak load is the peak force, x is the displacement, and F(x) is the force-displacement curve function expression. In step 1, the strength information of multiple joints can also be obtained through experiments, but the experimental workload can be greatly reduced through CAE simulation.

[0044] 3. Collect the layout feature information of the CAE fine model or experimental piece in step 1: Each group of models corresponds to a group of layout feature information, which includes the number of joints, the type of joints, the arrangement of joints (including but not limited to the physical distance between joints and the distance from the joints to the edge), and collect strength feature information: load peak (the point with the maximum force value in the stress-displacement curve).

[0045] 4. Preprocess the acquired joint layout feature information and the corresponding strength feature information, integrate the layout feature information and strength feature information of the preprocessed joint samples to form a joint database, and perform feature selection based on the random forest method.

[0046] Furthermore, the specific logic for feature selection of the layout feature information and the strength feature information of the multi-joint sample is as follows: collecting and preprocessing the layout feature information and the strength feature information data to ensure data quality, integrating the data into a multi-joint sample database, calibrating the layout feature information as a feature variable, and calibrating the strength feature information as a target variable;

[0047] The feature variables and target variables in the rock sample feature database are used to train the random forest. After the random forest model training is completed, the permutation importance of each feature variable is calculated, and the weight index of each feature is calculated based on the formula: Among them, P i is the permutation importance index, pdx is the permutation importance of the x-th feature in the random forest, and J features with the highest weights are selected according to the permutation importance of the features; by selecting features with higher weights for model training, it can ensure that the model focuses on the most relevant information without being disturbed by invalid data, improve the generalization ability of the model, and thus improve the prediction accuracy and efficiency of the model.

[0048] 5. Construct a strength feature model based on a convolutional neural network. The feature information data selected by the random forest method is used as the training set, and the corresponding joint strength feature information is used as the label to train the strength feature model of multiple joints.

[0049] After each feature selection, the top J feature data are integrated into a feature training set, and a multi-joint strength prediction model based on a convolutional neural network is constructed. The layout feature data in the feature training set is input, and the corresponding multi-joint strength features are used as output values.

[0050] 6. Integrate multiple convolutional neural network models to comprehensively obtain the predicted multi-joint strength feature information. The integration method is: after feature selection by random forest method, input the feature training set into the multi-joint strength feature prediction model, after m times of feature selection by random forest method, generate m multi-joint strength feature prediction models, and integrate the m prediction models by model fusion method. The specific method is: M(x) is the output of the model integration. is the prediction output of the nth model, is the weight index of the nth model, which is the permutation importance index mentioned in the third step.

[0051] 7. Model Evaluation

[0052] The mean absolute error and mean square error were selected as evaluation indicators, and the model was evaluated using the test set data samples. Different evaluation indicators were obtained by adjusting the number and size of convolutional layers, the number of bidirectional long short-term memory network layers, the number of neurons in each layer, the time step, and the attention score function. The evaluation indicators were compared and the parameter combination with the predicted value closer to the actual value and more stable prediction performance was selected.

[0053] 8. Model Application

[0054] The trained model is applied to the actual test project, and the multi-joint strength values ​​predicted by the model are compared with the multi-joint strength values ​​obtained from the actual test to evaluate the application prediction effect of the model. The model is retrained with new experimental data to update the model parameters.

[0055] Example 2

[0056] The present invention provides a prediction model based on convolutional neural networks for evaluating the mechanical properties of multiple joints in mechanical connections. By using historical data and simulation results as training data, the deep learning model can efficiently predict the mechanical properties when the number and position of joints change. The steps are as follows:

[0057] 1. Construct the CAE simulation mother model. First, perform fine model simulation of multiple joints of SPR, FDS, and solder joints, and compare the simulation with the experiment. The experimental loading conditions are kept the same. The peak simulation accuracy under shear load reaches 90%, which can be considered valid.

[0058] 2. Data collection: Use CAE simulation and data collection to collect historical data of mechanical connection multi-joints, randomly generate n multi-joint overlap models, and perform CAE simulation on the process of shear failure of the model, including the location parameters of the joints (recorded in the form of pictures, such as Figure 2 ), characteristic parameters of the joints (i.e. recording the type of each joint), and peak load data are used as databases for subsequent machine learning.

[0059] Data representation: First, each connection point is represented as a node in the graph. The characteristics of the node are the location information and characteristic parameters of the connection point (type of connection point, such as SPR, FDS, spot welding).

[0060] Graph construction: Next, you need to decide how the connection points are connected. This connection is based on the physical distance between the connection points. Save the arrangement of the connection points in the form of a graph.

[0061] Target variable: Ensure that the strength of each joint is part of the training data for supervised learning.

[0062] 3. The construction of the GNN neural network, the specific code is implemented by python, and the following is an implementation example.

[0063] Import necessary libraries and modules:

[0064] import torch;

[0065] from torch_geometric.nn import GCNConv, global_mean_pool;

[0066] from torch.nn import Linear;

[0067] Define the GNN model class:

[0068] Create a class that inherits from torch.nn.Module.

[0069] The layers of the model are defined in the initialization function. Typically, this includes one or more graph convolutional layers (such as GCNConv), and possibly pooling layers and fully connected layers (for output predictions).

[0070] class GNNModel(torch.nn.Module):

[0071] def __init__(self, num_node_features, num_classes):

[0072] super(GNNModel, self).__init__();

[0073] self.conv1 = GCNConv(num_node_features, 16);

[0074] self.conv2 = GCNConv(16, 32);

[0075] self.out = Linear(32, num_classes);

[0076] The forward method defines how data flows through the model, which typically includes activation functions (such as ReLU) and global pooling (aggregating the node representations of the graph into a graph-level representation).

[0077] def forward(self, data):

[0078] x, edge_index, batch = data.x, data.edge_index, data.batch;

[0079] x = self.conv1(x, edge_index);

[0080] x = torch.relu(x);

[0081] x = self.conv2(x, edge_index);

[0082] x = torch.relu(x);

[0083] x = global_mean_pool(x, batch) # Aggregate node features;

[0084] x = self.out(x);

[0085] return x;

[0086] Instantiate the model, define the loss function and optimizer:

[0087] Create a model instance, specifying the number of input node features (3 since the connection point locations are two-dimensional) and the number of output categories (1 in this example since we are predicting lap joint strength).

[0088] model = GNNModel(num_node_features=3, num_classes=1);

[0089] Select the loss function and optimizer. Since this is a regression task, you can use mean square error as the loss function and Adam as the optimizer.

[0090] criterion = torch.nn.MSELoss();

[0091] optimizer = torch.optim.Adam(model.parameters(), lr=0.01),

[0092] Model training:

[0093] Prepare data and load information such as connection point locations into the Data object.

[0094] Implement the model training loop, including forward propagation, loss calculation, backpropagation, and parameter update.

[0095] for epoch in range(200):

[0096] model.train();

[0097] optimizer.zero_grad() # clear gradient;

[0098] out = model(data) # forward propagation;

[0099] loss = criterion(out, target) # Calculate loss;

[0100] loss.backward() # Backward propagation;

[0101] optimizer.step() # Update parameters;

[0102] 4. Model evaluation and application:

[0103] (1) Evaluate model performance on the test set,

[0104] (2) Use the trained model to make predictions.

[0105] 5. Evaluate the accuracy of the shown convolutional neural network.

[0106] 6. Use the trained model to make predictions.

[0107] The deep learning-based mechanical multi-joint mechanical property prediction method provided by the present invention has significant technical advantages and practical application value.

[0108] (1) By building a CAE simulation master model and conducting detailed simulation and experimental benchmarking, the high accuracy and reliability of the prediction model are ensured;

[0109] (2) The random forest method is used for feature selection, which effectively improves the generalization ability and prediction efficiency of the model;

[0110] (3) The present invention uses deep learning technology, especially convolutional neural networks, to achieve rapid and accurate prediction of the mechanical properties of multiple joints in mechanical connections. This not only greatly reduces the number of physical prototypes to be made and tested, reduces R&D costs, but also significantly shortens the time to market and improves the market competitiveness of enterprises.

[0111] (4) The prediction method provided by the present invention is highly flexible and scalable. The number of joints, the type of joints, and the arrangement of joints can be changed at will in the overlap area according to actual needs, so as to quickly obtain the prediction results of mechanical properties under different configurations.

[0112] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting mechanical properties of mechanical multi-joints based on deep learning, characterized in that: include: S1: CAE simulation mother model is constructed to simulate the fine model of multiple joints of SPR, FDS and solder joints, and compare the simulation with the experiment, where SPR is self-piercing riveting and FDS is hot-melt self-tapping; S2: Based on the step of S1, CAE simulation is used to create several fine models of mechanical connection joints to achieve data construction; S3: Collect the layout feature information of the CAE fine model or the experimental sheet in steps S1 and S2: each group of models corresponds to a group of layout feature information, and the feature information includes the number of joints, the type of joints, and the arrangement of the joints; and collect the strength feature information: load peak value, the maximum force point in the stress-displacement curve; S4: Preprocessing the acquired joint layout feature information and the corresponding strength feature information, integrating the preprocessed layout feature information and strength feature information of the joint samples to form a joint database, and performing feature selection on the joint database based on the random forest method; S5: Building a strength feature model based on a convolutional neural network, using the feature information data selected by the random forest method as a training set, and using the corresponding joint strength feature information as a label to train the strength feature model of multiple joints; S6: Integrate multiple convolutional neural network models to comprehensively obtain the predicted multi-joint strength feature information; S7: Evaluation of the model. The mean absolute error and mean square error are selected as evaluation indicators. The model is evaluated using the test set data samples. Different evaluation indicators are obtained by adjusting the number and size of convolutional layers, the number of bidirectional long short-term memory network layers, the number of neurons in each layer, the time step, and the attention score function. The evaluation indicators are compared to select the parameter combination with a closer predicted value to the actual value and a more stable prediction performance. S8: Model application: Apply the trained model to the actual test project, compare the multi-joint strength values ​​predicted by the model with the multi-joint strength values ​​obtained from the actual test, evaluate the application prediction effect of the model, retrain the model with new experimental data, and update the model parameters.

2. A method for predicting mechanical properties of mechanical multiple joints based on deep learning according to claim 1, characterized in that: In step S1, the construction of the CAE simulation mother model must be consistent with the experimental loading conditions, and the peak simulation accuracy under shear load must reach more than 90% to be considered valid.

3. A method for predicting mechanical properties of mechanical multiple joints based on deep learning according to claim 1, characterized in that: In step S2, the number of joints, the type of joints, and the arrangement of joints are arbitrarily changed in the overlap region, and finally a number of CAE overlap specimens are obtained, and quasi-static tensile simulation is performed on them to obtain force-displacement curves.

4. A method for predicting mechanical properties of mechanical multiple joints based on deep learning according to claim 3, characterized in that: The types of connectors and the arrangement of connectors include but are not limited to the physical distance between connectors and the distance between connectors and the edge.

5. A method for predicting mechanical properties of mechanical multiple joints based on deep learning according to claim 4, characterized in that: In step S2, the peak force is obtained by the force-displacement curve obtained by CAE simulation or experiment through the function: Peak load=Max F(x),x∈(0,x max ),x max= max{x∣f(x)=0}; Peak load is the peak force, x is the displacement, F(x) is the function expression of the force-displacement curve, f(x) is the force value corresponding to a certain displacement, x max is the displacement when the force value is zero, that is, the maximum displacement.

6. A method for predicting mechanical properties of mechanical multiple joints based on deep learning according to claim 1, characterized in that: In step S4, the specific steps of performing feature selection on the layout feature information and the strength feature information of the multi-joint sample are as follows: (1) Collect and preprocess layout feature information and intensity feature information data to ensure data quality. (2) Integrate the data into a multi-joint sample database, (3) Calibrate the layout feature information as feature variables. (4) Calibrate the intensity feature information as the target variable; (5) Use the characteristic variables and target variables in the rock sample characteristic database to train the random forest; (6) After the random forest model training is completed, the permutation importance of each feature variable is calculated, and the weight index of each feature is calculated. The calculation formula is: ； Among them, P i is the permutation importance index, is the permutation importance of the xth feature in the random forest, and the J features with the highest weights are selected based on the permutation importance of the features.

7. The method for predicting mechanical properties of mechanical multiple joints based on deep learning according to claim 1, characterized in that: In step S5, after each feature selection, the top J feature data are integrated into a feature training set, a multi-joint strength prediction model based on a convolutional neural network is constructed, the layout feature data in the feature training set is input, and the corresponding multi-joint strength feature is used as the output value.

8. The method for predicting mechanical properties of mechanical multiple joints based on deep learning according to claim 1, characterized in that: In step S6, the integration method of multiple convolutional neural network models is: after feature selection by random forest method, input feature training set into multi-joint strength feature prediction model, after k times of feature selection by random forest method, generate k multi-joint strength feature prediction models, and integrate k prediction models by model fusion method, and the specific method is: , M(x) is the output of the model integration, m k (x) is the predicted output of the kth model, P ik is the weight index of the kth model, which is the permutation importance index mentioned in the third step.

Images