A physical information guided commercial vehicle frame performance prediction method and system

Abstract

This invention provides a physical information-guided method and system for predicting the performance of commercial vehicle frames, applied in the field of frame performance prediction technology. The method first acquires a three-dimensional voxelized multi-channel engineering data tensor corresponding to the commercial vehicle frame, including geometric topological features, material properties and cross-sectional structural features, connection attribute features, and load characteristic fields characterizing the frame. A prediction model is then constructed based on this engineering data tensor. Through multi-scale feature extraction, feature feedback, and joint decoding prediction of the model input data, a full-field physical response cloud map of the frame and global key performance scalars are simultaneously output. During the model training phase, a composite loss function including a physical consistency residual term is constructed, and combined with a virtual-real transfer strategy, the model is trained and calibrated using source domain simulation data and target domain experimental data. Through the above technical solution, the efficiency, accuracy, and physical consistency of commercial vehicle frame performance prediction can be improved.

Description

Technical Field

[0001] This invention belongs to the field of chassis performance prediction technology, specifically relating to a physical information-guided method and system for predicting the chassis performance of commercial vehicles. Background Technology

[0002] With the development trends of intelligentization, electrification, and decarbonization in commercial vehicles, the chassis, as the fundamental load-bearing component of the entire vehicle, faces the dual technical challenges of increased power battery weight and range anxiety. This inherent contradiction between structural load-bearing requirements and overall vehicle energy efficiency goals necessitates that chassis design achieve extreme lightweighting while ensuring high reliability. Currently, with the continuously shortening product update cycle in commercial vehicles, more stringent engineering requirements are being placed on the precision and efficiency of chassis structural design. As a core component of the chassis R&D process, high-confidence performance evaluation and lifespan prediction are crucial prerequisites for achieving forward structural design and lightweight optimization.

[0003] Currently, the prediction of vehicle frame static and dynamic properties and fatigue durability mainly relies on the finite element analysis (FEA) method. However, in practical engineering applications, this method faces a contradiction between computational efficiency and prediction accuracy. In numerical simulations, the pursuit of accurate capture of local stress gradients inevitably leads to a surge in the number of meshes, a significant increase in the scale of the equations to be solved, and high costs per calculation. While sparse meshes can improve computational efficiency, they cannot guarantee the prediction confidence of critical connection areas. Furthermore, in fatigue life prediction, the accuracy of existing mainstream methods for predicting vehicle frame fatigue life is often limited and exhibits significant prediction bias due to the strong nonlinearity of fatigue damage evolution, the discreteness of manufacturing processes, and the simplification of connection stiffness and neglect of residual stress in simulation models. This bias necessitates the use of large safety factors in design, resulting in excessive redundancy in the vehicle frame structure and limiting the realization of extreme lightweight design.

[0004] In recent years, deep learning technologies, represented by Convolutional Neural Networks (CNNs) and Graph Neural Networks (GNNs), have been introduced into the field of commercial vehicle frame performance prediction. For example, Chinese patent application CN202511445349.0 discloses a multimodal data-driven frame performance prediction method that uses a graph convolutional encoder to extract structural features and simultaneously predict multiple performance indicators. However, when such models are applied to complex frame designs, their training process often adopts a purely data-driven "black box" mapping paradigm, lacking explicit guidance from mechanical mechanisms. When the training samples are small, the prediction results often fail to strictly conform to physical laws, and their generalization ability and prediction accuracy in local high-stress areas are limited.

[0005] Furthermore, to achieve precise weight reduction in the chassis structure, it is necessary not only to obtain the overall performance scalar but also to identify local damage details through full-field response cloud maps. However, existing prediction models (such as those mentioned in the patent) only focus on the regression of global performance scalars, lacking the ability to spatially reconstruct the full-field physical response. This lack of information dimension makes it impossible for the model to guide precise optimization for local high-stress areas, and its prediction results are prone to deviations from physical logic, severely limiting the model's prediction accuracy and confidence in key fatigue failure points.

[0006] Finally, addressing the current situation in chassis development where "simulation data is abundant but experimental data is scarce," existing virtual-real calibration methods (such as the statistical alignment algorithm based on maximum mean difference and correlation alignment used in the aforementioned patent) primarily focus on mathematically narrowing the distribution characteristics of the two types of data. This purely statistical approach has not fully utilized physical mechanisms to guide the correction of virtual-real deviations. When dealing with systematic errors caused by the simplification of simulation models, it lacks the ability to constrain the laws of physical damage evolution, easily producing prediction results that violate mechanical common sense, thus limiting the confidence level of the calibration model under real complex working conditions.

[0007] In summary, there is an urgent need for a high-precision intelligent prediction method for vehicle frame performance that can deeply integrate prior knowledge of physical mechanisms, achieve synchronous prediction of full-field cloud maps and global scalars and bidirectional feature feedback, and possess virtual-real migration calibration capabilities based on physical mechanisms. Summary of the Invention

[0008] In view of the above-mentioned problems in the prior art, the purpose of this invention is to provide a physical information-guided method for predicting the performance of commercial vehicle frames. By uniformly representing the multi-source engineering information of commercial vehicle frames in three-dimensional voxels, and combining bi-branch collaborative prediction, physical consistency constraints, and virtual-real migration calibration, a fast, unified, and physically consistent prediction of the frame's full-field physical response cloud map and global performance scalar is achieved.

[0009] A method for predicting the chassis performance of commercial vehicles guided by physical information, characterized by comprising the following steps:

[0010] S1. Obtain the three-dimensional voxelized multi-channel engineering data tensor corresponding to the commercial vehicle frame, and perform classification preprocessing and dataset partitioning on the three-dimensional voxelized multi-channel engineering data tensor to obtain model input data; wherein, the three-dimensional voxelized multi-channel engineering data tensor is established in a unified three-dimensional spatial coordinate system and is used to characterize the geometric topological features, material properties and cross-sectional structural features, connection property features and load feature field of the frame;

[0011] S2. Construct a prediction model, and perform multi-scale feature extraction, feature feedback, and joint decoding prediction based on the model input data to obtain the frame's full-field physical response cloud map and global key performance scalars. Specifically, the model input data is encoded to extract the frame's local geometric and mechanical features and overall topological features at different spatial scales. Based on the extracted overall topological features and local field response features, bidirectional feature feedback is performed to establish a dynamic correlation between global performance features and local field response features. Finally, based on the feedback features, full-field response reconstruction and global scalar regression are performed respectively to output the frame's full-field physical response cloud map and global key performance scalars.

[0012] S3. Construct a composite loss function containing a physical consistency residual term, and train and calibrate the model based on the composite loss function; wherein, differentiable physical damage calculation is performed based on fatigue damage accumulation theory, and the physical consistency residual term is obtained according to the model output result; the training and calibration adopts a virtual-real transfer strategy, and optimizes and fine-tunes the model parameters using source domain simulation data and target domain experimental data.

[0013] Preferably, in step S1, the three-dimensional voxelized multi-channel engineering data tensor includes: a geometric topology channel for characterizing the spatial occupancy state of the vehicle frame; a material property and cross-sectional structure channel for characterizing the material grade of each beam segment and its corresponding PSN curve characteristics, cross-sectional shape, cross-sectional size, and plate thickness parameters; a connection attribute channel for characterizing the number of connectors, connector diameter, and connector spacing at the connection area between the crossbeam and the longitudinal beam; and a load characteristic channel for characterizing the load distribution characteristics of the vehicle frame under preset working conditions.

[0014] Preferably, in step S1, the process of classifying and preprocessing the three-dimensional voxelized multichannel engineering data tensor and partitioning the dataset includes:

[0015] One-hot encoding is performed on discrete features to transform them into numerical vectors with class distinguishability. The discrete features include at least material grade, cross-sectional shape category, number of beams, and connector specification category.

[0016] Normalization is performed on continuous features to map feature values ​​to a preset scaling range. The continuous features include at least cross-sectional dimensions, plate thickness parameters, connector spacing, and load feature values.

[0017] The preprocessed samples are divided into a source domain simulation dataset for model pretraining and a target domain test dataset for virtual-real transfer calibration. The topological similarity between samples is calculated based on the geometric topological features of the samples. The heterogeneous configuration samples in the target domain test dataset with a topological similarity lower than a preset threshold are divided into an independent test set for verifying the model's cross-topological generalization ability.

[0018] Preferably, in step S2, the model input data is encoded, including: using a hard parameter shared encoder with a symmetrical U-shaped encoding and decoding architecture, and downsampling the model input data step by step through multi-level feature extraction to extract the local geometric and mechanical features and overall topological features of the frame at different spatial scales.

[0019] Preferably, in step S2, bidirectional feature feedback is performed based on the extracted overall topological features and local field response features, including: setting up a bidirectional collaborative attention mechanism module between encoding processing and decoding prediction; in the global-to-local feedback path, applying the spatial weight distribution generated by the global performance features to the local field response features; in the local-to-global injection path, extracting hotspot extreme value features from the local field response features and injecting them into the global scalar regression path, so as to establish a dynamic correlation between the global performance features and the local field response features.

[0020] Preferably, in step S2, full-field response reconstruction and global scalar regression are performed based on the features after mutual feedback, including: using the full-field cloud map reconstruction branch to upsample the high-dimensional features compressed by the encoder step by step, and combining cross-layer skip connections to restore spatial resolution; after decoding step by step using the full-field cloud map reconstruction branch, the corresponding full-field physical response cloud map is output. The full-field physical response cloud map includes the stress response cloud map, deformation response cloud map, and damage degree cloud map or fatigue life response cloud map of the frame under the target working condition.

[0021] A global scalar regression branch is used to extract high-dimensional compressed features from the feature compression layer of the encoder, and the high-dimensional compressed features are input into a global average pooling layer to obtain a one-dimensional global feature vector. The one-dimensional global feature vector is then input into a multilayer perceptron structure for regression calculation, and the global key performance scalar is output through a fully connected layer of the output layer. The global key performance scalar includes frame mass, low-order modal frequency, bending stiffness, torsional stiffness, and the maximum stress value, maximum deformation value, and fatigue life value of the frame under bending, torsional, cornering, and braking conditions.

[0022] Preferably, the construction process of the physical consistency residual term of the composite loss function is as follows:

[0023] The material PSN curve features and load feature matrix from the input tensor channel are invoked. The current stress response contour map of the branch output is reconstructed based on the full-field contour map in the model. The theoretical life distribution of the whole field is calculated in combination with the Basquin equation. Then, the theoretical physical damage value is calculated according to the Miner fatigue damage accumulation rule. The predicted damage value of the branch output is reconstructed based on the overall cloud map. With the theoretical physical damage value The physical consistency residual term is obtained from the residual between them.

[0024] Preferably, the composite loss function further includes:

[0025] The data-driven error term, using mean squared error, is used to measure the numerical deviation between the full-field response contour plot output by the model and the corresponding true value contour plot, as well as the numerical deviation between the predicted performance scalar value output by the model and the corresponding true value scalar value.

[0026] The logical consistency constraint term is obtained by extracting the local hotspot extrema from the fatigue life response cloud map output by the global cloud map reconstruction branch and comparing them with the fatigue life scalar prediction value output by the global scalar regression branch; the weights corresponding to the logical consistency constraint term in the composite loss function are adjusted according to the relationship between the relative error between the two and the preset threshold.

[0027] Preferably, the training and calibration in step S3 adopts a two-stage training method, including a source domain pre-training stage and a target domain transfer calibration stage. In the source domain pre-training stage, the prediction model is trained using a source domain simulation dataset to obtain an initial mapping relationship between the structural topology and the physical response. In the target domain transfer calibration stage, the prediction model is fine-tuned using target domain experimental data with real physical properties. By correcting the deviation between the model prediction results and the experimental true values, the physical consistency calibration of the model parameters is achieved.

[0028] The second objective of this invention is to provide a physical information-guided commercial vehicle chassis performance prediction system, which executes the aforementioned physical information-guided commercial vehicle chassis performance prediction method, including:

[0029] The data tensor input and classification processing module is used to obtain the three-dimensional voxelized multi-channel engineering data tensor corresponding to the commercial vehicle frame, and to perform classification preprocessing and dataset partitioning on the three-dimensional voxelized multi-channel engineering data tensor to obtain model input data.

[0030] A multi-task 3D convolutional neural network construction module is used to build a prediction model based on the model input data and perform multi-scale feature extraction, feature feedback, and joint decoding prediction to simultaneously obtain the full-field physical response cloud map of the chassis and global key performance scalars. The multi-task 3D convolutional neural network construction module includes a symmetric U-shaped residual coding engine, a bidirectional collaborative attention unit, and a heterogeneous dual-branch decoding engine. The symmetric U-shaped residual coding engine uses cascaded 3D residual convolutional layers to extract features and construct a high-dimensional compressed feature space. The bidirectional collaborative attention unit is used to establish a dynamic feedback channel between global performance features and local field response features. The heterogeneous dual-branch decoding engine includes a full-field cloud map reconstruction branch and a global scalar regression branch.

[0031] A physics-guided training and virtual-real transfer calibration module is used to construct a composite loss function containing a physical consistency residual term, and to train and calibrate the prediction model based on the composite loss function. This module includes a differentiable physics loss engine, a logical consistency monitoring unit, and a virtual-real transfer learning engine. The differentiable physics loss engine performs differentiable physical damage calculations based on fatigue damage accumulation theory and obtains the physical consistency residual term from the model output. The logical consistency monitoring unit enforces the physical-logical consistency between the contour map extrema and the scalar prediction value. The virtual-real transfer learning engine optimizes and fine-tunes the model parameters using source domain simulation data and target domain experimental data.

[0032] The performance evaluation and result display module is used to restore, display, and evaluate the whole-field physical response cloud map of the chassis and the global key performance scalars output by the prediction model.

[0033] The beneficial effects of this invention are:

[0034] This invention constructs a composite loss function that includes a physical consistency residual term and performs differentiable physical damage calculations based on fatigue damage accumulation theory. This ensures that the model training process is driven not only by sample data but also constrained by the material fatigue damage evolution law. Compared with prediction methods that rely solely on data fitting, this technical solution can reduce the deviation of model output results from fundamental physical laws, improve the problem of insufficient physical consistency of prediction results under small sample conditions, and thus improve the reliability of fatigue life prediction for vehicle frames in local high stress concentration areas and complex working conditions.

[0035] This invention constructs a physical information-guided dual-branch, multi-task 3D convolutional neural network prediction model architecture. By coordinating the settings of the full-field contour map reconstruction branch and the global scalar regression branch, it achieves the synchronous output of the frame's full-field physical response results and global performance scalar results. By performing multi-scale feature extraction on the model input data and combining cross-layer feature transfer and joint decoding prediction, it can simultaneously obtain stress response contour maps, deformation response contour maps, damage degree contour maps or fatigue life response contour maps, as well as global performance parameters such as frame mass, low-order modal frequencies, bending stiffness, and torsional stiffness. This provides unified analysis results for the identification of local hazardous areas and the overall performance evaluation of the frame.

[0036] This invention establishes a bidirectional feature feedback relationship between global topological features and local field response features, creating a dynamic correlation between global performance features and local hotspot features. On one hand, global performance features can guide the distribution of local field response features; on the other hand, local hotspot features can feed back into the global scalar regression process. Through this bidirectional collaborative mechanism, inconsistencies between local extrema in the global contour map and the global performance scalar can be reduced, thereby improving the consistency between the model output results in local field distribution and overall performance evaluation.

[0037] This invention employs a two-stage training strategy combining a source domain pre-training stage and a target domain transfer calibration stage. It utilizes source domain simulation data and target domain experimental data to optimize and fine-tune model parameters. By first learning the fundamental mapping relationship between the chassis structure topology and load response using simulation samples, and then using experimental samples to perform transfer calibration on the model, the impact of systematic deviations between the simulation model and the real physical environment on the prediction results can be reduced, thereby improving the model's applicability to different chassis topologies and operating conditions. Attached Figure Description

[0038] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0039] Figure 1 This is a flowchart of the method of the present invention;

[0040] Figure 2 This is an architecture diagram of the prediction model of the present invention;

[0041] Figure 3 This is a logic block diagram of the physical information-guided loss function of the present invention;

[0042] Figure 4 This is a qualitative comparison diagram of the predicted value and the simulated true value of the chassis fatigue life response cloud map of the present invention, wherein (a) is the simulated fatigue life cloud map (true value) and (b) is the AI ​​prediction cloud map (predicted value) of the present invention.

[0043] Figure 5 This is a correlation analysis diagram of the predicted and true values ​​of the four key performance indicators of the present invention, where (a) is the first-order modal frequency, (b) is the bending stiffness, (c) is the fatigue life, and (d) is the maximum stress under torsional conditions. Detailed Implementation

[0044] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Equivalent substitutions, modifications, or improvements made by those skilled in the art based on the disclosure of this invention without departing from the inventive concept should all fall within the scope of protection of this invention.

[0045] It should be noted that the following embodiments use a side-beam commercial vehicle frame as an example, but the present invention is not limited to this type of structure. For skateboard chassis, truss space frames, and other irregular chassis frames, as long as the corresponding three-dimensional voxelized multi-channel engineering data tensors can be constructed, the method and system of the present invention can be used for performance prediction and evaluation.

[0046] Example 1

[0047] This embodiment selects a commercial vehicle frame with typical topological characteristics as the prediction object. The frame mainly includes two longitudinal beams and five transverse beams. The longitudinal beams can adopt a channel-shaped cross-section structure, while the transverse beams can adopt an I-shaped cross-section, a box-shaped cross-section, or a multi-cavity cross-section structure depending on the installation position and stress requirements. The transverse beams are connected to the longitudinal beams through bolt connection components of preset specifications. To achieve joint prediction of the full-field physical response and performance indicators of the frame under various working conditions, this embodiment uniformly represents the geometric topological information, material property information, connection property information, and load characteristic information of the frame as a three-dimensional voxelized multi-channel engineering data tensor in a preset format, and constructs a physical information-guided two-branch multi-task three-dimensional convolutional neural network prediction model based on this engineering data tensor.

[0048] Reference Figures 1 to 2 A method for predicting the chassis performance of commercial vehicles guided by physical information, specifically including the following steps:

[0049] Step S1: Acquisition and Preprocessing of 3D Engineering Data Tensors

[0050] This step is used to construct standardized input data containing multi-source engineering information of the vehicle frame, providing a data foundation for subsequent model building, training, and inference. The specific process is as follows:

[0051] S11: Obtain the 3D voxelized multichannel engineering data tensor in a preset format.

[0052] In this embodiment, the three-dimensional voxelized multi-channel engineering data tensor is established in a unified three-dimensional spatial coordinate system and can be represented as a four-dimensional matrix with dimensions C×D×H×W. Here, C represents the number of physical feature channels, and D, H, and W correspond to the voxel mesh resolution in the length, height, and width directions of the circumscribed cuboid space of the vehicle frame, respectively. For example, for a commercial vehicle frame sample with a length of 5500mm, a width of 948mm, and a height of 243mm, the voxel mesh resolution can be set to 512×128×32. In this case, the physical size of a single voxel unit is approximately 10.7mm×7.4mm×7.6mm, which characterizes the geometric details and stress characteristics of local areas of the vehicle frame. In other embodiments, other voxel mesh resolutions can be selected based on the vehicle frame size, computational resources, and accuracy requirements.

[0053] In this embodiment, the initial number of physical feature channels can be set to four, which are used to characterize the frame's geometric topology features, material properties and cross-sectional structure features, connection properties features, and load feature field, respectively. The geometric topology feature channel preferably uses binarization to characterize the spatial occupancy of the frame entity at each voxel position; the material properties and cross-sectional structure feature channel is used to characterize the material grade and corresponding PSN (probability-stress-life) curve features, cross-sectional shape, cross-sectional dimensions, and plate thickness parameters of each beam segment of the frame; the connection properties feature channel is used to characterize the number, diameter, spacing, and distribution of connectors at the connection areas of the transverse and longitudinal beams; and the load feature field channel is used to characterize the load action features and their spatial distribution information extracted from measured road spectra, simulated working conditions, or historical load data. Through this method, engineering information from different sources, in different forms, and with different dimensions can be uniformly mapped to the same three-dimensional spatial representation benchmark, achieving standardized integration.

[0054] S12: Classification preprocessing and dataset partitioning of 3D voxelized multichannel engineering data tensors

[0055] After obtaining the original 3D voxelized multichannel engineering data tensor, preprocessing operations are performed. For discrete features such as material grade, cross-sectional shape type, number of beams, and connector specifications, one-hot encoding can be used to convert them into high-bit sparse vectors with class discriminative power (e.g., encoding 3 candidate materials as...). To eliminate false numerical weights introduced by manual numbering,

[0056] For continuous features such as cross-sectional dimensions, plate thickness parameters, connector spacing, and load characteristic values, a normalization method can be adopted. The linear mapping method is used to map the values ​​of each channel to a preset [0,1] scale interval to reduce the impact of features of different dimensions and orders of magnitude on the network training process.

[0057] Normalization can be performed using the following formula:

[0058] (1)

[0059] in, Represents the original feature values. and These are the statistical minimum and maximum values ​​of the feature in the dataset, respectively. This represents the normalized eigenvalue. In other implementations, standardization or other numerical mapping methods may also be used.

[0060] After preprocessing, this embodiment performs dataset partitioning based on the data source and its physical confidence level: the samples generated in batches by global parameterized simulation are divided into source domain simulation datasets for the first stage of model pre-training; the high-value data obtained from open literature mining or real physical experiments are divided into target domain experimental datasets for the second stage of model virtual-real transfer calibration.

[0061] To further evaluate the applicability of the prediction model to different chassis topologies, the topological similarity between samples can be calculated based on the geometric topological features of the samples. The heterogeneous configuration samples of the target domain test dataset with a topological similarity lower than a preset threshold are divided into independent test sets to verify the model's cross-topological generalization ability.

[0062] The topological similarity can be calculated by expanding the geometric topological channels into vectors and then using cosine similarity, as shown in the following example:

[0063] (2)

[0064] in, This represents the flattened vector of sample A in the target domain test dataset. This represents the flattened vector of sample B in the source domain simulation dataset. This represents the topological similarity between the two. In this embodiment, when the similarity is lower than a preset threshold of 0.8, the corresponding target domain sample can be included in the independent test set; in other embodiments, this threshold can also be adjusted according to the actual data distribution.

[0065] Step S2: Construct a physically-guided, two-branch, multi-task 3D convolutional neural network (3D CNN) prediction model.

[0066] This step involves constructing a multi-task 3D convolutional neural network model that combines full-field physical response reconstruction capabilities with global performance scalar regression capabilities. The network employs a symmetrical U-shaped encoder-decoder structure, utilizing multi-task branches to collaboratively predict comprehensive chassis performance. For example... Figure 2As shown, the specific steps are as follows:

[0067] S21: Building a Hardware Parameter Shared Encoder

[0068] In this embodiment, the encoder can adopt a symmetrical U-shaped encoding / decoding architecture, which is composed of cascaded multi-layer three-dimensional residual convolutional modules. Each three-dimensional residual convolutional module may include two 3×3×3 three-dimensional convolutional layers, a batch normalization layer, and an activation function layer, and set cross-layer identity mapping branches to improve the gradient transfer stability during deep network training. As the network depth increases, the voxel feature map can be downsampled layer by layer through convolution operations with a stride of 2, so that the feature map resolution decreases layer by layer and the number of channels increases layer by layer, thereby extracting the local geometric and mechanical features and overall topological features of the frame at different spatial scales.

[0069] For example, in one implementation, the encoder is constructed from four cascaded 3D residual convolutional modules (3D ResNetBlocks), with the input feature map progressively compressed from 512×128×32 to 32×8×2, corresponding to feature channel numbers of 16, 32, 64, and 128 respectively. It should be understood that the specific number of layers, channels, and convolutional parameters can be adjusted according to the data scale, prediction task, and computational resources, without constituting a limitation on the scope of protection of this invention.

[0070] S22: Constructing a bidirectional collaborative attention mechanism module

[0071] A bidirectional collaborative attention mechanism module is set between the encoder and the decoder. The bidirectional collaborative attention mechanism module is used to establish a dynamic feedback channel between global performance characteristics and local field response characteristics.

[0072] Specifically, in the global-to-local feedback path, the global performance feature vector can be extracted from the model extraction branch B, the corresponding spatial weight distribution can be calculated through the channel attention mechanism, and the weight matrix can be multiplied with the feature map of branch A to suppress background noise in non-critical areas (such as low-stress areas).

[0073] In the local-to-global injection path, local high-stress hotspots in the global features of branch A can be identified from the model. The corresponding extreme feature vectors can be extracted using max pooling and injected into the scalar regression path, thereby overcoming the smoothing effect of subsequent global average pooling on extreme features and ensuring that the predicted minimum lifetime is strictly consistent with the failure location shown in the cloud map in terms of physical logic.

[0074] S23: Constructing the full-field cloud reconstruction branch in the dual-branch heterogeneous decoder (corresponding to...) Figure 1 Branch A shown

[0075] In this embodiment, branch A is used to spatially reconstruct the overall physical response of the chassis. Branch A can employ a combination of multi-layer 3D deconvolution layers and skip connections to progressively upsample the high-dimensional features compressed by the encoder, thereby achieving symmetrical restoration of spatial resolution. Preferably, branch A may include a 4-layer 3D deconvolution structure, and establish skip connections with the corresponding sampling layers in the encoder.

[0076] Among them, the cross-layer skip connection is used to directly transfer the shallow local geometric details and local mechanical features retained by the encoder at each sampling level to the corresponding decoding level of branch A through feature splicing or fusion. This preserves the detailed information of the frame rounded corner area, weight reduction hole area, and cross and longitudinal beam connection joint area during the spatial restoration process, reducing the loss or over-smoothing of local key area features during multi-layer downsampling and upsampling.

[0077] After completing the step-by-step decoding, the output layer of branch A can use a 1×1×1 convolutional layer to map the decoded multi-channel features back to the original physical space dimension, such as 512×128×32, to obtain a full-field physical response cloud map corresponding to the input voxel space. The full-field physical response cloud map can include the stress response cloud map, deformation response cloud map, and damage degree cloud map or fatigue life response cloud map of the frame under the target working condition, thereby providing spatial distribution basis for the identification of local dangerous areas and structural performance analysis.

[0078] S24: Constructing the global scalar regression branch in the dual-branch heterogeneous decoder (corresponding to...) Figure 1 Branch B shown

[0079] In this embodiment, branch B is used for regression prediction of the overall performance parameters of the chassis. Specifically, high-dimensional compressed features can be extracted from the feature compression layer at the bottom of the encoder, and these compressed features are input into a global average pooling layer (GAP) to compress the three-dimensional feature map into a one-dimensional global feature vector, thereby obtaining a compact representation of the overall topology, overall stress state, and global performance characteristics of the chassis. After obtaining the one-dimensional global feature vector, it can be further input into a multilayer perceptron structure for regression calculation. The multilayer perceptron structure may include multiple hidden layers, such as three hidden layers, and dropout layers can be set between adjacent hidden layers to reduce the risk of overfitting during model training. Finally, the global key performance scalar is output through the fully connected layer of the output layer.

[0080] The global key performance scalars can include overall performance parameters such as frame mass, low-order modal frequencies, bending stiffness, and torsional stiffness, as well as the maximum stress, maximum deformation, and fatigue life of the frame under bending, torsional, cornering, and braking conditions. By using branch B to regress and predict the global performance scalars, and combining this with branch A to spatially reconstruct the full-field response cloud map, the synchronous output of local and overall frame performance information can be achieved.

[0081] Step S3: Model training and calibration based on physical consistency constraints and virtual-real migration strategy

[0082] This step aims to improve the model's prediction confidence and generalization accuracy under small sample conditions by constraining physical mechanisms and evolving cross-domain data. For example... Figure 3 As shown, the specific steps are as follows:

[0083] S31: Constructing a composite loss function

[0084] In this embodiment, the composite loss function can be expressed as:

[0085] (3)

[0086] in, The data-driven error term can be represented by the mean squared error (MSE), which measures the numerical deviation between the full-field response contour plot of branch A and the corresponding true value contour plot, and the numerical deviation between the predicted performance scalar value of branch B and the corresponding true scalar value. The specific calculation process is as follows:

[0087] (4)

[0088] (5)

[0089] (6)

[0090] in, This indicates a data-driven error term; This represents the prediction error of the full-field response contour plot corresponding to branch A; This represents the performance scalar prediction error corresponding to branch B; Indicates the first The predicted value of the full-field response cloud map for each sample; Indicates the first The true value of the full-field response cloud plot for each sample; Indicates the first Scalar predicted values ​​of performance for each sample; Indicates the first The true value of the performance scalar for each sample; and These represent the number of samples corresponding to branch A and branch B, respectively. This represents the square norm 2.

[0091] The physical consistency residual term is obtained through the differentiable physical damage calculation process in step S311 below. It is used to introduce the material fatigue damage evolution mechanism into the model training process and constrain the neural network parameter update process from the mechanical mechanism level.

[0092] This represents a logical consistency constraint term, which is calculated through the following step S312 and is used to constrain the physical correspondence between the output results of branch A and branch B.

[0093] , , These represent the weight coefficients of the corresponding items.

[0094] By constructing the aforementioned composite loss function, the model training process can be simultaneously subject to data fitting requirements and physical mechanism constraints, thereby ensuring that the prediction results approximate the true label while remaining consistent with fundamental physical laws.

[0095] S311: Construct a differentiable physical damage calculation module based on Miner's fatigue damage accumulation law.

[0096] During the backpropagation process of model training, the differentiable physical damage calculation module calls upon the material PSN curve features and load feature matrix in the input tensor channel, and performs physical damage calculation on the model output based on preset differentiable theoretical operators. Specifically, based on the current stress response contour map output by branch A, combined with the Basquin equation:

[0097] (7)

[0098] Calculate the theoretical lifetime value under the corresponding load level, where, Indicates the first The theoretical life under level-2 load conditions, where C and m represent parameters related to the material's fatigue properties. This indicates the output of branch A. Predicted stress values ​​under high-level load conditions. After obtaining the theoretical life distribution, further calculations are based on Miner's fatigue damage accumulation rule:

[0099] (8)

[0100] Calculate the corresponding theoretical physical damage value ,in, Indicates the first The theoretical life under high load conditions. Indicates the first The actual number of cycles under the load condition.

[0101] Subsequently, the predicted damage value output by branch A is calculated. With respect to the theoretical physical damage distribution The residuals between them, and the residuals are used as physical consistency residual terms. Introduced into the composite loss function, it can be specifically expressed as:

[0102] (9)

[0103] By using the above settings, the evolution law of material fatigue damage can be embedded into the model training process in a differentiable manner, so that physical constraints can be transmitted to the network front-end parameters along the backpropagation path, thereby constraining the model by the fatigue damage mechanism while fitting the labeled data.

[0104] S312: Introduce logical consistency constraints

[0105] In this embodiment, by adding a logical consistency constraint term to the composite loss function, the consistency between the local hotspot extrema in the fatigue life response cloud map output by branch A and the fatigue life scalar output by branch B is constrained.

[0106] Specifically, local hotspot extrema can be extracted from the fatigue life response contour map of branch A, such as the minimum life point value. And compare it with the fatigue life scalar prediction value output by branch B. Compare them.

[0107] The system can set a preset deviation threshold, such as 5%; when the relative error between the two meets the threshold... At that time, the weight corresponding to the logical consistency constraint term is dynamically increased. This is to enhance the constraint on the consistency of the output results of the two branches.

[0108] This approach guides the network to adjust the feature feedback parameters in the bidirectional collaborative attention module, ensuring that the hotspot locations and their values ​​in the global fatigue life response cloud map are consistent with the global fatigue life scalar prediction results, thereby improving the physical and logical consistency of the multi-task output results.

[0109] S32: Perform two-stage training: "simulation pre-training and virtual-to-real migration calibration".

[0110] In this embodiment, model training and calibration adopt a two-stage training approach, including a source domain pre-training stage and a target domain transfer calibration stage.

[0111] In the first stage, the source domain pre-training stage, the multi-task 3D convolutional neural network is pre-trained using the source domain high-fidelity simulation dataset obtained in step S12. In this stage, the AdamW optimizer can be used, and a preset initial learning rate can be configured, for example, set to [value missing]. Simultaneously, the learning rate is adjusted using a cosine annealing decay strategy; backpropagation is performed based on the gradient generated by the composite loss function to update the convolutional kernel weights and related parameters in the network, so that the model establishes a mapping relationship between the frame structure topology, load characteristics and stress response.

[0112] In the second stage, the target domain transfer calibration stage, the pre-trained model parameters are fine-tuned using the target domain test dataset. In this stage, a relatively small learning rate (e.g., ...) can be used. The simulation model incorporates experimental samples with real physical properties, calculates the distribution of physical deviations between the experimental true values ​​and the model predictions, and further corrects the model parameters based on the gradient descent algorithm to reduce the impact of systematic errors between the simulation model and the real physical environment.

[0113] Through the two-stage training process of source domain pre-training and target domain transfer calibration described above, a transfer-calibrated prediction model can be obtained for performance prediction under different chassis topology conditions.

[0114] Example 2

[0115] This embodiment provides a physical information-guided commercial vehicle chassis performance prediction system for executing the method described in Embodiment 1. The system can be built on a computer hardware platform and includes at least a processor, a memory, and a graphics processing unit (GPU). The processor is communicatively connected to both the memory and the GPU. The memory stores executable program instructions, and when the processor invokes these instructions, it executes the following functional modules, specifically including:

[0116] Data Tensor Input and Classification Processing Module: This module is used to execute step S1 as described in Embodiment 1. It is equipped with a 3D data acquisition interface for retrieving 3D voxelized multi-channel engineering data tensors containing geometric, material, connection, and load features in a preset format; a classification preprocessing unit with built-in one-hot encoding and scaling normalization operators for numerical standardization of different types of physical channel features; and a dataset partitioning unit for automatically partitioning the data into source domain simulation datasets, target domain experimental datasets, and independent test sets based on the sample's source attributes and topological similarity, providing standardized material support for the model.

[0117] Multi-task 3D convolutional neural network construction module: used to execute step S2 as described in Example 1. This module is the core computing architecture of the system, and it includes: a symmetric U-shaped residual coding engine, which uses cascaded 3D residual convolutional layers to extract features and construct a high-dimensional compressed feature space; a bidirectional collaborative attention unit, used to establish a dynamic feedback channel between global performance features and local field response features; and a heterogeneous dual-branch decoding engine, which consists of a full-field cloud map reconstruction branch configured with cross-layer skip connections and a global scalar regression branch, used to achieve synchronous prediction output of the vehicle frame full-field physical response cloud map and the global performance scalar.

[0118] The physics-guided training and virtual-real transfer learning calibration module is used to execute step S3 as described in Example 1. This module integrates: a differentiable physics loss engine, which internally encapsulates a differentiable physics constraint operator based on Miner's fatigue damage accumulation theory, used to calculate the physical consistency residual between the predicted contour map and the mechanical theory value; a logical consistency monitoring unit, which forces the physical and logical consistency between the extreme values ​​of the constraint map and the scalar predicted value through a penalty operator; and a virtual-real transfer learning engine, used to schedule processor resources to execute a two-stage loop of simulation pre-training and experimental truth fine-tuning, to achieve physical-level calibration of model parameters and optimization of generalization ability.

[0119] Performance evaluation and result display module: This module is used to visualize the output of the neural network model. It includes a 3D field response reconstruction unit, which maps the voxelized field data output from the full-field contour map reconstruction branch back to the chassis geometry, generating intuitive chassis fatigue life contour maps or stress contour maps; and an accuracy statistical evaluation unit, which synchronously displays the key performance scalars predicted by the global scalar regression branch and generates correlation analysis charts between predicted and true values ​​based on the test set data (e.g., ...). Figure 5 As shown, it assists engineers in assessing the confidence level of model predictions and making intelligent design decisions for the chassis structure.

[0120] Example 3

[0121] To further verify the effectiveness and accuracy of the physical information-guided commercial vehicle frame performance prediction method proposed in this invention in practical engineering applications, this embodiment focuses on a typical commercial vehicle frame object consisting of 2 longitudinal beams and 5 transverse beams as described in Embodiment 1. The finite element simulation analysis results and bench physical test data are used as truth labels to qualitatively and quantitatively evaluate the predictive performance of the model of this invention.

[0122] First, a qualitative comparative analysis of the predictions of the overall physical response is conducted. For example... Figure 4 As shown, the distribution characteristics of the physical response of the chassis across the entire field are compared under typical torsional conditions. Among them, Figure 4The simulation true response cloud map shown in Figure (a) illustrates the discretized feature distribution based on finite element mesh calculation. The high damage areas are mainly concentrated at the edge of the opening in the web of the longitudinal beam and at the joint between the transverse and longitudinal beams. Figure 4 Figure (b) shows the AI ​​prediction cloud map of this invention, illustrating the continuous field feature distribution reconstructed by the model. The comparison reveals that the cloud map predicted by this invention highly matches the true values ​​of the finite element analysis (FEA) in terms of texture density distribution trends. It not only accurately reproduces the stress and lifespan distribution gradient of the entire frame but also precisely captures the local high stress concentration features in key areas such as fillets, joints, and holes. This indicates that thanks to the synergistic effect of the 3D Res-U-Net architecture and the physical information-guided loss function, the model successfully learned the strong nonlinear mapping mechanism between structural topology and local damage evolution, and the predicted field data possesses good spatial continuity and physical rationality.

[0123] Secondly, a quantitative correlation analysis of the predicted key performance indicators was conducted. To verify the actual engineering accuracy of the model after virtual-to-real migration calibration, this embodiment performed regression analysis on an independent test set containing 500 samples, using the experimental data of the target domain as the true benchmark. Figure 5 As shown, the correlation fitting results between the predicted values ​​and the experimental true values ​​of four key performance indicators are presented: Figure 5 Figure (a) in the middle and Figure 5 Figure (b) shows the prediction results for first-order modal frequencies and frame bending stiffness, respectively. Since these two indicators exhibit globally linear or weakly nonlinear characteristics, the scatter points are extremely closely clustered on the diagonal, and the coefficient of determination... All values ​​reached 0.99, proving that the model has extremely high predictive ability for global stiffness indices; Figure 5 Figure (c) illustrates the effectiveness of fatigue life prediction. Even with fatigue damage indicators exhibiting strong nonlinearity and discreteness, the model of this invention still demonstrates robust predictive performance, with a uniform scatter distribution and no obvious outliers. It reached 0.94; Figure 5 Figure (d) shows the prediction effect of the maximum stress under torsional conditions. Although the local stress is greatly affected by the sensitivity of the measurement location, the scatter plots still remain highly linear thanks to the weighted extraction of local hotspot features by the bidirectional collaborative attention mechanism. The score reached 0.97. This is mainly attributed to the strong constraint of the Miner cumulative damage operator in the physical information guided loss function, and the effective correction of systematic errors in the simulation by the virtual-real migration strategy.

[0124] Finally, regarding computational efficiency, traditional finite element fatigue analysis typically takes several hours for a single chassis sample. In contrast, the model trained in this invention, utilizing GPU acceleration during the inference phase, can complete a full set of predictions, including the entire field contour map and all performance scalars, in just 20 to 50 milliseconds. Data shows that this invention significantly improves design iteration efficiency while maintaining prediction confidence, effectively supporting the rapid selection and optimization of numerous solutions for commercial vehicle chassis during the conceptual design phase.

[0125] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the performance of a commercial vehicle chassis guided by physical information, characterized in that, Includes the following steps: S1. Obtain the three-dimensional voxelized multi-channel engineering data tensor corresponding to the commercial vehicle frame, and perform classification preprocessing and dataset partitioning on the three-dimensional voxelized multi-channel engineering data tensor to obtain model input data; wherein, the three-dimensional voxelized multi-channel engineering data tensor is established in a unified three-dimensional spatial coordinate system and is used to characterize the geometric topological features, material properties and cross-sectional structural features, connection property features and load feature field of the frame; S2. Construct a prediction model, and perform multi-scale feature extraction, feature feedback, and joint decoding prediction based on the model input data to obtain the frame's full-field physical response cloud map and global key performance scalars. Specifically, the model input data is encoded to extract the frame's local geometric and mechanical features and overall topological features at different spatial scales. Based on the extracted overall topological features and local field response features, bidirectional feature feedback is performed to establish a dynamic correlation between global performance features and local field response features. Finally, based on the feedback features, full-field response reconstruction and global scalar regression are performed respectively to output the frame's full-field physical response cloud map and global key performance scalars. S3. Construct a composite loss function containing a physical consistency residual term, and train and calibrate the model based on the composite loss function; wherein, differentiable physical damage calculation is performed based on fatigue damage accumulation theory, and the physical consistency residual term is obtained according to the model output result; the training and calibration adopts a virtual-real transfer strategy, and optimizes and fine-tunes the model parameters using source domain simulation data and target domain experimental data.

2. The method for predicting the performance of a commercial vehicle chassis guided by physical information according to claim 1, characterized in that, In step S1, the three-dimensional voxelized multi-channel engineering data tensor includes: a geometric topology channel for characterizing the spatial occupancy state of the vehicle frame; a material property and cross-sectional structure channel for characterizing the material grade of each beam segment and its corresponding PSN curve characteristics, cross-sectional shape, cross-sectional size, and plate thickness parameters; a connection attribute channel for characterizing the number of connectors, connector diameter, and connector spacing at the connection area between the crossbeam and the longitudinal beam; and a load characteristic channel for characterizing the load distribution characteristics of the vehicle frame under preset working conditions.

3. The method for predicting the performance of a commercial vehicle chassis guided by physical information according to claim 1, characterized in that, In step S1, the process of classifying and preprocessing the three-dimensional voxelized multi-channel engineering data tensor and partitioning the dataset includes: One-hot encoding is performed on discrete features to transform them into numerical vectors with class distinguishability. The discrete features include at least material grade, cross-sectional shape category, number of beams, and connector specification category. Normalization is performed on continuous features to map feature values ​​to a preset scaling range. The continuous features include at least cross-sectional dimensions, plate thickness parameters, connector spacing, and load feature values. The preprocessed samples are divided into a source domain simulation dataset for model pretraining and a target domain test dataset for virtual-real transfer calibration. The topological similarity between samples is calculated based on the geometric topological features of the samples. The heterogeneous configuration samples in the target domain test dataset with a topological similarity lower than a preset threshold are divided into an independent test set for verifying the model's cross-topological generalization ability.

4. The method for predicting the performance of a commercial vehicle chassis guided by physical information according to claim 1, characterized in that, In step S2, the model input data is encoded, including: using a hard parameter shared encoder with a symmetrical U-shaped encoding and decoding architecture, and downsampling the model input data step by step through multi-level feature extraction to extract the local geometric and mechanical features and overall topological features of the frame at different spatial scales.

5. The method for predicting the performance of a commercial vehicle chassis guided by physical information according to claim 4, characterized in that, In step S2, bidirectional feature feedback is performed based on the extracted overall topological features and local field response features. This includes: setting up a bidirectional collaborative attention mechanism module between encoding processing and decoding prediction; applying the spatial weight distribution generated by global performance features to local field response features in the global-to-local feedback path; and extracting hotspot extreme value features from local field response features and injecting them into the global scalar regression path in the local-to-global injection path to establish a dynamic correlation between global performance features and local field response features.

6. The method for predicting the performance of a commercial vehicle chassis guided by physical information according to claim 5, characterized in that, In step S2, full-field response reconstruction and global scalar regression are performed based on the features after mutual feedback, including: using the full-field cloud map reconstruction branch to upsample the high-dimensional features compressed by the encoder step by step, and combining cross-layer skip connections to restore spatial resolution; after decoding step by step using the full-field cloud map reconstruction branch, the corresponding full-field physical response cloud map is output. The full-field physical response cloud map includes the stress response cloud map, deformation response cloud map, and damage degree cloud map or fatigue life response cloud map of the frame under the target working condition. A global scalar regression branch is used to extract high-dimensional compressed features from the feature compression layer of the encoder, and the high-dimensional compressed features are input into a global average pooling layer to obtain a one-dimensional global feature vector. The one-dimensional global feature vector is then input into a multilayer perceptron structure for regression calculation, and the global key performance scalar is output through a fully connected layer of the output layer. The global key performance scalar includes frame mass, low-order modal frequency, bending stiffness, torsional stiffness, and the maximum stress value, maximum deformation value, and fatigue life value of the frame under bending, torsional, cornering, and braking conditions.

7. The method for predicting the performance of a commercial vehicle chassis guided by physical information according to claim 1, characterized in that, The construction process of the physical consistency residual term of the composite loss function is as follows: The material PSN curve features and load feature matrix from the input tensor channel are invoked. The current stress response contour map of the branch output is reconstructed based on the full-field contour map in the model. The theoretical life distribution of the whole field is calculated in combination with the Basquin equation. Then, the theoretical physical damage value is calculated according to the Miner fatigue damage accumulation rule. ; The predicted damage value of the branch output is reconstructed based on the overall field cloud map. With the theoretical physical damage value The physical consistency residual term is obtained from the residual between them.

8. The method for predicting the performance of a commercial vehicle chassis guided by physical information according to claim 1, characterized in that, The composite loss function also includes: The data-driven error term, using mean squared error, is used to measure the numerical deviation between the full-field response contour plot output by the model and the corresponding true value contour plot, as well as the numerical deviation between the predicted performance scalar value output by the model and the corresponding true value scalar value. The logical consistency constraint term is obtained by extracting the local hotspot extrema from the fatigue life response cloud map output by the global cloud map reconstruction branch and comparing them with the fatigue life scalar prediction value output by the global scalar regression branch; the weights corresponding to the logical consistency constraint term in the composite loss function are adjusted according to the relationship between the relative error between the two and the preset threshold.

9. The method for predicting the performance of a commercial vehicle chassis guided by physical information according to claim 1, characterized in that, The training and calibration in step S3 adopts a two-stage training method, including a source domain pre-training stage and a target domain transfer calibration stage. In the source domain pre-training stage, the prediction model is trained using a source domain simulation dataset to obtain the initial mapping relationship between the structural topology and the physical response. In the target domain transfer calibration stage, the prediction model is fine-tuned using target domain experimental data with real physical properties. By correcting the deviation between the model prediction results and the experimental true values, the physical consistency calibration of the model parameters is achieved.

10. A physical information-guided commercial vehicle chassis performance prediction system, characterized in that, A method for predicting the chassis performance of a commercial vehicle guided by physical information as described in any one of claims 1 to 9, comprising: The data tensor input and classification processing module is used to obtain the three-dimensional voxelized multi-channel engineering data tensor corresponding to the commercial vehicle frame, and to perform classification preprocessing and dataset partitioning on the three-dimensional voxelized multi-channel engineering data tensor to obtain model input data. A multi-task 3D convolutional neural network construction module is used to build a prediction model based on the model input data and perform multi-scale feature extraction, feature feedback, and joint decoding prediction to simultaneously obtain the full-field physical response cloud map of the chassis and global key performance scalars. The multi-task 3D convolutional neural network construction module includes a symmetric U-shaped residual coding engine, a bidirectional collaborative attention unit, and a heterogeneous dual-branch decoding engine. The symmetric U-shaped residual coding engine uses cascaded 3D residual convolutional layers to extract features and construct a high-dimensional compressed feature space. The bidirectional collaborative attention unit is used to establish a dynamic feedback channel between global performance features and local field response features. The heterogeneous dual-branch decoding engine includes a full-field cloud map reconstruction branch and a global scalar regression branch. A physics-guided training and virtual-real transfer calibration module is used to construct a composite loss function containing a physical consistency residual term, and to train and calibrate the prediction model based on the composite loss function. This module includes a differentiable physics loss engine, a logical consistency monitoring unit, and a virtual-real transfer learning engine. The differentiable physics loss engine performs differentiable physical damage calculations based on fatigue damage accumulation theory and obtains the physical consistency residual term from the model output. The logical consistency monitoring unit enforces the physical-logical consistency between the contour map extrema and the scalar prediction value. The virtual-real transfer learning engine optimizes and fine-tunes the model parameters using source domain simulation data and target domain experimental data. The performance evaluation and result display module is used to restore, display, and evaluate the whole-field physical response cloud map of the chassis and the global key performance scalars output by the prediction model.

Images