Method for predicting dynamic response of vehicle under settlement deformation driven by physical data
By constructing a vehicle-track coupled dynamics model and using a deep learning method that combines multi-scale feature extraction with attention fusion, the problems of low computational efficiency and insufficient prediction accuracy in existing technologies are solved, achieving fast and accurate prediction of vehicle dynamic response, which is applicable to complex settlement deformation scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING JIAOTONG UNIV
- Filing Date
- 2026-02-10
- Publication Date
- 2026-05-26
AI Technical Summary
Existing technologies in rail transit suffer from low computational efficiency in physical modeling, and data-driven methods ignore physical laws, making it impossible to quickly and accurately predict vehicle dynamic response. Furthermore, they cannot adapt to complex settlement and deformation scenarios, resulting in insufficient prediction accuracy and interpretability.
A vehicle-track coupled dynamics model is constructed, embedding the vehicle's buoyancy motion equation as a physical constraint. A deep learning method with multi-scale feature extraction and attention fusion modules is used to generate a vehicle dynamic response prediction model, which is then trained using measured data.
It achieves fast and accurate prediction of vehicle dynamic response, conforms to physical laws, improves the interpretability and prediction accuracy of the model, is applicable to complex settlement deformation scenarios, and meets practical engineering needs.
Smart Images
Figure CN122088083A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of rail transit engineering technology, specifically relating to a physical data-driven method for predicting vehicle dynamic response under settlement deformation. Background Technology
[0002] In recent years, the mileage of rail transit lines has continued to increase. However, during long-term operation, the track foundation is susceptible to uneven settlement caused by factors such as nearby construction, uneven soil conditions, and changes in groundwater levels, which compromises track smoothness. Track smoothness is directly related to the safety and comfort of vehicle operation. Therefore, under given track irregularities (including superimposed settlement deformation), the ability to quickly and accurately predict the vibration response of various vehicle structures has become a key requirement for ensuring the safe and efficient operation of rail transit.
[0003] However, existing technologies have the following shortcomings: physical modeling methods have extremely low computational efficiency, requiring tens of hours to solve the dynamic model of a track of tens of meters, which is too costly for long railways and may even be impossible to solve effectively, making it difficult to meet the needs of rapid prediction in actual engineering; data-driven methods only focus on minimizing data loss and ignore the consistency of physical laws among multiple output variables, which may lead to prediction results that violate the logic of vehicle-track dynamics and have poor interpretability; they are limited to ordinary random irregularity scenarios and lack research on composite irregularity conditions with superimposed settlement deformation, resulting in insufficient robustness and inability to adapt to complex track excitations in actual engineering; they simply combine existing algorithms without designing a core module specifically for identifying track irregularities, resulting in weak feature extraction and limited prediction accuracy.
[0004] Therefore, a new method is urgently needed. Summary of the Invention
[0005] The purpose of this invention is to provide a physical data-driven method for predicting vehicle dynamic response under settlement deformation. This method enables rapid prediction of vehicle dynamic response, closely matches actual working conditions, ensures that the prediction results conform to physical laws, improves model interpretability, and enhances prediction accuracy. At the same time, it can be extended to complex settlement deformation scenarios to meet engineering needs.
[0006] To achieve the above objectives, this invention provides a physical data-driven method for predicting vehicle dynamic response under settlement deformation, comprising the following steps: S1. Construct a vehicle-track coupled dynamic model, generate vertical track irregularities with superimposed uneven settlement, design multi-condition dynamic simulation and run the solution program to obtain the original dynamic dataset; S2. Perform sample splitting, dataset partitioning and normalization preprocessing on the original dynamics dataset in S1 to obtain the preprocessed sample set, including the preprocessed training set and the preprocessed test set. S3. Construct the model, embed the vehicle body's floating motion equation as a physical constraint, set up a multi-scale feature extraction and attention fusion module, and determine the model's total loss function. S4. Train the model in S3 based on the preprocessed sample set in S2, update the parameters iteratively through the optimizer and save the results to obtain the trained model. S5. Input the preprocessed sample set from S2 into the trained model in S4 to complete the inference and restore the response value to the actual physical dimensions.
[0007] Preferably, in S1, the vehicle system of the vehicle-track coupled dynamics model consists of wheelsets, bogies, car bodies, and suspensions, while the track system adopts a rigid structure, and vehicle-track coupling is achieved through wheel-rail interaction; The vertical track irregularity is formed by superimposing the measured value of vertical irregularity of a certain railway as the initial irregularity with the cosine-based uneven settlement. The settlement calculation formula is as follows: ; In the formula, This represents the uneven settlement value of the roadbed. The wavelength for uneven settlement of the roadbed; This represents the amplitude of uneven settlement of the roadbed. The independent variable is the spatial location of the orbit.
[0008] Preferably, in S1, the uneven settlement amplitude is 5mm, 10mm, 15mm, 20mm, or 30mm, the uneven settlement wavelength is 20m, 30m, 35m, 40m, 50m, or 60m, and the train speed is 20m / s, 25m / s, 40m / s, or 50m / s. The simulation conditions were combined to form 5×6×4=120 sets of dynamic simulation conditions, and the input-output corresponding data for each set of conditions were generated. The input was the track irregularity superimposed with the initial irregularity and uneven settlement, as well as the train running speed. The output was seven vehicle dynamic responses, namely front bogie displacement, front bogie speed, rear bogie displacement, rear bogie speed, car body displacement, car body speed, and car body vibration acceleration, resulting in 120 sets of original dynamic datasets.
[0009] Preferably, in S2, the sample splitting involves dividing the 120 sets of original data into track irregularity sequence input-dynamic response output sample pairs, where the track irregularity sequence input is the track irregularity sequence within a 75m range in the real world, and the dynamic response output is the dynamic response of the 7 vehicles corresponding to the midpoint of the 75m track; the dataset partitioning divides all samples into training set and test set in an 8:2 ratio; the normalization compresses the input and output of the training set and test set to the range [0,1].
[0010] Preferably, in S3, the total loss of the model The function is: ; In the formula, The weighting coefficient is set to... ; For data loss, This is a physical loss.
[0011] Preferably, in S3, the formula for calculating the data loss is: ; In the formula, The number of samples; For the index of the sample; For the first The true value corresponding to each sample; For the first The model prediction value corresponding to each sample.
[0012] Preferably, in S3, the formula for calculating the physical loss is: ; In the formula, For vehicle body mass; The vertical suspension damping coefficient of the car body-bogie; The vertical vibration acceleration of the vehicle body; The vertical suspension stiffness coefficient of the car body-bogie; This represents the vertical vibration displacement of the vehicle body; The vertical vibration velocity of the vehicle body; The vertical vibration acceleration of the vehicle body; This represents the vertical vibration displacement of the front bogie. The vertical vibration velocity of the front bogie; The vertical vibration velocity of the rear bogie; This refers to the vertical vibration displacement of the rear bogie.
[0013] Preferably, in S3, the multi-scale feature extraction and attention fusion module extracts features through three parallel Conv1d branches, calculates attention weights by combining global average pooling, shared MLP, and sigmoid activation function, and performs weighted fusion of multi-scale features, specifically as follows: Feature extraction is performed using three parallel Conv1d branches, as shown in the formula: ; In the formula, It is the ReLU activation function; The kernel size; Batch size; This represents the number of output channels for the convolution; The length of the track uneven sequence; It is the space of real numbers; This is a one-dimensional convolution operation; The formula for using Global Average Pooling (GAP) and shared MLP mapping scores is as follows: ; ; In the formula, To share the multilayer perceptron mapping function; This is a global average pooling operation; For the first Feature maps output by multi-scale convolutional branches; For the output of the first Global feature tensor of multiple scale convolutional branches; To indicate A real number space of B×C dimensions; This corresponds to the three convolutional kernel branches of different sizes in the multi-scale feature extraction module; For the first The feature importance score tensor corresponding to each multi-scale convolutional branch; The formula for calculating normalized weights and fusing features is as follows: ; ; In the formula, For the first The final normalized weight coefficients of the multi-scale convolutional branches; For the first The importance score corresponding to each multi-scale convolutional branch; Use the Sigmoid activation function; Sum the Sigmoid scores for the three multi-scale convolution branches; This is a smoothing term.
[0014] Therefore, the present invention employs the above-mentioned physical data-driven method for predicting vehicle dynamic response under settlement deformation. Compared with the prior art, the technical solution of the present invention has the following beneficial effects: (1) This invention uses deep learning technology, thus overcoming the technical problems of low computational efficiency of mechanism models and inability to combine with actual measurement data. In this way, it can directly and quickly output prediction results after training, avoid repeated modeling and high computational costs, and support the technical effect of optimizing models based on actual measurement data and conforming to actual working conditions. (2) The present invention adopts the technical means of embedding the vehicle-track coupled dynamics theory (vehicle body floating motion equation) into the model training and giving it physical constraints, thus overcoming the technical problems of deep learning models ignoring physical laws and having poor interpretability, thereby achieving the technical effect of ensuring that the prediction results conform to the dynamic laws while fitting the data and improving the interpretability of the model. (3) The present invention adopts the technical means of designing a multi-scale feature extraction convolution module and a corresponding attention mechanism, thus overcoming the technical problems of weak feature extraction and insufficient accuracy of deep learning models, thereby achieving the technical effect that the model can automatically extract track irregularity information of different wavelengths and assign corresponding weights, significantly improving the prediction accuracy. (4) The present invention adopts the technical means of establishing a database corresponding to track irregularities and dynamic responses with superimposed settlement deformation. Therefore, it overcomes the technical problem that the existing technology has a narrow application scenario and is only applicable to ordinary track irregularity scenarios. Thus, it achieves the technical effect of expanding the scope of application of the technology to composite track irregularity scenarios with superimposed settlement deformation and adapting to the core needs of actual engineering.
[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0016] Figure 1 This is a flowchart of an embodiment of a physical data-driven method for predicting vehicle dynamic response under settlement deformation according to the present invention; Figure 2 This is an architecture diagram of an embodiment of a physical data-driven method for predicting vehicle dynamic response under settlement deformation according to the present invention. Figure 3 This is a comparison of the loss curves during the training and validation phases of a model training process in an embodiment of a physical data-driven method for predicting vehicle dynamic response under settlement deformation according to the present invention; wherein... Figure 3 In the diagram, (a) represents the changes in physical loss for the three experimental groups during training (solid line) and validation (dashed line) within a large numerical range of 0-1200. Figure 3 (b) in the figure represents the changes in training and validation data loss for the three experimental groups within a small numerical range of 0-1.6; Figure 4 This is a bar chart of evaluation indexes for an embodiment of the physical data-driven method for predicting vehicle dynamic response under settlement deformation according to the present invention. Figure 5 This is a spatiotemporal comparison diagram of the actual and model predicted values of the vertical dynamic response of a vehicle under settlement deformation, according to an embodiment of the physical data-driven vehicle dynamic response prediction method of the present invention; wherein, Figure 5 (a) in the figure represents the curve of vertical displacement of the front bogie as a function of track distance; Figure 5(b) in the figure represents the curve of the vertical displacement of the rear bogie as a function of track distance; Figure 5 (c) in the figure represents the curve of the vertical displacement of the vehicle body as a function of the track distance; Figure 6 This is a spatiotemporal comparison diagram of the actual values and model prediction values of the multi-dimensional vertical dynamic response of a vehicle, based on an embodiment of a physical data-driven method for predicting vehicle dynamic response under settlement deformation according to the present invention; wherein, Figure 6 (a) in the figure represents the curve of the vertical velocity of the front bogie as a function of track distance; Figure 6 (b) in the figure represents the curve of the vertical velocity of the rear bogie as a function of track distance; Figure 6 (c) in the figure represents the curve of the vertical velocity of the vehicle body as a function of track distance; Figure 6 In the figure, (d) represents the curve of the vertical acceleration of the vehicle body as a function of track distance; Figure 7 This is a line graph showing the physical loss prediction results of an embodiment of the physical data-driven vehicle dynamic response prediction method under settlement deformation according to the present invention. Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used in the present invention should have the ordinary meaning understood by those skilled in the art.
[0018] Example 1 like Figures 1-2 As shown, this embodiment provides a physical data-driven method for predicting vehicle dynamic response under settlement deformation. It should be understood that the specific parameters, models and protocols mentioned in this embodiment are merely examples to help those skilled in the art understand the present invention, and are not intended to limit the present invention.
[0019] The present invention provides a physical data-driven method for predicting vehicle dynamic response under settlement deformation, comprising the following steps: S1. Construct a complete vehicle-track coupled dynamics model that includes the vehicle system and the track system, and realize vehicle-track coupling through wheel-track interaction; A vehicle-track coupled dynamics model is constructed; the vehicle system consists of wheelsets, bogies, car body, and suspension, while the track system adopts a rigid structure. Vertical track irregularities are generated. The measured values of vertical irregularities from a certain railway are used as the initial irregularity and superimposed with cosine-based uneven settlement. The settlement calculation formula is as follows: ; In the formula, This represents the uneven settlement value of the roadbed. The wavelength for uneven settlement of the roadbed; This represents the amplitude of uneven settlement of the roadbed. The independent variable of orbital spatial position; And according to the combination of 5 amplitudes and 6 wavelengths in Table 1, an uneven sequence of settlement conditions is generated; Table 1 Amplitude-Wavelength Combination Table
[0020] The dynamic simulation working conditions were designed, and 5×6×4=120 sets of working conditions were formed by combining four train speeds of 20m / s, 25m / s, 40m / s and 50m / s with different settlement conditions. Running the SIMPACK dynamic solver generates one set of input-output data for each working condition. The input consists of track irregularities superimposed with initial irregularities and uneven settlement, as well as the train speed. The output consists of seven vehicle dynamic responses: front bogie displacement, front bogie speed, rear bogie displacement, rear bogie speed, car body displacement, car body speed, and car body vibration acceleration. This yields 120 sets of raw dynamic datasets, each containing a track irregularity sequence (vertical), train speed, and seven vehicle dynamic response parameters. S2 receives the original dynamic dataset from S1; taking the real track unevenness sequence as the model input and the 7 dynamic responses at the corresponding positions as the output, the 120 sets of original data are split into several "input sequence-output response" sample pairs; All samples were divided into training and test sets in an 8:2 ratio; the inputs (track irregularity sequence + velocity) and outputs (7 dynamic responses) of the training and test sets were uniformly normalized to compress the data range to [0,1] and eliminate dimensional differences. Obtain the preprocessed sample set, wherein the preprocessed training set consists of normalized input samples and seven corresponding normalized dynamic response output samples; the preprocessed test set consists of normalized input-output samples in the same format. S3. A PI-MSANet model is constructed using a CNN-LSTM hybrid architecture. The CNN module performs feature engineering on the input track irregularity sequence to extract the spatial local features of the original signal. The LSTM module integrates the spatial features extracted by the CNN in the temporal dimension to capture the temporal dependency between track irregularity excitation and vehicle dynamic response. The equations of motion for vehicle body heave and buoyancy are selected as physical constraints and embedded into the training process of PI-MSANet model. This ensures that the prediction results conform to the dynamic laws of the vehicle-track system while minimizing data loss. The total loss function of the PI-MSANet model consists of two parts: data loss and physical loss. The sum of the mean squared errors (MSEs) of the seven vehicle dynamic response variables predicted by the model is calculated using the following formula: ; In the formula, The number of samples; For the index of the sample; For the first The true value corresponding to each sample; For the first The model prediction value corresponding to each sample; Physical loss The residual loss obtained by substituting each predicted variable (car body mass, damping, stiffness, wheel-rail force, etc.) into the car body heave motion equation is calculated using the following formula: ; In the formula, For vehicle body mass; The vertical suspension damping coefficient of the car body-bogie; The vertical vibration acceleration of the vehicle body; The vertical suspension stiffness coefficient of the car body-bogie; This represents the vertical vibration displacement of the vehicle body; The vertical vibration velocity of the vehicle body; The vertical vibration acceleration of the vehicle body; This represents the vertical vibration displacement of the front bogie. The vertical vibration velocity of the front bogie; The vertical vibration velocity of the rear bogie; This refers to the vertical vibration displacement of the rear bogie; The formula for calculating total loss is: ; In the formula, The weighting coefficient is set to... ; The PI-MSANet model extracts irregular activation features of different lengths through multi-scale convolution and endows it with the ability to learn the importance of these features. Specifically: Feature extraction is performed using three parallel Conv1d branches, as shown in the formula: ; In the formula, It is the ReLU activation function; The kernel size; Batch size; This represents the number of output channels for the convolution; The length of the track uneven sequence; It is the space of real numbers; This is a one-dimensional convolution operation; The formula for using Global Average Pooling (GAP) and shared MLP mapping scores is as follows: ; ; In the formula, To share the multilayer perceptron mapping function; This is a global average pooling operation; For the first Feature maps output by multi-scale convolutional branches; For the output of the first Global feature tensor of multiple scale convolutional branches; To indicate A real number space of B×C dimensions; This corresponds to the three convolutional kernel branches of different sizes in the multi-scale feature extraction module; For the first The feature importance score tensor corresponding to each multi-scale convolutional branch; The formula for calculating normalized weights and fusing features is as follows: ; ; In the formula, For the first The final normalized weight coefficients of the multi-scale convolutional branches; For the first The importance score corresponding to each multi-scale convolutional branch; Use the Sigmoid activation function; Sum the Sigmoid scores for the three multi-scale convolution branches; This is a smoothing term.
[0021] S4. Train the PI-MSANet model in S3 based on the preprocessed sample set in S2; read training set samples in batches, input them into the model, and sequentially complete multi-scale feature extraction, attention fusion, CNN-LSTM inference, and output 7 dynamic response prediction values; calculate the data loss and physical loss for each batch according to the total loss formula to obtain the total loss; calculate the gradient of the model parameters based on the total loss, and iteratively update the parameters through the Adam optimizer; validate with a test set after each round of training, and stop training if the validation loss does not decrease for 10 consecutive rounds, save the optimal model weights; output the trained PI-MSANet model. S5. Input the preprocessed sample set from S2 into the trained PI-MSANet model from S4; input the preprocessed sample set into the trained PI-MSANet model to complete multi-scale feature extraction, attention fusion, CNN-LSTM inference and physical law verification, and restore the response value output by the PI-MSANet model to the actual physical dimensions.
[0022] Ablation experiments were conducted to demonstrate the superiority of this embodiment, and the experimental arrangements are shown in Table 2. The model training loss and model performance under different experiments are as follows: Figures 3-4 As shown, the training and validation losses of the three experimental groups decreased rapidly with the increase of iterations. The loss reduction was significant in the early stage (around generation 0-20) and gradually stabilized thereafter. The training and validation loss curves showed a high degree of fit, indicating a low risk of model overfitting. The loss curves of different experimental groups showed consistent fluctuation trends and similar final convergence values, reflecting the stability and consistency of the training process and demonstrating good overall convergence performance. The addition of physical constraints not only improved the physical loss of the model but also effectively avoided the oscillation of model data loss during training. After adding the multi-scale attention module, the model showed significant improvement in various evaluation metrics, verifying the effectiveness and reliability of each module of the PI-MSANet model. Table 2 Ablation Experiment Design Table
[0023] The model's prediction performance on the complete line is as follows: Figures 5-7 As shown, the PI-MSANet model exhibits good performance in predicting vehicle dynamic response under settlement deformation: The predicted values of the vehicle's vertical displacement, velocity, acceleration, and other multi-dimensional dynamic responses have an extremely high spatiotemporal fit with the actual values. 2 The coefficients all reached above 0.997, and the RMSE and MAE values were extremely small, highlighting the ability of the multi-scale feature extraction and attention fusion module to accurately capture track irregularity information. The predicted results of the multi-dimensional vertical dynamic response also closely match the actual values, further verifying the model's accurate prediction capability for the dynamic response of different vehicle structures. The physical loss of the predicted results is effectively controlled in the settlement area, which intuitively reflects the effectiveness of the vehicle body heave motion equation as a physical constraint, ensuring that the predicted results are both consistent with the measured data and strictly follow the laws of vehicle-track dynamics.
[0024] Therefore, the present invention adopts the above-mentioned physical data-driven method for predicting vehicle dynamic response under settlement deformation. This method enables rapid prediction of vehicle dynamic response, conforms to actual working conditions, ensures that the prediction results conform to physical laws, improves model interpretability, and enhances prediction accuracy. At the same time, it can be extended to complex settlement deformation scenarios to meet engineering needs.
[0025] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0026] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A physical data-driven method for predicting vehicle dynamic response under settlement deformation, characterized in that, Includes the following steps: S1. Construct a vehicle-track coupled dynamic model, generate vertical track irregularities with superimposed uneven settlement, design multi-condition dynamic simulation and run the solution program to obtain the original dynamic dataset; S2. Perform sample splitting, dataset partitioning and normalization preprocessing on the original dynamics dataset in S1 to obtain the preprocessed sample set, including the preprocessed training set and the preprocessed test set. S3. Construct the model, embed the vehicle body's floating motion equation as a physical constraint, set up a multi-scale feature extraction and attention fusion module, and determine the model's total loss function. S4. Train the model in S3 based on the preprocessed sample set in S2, update the parameters iteratively through the optimizer and save the results to obtain the trained model. S5. Input the preprocessed sample set from S2 into the trained model in S4 to complete the inference and restore the response value to the actual physical dimensions.
2. The method for predicting vehicle dynamic response under settlement deformation driven by physical data according to claim 1, characterized in that, In S1, the vehicle system of the vehicle-track coupled dynamics model consists of wheelsets, bogies, car bodies, and suspensions, while the track system adopts a rigid structure, and vehicle-track coupling is achieved through wheel-rail interaction. The vertical track irregularity is formed by superimposing the measured value of vertical irregularity of a certain railway as the initial irregularity with the cosine-based uneven settlement. The settlement calculation formula is as follows: ； In the formula, This represents the uneven settlement value of the roadbed. The wavelength for uneven settlement of the roadbed; This represents the amplitude of uneven settlement of the roadbed. The independent variable is the spatial location of the orbit.
3. The method for predicting vehicle dynamic response under settlement deformation driven by physical data according to claim 1, characterized in that, In S1, the uneven settlement amplitude is 5mm, 10mm, 15mm, 20mm, and 30mm, and the uneven settlement wavelength is 20m, 30m, 35m, 40m, 50m, and 60m; the train speed is 20m / s, 25m / s, 40m / s, and 50m / s. The simulation conditions were combined to form 5×6×4=120 sets of dynamic simulation conditions, and the input-output corresponding data for each set of conditions were generated. The input was the track irregularity superimposed with the initial irregularity and uneven settlement, as well as the train running speed. The output was seven vehicle dynamic responses, namely front bogie displacement, front bogie speed, rear bogie displacement, rear bogie speed, car body displacement, car body speed, and car body vibration acceleration, resulting in 120 sets of original dynamic datasets.
4. The method for predicting vehicle dynamic response under settlement deformation driven by physical data according to claim 1, characterized in that, In S2, the sample splitting involves dividing 120 sets of original data into track irregularity sequence input-dynamic response output sample pairs. The track irregularity sequence input is the track irregularity sequence within a 75m range in the real world, and the dynamic response output is the dynamic response of 7 vehicles corresponding to the midpoint of the 75m track. The dataset partitioning divides all samples into training and test sets in an 8:2 ratio. The normalization compresses the input and output of the training and test sets to the range [0,1].
5. The method for predicting vehicle dynamic response under settlement deformation driven by physical data according to claim 1, characterized in that, In S3, the total loss of the model The function is: ； In the formula, The weighting coefficient is set to... ; For data loss, This is a physical loss.
6. The method for predicting vehicle dynamic response under settlement deformation driven by physical data according to claim 5, characterized in that, In S3, the formula for calculating the data loss is: ； In the formula, The number of samples; For the index of the sample; For the first The true value corresponding to each sample; For the first The model prediction value corresponding to each sample.
7. The method for predicting vehicle dynamic response under settlement deformation driven by physical data according to claim 6, characterized in that, In S3, the formula for calculating the physical loss is: ； In the formula, For vehicle body mass; The vertical suspension damping coefficient of the car body-bogie; The vertical vibration acceleration of the vehicle body; The vertical suspension stiffness coefficient of the car body-bogie; This represents the vertical vibration displacement of the vehicle body; The vertical vibration velocity of the vehicle body; The vertical vibration acceleration of the vehicle body; This represents the vertical vibration displacement of the front bogie. The vertical vibration velocity of the front bogie; The vertical vibration velocity of the rear bogie; This refers to the vertical vibration displacement of the rear bogie.
8. The method for predicting vehicle dynamic response under settlement deformation driven by physical data according to claim 7, characterized in that, In S3, the multi-scale feature extraction and attention fusion module extracts features through three parallel Conv1d branches, calculates attention weights by combining global average pooling, shared MLP, and sigmoid activation function, and performs weighted fusion of multi-scale features, specifically as follows: Feature extraction is performed using three parallel Conv1d branches, as shown in the formula: ； In the formula, It is the ReLU activation function; The kernel size; Batch size; This represents the number of output channels for the convolution; The length of the track uneven sequence; It is the space of real numbers; This is a one-dimensional convolution operation; The formula for using global average pooling and shared MLP mapping scores is as follows: ； ； In the formula, To share the multilayer perceptron mapping function; This is a global average pooling operation; For the first Feature maps output by multi-scale convolutional branches; For the output of the first Global feature tensor of multiple scale convolutional branches; To indicate A real number space of B×C dimensions; This corresponds to the three convolutional kernel branches of different sizes in the multi-scale feature extraction module; For the first The feature importance score tensor corresponding to each multi-scale convolutional branch; The formula for calculating normalized weights and fusing features is as follows: ； ； In the formula, For the first The final normalized weight coefficients of the multi-scale convolutional branches; For the first The importance score corresponding to each multi-scale convolutional branch; Use the Sigmoid activation function; Sum the Sigmoid scores for the three multi-scale convolution branches; This is a smoothing term.
9. A computer device, characterized in that, include: A processor configured to be coupled to memory, read and execute instructions and / or program code in the memory to perform the method as described in any one of claims 1-8.
10. A computer-readable medium, characterized in that, The computer-readable medium stores computer program code that, when executed on a computer, causes the computer to perform the method as described in any one of claims 1-8.