A method for predicting impact energy of automotive CFRP based on deep learning and damage images

Abstract

The present invention discloses a method for predicting the impact energy of automotive CFRP based on deep learning and damage images, comprising: step 1, establishing a simulation model of automotive CFRP laminates and impact hammers, and performing drop hammer impact simulation; step 2, obtaining fiber tensile damage images, fiber compression damage images, body tensile damage images, body compression damage images, and layered damage images from the simulation results, and integrating and processing the simulated damage images into corresponding data sets according to different impact energies and different categories of damage images; step 3, transferring different categories of damage image data sets to multiple deep neural networks, and training based on multiple optimization methods; comparing by accuracy, loss value, and visualized heat map, analyzing and obtaining the damage image category and optimal model with the highest correlation with the initial impact energy. Generate corresponding models for damage images based on deep learning methods to predict the initial impact energy of vehicles, thereby improving prediction efficiency and accuracy.

Description

Technical Field

[0001] The invention relates to a method for predicting the impact energy of automotive CFRP based on deep learning and damage images, and belongs to the research field of carbon fiber composite materials. Background Art

[0002] Carbon fiber reinforced plastics (CFRP) have been used in automobile structures due to their high specific stiffness and specific strength, excellent fatigue performance and corrosion resistance. At present, composite materials are gradually used in automobile structural parts, such as automobile seats, B-pillars and other collision safety parts. The density of carbon fiber reinforced composite materials is only 16g / cm3. Applying them to the design of the body and other parts can reduce the weight of the whole vehicle by about 35%, achieve the purpose of reducing fuel consumption, and realize the lightweight of the vehicle. At the same time, the main components of carbon fiber reinforced composite materials are carbon fiber tows and resin materials, which have stable chemical properties, good durability, long life and high fatigue strength. In terms of safety issues, the tensile strength of carbon fiber reinforced composite materials is five times that of ordinary steel, generally above 3500MPa, which can absorb a large amount of energy in the event of a collision, thereby effectively improving the passive safety of the vehicle. Carbon fiber reinforced composite materials are mainly used in automobile bodies, interior and exterior decoration, chassis systems, power systems, etc. They have high vibration damping and make a great contribution to improving the NVH (noise, vibration and harshness) of the entire vehicle. They can effectively play a shock-absorbing role and greatly improve the comfort of the car during driving.

[0003] Research by researchers at home and abroad has found that matrix cracking is the earliest impact damage to automotive CFRP materials. The matrix crack direction is parallel to the fiber direction and is at a certain angle to the thickness direction of the composite material. As the impact process progresses, the matrix cracks will gradually spread to the interface between the layers, causing the appearance and spread of delamination damage. For automotive CFRP laminates, fiber breakage is the last type of impact damage to occur. The area where the laminate surface contacts the impact object will cause compression and shear damage to the fibers due to the contact force and bending load of the laminate; on the back of the laminate, the tensile stress generated will cause local fiber tensile failure in the laminate. At the same time, the fibers in the damaged area will cause local stress concentration, which leads to a significant loss of compressive strength.

[0004] Impact damage will greatly reduce the compressive strength of automotive CFRP composite materials, and its complex failure mechanism will also bring great difficulties to the prediction of various physical parameters. With the development of computers and the continuous improvement of computing power, researchers have proposed a progressive damage model to simulate the impact and compression failure process of laminates. This type of research usually simplifies the impact damage, and the results are more sensitive to boundary conditions, failure criteria and other parameters. Researchers have also proposed an empirical formula for predicting the mechanical properties of CFRP composite laminates based on impact energy and delamination area. The empirical formula can quickly solve engineering problems, but the error may be large, and some details of impact damage, such as the shape and position of various damage images, may be ignored, affecting the accuracy. Therefore, it is a new and feasible development direction to generate corresponding models for damage images based on deep learning methods to predict the required parameters. Summary of the invention

[0005] The present invention designs and develops a method for predicting the impact energy of automotive CFRP based on deep learning and damage images. The method generates a corresponding model for the damage image based on the deep learning method to predict the initial impact energy of the vehicle, thereby improving the prediction efficiency and accuracy.

[0006] The technical solution provided by the present invention is:

[0007] A method for predicting the impact energy of automotive CFRP based on deep learning and damage images, comprising:

[0008] Step 1: Establish a simulation model of a CFRP laminate and an impact hammer head for a vehicle, and perform a drop hammer impact simulation;

[0009] Step 2: Obtain fiber tensile damage images, fiber compression damage images, body tensile damage images, body compression damage images, and layered damage images from the simulation results, and integrate and process the damage images obtained by simulation into corresponding data sets according to different impact energies and different types of damage images;

[0010] Step 3: Pass different categories of damage image data sets into multiple deep neural networks and train them based on multiple optimization methods; compare them through accuracy, loss value and visualized heat map to obtain the damage image category with the highest correlation with the initial impact energy and the optimal model, and determine the initial impact energy of the vehicle.

[0011] Preferably,

[0012] In the step 1, the laminate model is composed of laminates;

[0013] Among them, the odd-numbered layers are arranged in the same way, and the even-numbered layers are arranged in the same way;

[0014] The shape of the impact hammer head is bullet-shaped.

[0015] Preferably, the initial energy of the falling hammer is obtained by the kinetic energy theorem:

[0016]

[0017] Where W 外 is the work done by the resultant force on the object, v 初 is the initial velocity of the object, v 末 is the final velocity of the object, m is the total mass of the object;

[0018] The drop weight mass is indirectly modified by modifying the density parameter of the drop weight. The formula is:

[0019] m=ρV;

[0020] Where m is the mass of the object, ρ is the density of the object, and V is the volume of the object.

[0021] Preferably, the step 2 comprises:

[0022] Store damage images with different impact energies into different folders to create the data set required for deep learning training;

[0023] Before the data set is passed into the deep learning network, it needs to be preprocessed, including:

[0024] Crop the original image, expand and transform the data, and randomly reverse or rotate the image;

[0025] The processed data set is used to obtain damage image data sets under different impact energies;

[0026] After data enhancement, it is divided into training set, validation set and test set in the ratio of 7:1.5:1.5.

[0027] Preferably, the step three comprises:

[0028] Iteratively train the data set based on different deep learning networks;

[0029] Based on the classification cross entropy loss function of the classification task, the training learning rate is 0.001;

[0030] The loss function image dataset obtained through the drop hammer low-speed impact simulation experiment was used as the training set, validation set, and test set in a ratio of 7:1.5:1.5;

[0031] The training is performed with the damage image as input and the initial impact energy as output;

[0032] During the training process, a small batch learning method is used. Model training is terminated when the loss value tends to be stable and no longer decreases, the accuracy tends to be stable and no longer increases, or the model fails to converge.

[0033] During the training iteration, the value of the hyperparameter is the epoch parameter data with the highest accuracy during the training iteration.

[0034] The training and testing process was repeated three times on each model. The training set, validation set, and test set were randomly redistributed each time, and the best trained model was selected.

[0035] Preferably, four different damage images are trained iteratively based on the three models of ResNet50, VAN, and ViT, and 12 different training methods are obtained. Then, the data set is redistributed into training set, validation set, and test set, and the 12 training methods are repeated three times, and the prediction results are averaged;

[0036] Among them, the four different damage images are: fiber compression damage image, fiber tension damage image, body compression damage image, and body tension damage image.

[0037] Preferably, each neural network model uses Accuracy to calculate the accuracy, and the calculation formula is:

[0038]

[0039] In the formula, TP means making a correct judgment and the judgment is correct, FP means making an incorrect judgment but the judgment is incorrect, TN means making an incorrect judgment but the judgment is correct; FN means making an incorrect judgment but the judgment is incorrect.

[0040] The beneficial effects of the present invention are:

[0041] The method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention is different from the prior art in that damage images are learned and predicted through a deep learning neural network, and the initial impact energy can be accurately predicted without complex empirical formulas, thereby improving the efficiency and accuracy of the prediction.

[0042] The method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention is advanced in that the present invention introduces deep learning into the prediction of impact energy, bypassing the empirical formula and directly training the model through a large number of experiments and simulation results, thereby obtaining the optimal model suitable for predicting the initial impact energy of automotive CFRP. By submitting the damage image to the trained model, the initial impact energy of automotive CFRP can be directly, accurately and efficiently predicted. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a technical route flow chart of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0044] Figure 2 This is a schematic diagram of Abaqus software simulation modeling of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0045] FIG. 3( a ) is a density parameter in the inp file of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0046] FIG3( b ) is a speed parameter in the inp file of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0047] FIG4( a ) is a model diagram of the simulation result of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention, with the animation frame number being 0.

[0048] FIG4( b ) is a model diagram of the simulation result of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention, with an animation frame number of 150 frames.

[0049] FIG4( c ) is a model diagram of the simulation result of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention, with an animation frame number of 300 frames.

[0050] Figure 5 This is a curve diagram of the displacement and acceleration of the laminate over time in the simulation results of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0051] FIG6( a ) is a damage image of the FC of the 14th layer of the laminate model of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0052] FIG6( b ) is the FT damage image of the 14th layer of the laminate model of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0053] FIG6( c ) is a damage image of the MC of the 14th layer of the laminate model of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0054] FIG6( d ) is a damage image of the MT of the 14th layer of the laminate model of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0055] FIG6( e ) is a damage image of the FC of the 13th layer of the laminate model of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0056] Figure 6(f) is the FT damage image of the 13th layer of the laminate model of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0057] Figure 6(g) is the MC damage image of the 13th layer of the laminate model of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0058] FIG6( h ) is a damage image of the MT layer 13 of the laminate model of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0059] Figure 7 Schematic diagram of partial data set preparation for the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0060] Figure 8 This is the original damage image obtained from the Abaqus software simulation result of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0061] Fig. 9 This is a damaged image with edge noise removed after shear preprocessing in the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0062] FIG10( a ) is a randomly rotated damage image of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0063] FIG10( b ) is a vertically inverted damage image of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0064] FIG11( a ) is a graph showing the FC model result of ResNet50 training of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0065] FIG11( b ) is a graph showing the FT model result of ResNet50 training of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0066] FIG11( c ) is a graph showing the MC model result of ResNet50 training of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0067] FIG11( d ) is a graph showing the MT model result of ResNet50 training of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0068] FIG. 12( a ) is a diagram showing the FC model result of VAN training of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0069] FIG12( b ) is a diagram showing the FT model result of VAN training of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0070] FIG12( c ) is a diagram showing the MC model result of VAN training of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0071] FIG12( d ) is a graph showing the MT model result of VAN training of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0072] FIG. 13( a ) is a graph showing the FC model results of ViT training of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0073] FIG13( b ) is a diagram showing the FT model results of ViT training for the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0074] FIG13( c ) is a diagram showing the MC model result of ViT training of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0075] FIG13( d ) is a graph showing the MT model result of ViT training of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0076] Fig.14 This is the original MT damage image of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0077] FIG15( a ) is a thermal diagram of the MT model trained with ResNet50 of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0078] FIG15( b ) is a thermal diagram of the MT model trained by VAN of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0079] FIG15( c ) is a thermal diagram of the MT model trained by ViT of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0080] FIG16( a ) is a graph showing the result of the MT model trained using SGD using ViT of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0081] FIG16( b ) is a graph showing the result of the MT model trained by Adam using ViT of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0082] FIG16( c ) is a graph showing the MT model result of ViT using Adagrad training of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0083] FIG. 17( a ) is a 4J impact energy prediction diagram of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0084] FIG17( b ) is a 6J impact energy prediction diagram of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0085] FIG17( c ) is a 9J impact energy prediction diagram of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0086] Figure 17(d) is a 13J impact energy prediction diagram of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention.

[0087] Figure 17(e) is a 18J impact energy prediction diagram of the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention. DETAILED DESCRIPTION

[0088] The present invention is further described in detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0089] like Figure 1-1 As shown in FIG. 7 , the present invention provides a method for predicting the impact energy of automotive CFRP based on deep learning and damage images, which generates a corresponding model for the damage image based on the deep learning method to predict the initial impact energy of the vehicle, thereby improving the prediction efficiency and accuracy, including:

[0090] Step 1: Establish a simulation model of a CFRP laminate and an impact hammer head for a vehicle, and perform a drop hammer impact simulation;

[0091] Step 2: Obtain fiber tensile damage images, fiber compression damage images, body tensile damage images, body compression damage images, and layered damage images from the simulation results, and integrate and process the damage images obtained by simulation into corresponding data sets according to different impact energies and different types of damage images;

[0092] Step 3: Pass different categories of damage image data sets into multiple deep neural networks and train them based on multiple optimization methods; compare them through accuracy, loss value and visual heat map, and analyze the damage image category with the highest correlation with the initial impact energy and the optimal model.

[0093] Step 1: Use Abaqus software to establish the simulation model of automotive CFRP laminate and impact hammer:

[0094] Establish the laminate and impact hammer model. The CFRP used in the car is composed of laminates, and the arrangement of every two layers is the same, that is, the arrangement of odd-numbered layers is the same, and the arrangement of even-numbered layers is the same. The impact hammer is in the shape of a bullet, which can cause more serious internal damage to CFRP under the same impact energy conditions, and is used to simulate more extreme situations in real situations, such as Figure 2 shown.

[0095] Use Abaqus software to simulate the drop hammer impact of the automotive CFRP laminate model:

[0096] (1) Setting different initial working conditions for the CFRP laminate and the impact hammer:

[0097] In the relationship between the different loss images of each layer of the automotive CFRP laminate and the initial impact energy of the drop hammer, a large number of simulation models with different initial impact energies of the drop hammer need to be prepared. The Abaqus simulation software solver supports the direct submission of inp files for calculation, so it is necessary to generate the corresponding inp file after the CFRP laminate impact model is first established, and the initial parameters therein can be directly modified to change the required different initial impact energy working condition models to achieve the damage images of different simulation working conditions. The initial energy of the drop hammer can be obtained through the kinetic energy theorem:

[0098]

[0099] Where W 外 is the work done by the resultant force on the object, v 初 is the initial velocity of the object, v 末 is the final velocity of the object, m is the total mass of the object;

[0100] The initial velocity of the falling hammer is 0, that is, it follows the law of free fall, so when it just hits the CFRP laminate, the energy it has is It is only determined by the mass of the falling hammer and the speed when it impacts the laminate. Both parameters can be changed by modifying the inp file.

[0101] In the inp file, the drop weight mass is modified indirectly by modifying the density parameter of the drop weight:

[0102] m=ρV

[0103] Where m is the mass of the object, ρ is the density of the object, and V is the volume of the object.

[0104] As shown in FIG3(a), the density parameter of the impact hammer can be modified at *Density in the inp file, thereby modifying the mass parameter of the impact hammer.

[0105] In the velocity parameters, the velocities in the x and y directions are 0, and there is velocity only in the z direction. The velocity value of this model is 1986.66 mm / s. The negative sign indicates the direction, which is vertically downward, as shown in Figure 3(b). This parameter can be directly modified in the inp file.

[0106] (2) Obtain simulation results:

[0107] After submitting the parameter-set inp file to the Abaqus solver, the computer performs calculation simulation and the result file is saved in odb format. Use the Abaqus visualization panel to open the odb file after the solution is completed to obtain a simulation animation. The animation process mainly shows that the hammer contacts the automotive CFRP laminate, and then gradually squeezes and impacts the laminate to cause it to deform and the contact part moves downward. When the impact energy of the hammer gradually drops to 0, due to the inertia of the laminate, a reverse upward force will be applied to the hammer. At this time, the hammer will obtain an acceleration in the vertical upward direction and move upward with the laminate until the acceleration of the laminate becomes 0, but the laminate cannot be completely restored to its initial state because there is obvious damage inside after being impacted by the hammer.

[0108] Take three different frame numbers of effect pictures respectively, such as Figure 4(a) , 4(b) and 4(c). It can be seen from the three images that the drop hammer is close to the CFRP laminate at the initial moment. As time goes by, the drop hammer gradually squeezes the laminate, causing the latter to deform and absorb energy. Finally, the laminate has basically returned to its original state and the drop hammer has also detached from the laminate, indicating that the laminate has successfully absorbed the impact energy brought by the drop hammer. By analyzing the damage and deformation of each layer of the laminate, it can be known that the compressive strength of the laminate is relatively strong. The CFRP laminate is only impacted by the hammer in the vertical direction, so it is only necessary to pay attention to the result parameters of the laminate and the hammer in the vertical direction. Therefore, the simulation test data is used to draw the part of the laminate impacted by the drop hammer. Figure 5The curves of displacement and acceleration changing with time are shown in . It can be seen from the curves that the whole process of the laminate being impacted by the drop hammer was tested, and the simulation did not end when the hammer was still impacting the laminate. The displacement of the laminate shifted upward compared to the initial situation, indicating that damage occurred inside the laminate and could not be restored to its original state. Since the impact time is very small, this simulation method can be used to observe the entire process of the hammer impacting the laminate in more detail. At the same time, the simulation results can also observe the condition of each layer, and intuitively see various types of damage images, providing a basis for subsequent data analysis. The displacement-time and acceleration-time curves of the laminate obtained from the simulation results show that the simulation data is valid, and its damage image can be used as a data set for image recognition prediction.

[0109] Step 2: Obtain fiber tensile damage image, fiber compression damage image, body tensile damage image, body compression damage image and layered damage image from the simulation results, including:

[0110] From a physical perspective, the FC (fiber compression damage image), FT (fiber tensile damage image), MC (body compression damage image), and MT (body tensile damage image) of the laminate are highly correlated with the initial energy before the collision. Through the Abaqus software visualization processing module, open the odb file obtained by the simulation, display the damage image at the last moment, and then hide the hammer head model, only displaying the simulation result image of the automotive CFRP laminate. In the modeling process, the CFRP laminate is multi-layered, and the arrangement between each two layers of the laminate is different, so it is necessary to select different damage images of all layers plus the layered damage image. Here, select some damage images such as Figures 6(a) to 6(h) As shown, display is performed.

[0111] According to different impact energies and different types of damage images, the simulated damage images are integrated and processed into corresponding data sets:

[0112] In the analysis results obtained in the previous step, the damage images with different impact energies are stored in different folders to create the data set required for deep learning training, such as Figure 7 As shown in the figure, part of the created data set is shown. Before the data set is passed into the deep learning neural network, it is necessary to preprocess the data set to prevent overfitting of the model. Figure 8 As shown in , in the original damage image, the location where internal damage occurs is only in the middle part, and the other locations near the boundary are not impacted by the falling hammer and no damage occurs. This part of the information points is redundant in the neural network, as shown in Fig. 9As shown, the noise at the boundary can be removed by cropping the original image. In addition, in order to enhance the robustness of the neural network, the data needs to be expanded and transformed. For the same damaged image, the result should be rotation and inversion invariant, so the image can be randomly inverted or rotated, such as Figure 10(a) and 10(b) This is the image obtained after data expansion. The Python script can be used to efficiently save the different damage images of each layer obtained from the Abaqus simulation results, which are used to construct damage image data sets of FC, FT, MC, and MT under different impact energies. After data enhancement, it is divided into training set, validation set, and test set in a ratio of 7:1.5:1.5. The training set and validation set are used for model iteration and parameter update during the training process, while the test set does not participate in the training process. It is a tool for the final evaluation of model quality.

[0113] Step 3: Pass different types of damage image data sets into multiple deep learning neural networks, and then train them using multiple optimization methods. Compare and analyze the damage image categories with the highest correlation with the initial impact energy and the optimal model method through accuracy, loss value, and visualized heat map:

[0114] (1) In the present invention, as a preferred embodiment, three different deep learning neural networks ResNet, VAN, and ViT are selected to iteratively train the data set, and the model must be configured first.

[0115] In the present invention, as a preference, all neural network models used are implemented based on pytorch version 2.1.2, and the cuda version used is 12.4; the present invention uses a classification cross entropy loss function that is mostly used for classification tasks, and the training learning rate is 0.001. The loss function image data set obtained by the drop hammer low-speed impact simulation experiment is used as a training set, a validation set, and a test set in a ratio of 7:1.5:1.5, wherein the test set data consists of damage images that are not included in the training process, ensuring a strict evaluation of the model's generalization ability to unseen damage image samples.

[0116] Then the model is trained with the damage image as input and the initial impact energy as output, that is, the image recognition and classification functions are realized. The small batch learning method is used in the training process. The model training is terminated when the loss value decreases to a certain value, the accuracy rate tends to be stable and no longer increases, or the model cannot converge. The value of the hyperparameter is the epoch parameter data with the highest accuracy during the training iteration. The training and testing process is repeated 3 times on each model. Each repetition will randomly re-assign the training set, validation set, and test set, and take the best training set of models.

[0117] Each neural network model mainly uses Accuracy to calculate the accuracy, so the Accuracy(Top[1,2]) method is used for calculation, where Top1 refers to the accuracy of the predicted first-ranked category and the actual result, and Top2 refers to the accuracy of the predicted top two categories containing the actual result.

[0118]

[0119] In the formula, TP (True Positive) means making a positive judgment, and the judgment is correct;

[0120] FP (False Positive) is to make a positive (correct) judgment, but the judgment is wrong;

[0121] TN (True Negative) is a Negative (wrong) judgment, but the judgment is correct;

[0122] FN (False Negative) is a Negative (wrong) judgment, but the judgment is wrong.

[0123] When evaluating a good model, its optimal accuracy will be above 90%, and there will be no gradient explosion or gradient disappearance. The loss value will continue to decrease with the iteration process. When it is close to 0, the hyperparameters in the model will hardly change, the model tends to be stable, and the accuracy will not change significantly. At this time, there is no need to train the model anymore. Therefore, during training, the number of iterations may stop before reaching the set value of epoch. This does not affect the data and results of the experiment, but improves the training efficiency and saves time costs.

[0124] (2) In the present invention, as a preferred embodiment, three models, ResNet50, VAN, and ViT, are used to iterate the training of four different damaged images, and 12 different training methods are obtained. Then, the data set is redistributed into training set, validation set, and test set, and these 12 training methods are repeated three times, and the prediction results are averaged.

[0125] The training results of the ResNet50 model for 4 different damage images are as follows Figure 11(a) , 11(b) , 11(c) and 11(d). From the training effect of ResNet50 on FT damaged images in Figure 11(c), although the final accuracy rate reaches 100%, there are still cases of gradient vanishing and non-convergence, so the result of 100% accuracy is not credible.

[0126] The training results of the VAN model on four different damaged images are as follows: Figure 12(a) , 12(b) , 12(c) and 12(d).

[0127] The training results of the ViT model on four different damaged images are as follows: Figure 13(a) , 13(b) , 13(c) and 13(d).

[0128] By comparing different damage images under the same network, it can be obtained that the priority of the damage image most suitable for predicting the initial impact energy is: MT>MC>FT≈FC.

[0129] Select as Fig.14 The MT damage images shown in the figure are used as the dataset and Figure 15(a) , 15(b) By comparing the visualized heat maps obtained with the three different models in 15(c), we can get the model prediction effect: VIT>VAN>ResNet50.

[0130] (3) The VIT model and MT damage images were selected as the deep learning neural network and data set, respectively, and then the three different optimization methods were compared. The three mainstream optimization methods SGD, Adam, and Adagrad were added to VIT for training, repeated 3 times, and the training results were averaged. Figure 16(a) , 16(b) Comparing the Accuracy obtained from the three training results shown in 16(c), the effects of the three optimization methods are: Adagrad>SGD>Adam.

[0131] Step 6: Use the damage image test set with the highest correlation with the initial impact energy and the optimal prediction model method to perform effect testing:

[0132] The MT test set of the damage image category with the highest correlation with the initial impact energy obtained by comparison is used to test the optimal solutions evaluated: VIT neural network model and Adagrad optimization method, such as Figure 17(a) , 17(b), 17(c), 17(d) and 17(e) respectively show the prediction effect of the model for different impact energies. In order to detect the prediction effect of this model on damage images other than the damage image dataset used for training, this method re-simulates by modifying the inp file to obtain a large number of damage image datasets for testing, and processes part of the image data through code, such as Gaussian blur, random rotation, random flipping, changing the image size ratio, etc., to obtain multiple MT damage images belonging to different initial impact energies. At the same time, when testing the effect of the model, only the size ratio of the image is adjusted to reduce the number of tensors in the model calculation and save calculation time. The damage image used for testing is submitted to the optimal solution for detection in the trained model, and it can successfully identify and predict the damage image dataset after data enhancement, proving that the prediction method of the present invention is successful and feasible.

[0133] Example:

[0134] The present invention then introduces the process of using the method for predicting the impact energy of automotive CFRP based on deep learning and damage images described in the present invention with examples.

[0135] 1. Use Abaqus software to establish the simulation model of automotive CFRP laminate and impact hammer

[0136] (1) Establishing the laminate model

[0137] Abaqus software was used to establish a simulation test model of the laminate. The automotive CFRP is composed of 14 layers of laminates, and the arrangement of every two layers is consistent, that is, the odd-numbered layers are arranged in the same direction, and the even-numbered layers are arranged in the same direction.

[0138] (2) Establishing the impact hammer model

[0139] Abaqus software was used to establish a simulation model of the impact hammer head. The shape of the impact hammer head is bullet-shaped, which can cause more serious internal damage to CFRP under the same impact energy conditions, and is used to simulate more extreme situations in real situations.

[0140] 2. Use Abaqus software to simulate the drop hammer impact of the automotive CFRP laminate model

[0141] (1) Setting different initial working conditions for CFRP laminate and impact hammer

[0142] In the relationship between the different loss images of each layer of the automotive CFRP laminate and the initial impact energy of the drop hammer, a large number of simulation models with different initial impact energies of the drop hammer need to be prepared. The Abaqus simulation software solver supports the direct submission of inp files for calculation, so it is necessary to generate the corresponding inp file after the CFRP laminate impact model is first established, and the initial parameters therein can be directly modified to change the required different initial impact energy condition models to achieve the damage images of different simulation conditions. The initial energy of the drop hammer can be obtained through the kinetic energy theorem:

[0143]

[0144] Where W 外 is the work done by the resultant force on the object, v 初 is the initial velocity of the object, v 末 is the final velocity of the object, m is the total mass of the object;

[0145] The initial velocity of the falling hammer is 0, that is, it follows the law of free fall, so when it just hits the CFRP laminate, the energy it has is It is only determined by the mass of the falling hammer and the speed at which it impacts the laminate, both of which can be changed by modifying the inp file.

[0146] In the inp file, the drop weight mass is modified indirectly by modifying the density parameter of the drop weight:

[0147] m=pV

[0148] Where m is the mass of the object, ρ is the density of the object, and V is the volume of the object.

[0149] As shown in FIG3(a), the density parameter of the impact hammer can be modified at *Density in the inp file, thereby modifying the mass parameter of the impact hammer. The density parameter of 1.24674e-06 corresponds to a drop hammer mass of 5.097 kg.

[0150] In the velocity parameters, the velocities in the x and y directions are 0, and there is velocity only in the z direction. The velocity value of this model is 1986.66 mm / s. The negative sign indicates the direction, which is vertically downward, as shown in Figure 3(b). This parameter can be directly modified in the inp file.

[0151] The initial impact energy of CFRP can be changed by modifying the density and speed parameters of the impact hammer in the inp file. The present invention conducts simulation experiments with initial impact energies of 4J, 6J, 9J, 13J and 18J, and uses these five initial impact energies as classifications for prediction.

[0152] (2) Obtaining simulation results

[0153] After submitting the parameter-set inp file to the Abaqus solver, the computer performs calculation simulation, and the result file is saved in odb format. The Abaqus visualization panel is used to open the odb file after the solution is completed, and a 300-frame simulation animation is obtained. The animation process mainly shows that the hammer head contacts the automotive CFRP laminate, and then gradually squeezes and impacts the laminate to cause it to deform, and the contact part moves downward. When the impact energy of the hammer head gradually drops to 0, due to the inertia of the laminate, a reverse upward force will be applied to the hammer head. At this time, the hammer head will obtain an acceleration in the vertical upward direction and move upward with the laminate until the acceleration of the laminate becomes 0, but the laminate cannot be completely restored to its initial state because there is obvious damage inside after being impacted by the hammer head.

[0154] Take the effect pictures of the 0th frame, the 150th frame and the 300th frame respectively, as follows Figure 4(a) , 4(b) As shown in Figure 4(c), the three images show that the drop hammer is close to the CFRP laminate at the initial moment. As time goes by, the drop hammer gradually presses on the laminate, causing the latter to deform and absorb energy. Finally, the laminate has basically returned to its original state and the drop hammer has also detached from the laminate, indicating that the laminate has successfully absorbed the impact energy brought by the drop hammer. By analyzing the damage and deformation of each layer of the laminate, it can be seen that the compressive strength of the laminate is relatively strong. The CFRP laminate is only impacted by the hammer in the vertical direction, so it is only necessary to pay attention to the result parameters of the laminate and the hammer in the vertical direction. Therefore, the simulation test data is used to draw the part of the laminate impacted by the drop hammer. Figure 5 The curves of displacement and acceleration changing with time are shown in the figure. It can be seen from the curves that the whole process of the laminate being impacted by the drop hammer was tested, and there was no situation where the simulation ended while the hammer was still impacting the laminate. The displacement of the laminate is upward compared to the initial situation, indicating that damage has occurred inside the laminate and it cannot be restored to its original state. Since the impact time is very small, this simulation method can be used to observe the entire process of the hammer impacting the laminate in more detail. At the same time, the simulation results can also observe the condition of each layer, and intuitively see various types of damage images, providing a basis for subsequent data analysis. The displacement-time and acceleration-time curves of the laminate obtained from the simulation results show that the simulation data is valid, and its damage image can be used as a data set for image recognition prediction.

[0155] 3. Obtain fiber tensile damage image, fiber compression damage image, body tensile damage image, body compression damage image and layered damage image from the simulation results;

[0156] From a physical perspective, the FC (fiber compression damage image), FT (fiber tensile damage image), MC (body compression damage image), and MT (body tensile damage image) of the laminate are highly correlated with the initial energy before the collision. Through the Abaqus software visualization processing module, open the odb file obtained by the simulation, display the damage image at the last moment, and then hide the hammer head model, and only display the simulation result image of the automotive CFRP laminate. In the modeling process, the CFRP laminate has 14 layers, and the arrangement between each two layers of the laminate is different, so it is necessary to select all 14 layers of different damage images plus 13 layered damage images, and a total of 14×4+13=69 damage images can be obtained. Some damage images are selected here, such as Figures 6(a) to 6(h) As shown in the figure, the same damage type presents completely different images in the 14th layer and the 13th layer because the CFRP simulation model established by the present invention has different arrangements for every two layers, the arrangement of odd-numbered layers is consistent, and the arrangement of even-numbered layers is consistent. In addition, in the present invention, the 14th layer is the layer closest to the impact hammer head, that is, the top layer, and the 13th layer is the layer below the 14th layer, and so on.

[0157] 4. According to different impact energies and different types of damage images, the simulated damage images are integrated and processed into corresponding data sets

[0158] In the analysis results obtained in the previous step, the damage images with different impact energies are stored in different folders to create the data set required for deep learning training, such as Figure 7 As shown, part of the created data set is displayed. Since the data source in the present invention is a simulation test, it takes 5 to 6 hours to conduct a drop hammer impact simulation experiment in the Abaqus software. If only the 300th frame image of each simulation result is used as the data set, a large number of simulation experiments need to be solved, which will increase a lot of time cost. Therefore, in the present invention, as a preferred embodiment, a data expansion method is adopted, that is, in the images of the 200th to 300th frames of a simulation model, a set of damage image data sets is taken out every 10 frames, so that 11×69=759 image data can be obtained in a simulation experiment. The present invention carried out a total of 50 simulations, and the amount of data obtained is completely sufficient for deep learning training. Before the data set is passed into the deep learning neural network, the data set needs to be preprocessed to prevent overfitting of the model. As Figure 8 As shown in the figure, the resolution of the original damage image obtained by the Abaqus software simulation result is 1318 pixels × 547 pixels. In the original damage image, only the middle part of the image has internal damage, and the other parts near the boundary are not hit by the falling hammer and no damage occurs. In addition, there is also information such as legend on the left side of the image. This part of the information points is redundant in the neural network, such as Fig. 9 As shown in , the noise at the boundary can be removed by cropping the original image to 0.4 times, that is, 312 pixels × 208 pixels. Images of this size can be used for the ResNet50 network, but for the VAN and ViT networks, due to the large dimension of the model's computational tensor, the damaged image used for training can be reduced by 1 / 2, that is, 156 pixels × 104 pixels. In addition, in order to enhance the robustness of the neural network, the data needs to be expanded and transformed. For the same damaged image, the result should be rotation and inversion invariant, so the image can be randomly inverted or rotated, such as Figure 10(a) and 10(b) The image shown in the figure is obtained after data expansion. The Python script can be used to efficiently save the different damage images of each layer obtained from the Abaqus simulation results, which are used to construct the damage image data sets of FC, FT, MC, and MT under different impact energies. After data enhancement, it is divided into training set, validation set, and test set in a ratio of 7:1.5:1.5. The training set and validation set are used for model iteration and parameter update during the training process, while the test set does not participate in the training process. It is a tool for the final evaluation of model quality.

[0159] 5. Different types of damage image data sets are passed into various deep learning neural networks, and then trained using various optimization methods. The damage image categories with the highest correlation with the initial impact energy and the optimal model method are compared and analyzed through accuracy, loss value, and visualized heat map:

[0160] (1) Neural network configuration

[0161] Three different deep learning neural networks, ResNet, VAN, and ViT, are used to iteratively train the data set. The model must be configured first. In the present invention, as a preferred embodiment, all neural network models used are implemented based on the 2.1.2 version of pytorch, and the cuda version used is 12.4.

[0162] In the present invention, as a preferred embodiment, the classification cross entropy loss function of the classification task is used, and the training learning rate is 0.001. The loss function image data set obtained by the drop hammer low-speed impact simulation experiment is used as a training set according to 70%, a validation set according to 15%, and a test set according to 15%, wherein the data of the test set is composed of damage images not included in the training process, ensuring a strict evaluation of the model's generalization ability to unseen damage image samples.

[0163] Then the model is trained with the damage image as input and the initial impact energy as output. The initial impact energies are 4J, 6J, 9J, 13J and 18J, respectively, to achieve the image recognition and classification function. The small batch learning method is used in the training process, and the batch size of the training sample data is set to 32. Each round of training is up to 150 epochs. The model training is terminated when the loss value decreases to a certain value, the accuracy tends to be stable and no longer increases, or the model cannot converge. The value of the hyperparameter is the epoch parameter data with the highest accuracy during the training iteration. The training and testing process is repeated 3 times on each model. Each repetition will randomly reassign the training set, validation set and test set, and take the set of models with the best training effect.

[0164] Each neural network model mainly uses Accuracy to calculate the accuracy, so the Accuracy(Top[1,2]) method is used for calculation, where Top1 refers to the accuracy of the predicted first-ranked category and the actual result, and Top2 refers to the accuracy of the predicted top two categories containing the actual result.

[0165]

[0166] In the formula, TP (True Positive) means making a positive judgment, and the judgment is correct;

[0167] FP (False Positive) is to make a positive (correct) judgment, but the judgment is wrong;

[0168] TN (True Negative) is a Negative (wrong) judgment, but the judgment is correct;

[0169] FN (False Negative) is a Negative (wrong) judgment, but the judgment is wrong.

[0170] When evaluating a good model, its optimal accuracy will be above 90%, and there will be no gradient explosion or gradient disappearance. The loss value will continue to decrease with the iteration process. When it is close to 0, the hyperparameters in the model will hardly change, the model tends to be stable, and the accuracy will not change significantly. At this time, there is no need to train the model anymore. Therefore, during training, the number of iterations may stop before reaching the set value of epoch. This does not affect the data and results of the experiment, but improves the training efficiency and saves time costs.

[0171] (2) Comparison of different types of damage image training sets and different neural network models

[0172] In the present invention, as a preferred embodiment, three different deep learning neural networks ResNet, VAN, and ViT are used to iteratively train four different damage image data sets, and 12 different training methods are obtained. The data sets are then redistributed into training sets, validation sets, and test sets, and these 12 trainings are repeated three times, and the prediction results are averaged.

[0173] The training results of the ResNet50 model for 4 different damage images are as follows Figure 11(a) , 11(b) As shown in Figure 11(c) and 11(d), from the training effect of ResNet50 on FT damaged images in Figure 11(c), although the final accuracy reached 100%, there are still cases of gradient vanishing and failure to converge, so the result of 100% accuracy is not credible. The accuracy of the MT model rises smoothly during the iteration process, and the best accuracy reaches 99.6%. The best accuracy of the MC model is 94.6%, and the accuracy of the FC model is 91.8%. In addition, the FC model occasionally fails to converge.

[0174] The training results of the VAN model on four different damaged images are as follows: Figure 12(a) , 12(b) , 12(c) and 12(d), the best accuracy of the MT model is still the highest at 99.6%, followed by the MC model, which has a best accuracy of 95.8%, while the FC and FT models are slightly worse, at 91.3% and 91.5%, respectively.

[0175] The training results of the ViT model on four different damaged images are as follows: Figure 13(a) , 13(b) , 13(c) and 13(d), the FC model is not very effective, with the accuracy less than 70% after 70 iterations, while the FT model on ViT is much better than the previous two networks, the model converges normally, and the best accuracy reaches 91.3%, but the MT model is still the best performer, with the best accuracy of 99.6%, and the best accuracy of the MC model is 92.6%.

[0176] By comparing different damage images under the same network, it can be obtained that the priority of the damage image most suitable for predicting the initial impact energy is: MT>MC>FT≈FC.

[0177] When the mass of the impact hammer is 5.097 kg and the initial impact energy is 6 J, Fig.14 The MT damage images are used as the dataset, such as Figure 15(a) , 15(b)As shown in Figure 15(c), a horizontal comparison of the visualized heat maps obtained by the three different models shows that the model prediction effect is: VIT>VAN>ResNet50.

[0178] (3) Comparison of different optimization methods

[0179] The VIT model and MT damage images were selected as the deep learning neural network and data set respectively, and then three different optimization methods were compared. The three mainstream optimization methods SGD, Adam, and Adagrad were added to ViT for training, and the training was repeated 3 times, and the training results were averaged. Figure 16(a) , 16(b) Comparing the Accuracy obtained from the three training results in 16(c), it can be seen that the Adam optimization method has a gradient explosion on the MT model trained with ViT, and the effect of SGD is not as good as that of Adagrad. Therefore, the effects of the three optimization methods are: Adagrad>SGD>Adam.

[0180] 6. Use the damage image test set with the highest correlation with the initial impact energy and the optimal prediction model method to conduct effect testing

[0181] The best solution evaluated: ViT neural network model and Adagrad optimization method were tested using the damage image category that is most suitable for predicting the initial impact energy: the body compression damage image (MT) test set. Figure 17(a) , 17(b) , 17(c), 17(d) and 17(e) respectively show the prediction effect of the model for different impact energies. In the prediction graph, the first number after Prediction is the classification group, where 4J is the first group, group 0, 18J is the last group, group 4, and so on. The second number represents the possibility of prediction, that is, how likely it is to be a certain group. From the displayed prediction graph, it can be seen that the prediction probability is above 90%. In fact, the comprehensive accuracy of all damage images is 95.3%, which shows that the model can predict the damage image data in most cases, and it can be inferred that the prediction effect of this model is relatively good.

[0182] In order to detect the prediction effect of this model on damage images other than the damage image dataset used for training, this method re-simulates by modifying the inp file to obtain a large number of damage image datasets for testing, and processes part of the image data through code, such as Gaussian blur, random rotation, random flipping, changing the image size ratio, etc., to obtain multiple MT damage images with different initial impact energies. At the same time, when testing the effect of the model, only the size ratio of the image is adjusted to reduce the number of tensors in the model calculation and save calculation time. The damage images used for testing are submitted to the optimal solution for detection in the trained model, and the damage image dataset after data enhancement can be successfully identified and predicted, proving that the prediction method of the present invention is successful and feasible.

[0183] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation modes, and they can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. A method for predicting the impact energy of automotive CFRP based on deep learning and damage images, characterized in that: include: Step 1: Establish a simulation model of a CFRP laminate and an impact hammer head for a vehicle, and perform a drop hammer impact simulation; Step 2: Obtain fiber tensile damage images, fiber compression damage images, matrix tensile damage images, matrix compression damage images and delamination damage images from the simulation results, and integrate the damage images obtained by simulation into corresponding data sets according to different impact energies and different types of damage images; Step 3: Pass different types of damage image data sets into multiple deep neural networks and train them based on multiple optimization methods; compare the accuracy, loss value, and visualized heat map to obtain the damage image category with the highest correlation with the initial impact energy and the optimal model, and determine the initial impact energy of the vehicle, including: Based on the three models of ResNet50, VAN, and ViT, four different damage images were trained iteratively, and 12 different training methods were obtained. Then the data set was redistributed into training set, validation set, and test set, and these 12 trainings were repeated three times, and the prediction results were averaged; Among them, the four different damage images are: fiber compression damage image, fiber tension damage image, matrix compression damage image, and matrix tension damage image.

2. The method for predicting the impact energy of automotive CFRP based on deep learning and damage images according to claim 1 is characterized in that: In the step 1, the laminate model is composed of laminates; Among them, the odd-numbered layers are arranged in the same way, and the even-numbered layers are arranged in the same way; The shape of the impact hammer head is bullet-shaped.

3. The method for predicting the impact energy of automotive CFRP based on deep learning and damage images according to claim 2 is characterized in that: The initial energy of the falling hammer is obtained by the kinetic energy theorem: Where W 外 is the work done by the resultant force on the object, v 初 is the initial velocity of the object, v 末 is the final velocity of the object, m is the total mass of the object; The drop weight mass is indirectly modified by modifying the density parameter of the drop weight. The formula is: m=ρV; Where m is the mass of the object, ρ is the density of the object, and V is the volume of the object.

4. The method for predicting the impact energy of automotive CFRP based on deep learning and damage images according to claim 3 is characterized in that: The second step comprises: Store damage images with different impact energies into different folders to create the data set required for deep learning training; Before the data set is passed into the deep learning network, it needs to be preprocessed, including: Crop the original image, expand and transform the data, and randomly reverse or rotate the image; The processed data set is used to obtain damage image data sets under different impact energies; After data enhancement, it is divided into training set, validation set and test set in the ratio of 7:1.5:1.

5.

5. The method for predicting the impact energy of automotive CFRP based on deep learning and damage images according to claim 4 is characterized in that: The step three comprises: Iteratively train the data set based on different deep learning networks; Based on the classification cross entropy loss function of the classification task, the training learning rate is 0.001; The loss function image dataset obtained through the drop hammer low-speed impact simulation experiment was used as the training set, validation set, and test set in a ratio of 7:1.5:1.5; The training is performed with the damage image as input and the initial impact energy as output; During the training process, a small batch learning method is used. Model training is terminated when the loss value tends to be stable and no longer decreases, the accuracy tends to be stable and no longer increases, or the model fails to converge. During the training iteration, the value of the hyperparameter is the epoch parameter data with the highest accuracy during the training iteration. The training and testing process was repeated three times on each model. The training set, validation set, and test set were randomly redistributed each time, and the best trained model was selected.

6. The method for predicting the impact energy of automotive CFRP based on deep learning and damage images according to claim 5, characterized in that: Each neural network model uses Accuracy to calculate the accuracy, and the calculation formula is: In the formula, TP means making a correct judgment and the judgment is correct, FP means making an incorrect judgment but the judgment is wrong, and TN means making an incorrect judgment but the judgment is correct; FN is to make a wrong judgment, but the judgment is wrong.

Images