Vehicle load spectrum prediction method, device and medium based on physical parameter embedding

Abstract

The application discloses a vehicle load spectrum prediction method and device based on physical parameter embedding and a medium, and belongs to the technical field of vehicle structure durability test and load spectrum prediction. The method comprises the following steps: constructing a vehicle chassis topology graph, defining initial feature vectors of nodes and edges; splicing a road surface input vector and a vehicle operating state vector to form a global time sequence, and then inputting the global time sequence into a space-time coupling deep prediction model to predict a vehicle load spectrum; the space-time coupling deep prediction model comprises a Transformer encoder and a GNN encoder; the Transformer encoder extracts a road surface excitation feature vector; the GNN encoder injects the road surface excitation feature vector into the vehicle chassis topology graph; and after the feature vectors of all nodes are updated through a multilayer graph neural network, the vehicle load spectrum is predicted. The application directly mines potential laws of vehicle dynamics from data, avoids regarding the vehicle as a black box, and realizes rapid generation of a load spectrum under any combination of design parameters.

Description

Technical Field

[0001] This invention belongs to the field of vehicle structure durability testing and load spectrum prediction technology, specifically involving a vehicle load spectrum prediction method, device and medium based on physical parameter embedding. Background Technology

[0002] Vehicle load spectrum is the collection of forces, moments, or strains that a vehicle experiences over time during actual driving, affecting key load-bearing components such as the frame, suspension, and axles. It serves as the fundamental basis for vehicle fatigue durability analysis, remaining life prediction, and reliability design. In existing automotive R&D processes, full-vehicle road durability testing is required early in the project to obtain accurate and reliable road load data.

[0003] However, the traditional method for obtaining vehicle load spectra currently relies primarily on real-vehicle road tests. This "physical measurement" paradigm faces several severe challenges in practical applications: First, the extremely high economic cost; companies not only need to manufacture expensive early prototype vehicles, but also purchase and install a large number of precision sensors such as strain gauges and wheel force analyzers, while also paying huge rents for professional test tracks and the manpower costs of testing teams. Second, the lengthy development cycle; the entire process, encompassing prototype modification and preparation, complex sensor calibration, road test data collection, and tedious post-processing data, often takes several months, severely restricting the overall vehicle development efficiency and market launch speed. Finally, the lack of flexibility; frequent changes to design parameters are commonplace during vehicle development. Once a change occurs, the original test data may become invalid, requiring the entire testing process to be restarted, making it difficult to adapt to the rapid iteration development needs of the current automotive industry. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a vehicle load spectrum prediction method, device, and medium based on physical parameter embedding. It adopts a purely data-driven approach, abandoning the dependence on solving physical equations. It utilizes the powerful nonlinear mapping capabilities of deep learning to directly mine the potential laws of vehicle dynamics from the data. At the same time, it avoids treating the vehicle as a black box, enabling rapid generation of load spectra under arbitrary combinations of design parameters.

[0005] This invention provides the following technical solution:

[0006] Firstly, a vehicle load spectrum prediction method based on physical parameter embedding is provided, comprising the following steps:

[0007] Construct a vehicle chassis topology graph and define initial feature vectors for nodes and edges based on the vehicle's physical parameters; where nodes in the vehicle chassis topology graph are the key hard points of the vehicle chassis, and edges are the physical components connecting the key hard points.

[0008] The road input vector and the vehicle running state vector are generated by sampling the vehicle driving process, and the two are concatenated to form a global time series sequence.

[0009] The global time series is input into the spatiotemporal coupling deep prediction model, which outputs the predicted vehicle load spectrum. The spatiotemporal coupling deep prediction model includes a Transformer encoder and a GNN encoder. The Transformer encoder is used to extract the road surface excitation feature vector from the global time series. The GNN encoder injects the road surface excitation feature vector into the vehicle chassis topology graph and predicts the vehicle load spectrum after updating the feature vectors of all nodes through a multi-layer graph neural network based on the initial feature vectors of nodes and edges.

[0010] Optionally, the initial feature vectors of the nodes and edges are respectively:

[0011] ; ;

[0012] in, For nodes Physical characteristics For nodes The geometric position features of the nodes Static three-dimensional coordinates in the vehicle coordinate system For nodes The quality characteristics of the node represent the node's quality characteristics. The quality of the components For nodes The rotational inertia characteristics of the nodes The moment of inertia of the rigid body Indicates transpose; For nodes The initial feature vector, This represents the weights of the multilayer perceptron network used to process node features. Indicates normalization;

[0013] ; ;

[0014] in, For connecting nodes and its neighboring nodes edge Physical characteristics For the edge The equivalent stiffness, For the edge The equivalent damping, For the edge Length, For the edge The type coding includes rigid links, elastic elements, active elements, or semi-active elements; For the edge The direction vector; For the edge The initial feature vector, This represents the weights of the multilayer perceptron network used to process edge features.

[0015] Optionally, the road surface input vector, the vehicle operating state vector, and the global time series sequence are respectively:

[0016] ;

[0017] ;

[0018] ; ;

[0019] in, and They are respectively The road surface input vector and the vehicle operating state vector at each time step. , , and These are the road surface elevations at the left front wheel, right front wheel, left rear wheel, and right rear wheel, respectively. , , and They are respectively , , and The vertical rate of change for longitudinal speed at any given moment for Steering wheel angle at any moment and These are longitudinal acceleration and lateral acceleration, respectively. Indicates transpose. for The total input vector at time t, This indicates a splicing operation. To set the global time series sequence within the time window, To set the step size of the time window.

[0020] Optionally, the Transformer encoder is used to extract road surface excitation feature vectors from the global time-series sequence, specifically including:

[0021] The sine-cosine positional encoding method is used to encode the position of the global time series sequence and use it as the input sequence. The global dependency features across time steps are extracted through a multi-layer multi-head self-attention mechanism to obtain the road surface excitation feature vector.

[0022] Optionally, the GNN encoder injects the road surface excitation feature vector into the vehicle chassis topology graph, and predicts the vehicle load spectrum after updating the feature vectors of all nodes through a multi-layer graph neural network based on the initial feature vectors of nodes and edges, specifically:

[0023] The road surface excitation feature vector is concatenated to the wheel center node of the vehicle chassis topology to obtain the input feature vector of all nodes;

[0024] ;

[0025] in, For nodes The input feature vector, For nodes The initial feature vector, The road surface excitation feature vector, This indicates a splicing operation. Represents a vector of all zeros, with dimensions ANDed. The dimensions are the same;

[0026] The input feature vectors of the nodes and the initial feature vectors of the edges are input into the multilayer graph neural network to update the feature vectors of all nodes.

[0027] ; ; ;

[0028] in, For the first In a layered graph neural network, from neighboring nodes Passed to the node Information, Indicates from neighboring nodes Passed to the node , For the first Multilayer perceptron of layered graph neural networks and After the first After updating the nodes in the layered graph neural network Feature vectors and neighboring nodes eigenvectors, when hour, and They are nodes and neighboring nodes The input feature vector, For connecting nodes and its neighboring nodes edge The initial feature vector, For nodes The sum of all messages received, For nodes The set of all neighboring nodes, This represents the vector composition of forces. For the first After updating the nodes in the layered graph neural network eigenvectors, For the first Node feature update function of layered graph neural network;

[0029] Based on the updated feature vectors of all nodes and the initial feature vectors of the edges, the vehicle load spectrum is predicted through the output layer of the GNN encoder.

[0030] ;

[0031] ;

[0032] in, for Edge of time prediction The load vector, For the output layer of the GNN encoder, a multilayer perceptron, For the first After updating the nodes in the layered graph neural network eigenvectors, For the first After the layered graph neural network updates the neighbor nodes eigenvectors, This represents the number of layers in the graph neural network of the GNN encoder. , and They are the edges The forces acting on the corresponding physical components along the X, Y, and Z directions of the vehicle coordinate system. , and They are the edges The torques of the corresponding physical components rotating around the X, Y, and Z axes of the vehicle coordinate system. Indicates transpose. for Predicted vehicle load spectrum at any time express The real-valued feature space of dimension , This represents the number of nodes in the vehicle chassis topology graph.

[0033] Optionally, the loss function during training of the spatiotemporal coupled deep prediction model includes temporal reconstruction loss, frequency domain consistency loss, and peak weighting loss; the temporal reconstruction loss is used to measure the mean square error between the predicted vehicle load spectrum waveform and the actual vehicle load spectrum waveform, the frequency domain consistency loss is used to measure the frequency domain consistency between the predicted vehicle load spectrum waveform and the actual vehicle load spectrum waveform, and the peak weighting loss is used to apply greater weight to training samples whose actual vehicle load spectrum is greater than a set threshold.

[0034] Optionally, the loss function during training of the spatiotemporal coupled deep prediction model... for:

[0035] ;

[0036] ; ;

[0037] in, For temporal reconstruction loss, For frequency domain consistency loss, Peak-weighted loss, , and They are respectively , and The corresponding weighting coefficients, and They are respectively The predicted vehicle load spectrum and the actual vehicle load spectrum at any given time. For dynamic weighting coefficients, These are weighting coefficients. This is the steepness coefficient. for The absolute value, To set a threshold, This is the activation function.

[0038] Optionally, the training of the spatiotemporal coupled deep prediction model includes two stages. In the first stage, a virtual dataset is used for training, and the data in the virtual dataset is generated in batches using dynamic simulation software. In the second stage, a real-world dataset is used for training.

[0039] In a second aspect, a computer device is provided, including a processor and a memory; wherein, when the processor executes a computer program stored in the memory, it implements the steps of the vehicle load spectrum prediction method based on physical parameter embedding as described in any one of the first aspects.

[0040] Thirdly, a computer-readable storage medium is provided for storing a computer program; when executed by a processor, the computer program implements the steps of the vehicle load spectrum prediction method based on physical parameter embedding as described in any one of the first aspects.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] This invention embeds the physical parameters of the vehicle chassis into the vehicle chassis topology diagram and, after representing the road excitation and vehicle state temporally, inputs them into a spatiotemporally coupled deep prediction model including a Transformer encoder and a GNN encoder. This enables the spatiotemporally coupled deep prediction model to understand the influence mechanism of physical parameters on mechanical response, ensuring that the force and torque transmission paths conform to real mechanical logic. This invention abandons the traditional reliance on solving explicit physical equations, instead utilizing the nonlinear mapping capabilities of deep learning to directly learn the underlying laws of vehicle dynamics from data. In this way, the model achieves powerful predictive capabilities while avoiding the black-box problem; its predictive behavior is dually constrained by physical parameters and topological relationships. Once trained, the model can rapidly generate full-condition load spectra under any given combination of physical parameters in a very short time, thereby greatly accelerating the design iteration and fatigue durability analysis process of the chassis system. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the vehicle load spectrum prediction method based on physical parameter embedding of the present invention;

[0044] Figure 2 This is a flowchart of the vehicle load spectrum prediction method based on physical parameter embedding of the present invention. Detailed Implementation

[0045] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be used to limit the scope of protection of the present invention. It should be noted that the term "comprising" and any variations thereof in the specification, claims and the above-mentioned drawings of the present invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products or devices.

[0046] Example 1:

[0047] like Figure 1 and Figure 2 As shown, a vehicle load spectrum prediction method based on physical parameter embedding includes the following steps:

[0048] Step S1: Construct a vehicle chassis topology graph and define the initial feature vectors of nodes and edges according to the vehicle's physical parameters; where the nodes of the vehicle chassis topology graph are the key hard points of the vehicle chassis, and the edges are the physical components connecting the key hard points.

[0049] Step S2: Generate road surface input vector and vehicle operating state vector by sampling the vehicle driving process, and concatenate the two to form a global time series sequence;

[0050] Step S3: Input the global time series into the spatiotemporal coupling deep prediction model and output the predicted vehicle load spectrum.

[0051] The spatiotemporal coupled deep prediction model includes a Transformer encoder and a GNN encoder. The Transformer encoder is used to extract road surface excitation feature vectors from the global time series. The Transformer encoder is a neural network layer based on a self-attention mechanism. The GNN (Graph Neural Network) encoder injects the road surface excitation feature vectors into the vehicle chassis topology graph, and predicts the vehicle load spectrum after updating the feature vectors of all nodes through a multi-layer graph neural network based on the initial feature vectors of nodes and edges.

[0052] In this embodiment, step S1 specifically includes the following steps:

[0053] S1.1: Chassis topology diagram construction.

[0054] Vehicle chassis topology diagram Let be an undirected graph as defined, where the set of nodes is... Nodes represent key hard points of the vehicle chassis, specifically including: wheel center, wheelbase, track width, ground clearance, front / rear overhang length, control arm joints, shock absorber mounts, and subframe joints. This represents the number of nodes in the vehicle chassis topology graph. Let be the set of edges. The edge is a physical component connecting key hard points. For connecting nodes and its neighboring nodes The physical components are represented by the sides, which indicate the force transmission channels. According to the vehicle structure, the sides are divided into the following categories: rigid linkages: such as the control arm body, which mainly transmits axial force and bending moment; elastic elements: such as rubber bushings, which have nonlinear stiffness and damping; active / semi-active elements: such as coil springs and shock absorbers, which have complex nonlinear force-velocity characteristics.

[0055] S1.2: Define node characteristics.

[0056] Define nodes physical characteristics : ;

[0057] in, For nodes The geometric position features of the nodes Static three-dimensional coordinates in the vehicle coordinate system , Represents a 3-dimensional feature space of real numbers. For nodes The quality characteristics of the node represent the node's quality characteristics. The quality of the components Represents a 1-dimensional feature space of real numbers. For nodes The rotational inertia characteristics of the nodes Moment of inertia of the rigid body , , and They are nodes The moments of inertia of the rigid body about the X, Y, and Z axes of the vehicle coordinate system. , and These represent the products of inertia in the XY, YZ, and XZ planes, respectively. Indicates transpose. It represents a 6-dimensional real number feature space.

[0058] S1.3: Nonlinear embedding of node features.

[0059] The original nodes have diverse physical feature units and large numerical ranges. To enable the spatiotemporally coupled depth prediction model to better handle these data, a node encoder is introduced. The structure of the node encoder can refer to existing technologies, specifically it can be composed of a multilayer perceptron (MLP). Physical characteristics mapped to Real-valued feature space of dimension .

[0060] ;

[0061] in, For nodes The initial feature vector, This represents the weights of the multilayer perceptron network used to process node features. Indicates normalization; express The real-valued feature space of dimension.

[0062] S1.4: Define edge features.

[0063] For connection nodes and its neighboring nodes edge Define its physical characteristics :

[0064] ;

[0065] in, For the edge The equivalent stiffness, For the edge The equivalent damping, For the edge Length, For the edge The type coding includes rigid links, elastic elements, active elements, or semi-active elements; For the edge The direction vector.

[0066] S1.5: Nonlinearly embed edge features.

[0067] ;

[0068] in, For the edge The initial feature vector enables the spatiotemporal coupling depth prediction model to distinguish the essential differences between different connection methods in force transmission. This represents the weights of the multilayer perceptron network used to process edge features.

[0069] In this embodiment, step S2 specifically includes the following steps:

[0070] S2.1: Generate the road surface input vector.

[0071] Specifically, sampling is performed along the vehicle's travel trajectory to generate... The road surface input vector at time t.

[0072] ;

[0073] in, for The road surface input vector at time t, , , and These are the road surface elevations at the left front wheel, right front wheel, left rear wheel, and right rear wheel, respectively. , , and They are respectively , , and The vertical rate of change.

[0074] S2.2: Generate the vehicle operating state vector.

[0075] The driver's actions determine the vehicle's movement, thereby generating load transfer, which defines the vehicle's operating state vector.

[0076] ;

[0077] in, for The vehicle's operating state vector at time 1. for longitudinal speed at any given moment for Steering wheel angle at any moment and These are longitudinal acceleration and lateral acceleration, respectively. This indicates transpose.

[0078] S2.3: Global time series input fusion to obtain a global time series sequence.

[0079] First, the road input vector and the vehicle operating state vector are concatenated to form... Total input vector at time step for: , This indicates a splicing operation; then, in order to capture the hysteresis of the vehicle dynamics system, a global time series sequence within a defined time window is constructed. : ; express The total input vector at time t, To set the step size of the time window, the value is usually taken as a time interval of 0.5 seconds to 1.0 seconds corresponding to physical time.

[0080] In this embodiment, step S3 specifically includes the following steps:

[0081] S3.1: The Transformer encoder is used to extract the road surface excitation feature vector from the global time series.

[0082] S3.2: The GNN encoder injects the road excitation feature vector into the vehicle chassis topology graph, and predicts the vehicle load spectrum after updating the feature vectors of all nodes through a multi-layer graph neural network based on the initial feature vectors of nodes and edges.

[0083] Step S3.1 specifically includes the following steps:

[0084] S3.1.1: The sine-cosine position coding method is used to perform position coding on the global time series.

[0085] ;

[0086] ;

[0087] in, For the position encoding matrix, Represents a linear transformation. This serves as the initial input for the Transformer encoder. This is the index of the current time step data within the global time series. This refers to the dimensional index within the global time series. To encode the position matrix Location and The specific value of the dimension, To encode the position matrix Location and The specific value of the dimension, This represents the hidden layer dimension of the Transformer encoder.

[0088] S3.1.2: Stacking The Transformer encoding layer has multiple layers, and each encoding layer uses a multi-head attention mechanism for feature extraction.

[0089] The specific method for feature extraction using a multi-head attention mechanism in each coding layer can refer to existing technologies. Specifically, the input vector is first divided into multiple heads, generating a query, key, and value matrix for each head. Then, the attention weights of each head are calculated, and all heads are concatenated and linearly transformed. Finally, residual connections and layer normalization are introduced to obtain the output features of the current coding layer.

[0090] S3.1.3: The output features of the last coding layer are obtained by global average pooling to obtain the road surface excitation feature vector.

[0091] Step S3.2 specifically includes the following steps:

[0092] S3.2.1: Incentive injection.

[0093] The road surface excitation feature vector is concatenated to the wheel center node of the vehicle chassis topology to obtain the input feature vector of all nodes.

[0094] ;

[0095] in, For nodes The input feature vector, For nodes The initial feature vector, The road surface excitation feature vector, This indicates a splicing operation, which splices the road surface excitation feature vector to the wheel center node of the vehicle chassis topology, ensuring that the road surface excitation can only enter the vehicle chassis system through the wheel center. Represents a vector of all zeros, with dimensions ANDed. They have the same dimensions.

[0096] S3.2.2: Message interaction is performed by a multi-layer GNN network.

[0097] The input feature vectors of the nodes and the initial feature vectors of the edges are input into a multilayer graph neural network to update the feature vectors of all nodes.

[0098] In this embodiment, the specific process of message interaction is as follows:

[0099] Message computation: Simulated edge The transmitted force is achieved using a multilayer perceptron based on nodes. and neighboring nodes The input feature vector, combined with the edge between the two initial feature vector Together, calculate from neighboring nodes Passed to the node Information:

[0100] ;

[0101] in, For the first In a layered graph neural network, from neighboring nodes Passed to the node Information, Indicates from neighboring nodes Passed to the node , For the first Multilayer perceptron of layered graph neural networks and After the first After updating the nodes in the layered graph neural network Feature vectors and neighboring nodes eigenvectors, when hour, and They are nodes and neighboring nodes The input feature vector, For connecting nodes and its neighboring nodes edge The initial eigenvectors.

[0102] Message aggregation: Simulated nodes The forces at the node are in equilibrium. According to Newton's laws, the resultant force on the node is the vector sum of the forces acting on all the connecting parts. Summation is used as the aggregation function:

[0103] ;

[0104] in, For nodes The sum of all messages received, For nodes The set of all neighboring nodes, This represents the vector composition of forces.

[0105] Node state update: Simulate the motion response of nodes under the action of resultant force:

[0106] ;

[0107] in, For the first After updating the nodes in the layered graph neural network eigenvectors, For the first The node feature update function of the layered graph neural network is obtained, for example, through GRU or ResNet blocks, to ensure gradient propagation in the GNN network.

[0108] S3.2.3: Load prediction.

[0109] Based on the updated feature vectors of all nodes and the initial feature vectors of the edges, the vehicle load spectrum is predicted through the output layer of the GNN encoder.

[0110] After message passing through all layers of the graph neural network, the updated feature vector of each node has incorporated the excitation information from the road surface and the structural response information of the entire vehicle. Therefore, for each edge that needs to output vehicle load, the final features of its two endpoints are concatenated, and the load time series is output through the output layer; the specific formula is as follows:

[0111] ;

[0112] ;

[0113] in, for Edge of time prediction The load vector, The multilayer perceptron in the output layer of the GNN encoder is responsible for decoding. For the first After updating the nodes in the layered graph neural network eigenvectors, For the first After the layered graph neural network updates the neighbor nodes eigenvectors, This represents the number of layers in the graph neural network of the GNN encoder. , and They are the edges The forces acting on the corresponding physical components along the X, Y, and Z directions of the vehicle coordinate system. , and They are the edges The torques of the corresponding physical components rotating around the X, Y, and Z axes of the vehicle coordinate system. Indicates transpose. for Predicted vehicle load spectrum at any time express The real-valued feature space of dimension , This represents the number of nodes in the vehicle chassis topology graph.

[0114] In this embodiment, the spatiotemporal coupled depth prediction model adopts a simulation-experiment hybrid-driven transfer training strategy. Specifically, the training of the spatiotemporal coupled depth prediction model includes two stages. In the first stage, a virtual dataset is used for training, and the data in the virtual dataset is generated in batches using dynamic simulation software. In the second stage, an experimental dataset is used for training.

[0115] The virtual dataset was generated in batches using dynamic simulation software, with parameters such as hard point coordinates, stiffness, and mass randomly changed to collect simulation data. The real-world dataset was collected from a real test track and included vehicle configurations. The data contained characteristics that are difficult to model in simulations, such as noise, nonlinear friction, and gaps. In the first training phase, a large learning rate was used to allow the model to learn the force transmission mechanism. In the second training phase, the real-world dataset was used, and the Transformer encoding layer and the output layer of the GNN encoder itself were fine-tuned using a smaller learning rate. The GNN layer of the GNN encoder was frozen to keep the underlying force transmission logic unchanged.

[0116] In this embodiment, in order to ensure that the generated load spectra are not only numerically close, but also consistent in fatigue damage characteristics, a loss function considering the combination of time domain, frequency domain, and peak values ​​is designed.

[0117] Loss function during training of spatiotemporally coupled deep prediction models for:

[0118] ;

[0119] in, For temporal reconstruction loss, For frequency domain consistency loss, Peak-weighted loss, , and They are respectively , and The corresponding weighting coefficients.

[0120] The time-domain reconstruction loss measures the mean square error between the predicted vehicle load spectrum waveform and the actual vehicle load spectrum waveform. The formula for the time-domain reconstruction loss is:

[0121] ;

[0122] In the formula, The total number of training samples or the length of the time series, i.e., one sample per time step; and They are respectively The predicted vehicle load spectrum and the actual vehicle load spectrum at any given time.

[0123] Frequency domain consistency loss is used to measure the frequency domain consistency between the predicted vehicle load spectrum waveform and the actual vehicle load spectrum waveform. Specifically, it can be achieved using the Fast Fourier Transform. The formula for frequency domain consistency loss is as follows:

[0124] ;

[0125] in, The vehicle load spectrum predicted for all training samples or all time points in a time series. For all training samples or for all moments in a time series, the true vehicle load spectrum is obtained. For Fast Fourier Transform (FFT) operations.

[0126] Peak-weighted loss is used to assign greater weights to training samples whose real vehicle load spectrum exceeds a set threshold. The formula for peak-weighted loss is:

[0127] ; ;

[0128] in, and They are respectively The predicted vehicle load spectrum and the actual vehicle load spectrum at any given time. For dynamic weighting coefficients, These are weighting coefficients. This is the steepness coefficient. for The absolute value, To set a threshold, As an activation function, it maps the input to... Interval.

[0129] Example 2:

[0130] The present invention provides a computer device, including a processor and a memory; wherein, when the processor executes a computer program stored in the memory, it implements the steps of the above-described vehicle load spectrum prediction method based on physical parameter embedding.

[0131] For more detailed information on the above methods, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.

[0132] Example 3:

[0133] The present invention provides a computer-readable storage medium for storing a computer program; when the computer program is executed by a processor, it implements the steps of the above-described vehicle load spectrum prediction method based on physical parameter embedding.

[0134] For more detailed information on the above methods, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.

[0135] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Regarding the devices and storage media disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant details can be found in the method section. Those skilled in the art will clearly understand that the technologies in the embodiments of this invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of this invention, in essence or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or certain parts of the embodiments of this invention.

[0136] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A vehicle load spectrum prediction method based on physical parameter embedding, characterized in that, Includes the following steps: Construct a vehicle chassis topology graph and define initial feature vectors for nodes and edges based on the vehicle's physical parameters; where nodes in the vehicle chassis topology graph are the key hard points of the vehicle chassis, and edges are the physical components connecting the key hard points. The road input vector and the vehicle running state vector are generated by sampling the vehicle driving process, and the two are concatenated to form a global time series sequence. The global time series is input into the spatiotemporal coupling depth prediction model, and the predicted vehicle load spectrum is output. The spatiotemporal coupling depth prediction model includes a Transformer encoder and a GNN encoder. The Transformer encoder is used to extract the road surface excitation feature vector from the global time series. The GNN encoder injects road surface excitation feature vectors into the vehicle chassis topology graph, and predicts the vehicle load spectrum after updating the feature vectors of all nodes through a multi-layer graph neural network based on the initial feature vectors of nodes and edges. Specifically: The road surface excitation feature vector is concatenated to the wheel center node of the vehicle chassis topology to obtain the input feature vector of all nodes; ； in, For nodes The input feature vector, For nodes The initial feature vector, The road surface excitation feature vector, This indicates a splicing operation. Represents a vector of all zeros, with dimensions ANDed. The dimensions are the same; The input feature vectors of the nodes and the initial feature vectors of the edges are input into the multilayer graph neural network to update the feature vectors of all nodes. ； ； ； in, For the first In a layered graph neural network, from neighboring nodes Passed to the node Information, Indicates from neighboring nodes Passed to the node , For the first Multilayer perceptron of layered graph neural networks and After the first After updating the nodes in the layered graph neural network Feature vectors and neighboring nodes eigenvectors, when hour, and They are nodes and neighboring nodes The input feature vector, For connecting nodes and its neighboring nodes edge The initial feature vector, For nodes The sum of all messages received, For nodes The set of all neighboring nodes, This represents the vector composition of forces. For the first After updating the nodes in the layered graph neural network eigenvectors, For the first The node feature update function of a layered graph neural network; Based on the updated feature vectors of all nodes and the initial feature vectors of the edges, the vehicle load spectrum is predicted through the output layer of the GNN encoder. ； ； in, for Edge of time prediction The load vector, For the output layer of the GNN encoder, a multilayer perceptron, For the first After updating the nodes in the layered graph neural network eigenvectors, For the first After the layered graph neural network updates the neighbor nodes eigenvectors, This represents the number of layers in the graph neural network of the GNN encoder. , and They are the edges The forces acting on the corresponding physical components along the X, Y, and Z directions of the vehicle coordinate system. , and They are the edges The torques of the corresponding physical components rotating around the X, Y, and Z axes of the vehicle coordinate system. Indicates transpose. for Predicted vehicle load spectrum at any time express The real-valued feature space of dimension , This represents the number of nodes in the vehicle chassis topology graph.

2. The vehicle load spectrum prediction method based on physical parameter embedding according to claim 1, characterized in that, The initial feature vectors of the nodes and edges are as follows: ； ； in, For nodes Physical characteristics For nodes The geometric position features of the nodes Static three-dimensional coordinates in the vehicle coordinate system For nodes The quality characteristics of the node represent the node's quality characteristics. The quality of the components For nodes The rotational inertia characteristics of the nodes The moment of inertia of the rigid body Indicates transpose; For nodes The initial feature vector, This represents the weights of the multilayer perceptron network used to process node features. Indicates normalization; ； ； in, For connecting nodes and its neighboring nodes edge Physical characteristics For the edge The equivalent stiffness, For the edge The equivalent damping, For the edge Length, For the edge The type coding includes rigid links, elastic elements, active elements, or semi-active elements; For the edge The direction vector; For the edge The initial feature vector, This represents the weights of the multilayer perceptron network used to process edge features.

3. The vehicle load spectrum prediction method based on physical parameter embedding according to claim 1, characterized in that, The road surface input vector, vehicle operating state vector, and global time series sequence are respectively: ； ； ； ； in, and They are respectively The road surface input vector and the vehicle operating state vector at each time step. , , and These are the road surface elevations at the left front wheel, right front wheel, left rear wheel, and right rear wheel, respectively. , , and They are respectively , , and The vertical rate of change for longitudinal speed at any given moment for Steering wheel angle at any moment and These are longitudinal acceleration and lateral acceleration, respectively. Indicates transpose. for The total input vector at time t, This indicates a splicing operation. To set the global time series sequence within the time window, To set the step size of the time window.

4. The vehicle load spectrum prediction method based on physical parameter embedding according to claim 1, characterized in that, The Transformer encoder is used to extract road surface excitation feature vectors from the global time series, specifically including: The sine-cosine positional encoding method is used to encode the position of the global time series sequence and use it as the input sequence. The global dependency features across time steps are extracted through a multi-layer multi-head self-attention mechanism to obtain the road surface excitation feature vector.

5. The vehicle load spectrum prediction method based on physical parameter embedding according to claim 1, characterized in that, The loss function during training of the spatiotemporal coupled deep prediction model includes temporal reconstruction loss, frequency domain consistency loss, and peak weighting loss. The temporal reconstruction loss is used to measure the mean square error between the predicted vehicle load spectrum waveform and the actual vehicle load spectrum waveform. The frequency domain consistency loss is used to measure the frequency domain consistency between the predicted vehicle load spectrum waveform and the actual vehicle load spectrum waveform. The peak weighting loss is used to apply greater weight to training samples whose actual vehicle load spectrum is greater than a set threshold.

6. The vehicle load spectrum prediction method based on physical parameter embedding according to claim 5, characterized in that, The loss function during training of the spatiotemporal coupled deep prediction model for: ； ； ； in, For temporal reconstruction loss, For frequency domain consistency loss, Peak-weighted loss, , and They are respectively , and The corresponding weighting coefficients, and They are respectively The predicted vehicle load spectrum and the actual vehicle load spectrum at any given time. For dynamic weighting coefficients, These are weighting coefficients. This is the steepness coefficient. for The absolute value, To set a threshold, This is the activation function.

7. The vehicle load spectrum prediction method based on physical parameter embedding according to claim 1, characterized in that, The training of the spatiotemporal coupled deep prediction model includes two stages. In the first stage, a virtual dataset is used for training, and the data in the virtual dataset is generated in batches using dynamic simulation software. In the second stage, a real-world dataset is used for training.

8. A computer device, characterized in that, It includes a processor and a memory; wherein, when the processor executes the computer program stored in the memory, it implements the steps of the vehicle load spectrum prediction method based on physical parameter embedding as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, Used to store computer programs; when executed by a processor, the computer programs implement the steps of the vehicle load spectrum prediction method based on physical parameter embedding as described in any one of claims 1-7.

Images