Prediction method of geometric parameters and crashworthiness performance of rail vehicle composite structure based on transfer learning and image recognition

Abstract

The present invention provides a method for predicting the geometric parameters and crashworthiness of a composite structure of a rail vehicle based on transfer learning and image recognition, establishes a finite element model of a composite energy absorption structure of a subway train, verifies the finite element model in combination with a real vehicle collision test, and calculates the collision results of the finite element model of the composite energy absorption structure under different design parameters through an experimental design method; based on the calculated collision results, the crashworthiness characteristic indicators and structural parameters are inversely calculated from the deformed image to ensure that the energy absorption structure has a better deformation pattern during the crashworthiness optimization process, broadens the application of image recognition in the crashworthiness design of the composite energy absorption structure of subway trains, and enables big data.

Description

Technical Field

[0001] The present invention relates to the technical field of train crashworthiness image recognition, and in particular to a method for predicting geometric parameters and crashworthiness of a rail vehicle composite structure based on transfer learning and image recognition. Background Art

[0002] The protective devices of subway trains are mainly divided into active protection and passive protection. With the continuous improvement of technology, the technology of active protection has been significantly developed. However, due to extreme weather or human factors, train collision accidents still occur frequently. When a train collides, the energy-absorbing structure becomes the last line of defense for subway trains. The research on its crashworthiness is of great research significance for protecting the safety of train passengers and the train body structure.

[0003] At present, the method of studying the crashworthiness of subway train energy-absorbing structures is mainly to establish a finite element model, verify the finite element model through a full-scale collision test of a real vehicle, and then design and optimize the crashworthiness of the energy-absorbing structure. However, due to the complexity of the full-scale collision test of a real vehicle, some test data may be lost during the actual experiment, and only the deformation image of the energy-absorbing structure after the collision can be obtained. Therefore, it is an effective way to inversely obtain the structural parameters and crashworthiness characteristic indicators through the deformation image, and then verify the finite element model.

[0004] In the process of optimizing the crashworthiness of subway train energy-absorbing structures, it is necessary to optimize the crashworthiness characteristic indexes while ensuring an orderly, stable and reliable deformation image. The traditional method for optimizing the composite energy-absorbing structure mainly uses the crashworthiness index as the optimization target to optimize the crashworthiness of the structure, ignoring the influence of the deformation image of the energy-absorbing structure and failing to ensure the robustness of the optimization results. Therefore, it is necessary to combine the image recognition method to inversely calculate the crashworthiness characteristic parameter indexes and structural parameters from the deformation image of the energy-absorbing structure, providing a new way to optimize the crashworthiness of the energy-absorbing structure. Summary of the invention

[0005] The purpose of the present invention is to provide a method for predicting the geometric parameters and crashworthiness of a rail vehicle composite structure based on transfer learning and image recognition, so as to solve the technical problems of the traditional method, such as the complexity of the test method and the easy loss of test data.

[0006] To achieve the above object, the present invention provides a method for predicting geometric parameters and crashworthiness of a rail vehicle composite structure based on transfer learning and image recognition, comprising the following steps:

[0007] S1. Establish a finite element model of the composite energy-absorbing structure of a subway train and use a full-scale real vehicle collision test method to verify the accuracy of the established finite element model;

[0008] S2. Based on Hammersley's sampling method, the finite element model corresponding to the scope of the structural parameter variables is subjected to virtual experimental design DOE to obtain the deformation image and crashworthiness characteristic index data corresponding to the scope of the energy-absorbing structure;

[0009] S3. TCL language programming to obtain the deformation image of the energy absorbing structure:

[0010] S4. Based on the results obtained by DOE calculation, the structural parameters and crashworthiness characteristic indicators corresponding to the deformed image are integrated using the TFRecord protocol method to generate the TFRecord data required by the network;

[0011] S5. Decode data based on single_example_parser protocol;

[0012] S6. Based on the Keras library, the classic image recognition network is called, and the classic network is applied to the image recognition of the composite energy absorption structure of the subway train by transfer learning. The transfer learning model for inverse structural parameters and crashworthiness characteristic indicators is obtained to realize the construction of the learning framework;

[0013] S7. Set the number of Epoch cycles for training, and determine whether the number of training cycles is less than Epoch. If so, call the Adam optimization function, call the custom loss function and evaluation function, use the training set and the validation set for cross-validation, and use the tensorflow training mode to train the network until the training coefficient is greater than Epoch, then end; otherwise, end directly.

[0014] As one of the preferred technical solutions, in step S1, based on the standard of "EN15227:2020 Railway Facilities-Crashworthiness Requirements for Railway Vehicles", a full-scale dynamic impact test of a real vehicle is adopted, and the boundary conditions of the finite element simulation model are set consistent with the test conditions. By combining the experiment with the simulation, the force-displacement curves and deformation images between the experiment and the simulation are compared to verify the accuracy of the established finite element model.

[0015] As one of the preferred technical solutions, in step S2, the structural parameter variables include the thickness of the thin wall A of the energy absorbing structure (T A ), thin wall B thickness (T B ), partition thickness (T gb ), aluminum honeycomb A strength (δ A ) and aluminum honeycomb B strength (δ B ); The crashworthiness characteristic indicators include energy absorption (EA) and initial peak force (IPIF).

[0016] As one of the preferred technical solutions, the specific method of step S3 is as follows: using TCL language programming, using Hyperview software to read finite element calculation results and screenshot functions, combined with the HWC handle of Hyperview software secondary development, setting the pixel frame size (length: 300 pixels, width: 300 pixels) for reading the deformation picture of the finite element calculation results of the energy-absorbing structure, opening the finite element results calculated by DOE in a loop, and automatically adjusting the view to a given angle, intercepting the deformation picture according to the set pixel frame size, and naming the screenshots in a loop according to the result file sequence corresponding to DOE and the numbers 1-1000 and saving them in the set path. Using the method of randomly extracting pictures, randomly select the pictures under the screenshot saving path according to a certain ratio, combine the absolute path and the saving path corresponding to the randomly selected pictures, store the test set, training set and verification set in different folders, and realize the processing and saving of the deformation picture set.

[0017] As one of the preferred technical solutions, the specific method of step S4 is as follows: in order to improve the data reading speed of the transfer learning network, the images and corresponding structural parameters captured in the DOE calculation results are saved in the TFRecord format, the names of the corresponding images in the folders storing the test set, validation set and training set are extracted, and then the corresponding structural parameters (TFRecord) are indexed according to the serial numbers corresponding to the names. A ,T B ,T gb ,δ A and δ B ) and crashworthiness characteristic indicators (EA and IPIF), and the deformation images and true structural parameters corresponding to the test set, validation set, and training set are combined and packaged according to the TFRecord protocol to generate the TFRecord data input required for network training.

[0018] As one of the preferred technical solutions, the specific method of step S5 is: using the decoding protocol of single_example_parser, resizing the deformed image according to the input dimension required by the transfer learning network, decompressing it to obtain the data sets of deformed image-structural parameters and deformed image-crashworthiness feature indicators for model training of transfer learning.

[0019] As one of the further preferred technical solutions, the structural parameters and collision resistance characteristic parameters are sequence tag values ​​with a certain length, and the intercepted pictures are non-serialized data. Therefore, the context_features and sequence_features features are used to save the non-serialized data (deformed pictures) and sequence tag values ​​with a certain length (structural parameters and collision resistance characteristic parameters). In the process of data decoding, the single_example_parser protocol method is used to decode the data corresponding to the context_features and sequence_features features, where the data of the sequence_features feature needs to set the length corresponding to the data.

[0020] As one of the preferred technical solutions, in step S6, the corresponding classic image recognition networks: VGG16, ResNet50, LetNet and AlexNet are imported through the Keras library; in order to facilitate the observation of the convergence and divergence of the network training process, the loss function and evaluation function are defined based on the K function of the backend function library of Keras; the mean square error (MSE) is used as the loss function of the optimization process, and the mean absolute error (MAE), root mean square error (RMSE), regression coefficient (R2) and mean relative absolute error (RAAE) are used as evaluation functions.

[0021] As one of the preferred technical solutions, in step S6, when building the transfer learning framework, the output layer of the classic image recognition network is deleted, the original weights of the network are used, and the middle layer of the network is set to a non-trainable mode, the middle layer is frozen, and a fully connected layer and a flatten layer are added to the output layer.

[0022] The present invention has the following beneficial effects:

[0023] In order to ensure the improvement of calculation time and deformation image pattern recognition, the present invention proposes a method for predicting the geometric parameters and crashworthiness of rail vehicle composite structures based on transfer learning and image recognition. The method of the present invention can not only invert the crashworthiness characteristic indicators and structural parameters from the deformed image, but also ensure that the energy-absorbing structure has a better deformation pattern during the crashworthiness optimization process, broaden the application of image recognition in the crashworthiness design of composite energy-absorbing structures of subway trains, and enable big data.

[0024] The present invention mainly performs parameter inverse calculation on structural parameters and crashworthiness indexes through collision pictures of composite energy-absorbing structures of subway trains, and specifically comprises the following steps: establishing a finite element model of composite energy absorption of subway trains, verifying the finite element model in combination with a real vehicle collision test, and calculating the collision results of the finite element model of the composite energy-absorbing structure under different design parameters through a design of experiments (DOE) method; based on the calculated collision results, extracting displacement deformation pictures of the composite energy-absorbing structure after collision by means of TCL language programming, and dividing the displacement deformation pictures into a test set, a test set and a verification set according to a random sampling method; packaging the displacement deformation pictures, structural parameters and crashworthiness indexes based on the TFRecord protocol to improve the data reading and storage speed, and decoding the data according to the single_example_parser decoding method to obtain the sample input set required for transfer learning; freezing the middle layer of the classic network model of transfer learning and saving the initial weights, deleting the bottom output layer of the network, and adding corresponding expansion layers and fully connected layers, and training the network by using the network architecture of tensorflow to obtain a transfer learning model for predicting structural parameters and crashworthiness indexes.

[0025] The specific analysis is as follows:

[0026] 1. This invention proposes a method for predicting the geometric parameters and crashworthiness of rail vehicle composite structures based on transfer learning and image recognition. Based on the verified finite element model, the DOE design method is adopted to obtain the collision deformation image of the composite energy-absorbing structure. Using the transfer learning method, the deformation image and structural parameters (T A ,T B ,T gb ,δ A and δ B ), crashworthiness characteristic indicators (EA, IPIF). During the test, the trained transfer learning network can be used to inversely search for crashworthiness parameters such as EA and IPIF during the test, thereby verifying the accuracy of the finite element model. At the same time, during the verification of the finite element model, the accuracy of the crashworthiness characteristic indicators collected by the test equipment can be verified based on the inversely searched parameters, thereby verifying the matching of the structural parameters corresponding to the finite element model and the test model, thereby achieving the effect of auxiliary testing.

[0027] 2. The traditional way to optimize the crashworthiness of energy-absorbing structures is mainly to verify the finite element model through a combination of experiments and simulations through the finite element method. Based on the verified finite element model, the design domain of the structural parameters is set, a large number of calculations are performed on the model through the DOE method, and the crashworthiness of the energy-absorbing structure is designed and optimized based on the response surface and optimization algorithms. In the process of improving the crashworthiness of the energy-absorbing structure parameters, a reasonable deformation mode cannot be guaranteed.

[0028] The present invention proposes a method for predicting geometric parameters and crashworthiness of rail vehicle composite structures based on transfer learning and image recognition. By adopting a transfer learning network, the structural parameters and crashworthiness characteristic indicators can be inversely calculated by setting a reasonable deformation mode, and a more effective structural parameter design domain is provided for the crashworthiness optimization design problem of the energy-absorbing structure, which provides a reliable guarantee for the crashworthiness design of the energy-absorbing structure. The present invention can not only ensure the optimization of the structural parameter indicators of the energy-absorbing structure, but also meet the requirements of stable and orderly deformation modes.

[0029] 3. The present invention proposes a method for predicting the geometric parameters and crashworthiness of rail vehicle composite structures based on transfer learning and image recognition, which can quickly predict structural parameters and crashworthiness, and introduces image judgment into the method for predicting crashworthiness indicators using structural parameters. It can be used to build a deep learning framework that includes curve prediction and image recognition, and provide image pattern discrimination for the intermediate link between structural parameter prediction and crashworthiness design, thereby improving the robustness of crashworthiness design.

[0030] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0032] Figure 1 It is a technical flow chart of the present invention;

[0033] Figure 2 It is a full-scale real car crash test diagram;

[0034] Figure 3 It is the comparison chart of impact force-displacement between test and simulation;

[0035] Figure 4 It is a comparison chart of the experimental and simulated deformation image sequences;

[0036] Figure 5 It is a flowchart of the screenshot and saving method of TCL language programming;

[0037] Figure 6 It is the LetNet network flow chart;

[0038] Figure 7 This is the VGG16 network flow chart;

[0039] Figure 8 It is the ResNet50 network flow chart;

[0040] Fig. 9 This is the AlexNet network flow chart;

[0041] Fig.10 It is a flow chart of transfer learning network training;

[0042] Fig.11 This is a comparison chart of the computational efficiency of transfer learning and finite element model;

[0043] Fig.12 It is a radar chart comparing the predicted structural parameters and the real structural parameters corresponding to the deformation image;

[0044] Fig.13 It is the impact force-displacement diagram of the predicted structural parameters and the real structural parameters calculated by finite element method;

[0045] Fig.14 It is the energy-displacement diagram of the finite element calculation of the predicted structural parameters and the real structural parameters;

[0046] Fig.15 It is the displacement change diagram of the predicted structural parameters and the actual structural parameters calculated by finite element calculation. DETAILED DESCRIPTION

[0047] The embodiments of the present invention are described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0048] The process of the present invention is as follows Figure 1As shown in the figure: First, the accuracy of the finite element model is verified by combining experiments and finite elements. Then, by changing the structural parameter scope of the energy-absorbing structure, a large number of finite element calculation results are generated using the verified finite model and DOE method. Combined with TCL language programming and the function of Hyperview software to read finite element results and deformation images, the finite element results of DOE calculation are read cyclically, and the deformation images corresponding to the structural parameters and the data corresponding to the crashworthiness characteristic indicators in the DOE calculation results are read. In order to improve the data reading speed of the transfer learning network, the captured images, corresponding structural parameters and crashworthiness characteristic parameters are saved in the TFRecord format. Using the transfer learning method, the Keras library is imported into the corresponding classic image recognition networks (VGG16, ResNet50, LetNet and AlexNet). In order to facilitate the observation of the convergence and divergence of the network training process, the loss function and evaluation function are defined based on the K function of the backend function library of Keras. The single_example_parser protocol method is used to decode the data, and the decoded data is used as the input of the transfer learning network. The middle layer of the network is frozen, and the corresponding fully expanded layer and fully connected layer are added to the output layer of the network. The network is trained using the Adam optimization algorithm, and the training effects of VGG16, ResNet50, LetNet and AlexNet networks are compared. Then, a transfer learning network model suitable for image recognition of energy-absorbing structures is obtained.

[0049] Specifically, the present invention proposes a method for predicting geometric parameters and crashworthiness of a rail vehicle composite structure based on transfer learning and image recognition, comprising the following steps:

[0050] Step 1: Establish a finite element simulation model of the subway train composite energy absorption structure, such as Figure 2 As shown, a full-scale real vehicle collision test (such as Figure 3 The finite element simulation model was verified by comparing the force-displacement curve and deformation image between the experiment and the simulation to verify the accuracy of the established finite element model. The deformation image is shown in Figure 4 shown.

[0051] Step 2: The thickness of the thin wall A of the composite energy absorbing structure of the subway train (T A ), thin wall B thickness (T B ), partition thickness (T gb ), aluminum honeycomb A strength (δ A ) and aluminum honeycomb B strength (δ B ) as the structural parameter variable, the Hammersley sampling method was used to perform DOE design on the finite element model. The number of DOEs was set to 1000 groups, and the scope of the structural parameter design variables is shown in Table 1.

[0052] Table 1. Structural parameter variable scope.

[0053]

[0054] Step 3. Use TCL language programming to use the Hyperview software to read the finite element calculation results and screenshot functions. Combined with the HWC handle of the secondary development of Hyperview software, set the pixel frame size (length: 300 pixels, width: 300 pixels) for reading the deformation image of the finite element calculation results of the energy-absorbing structure. Open the finite element results calculated by DOE in a loop, and automatically adjust the view to a given angle. Capture the deformation image according to the set pixel frame size, and name the screenshots in a loop according to the result file sequence corresponding to DOE, according to the numbers 1-1000, and save them in the set path. Use the method of randomly extracting pictures, randomly select the pictures under the screenshot saving path according to a certain ratio, combine the absolute path and the saving path corresponding to the randomly selected pictures, and store the test set, training set, and verification set in different folders to realize the processing and saving of the deformation picture set. The process of saving deformation screenshots by TCL language programming is as follows: Figure 5 shown.

[0055] Step 4: Save the images and corresponding structural parameters captured in the DOE calculation results in the TFRecord format, extract the names of the corresponding images in the folders storing the test set, validation set, and training set, and then index the corresponding structural parameters (T A ,T B ,T gb ,δ A and δ B ) and crashworthiness characteristic indicators (EA and IPIF), and the deformed images and true structural parameters corresponding to the test set, validation set, and training set are combined and packaged according to the TFRecord protocol to generate the TFRecord data input required for network training. Among them, the structural parameters and crashworthiness characteristic parameters are sequence label values ​​with a certain length, and the intercepted images are non-serialized data. Therefore, it is necessary to use context_features and sequence_features features to save non-serialized data (deformed images) and sequence label values ​​with a certain length (structural parameters and crashworthiness characteristic parameters). The read structural parameters (T A ,T B ,T gb ,δ A and δ B ) and crashworthiness characteristic indicators (EA and IPIF) need to be normalized, and the calculation equation for normalization is as follows:

[0056]

[0057]

[0058] Where i in the formula is the i-th structural parameter, i = 1, ..., 5 (corresponding to T A ,T B ,T gb ,δ A and δ B ), d represents the number of samples, d = 1, ..., 1000, t represents the t-th crashworthiness characteristic parameter, t = 1, 2 (corresponding to EA and IPIF respectively); is the regularized value of the i-th input feature of the d-th sample; X d,i The value of the i-th structural parameter of the d-th sample; is the regularized value of the output corresponding to the t-th crashworthiness characteristic parameter of the d-th sample; Y d,i is the value of the t-th crashworthiness characteristic parameter of the d-th sample; X and Y represent the structural parameters and the crashworthiness characteristic parameters; min(X) and min(Y) are the minimum values ​​of the structural parameters and the crashworthiness characteristic parameters, respectively; max(X) and max(Y) are the maximum values ​​of the structural parameters and the crashworthiness characteristic parameters, respectively.

[0059] Step 5. During data decoding, due to the different dimensions and sizes of the network input corresponding to transfer learning, the dimensions of the image size need to be changed during reading TFRecord data. Therefore, the single_example_parser protocol method is used to decode the data corresponding to the context_features and sequence_features features, where the data corresponding to the sequence_features feature needs to set the length of the data, the data length of the structure parameter is set to 5, and the length of the collision resistance feature parameter is set to 2.

[0060] Step 6: Import the corresponding classic image recognition networks through the Keras library: VGG16, ResNet50, LetNet and AlexNet. The corresponding network process analysis diagram is as follows Figure 6-Figure 9 As shown in the figure; In order to facilitate the observation of the convergence and divergence of the network training process, the loss function and evaluation function are defined based on the K function of the backend function library of Keras. The mean square error (MSE) is used as the loss function of the optimization process, and the mean absolute error (MAE), root mean square error (RMSE), and regression coefficient (R 2 ) and the average relative absolute error (RAAE) are used as the evaluation function. The calculation formulas of the loss function and the evaluation function are as follows:

[0061]

[0062]

[0063]

[0064]

[0065]

[0066] Where n is the number of samples; is the prediction output of the neural network corresponding to machine learning; y i is the actual output of the sample; and is the standard deviation between the predicted output and the actual output; It is the convolution of actual output and predicted output.

[0067] Step 7: In the process of building the energy-absorbing structure image recognition network, it is necessary to define the corresponding loss function and evaluation function, delete the output layer of the classic image recognition network, use the original weights of the network, set the middle layer of the network to a non-trainable mode, and add a flatten layer and a fully connected layer to the output layer; in the process of transfer learning network training, it is necessary to set the number of Epoch cycles for training. When the number of network training times is less than the Epoch value, the network will call the Adam optimization algorithm, use the training set and the validation set for cross-validation, and use the Tensorflow training mode to train the network model until the number of training times is greater than the number of Epochs. The flowchart of the network training process is as follows: Fig.10 shown.

[0068] The finite element model is able to forward calculate deformation patterns and crashworthiness characteristics (e.g., IPIF and SEA). However, the transfer learning model can not only realize the functions of the finite element model, but also reversely invert the structural parameters and crashworthiness characteristics from the deformation image. In order to further illustrate the efficiency of transfer learning in processing the image recognition of energy-absorbing structures, the ResNet50 classic network and the finite element model are used to compare the efficiency of the prediction process. The calculation time of the finite element model and the transfer learning model are 1h and 0.5s, respectively. During the training process, when the number of DOE samples is set to 1000, ResNet50 is used for prediction, and the corresponding training time is 1025h; however, when the network training is completed, the time consumed for a single prediction is only 0.5s. In order to further quantify the calculation time of the finite element model and the transfer learning model, the following computational efficiency indicators are used to evaluate the computational efficiency:

[0069] T ML=t train +n×t ml (4)

[0070] Among them, T ML The total computation time required to calculate DOE for transfer learning, t train is the time required for training the transfer learning model, t ml It is the time required to calculate each sample after the model is successfully trained, and n is the number of DOE samples.

[0071] T FE =n×t FE (5)

[0072] Among them, T FE is the total computation time required for DOE calculation of the finite element model, t FE is the time required to calculate each sample of the finite element model, and n is the number of samples in DOE.

[0073]

[0074] Among them, T ML The total computation time required to calculate DOE for the transfer learning model; T FE It is the total computational time required to calculate DOE of the finite element model; TIEF is the ratio of the time required to calculate DOE of the transfer learning model to that of the finite element model.

[0075] Comparison of computational efficiency indicators of transfer learning and finite element model under different DOE samples Fig.11 As shown. Fig.11 It can be seen that when the number of DOE samples is small, the calculation time of the transfer learning model is longer than that of the finite element model. As n increases, the calculation time of the transfer learning model remains basically unchanged, while the calculation speed of the finite element model increases significantly. For example, when the number of DOE samples is set to 500, T ML and T FE 1000.07 and 500 respectively, and the TIFE is 200.01%, which means that in this case, the time required for the transfer learning model is greater than that for the finite element model. However, when the number of DOE samples is 5000, T ML 、T FE and TIFE are 1000.69%, 5000 and 20.01% respectively, and the time required for the transfer learning model is less than that of the finite element model. At the same time, structural parameters, crashworthiness characteristics and deformation images can be obtained through the finite element and transfer learning models, while the transfer learning model can reversely predict structural parameters and crashworthiness from the deformation images, which is impossible with the finite element model.

[0076] The present invention is explained and illustrated below in conjunction with specific embodiments.

[0077] In order to further analyze the prediction effect, a group of deformation images of energy-absorbing structures are randomly used as samples, and the structural parameters of the energy-absorbing structure are predicted by the trained migration model. The prediction results and actual results corresponding to the deformation images are shown in Figure 2. Fig.12 As shown. Fig.12 As shown, the predicted structural parameters corresponding to the same deformation picture are very close to the actual structural parameters, indicating that machine learning is effective in predicting the structural parameters of the energy absorber.

[0078] In order to further verify the prediction effect of deformation mode and crashworthiness index, the energy absorber with predicted structural parameters was substituted into the finite element model for recalculation, and the actual calculation results and predicted values ​​are listed in Table 2. The difference between the real deformation image and the predicted deformation image is very small, and the deformation area is relatively consistent; at the same time, by comparing the errors of crashworthiness index, it can be found that the errors of EA and IPIF are 5.47% and 4.17% respectively, and the error of structural parameters is also less than 10%, indicating that the prediction effect is acceptable.

[0079] Table 2. Finite element calculation results corresponding to the actual structural parameters and predicted structural parameters

[0080]

[0081] In order to verify the prediction effect of the historical response of the crashworthiness index, Fig.13 and Fig.14 Compare the predicted results with the actual results. Fig.13 As shown in Figure 2, the predicted impact force-displacement curve is similar to the true curve, although the predicted force-displacement peak and peak are slightly lower than the true values. Fig.14 It can be seen that the predicted energy absorption curve is basically consistent with the true curve, and the predicted value is also slightly lower than the true value.

[0082] In order to further verify the robustness of the network model, Fig.15 The deformation process between the predicted data and the actual data was compared and analyzed. Fig.15 As shown, at the same time, the folding numbers of the energy absorbing structure corresponding to the predicted data and the real data are consistent. When T = 140ms, the deformation corresponding to the real data is slightly different from the predicted deformation.

[0083] In general, the deformation patterns, crashworthiness indicators, structural parameters, and historical responses corresponding to the predicted results are almost the same as the actual results, showing a high degree of consistency. Therefore, the trained transfer learning model has a good prediction effect and can be used in the image recognition process of energy absorbing structures.

[0084] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting geometric parameters and crashworthiness of rail vehicle composite structures based on transfer learning and image recognition, characterized in that: The following steps are involved: S1. Establish a finite element model of the composite energy absorption structure of a subway train and use a full-scale real vehicle collision test method to verify the accuracy of the established finite element model; S2. Based on Hammersley's sampling method, the finite element model corresponding to the scope of the structural parameter variables is subjected to virtual experimental design DOE to obtain the deformation image and crashworthiness characteristic index data corresponding to the scope of the energy-absorbing structure; S3. Obtain the deformation image of the energy-absorbing structure by programming in TCL language; use the Hyperview software to read the finite element calculation results and screenshot functions by TCL language programming, combine the HWC handle of the secondary development of the Hyperview software, set the pixel frame size of the deformation image of the finite element calculation results of the energy-absorbing structure, open the finite element results calculated by DOE in a loop, automatically adjust the view to a given angle, capture the deformation image according to the set pixel frame size, and name the screenshots in a loop according to the result file sequence corresponding to DOE and the numbers from 1 to 1000 and save them in the set path; use the method of randomly extracting pictures to randomly select the pictures under the screenshot saving path according to a certain ratio, combine the absolute path and the saving path corresponding to the randomly selected pictures, store the test set, training set and validation set in different folders, and realize the processing and saving of the deformation picture set; S4. Based on the results obtained by DOE calculation, the structural parameters and crashworthiness characteristic indicators corresponding to the deformed images are integrated using the TFRecord protocol method to generate the TFRecord data required by the network; in order to improve the data reading speed of the transfer learning network, the images intercepted from the DOE calculation results and the corresponding structural parameters are saved in the TFRecord format, the names of the corresponding images in the folders storing the test set, validation set, and training set are extracted, and then the corresponding structural parameters and crashworthiness characteristic indicators are indexed according to the serial numbers corresponding to the names, and the deformed images and true structural parameters corresponding to the test set, validation set, and training set are combined and packaged according to the TFRecord protocol to generate the TFRecord data input required for network training; S5. Decode data based on single_example_parser protocol; Using the decoding protocol of single_example_parser, the deformed image is resized according to the input dimension required by the transfer learning network, and decompressed to obtain the datasets of deformed image-structural parameters and deformed image-crashworthiness feature indicators for model training of transfer learning; S6. Based on the Keras library, the classic image recognition network is called, and the classic network is applied to the image recognition of the composite energy absorption structure of subway trains by transfer learning. The transfer learning model for inverse structural parameters and crashworthiness characteristic indicators is obtained, and the learning framework is built; When building the transfer learning framework, the output layer of the classic image recognition network is deleted, the original weights of the network are used, and the middle layer of the network is set to a non-trainable mode, the middle layer is frozen, and a fully connected layer and an expanded layer are added to the output layer; S7. Set the number of Epoch cycles for training and determine whether the number of training cycles is less than Epoch. If so, call the Adam optimization function, call the custom loss function and evaluation function, use the training set and validation set for cross-validation, and use the tensorflow training mode to train the network until the training coefficient is greater than Epoch, then end; otherwise, end directly.

2. The method for predicting geometric parameters and crashworthiness of composite structures of rail vehicles based on transfer learning and image recognition according to claim 1, characterized in that: In step S1, a full-scale dynamic impact test of a real vehicle is used, and the boundary conditions of the finite element simulation model are set to be consistent with the test conditions. By combining the test with the simulation, the force-displacement curves and deformation images between the experiment and the simulation are compared to verify the accuracy of the established finite element model.

3. The method for predicting geometric parameters and crashworthiness of composite structures of rail vehicles based on transfer learning and image recognition according to claim 1, characterized in that: In step S2, the structural parameter variables include the thickness of thin wall A, the thickness of thin wall B, the thickness of the partition, the strength of aluminum honeycomb A and the strength of aluminum honeycomb B of the energy absorbing structure; and the crashworthiness characteristic indicators include energy absorption and initial peak force.

4. The method for predicting geometric parameters and crashworthiness of composite structures of rail vehicles based on transfer learning and image recognition according to claim 1, characterized in that: In step S6, the corresponding classic image recognition networks: VGG16, ResNet50, LetNet and AlexNet are imported through the Keras library; in order to facilitate the observation of the convergence and divergence of the network training process, the loss function and evaluation function are defined based on the K function of the backend function library of Keras; the mean square error is used as the loss function of the optimization process, and the mean absolute error, root mean square error, regression coefficient and mean relative error absolute value are used as evaluation functions.

Images