Automobile collision passenger injury curve prediction method based on self-attention enhancement
Through the 1D-CNN model with self-attention enhancement, the problem of long and low accuracy of automobile collision safety testing in the existing technology is solved, efficient and accurate prediction of occupant injury curves is achieved, and the efficiency of automobile research and development is improved.
Patent Information
- Application Number
- CN202510486382.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-17
- Publication Date
- 2025-08-01
AI Technical Summary
The existing car collision safety testing methods are costly, time-consuming and rely on engineer experience, and cannot efficiently utilize historical data, resulting in inaccurate prediction of occupant injury curves and difficult to meet the needs of efficient R&D.
Using a 1D-CNN model based on self-attention enhancement, the training set is constructed and a sliding self-attention decoding layer and variational dropout are set in the network structure to improve the key feature capture ability and model generalization ability, and achieve high-precision occupant injury curve prediction.
It achieves fast and accurate occupant injury curve prediction, shortens the R&D cycle, improves the efficiency of automobile development, and has an ISO score of 0.9436. The training time is only 82 seconds and the prediction time is less than 1 second, which is more than 10% higher than the existing methods.
Smart Images
Figure CN120409222A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a technology in the field of automotive crash safety, specifically a method for predicting the occupant injury curve in vehicle collisions based on a self-attention enhanced one-dimensional convolutional neural network (1D-CNN). Background Art
[0002] Existing crash safety tests usually obtain the injury curves of various parts of the occupants in vehicle collisions based on experiments or finite element simulations. With the increasingly fierce competition in the automotive industry, the market's requirements for the R & D efficiency of automobiles are constantly increasing. In this context, the deficiencies of high cost, long time consumption, high dependence on engineers' experience, and low utilization of the data accumulated in the historical R & D process in conducting experiments and finite element simulations are continuously magnified, and it gradually becomes difficult to meet the requirements of high-efficiency R & D. With the development of artificial intelligence, deep learning technologies represented by neural networks can build and train prediction models based on historical data, and then accurately and quickly obtain the occupant injury curves. Therefore, it can be used as an important tool to improve the R & D efficiency of crash safety. Summary of the Invention
[0003] Aiming at multiple deficiencies of the prior art, such as the inability to focus on key features and the prediction results not being able to reflect the entire collision process, the present invention proposes a method for predicting the occupant injury curve in vehicle collisions based on self-attention enhancement. Driven by historical collision R & D data, deep learning technology is used to predict the injury curve of the occupant during vehicle collisions with high efficiency and high precision, shorten the vehicle collision safety R & D cycle, and improve the vehicle development efficiency.
[0004] The present invention is realized through the following technical solutions:
[0005] The present invention relates to a method for predicting the occupant injury curve in vehicle collisions based on self-attention enhancement. After preprocessing historical crash safety data to construct a curve prediction data set, it is used to train and construct the self-attention enhanced 1D-CNN. Then, real-time injury curve prediction is performed through the trained occupant injury curve prediction model of the self-attention enhanced 1D-CNN.
[0006] The historical crash safety data includes: the design variable values of various parts of the vehicle occupant protection system such as airbags, seat belts, and steering columns in different vehicle safety design schemes accumulated in the historical R & D process, as well as the relative positions and posture parameters of the occupant dummies, and the occupant injury curves corresponding to each design scheme obtained through crash tests or finite element simulations.
[0007] The occupant injury curve includes, but is not limited to, the chest compression deformation curve.
[0008] The preprocessing mentioned above refers to: performing maximum-minimum normalization on various automotive design parameters (and data inputs), scaling the values of each column of parameters to the interval [0, 1] and eliminating the dimension. Specifically: Among them, represents the value of the i-th row and d-th column of the parameter matrix, represents the result after normalizing , and represent the maximum and minimum values in the d-th column of the parameter matrix respectively.
[0009] The curve prediction data set mentioned above is preferably divided into a training set and a test set in a ratio of 4:1.
[0010] The 1D-CNN mentioned above, for one-dimensional data, slides the convolutional kernel along a specific direction to extract data features, including an input layer, a convolutional layer, a pooling layer, and a fully connected layer.
[0011] The self-attention reinforcement mentioned above refers to: aiming at the problem that in the process of model training and prediction of 1D-CNN, the attention degrees to all parts of the prediction output are the same, and it is impossible to focus on the key learning parts, resulting in difficulty in capturing key features. A sliding self-attention output layer is set in the network structure. Through further self-attention decoding and reconstruction of the preliminary output of the network, different attention weights are applied to each part of the output curve, fully evaluating the importance of each part of the curve and focusing on learning the characteristics of the key parts, thereby improving the prediction accuracy. Among them: The sliding self-attention mechanism is an improved method of self-attention, aiming to strengthen the focus on key local parts by restricting the attention range of each window, improving the local feature capture ability, and reducing the computational complexity. Specifically, for a curve output sequence, each self-attention window only focuses on a specific point and other data elements within a fixed window size around it.
[0012] For example, when the window size is l, the i-th self-attention window only calculates the self-attention weights of the data elements in the range [i - l, i + l] adjacent to the data element i. Further, the calculation of the sliding self-attention weights means that by taking all the data elements in the window itself as keys, values, and queries at the same time, the self-attention weights of the data points corresponding to this window are calculated based on the softmax function, and then the results of each point are combined into the attention weights of the complete data. Specifically: For a curve sequence of length d, the weight corresponding to the i-th data element, where: Q, K, and V are the query, key, and value respectively. After calculating the weights of all points, the attention weights of the complete curve sequence can be obtained.
[0013] The described model training means: Based on a deep learning framework, an initial 1D-CNN reinforced by self-attention is constructed, and training parameters are set, such as the optimizer, loss function, learning rate, batch size, number of epochs, etc. Collision safety data used as the training set is introduced into the initial network for iterative training. After the training is completed, the best network parameters of the 1D-CNN reinforced by self-attention applicable to injury curve prediction are obtained. During the training process, variational dropout is carried out, and the generalization ability of the model is enhanced by dynamically adjusting the intensity of dropout. Variational dropout regards dropout as an approximation of the Bayesian process, dynamically adjusts the dropout probability of each neuron, enabling the model to adaptively learn, and thus more effectively process long sequences and different prediction tasks.
[0014] The described model testing means: For the 1D-CNN reinforced by self-attention obtained through training, the input parameters of the test set are input into the model to obtain the corresponding predicted curve, and then the prediction effect of the model is evaluated according to the similarity between the predicted curve and the real curve.
[0015] The described injury curve prediction means: Based on the prediction model that has passed the test, new vehicle design parameters are used as the input to quickly predict the corresponding vehicle collision occupant injury curve. Technical effects
[0016] Through the novel 1D-CNN based on sliding self-attention reinforcement and variational dropout, the present invention sets a sliding self-attention decoding and reconstruction layer in the network structure and uses the variational dropout learning method during the training process. The sliding self-attention reinforcement mechanism of the present invention further performs self-attention decoding and reconstruction on the preliminary output of the network, applies different attention weights to each part of the output curve, fully evaluates the importance of each part of the curve and focuses on learning the characteristics of the key parts, thereby improving the prediction accuracy; variational dropout regards dropout as an approximation of the Bayesian process, dynamically adjusts the dropout probability of each neuron to enhance the generalization ability of the model, enables the model to adaptively learn, and thus more effectively process long sequences and better adapt to different prediction tasks. Description of the drawings
[0017] Figure 1 It is a flowchart of the present invention;
[0018] Figure 2 It is a schematic diagram of the simulation model for obtaining collision safety data in the embodiment;
[0019] Figure 3 It is a schematic diagram of the structure of the 1D-CNN reinforced by self-attention in the embodiment;
[0020] Figure 4It is a comparison graph of the predicted curve and the true curve in the embodiment. Detailed implementation manner
[0021] As Figure 1 shown, this embodiment relates to a method for predicting the occupant injury curve in vehicle collisions based on 1D-CNN, including:
[0022] S1. Obtain a vehicle collision dataset based on finite element simulation: The input variables are the design parameters of each component of the occupant restraint system and the relative positions and postures of the dummy, and the output response is the dummy chest injury curve corresponding to the input, that is, the chest compression deformation curve.
[0023] For the finite element simulation mentioned above, use LS-DYNA software to construct a vehicle occupant restraint system simulation model as Figure 2 shown, and carry out 100% frontal collision simulation at a speed of 56 kph. The dummy used is the 50th male dummy in the driver's position.
[0024] Examples of the data input variables are shown in Table 1.
[0025] The dataset contains 180 samples, including 29 input variables and 1 output response.
[0026] Table 1 Examples of data input variables
[0027] S2. Data preprocessing: Perform min-max normalization on the input of the vehicle collision dataset, and then randomly and evenly divide the vehicle collision dataset into a training set and a test set according to a ratio of 4:1.
[0028] S3. Model construction: Based on the Python programming language and the PyTorch framework, construct an initial self-attention enhanced 1D-CNN as Figure 3 shown, including: an input layer, three convolutional layers, three average pooling layers, one sliding self-attention decoding and reconstruction layer, and one fully connected output layer. The activation function used is ReLU, and the convolutional kernel size used is 3.
[0029] S4. Model Training: The 1D-CNN based on self-attention reinforcement is trained using the input parameters and output curves in the training set. Among them, the number of epochs is 1000, the batch size is 16, the learning rate is 0.001, the optimizer is Adam, and the loss function is Mean Square Error (MSE). During the training process, variational dropout is used to dynamically and adaptively discard some neurons to improve the model's ability. The model parameters involved in training include network parameters, neuron weights, and sliding self-attention weights.
[0030] S5. Model Testing: The model is tested using the test set. Specifically, the input parameters in the test set are input into the trained 1D-CNN based on self-attention reinforcement to obtain the corresponding predicted curve. Then, the ISO (International Organization for Standardization) score is used as the evaluation index for curve similarity to calculate the similarity between the predicted chest injury curve and the corresponding real curve in the test set. The ISO score comprehensively evaluates the similarity between the two curves from four aspects: corridor score, phase score, amplitude score, and slope score. It is a value in the range of [0, 1]. The higher the score, the more similar the two curves are. In engineering, an ISO score in the range of (0.80, 0.94] indicates a relatively high similarity between the two curves, and a score greater than 0.94 indicates an extremely high similarity between the two curves. During the training and prediction processes, the time spent is recorded respectively.
[0031] The average ISO score between the chest injury curve predicted by the 1D-CNN based on self-attention reinforcement trained in this embodiment and the corresponding real curve can reach 0.9436, which is greater than 0.94, indicating that the prediction accuracy of the model can reach a very high level. Model Efficiency Evaluation: In this embodiment, the training duration of the model is 82 seconds, and the prediction time is less than 1 second. In contrast, it usually takes more than 26 hours to obtain a set of dummy chest injury curves through finite element simulation.
[0032] Visualization of Prediction Results: To more intuitively display the prediction results of the model, a sample in the test set is randomly selected in this embodiment, and the real curve and the predicted curve are plotted for visualization. The results are as Figure 4 shown. It can be seen that the predicted curve is very close to the real curve.
[0033] Compared with the prior art, it takes more than 26 hours to obtain a set of chest injury curves through finite element simulation in the present method. In this embodiment, the training duration of the model is 82 seconds, and the prediction time is less than 1 second. The ISO score between the chest injury curve predicted by the 1D-CNN based on self-attention reinforcement trained in this embodiment and the corresponding true curve can reach 0.9436, and the accuracy is improved by more than 10% compared with using other machine learning methods.
[0034] The above specific implementation can be locally adjusted in different ways by those skilled in the art without departing from the principles and purposes of the present invention. The protection scope of the present invention is subject to the claims and is not limited by the above specific implementation, and all implementation solutions within its scope are subject to the present invention.
Claims
1. A method for predicting the occupant injury curve in vehicle collisions based on self-attention reinforcement, characterized in that After preprocessing historical crash safety data to construct a curve prediction dataset, it is used to train and construct a 1D-CNN based on self-attention reinforcement. During the online stage, real-time injury curve prediction is performed through the trained 1D-CNN occupant injury curve prediction model based on self-attention reinforcement. The historical crash safety data includes: the design variable values of various parts of the vehicle occupant protection system such as airbags, seat belts, and steering columns in different vehicle safety design schemes accumulated during the historical R & D process, the relative positions and posture parameters of the occupant dummy, and the occupant injury curves corresponding to each design scheme obtained through crash tests or finite element simulations.
2. The method for predicting the occupant injury curve of a vehicle collision based on self-attention reinforcement according to claim 1, wherein The preprocessing mentioned above refers to: performing maximum-minimum normalization on various automotive design parameters (and data inputs), scaling the values of each column of parameters to the interval [0, 1] and eliminating the dimension, specifically: Among them, represents the value of the element in the i-th row and d-th column of the parameter matrix, represents the result after normalizing and represent the maximum and minimum values in the d-th column of the parameter matrix respectively.
3. The method for predicting the occupant injury curve in a vehicle collision based on self-attention reinforcement according to claim 1, characterized in that, The 1D-CNN slides the convolutional kernel along a specific direction for one-dimensional data to extract data features, including an input layer, a convolutional layer, a pooling layer, and a fully connected layer.
4. The method for predicting the occupant injury curve in vehicle collision based on self-attention reinforcement according to claim 1, characterized in that The self-attention reinforcement means that, aiming at the problem that in the model training and prediction process of the 1D-CNN, the attention degrees to each part of the prediction output are the same, making it difficult to focus on learning key parts and resulting in difficulty in capturing key features, a sliding self-attention output layer is set in the network structure. Through further self-attention decoding and reconstruction of the preliminary output of the network, different attention weights are applied to each part of the output curve, fully evaluating the importance of each part of the curve and focusing on learning the characteristics of the key parts, thereby improving the prediction accuracy. Among them, the sliding self-attention mechanism is an improved method of self-attention, aiming to enhance the focus on key local parts by restricting the attention range of each window, improving the local feature capture ability, and reducing the computational complexity. Specifically, for a curve output sequence, each self-attention window only focuses on a specific point and other data elements within a fixed window size around it.
5. The method for predicting the occupant injury curve of a vehicle collision based on self-attention reinforcement according to claim 1, wherein The model training refers to: based on the deep learning framework, constructing an initial 1D-CNN based on self-attention reinforcement, setting training parameters such as optimizer, loss function, learning rate, batch size, number of epochs, etc., introducing the crash safety data as the training set into the initial network for iterative training. After training, the best network parameters of the 1D-CNN based on self-attention reinforcement suitable for injury curve prediction are obtained. During the training process, variational dropout is carried out, and the generalization ability of the model is enhanced by dynamically adjusting the intensity of dropout. Variational dropout regards dropout as an approximation of the Bayesian process, dynamically adjusting the dropout probability of each neuron, enabling the model to adaptively learn, and thus more effectively processing long sequences and different prediction tasks.
6. The method for predicting the occupant injury curve of a vehicle collision based on self-attention reinforcement according to claim 1, wherein The model testing refers to: for the trained 1D-CNN based on self-attention reinforcement, inputting the input parameters of the test set into the model to obtain the corresponding prediction curve, and then evaluating the prediction effect of the model according to the similarity between the prediction curve and the real curve.
7. The method for predicting the occupant injury curve of a vehicle collision based on self-attention reinforcement according to claim 1, characterized in that, The injury curve prediction refers to: based on the prediction model that passes the test, using new vehicle design parameters as input to quickly predict the corresponding vehicle crash occupant injury curve.