Multi-modal data driven commercial vehicle frame performance prediction method and system

Abstract

The invention discloses a multi-modal data driven commercial vehicle frame performance prediction method and system. The method comprises the following steps: uniformly coding design variables of a frame; constructing a multi-modal performance response data set based on finite element simulation and test results of mass production vehicle models, and realizing fusion of simulation and test data by adopting a maximum mean difference and related alignment algorithm; constructing a graph perception Transform multi-task prediction model fusing the structure topology and the physical position features, and predicting key performance indexes of the frame under a plurality of typical working conditions; through weighted multi-task loss function joint training, an uncertainty mechanism is introduced to dynamically adjust task weights; and after training is completed, deploying to an inference engine to realize second-level prediction and support increment fine adjustment updating. According to the method, repeated modeling and solving processes are avoided, the frame performance prediction efficiency and the adaptive capacity are remarkably improved, and the method is suitable for rapid evaluation of the frame performance of commercial vehicles of various structural configurations and material types.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of automobile engineering and artificial intelligence, and particularly relates to a multi-modal data driven commercial vehicle frame performance prediction method and system, which is used to replace finite element analysis to realize rapid prediction of key performance indicators of various types of frames. BACKGROUND

[0002] As the basic load-bearing structure of the whole vehicle, the performance of the commercial vehicle frame is directly related to the safety, reliability and service life of the whole vehicle. Therefore, in the whole vehicle development process, accurate evaluation of the frame performance is always a core task. At present, the acquisition of frame performance mainly depends on finite element simulation analysis and bench / real vehicle test methods. The finite element method calculates the stress, deformation, modal, fatigue life, etc. of the frame under different load conditions through the establishment of a high-precision model; and the test verification obtains the corresponding performance parameters through actual loading test. Although this combination of simulation and test is widely used in engineering practice, it still has many shortcomings.

[0003] Firstly, the finite element modeling process is complex, the parameter definition is fine, and the professional ability and experience of the engineering personnel are required, and the calculation amount is huge, the solving period is long, especially when dealing with various structural arrangements and material type changes, the efficiency is obviously limited. Secondly, bench and real vehicle tests are not only costly, but often subject to test conditions, resource allocation and cycle arrangement, making it difficult to support rapid iterative design requirements. More importantly, whenever the frame structure or material parameters are adjusted, modeling, simulation and verification must be re-performed, greatly restricting the flexibility and response speed of design optimization. This process relies heavily on professional software and hardware resources, resulting in high research and development costs, which is not conducive to efficient product updates and platform development.

[0004] With the development of industrial intelligent technology, data-driven prediction methods have gradually attracted attention in engineering structure design. Researchers attempt to build response surface models or approximate models based on machine learning to replace part of the traditional simulation process, thereby accelerating the prediction efficiency and reducing the dependence on high-performance computing resources. However, existing methods often face the problems of low accuracy and poor adaptability when dealing with high-dimensional input features, complex structure forms and nonlinear relationships between multiple performance indicators. Especially in the joint prediction task involving stress, deformation, modal frequency and fatigue life, the model can only cover local indicators, making it difficult to meet the demand for systematic evaluation in actual engineering.

[0005] In addition, most of the existing researches construct prediction models based on a single vehicle model or a specific structural platform as a sample, which directly leads to insufficient adaptability of the model in dealing with multiple vehicle models and multiple structural commercial vehicle frames. In actual engineering development, commercial vehicle frames of different platforms have significant differences in the layout mode, cross-sectional type, connection structure, material properties, etc. of the longitudinal beam and cross beam. Even within the same platform, the frame design also presents diversification due to different vehicle purposes (such as cargo trucks, tractor trucks, light passenger cars, etc.). If the model is trained only on a single structural configuration, the resulting model can only have good performance on that structure or its slightly modified variants. Once it encounters a significantly different structural design, the prediction deviation may increase significantly, seriously affecting its practicality in real design processes.

[0006] More importantly, the existing methods generally have the problem of singularity in data sources, lacking systematic mining and integration of a large amount of historical design data, simulation data and test data of mass-produced vehicle frames accumulated within the enterprise. These data have high engineering value and representativeness, covering structural performance response characteristics under multiple platforms, multiple batches and multiple working conditions. If they can be systematically integrated and incorporated into a unified modeling framework, not only the robustness and generalization ability of the model can be improved, but also the cycle of modeling for new vehicle models can be significantly shortened. However, due to the lack of unified structural coding mechanism, performance data standardization process and simulation / measured data fusion algorithm, most current researches cannot effectively integrate these heterogeneous historical data, resulting in that the prediction model still stays at the local modeling level with small sample and weak generalization. SUMMARY

[0007] To solve the above technical problems, the present application provides a multi-modal data-driven commercial vehicle frame performance prediction method and system, which integrates multi-source heterogeneous data, has multi-task prediction capability, and is suitable for intelligent prediction of performance of multiple types of commercial vehicle frames. It can break through the technical bottlenecks of traditional processes in modeling cost, prediction efficiency and model generalization, provide more efficient, more economical and more reliable technical support for frame design optimization, and accelerate the development and iteration process of new vehicle platforms.

[0008] Specifically, the technical solutions provided by the present application are as follows: A multi-modal data-driven commercial vehicle frame performance prediction method, comprising the steps of: constructing a frame structure design graph for reflecting a frame topology structure; inputting the frame structure design graph into a graph perception multi-task prediction model to obtain predicted values of each performance index of the frame; the frame structure design graph comprises a plurality of nodes with feature vectors and a node adjacency matrix for representing connection relationships of the nodes, and the feature vectors of the nodes collectively constitute a node feature matrix; The graph-aware multi-task prediction model is trained by a multi-modal performance response dataset, and the prediction model comprises an input embedding module, a graph-aware encoding module and a multi-task decoding module; wherein the input embedding module is used for inputting the node feature matrix of the frame structure design drawing and the node adjacency matrix after self-loop normalization processing; the graph-aware encoding module comprises a graph convolution encoder and a Transformer encoder, the graph convolution encoder is used for extracting the graph convolution features of each node in the frame structure design drawing, and the Transformer encoder is used for deep modeling of the graph convolution features; the multi-task decoding module comprises a plurality of prediction branches, and is used for predicting each performance index of the frame according to the deep features output by the Transformer encoder.

[0009] Further, the construction of the frame structure design drawing comprises: The circumscribed rectangle of the entire frame structure is subjected to grid processing, and is uniformly divided into a plurality of grids of the same size, each grid represents a node, and if there is a mutual connection or an integral longitudinal or transverse beam between the frame regions corresponding to adjacent grids, the corresponding two nodes are regarded as existing a connection edge, and the number or order of the connected longitudinal or transverse beam is taken as the weight of the edge; Each grid corresponding to the frame region in the frame design variable is quantitatively coded and spliced into a feature vector of the corresponding node; the frame design variable comprises the number, geometric size, cross-section type and material type of the longitudinal or transverse beam, for the numerical continuous variable in the design variable, a standardization processing method is used for normalization processing, and for the non-numerical discrete variable in the design variable, a one-hot encoding is used for numerical coding; The node adjacency matrix A is used to represent the connection relationship of each node, and the matrix element A ij is taken as the weight of the connection edge between the node i and the node j, and the self-loop normalization processing is performed on the node adjacency matrix A. ij

[0010] ​Further, the multi-modal performance response dataset includes a plurality of training samples, each training sample including a frame structure design drawing and corresponding performance index values, the performance indexes including maximum stress and maximum deformation under each typical working condition, low-order modal frequency and fatigue life; the multi-modal refers to the diversity of performance data sources and data types, including both data generated by numerical simulation and data collected in real bench or vehicle tests, for unified data representation, the multi-modal data needs to be preprocessed; the preprocessing includes: firstly, performing anomaly detection and cleaning on all original data, eliminating outliers, and oversampling and expanding a small amount of data, then unifying the dimensions and coordinate systems of simulation data and test data, and aligning the distributions of the same performance indexes in the two types of data sources, then performing standardization processing, and finally fusing the simulation data and test data of the same frame according to the weight coefficients determined by the Bayesian optimization algorithm.

[0011] Preferably, the SMOTE-NC method is used to alleviate the class imbalance problem and enhance the representativeness of boundary samples during the oversampling and expansion processing, and the rescaling and sample weighting methods are used to optimize the balance of the dataset samples; the MMD algorithm is used to align the distributions of the same performance indexes in the two types of data sources during the distribution alignment, minimizing the distribution difference caused by different data sources, and the CORAL algorithm is used to optimize the statistical consistency between the data sources, ensuring the alignment of simulation data and test data in high-order statistical distribution, and improving the quality and stability of data fusion.

[0012] Further, the graph convolutional encoder includes a plurality of front and rear connected graph convolutional layers, the graph convolutional layers adopt GCN based on mean aggregation or GAT or GIN based on attention mechanism.

[0013] Further, the Transformer encoder includes a plurality of stacked encoder layers, each encoder layer including a multi-head self-attention mechanism and a feedforward neural network two sub-layers, the multi-head self-attention mechanism calculates the attention weight of each node representation as a query, key and value vector with other nodes, thereby realizing global interaction and fusion of features, each node after the attention layer, independently passes through a feedforward neural network to further enhance the nonlinear modeling ability, each sub-layer has a residual connection and layer normalization to ensure gradient stability and training efficiency.

[0014] Preferably, the Transformer encoder adopts Graph Transformer, and the node adjacency matrix is injected into the attention calculation as a bias term.

[0015] Further, the multi-task decoding module comprises three prediction branches: a static performance branch, a modal frequency branch and a fatigue life branch, each branch adopts a full connection network, including an input layer, an output layer and a plurality of hidden layers between the input layer and the output layer, the input layer is used for receiving the deep features output by the graph perception coding module; the static performance branch is used for predicting the static performance indexes under typical working conditions, including a total of 8 indexes of maximum stress and maximum deformation under four typical working conditions of bending, torsion, braking and turning, the modal frequency branch is used for predicting the low-order modal frequencies of the frame, including the first six-order modal frequencies.

[0016] A commercial vehicle frame performance prediction system based on the above method, comprising a frame structure coding module and a multi-task intelligent prediction module, the frame structure coding module is used for constructing a frame structure design drawing of a to-be-predicted frame, and the multi-task intelligent prediction module is deployed with a trained graph perception multi-task prediction network, which is used for outputting corresponding performance prediction values according to an input frame structure design drawing of the to-be-predicted frame.

[0017] Further, the system is also equipped with a performance display and human-computer interaction module, and a user can view the structure modeling process, the graphical layout and the connection path of the design variables in a visual interface in real time, and after model reasoning is completed, a visual chart of performance prediction results is automatically generated; The system also supports comparative analysis of different frame design schemes, and parallel performance prediction of multiple frame design schemes input by a user, and outputs a horizontal comparison chart and a performance score report; the system also supports a prediction result export function, and prediction values and chart results are exported in the form of a report as a PDF or Excel file.

[0018] The present application significantly improves the many limitations of the existing frame performance evaluation method in terms of efficiency, accuracy and adaptability. Compared with the traditional method which relies heavily on finite element simulation and test verification, the present application fully excavates and utilizes a large amount of simulation and test data resources accumulated in the development process of mass-produced vehicles, and through unified coding and fusion of frame structure characteristics and multi-working-condition response data, an intelligent model with structure perception ability and multi-task prediction ability is constructed, realizing rapid and accurate prediction of key performance indexes of the frame. Thus, time and cost consumption caused by repeated modeling and solving are avoided, and the prediction efficiency and design response speed are effectively improved, providing an online evaluation capability with high integration, high concurrency and low delay for new vehicle development.

[0019] In terms of model capability, the prediction framework constructed by the present application fuses the graph perception Transformer network structure and the multi-task decoding mechanism, and can simultaneously complete the prediction tasks of multiple performance dimensions such as stress, deformation, modal frequency and fatigue life in the same model. With the collaborative design of structure position coding, graph convolution mechanism and task-specific decoder, the model has strong structure recognition capability and task decoupling capability, which significantly improves the modeling adaptability to complex structure forms and multiple load conditions. Compared with the traditional single-task modeling path, the end-to-end training strategy and the multi-task weighted loss function design used in the present application not only enhance the stability of model training, but also further optimize the collaborative prediction effect between various performance indicators.

[0020] The present application also introduces a simulation-test data fusion mechanism and an uncertainty adaptive weight adjustment method, which significantly enhances the robustness of the model to data distribution differences and task difficulty differences, effectively improves the generalization ability and prediction stability of the model under real working conditions. In addition, the present application also introduces an incremental fine-tuning mechanism, which allows the model to perform efficient parameter updating after new vehicle data arrives, avoids the waste of resources caused by repeated training, and maintains the continuous learning and long-term adaptation ability of the model.

[0021] By introducing a standardized vehicle frame structure design diagram construction method and a graph perception multi-task prediction model, the present application realizes the unified coding and efficient modeling of vehicle frame structures of different configurations, different materials and different platforms, breaks through the dependence restriction of traditional methods on single structure models, and has good structure adaptability and task migration ability. Through the multi-modal performance response data set constructed by fusing multi-source simulation and test data, the identification and prediction ability of the model for the performance differences of diversified vehicle frame structures under different typical working conditions is effectively improved, and the generalization performance and engineering practicability of the model in a multi-vehicle and multi-platform environment are significantly enhanced. Therefore, the present application is not only suitable for performance prediction of traditional single vehicle, but also can be widely adapted to various commercial vehicle frame platforms, and has the ability of rapid deployment across vehicle types and working conditions, providing efficient and reliable technical support for new vehicle development and platformized architecture design. BRIEF DESCRIPTION OF DRAWINGS

[0022] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, which together with the embodiments of the present application, serve to explain the present application, and do not constitute a limitation of the present application.

[0023] Figure 1 is a network architecture schematic diagram of the graph perception multi-task prediction model provided by the embodiment of the present application; Figure 2 is an R 2Fitting comparison chart, wherein (a) is a comparison chart of maximum stress under bending condition, (b) is a comparison chart of maximum displacement under torsion condition, (c) is a comparison chart of first-order modal frequency, and (d) is a comparison chart of fatigue life. DETAILED DESCRIPTION

[0024] To make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0025] Embodiment one The embodiment provides a multi-modal data-driven commercial vehicle frame performance prediction method, which is suitable for performance evaluation of commercial vehicle frames of various material types and structural configurations. The method comprehensively considers frame structure parameters, material properties and arrangement forms, and realizes rapid prediction of key performance indicators (including stress, deformation, modal frequency and fatigue life) under typical working conditions by using deep learning means. The method mainly includes the following steps: Step one: parameterization and coding processing of frame design variables Firstly, a frame structure design diagram for reflecting the topology of the frame is constructed. The diagram here is a data structure for representing a set of objects and their relationships. In the diagram, objects can be represented by nodes, each node has its own attribute characteristics, and the relationship between objects can be represented by edges, and each edge can have a weight.

[0026] In the embodiment, the circumscribed rectangle of the entire frame structure is subjected to MxN grid processing, and is uniformly divided into a plurality of grids of the same size. Each grid represents a graph node. If there is a mutual connection or one-piece forming longitudinal beam or transverse beam between the frame areas corresponding to adjacent grids, the corresponding two nodes are regarded as having a connection edge. The number / number of longitudinal beams or transverse beams is taken as the weight of the edge. The nodes represented by non-adjacent grids are by default not connected.

[0027] For each node represented by the grid, a feature vector of the node is constructed. The design variables of the frame mainly include: the number of longitudinal / lateral beams, geometric dimensions, cross-section types (such as channel type, I type, box type or U type, etc.), material types (such as high-strength steel, aluminum alloy, etc.), so the above design values in the frame area corresponding to each grid can be quantitatively spliced, thereby serving as the feature vector of the corresponding node. For numerical continuous variables in the design variables, such as geometric dimensions, directly using the original values will cause too large a dimensional difference between different dimensions, thereby affecting the training efficiency and stability of the model. Therefore, a standardization processing method is used for normalization processing, such as Z-Score standardization. For non-numerical discrete variables in the design variables, such as cross-section types, etc., One Hot Enconding is used for numerical expression, preserving the independence and separability of the category information. The above standardized / encoded continuous variables and discrete variables jointly constitute a structured vector representation as the feature vector of the corresponding node. The dimension of the vector is generally controlled between 80-120, depending on the number of design variables involved in the actual project.

[0028] And for the connection (edge) relationship between nodes, it can be represented by an adjacency matrix. For example, for MxN=S graph nodes, an SxS adjacency matrix A can be constructed. If the matrix element A ij is non-zero, it indicates that there is a connection edge between node i and node j. Further, the weight of the above-mentioned edge (the number / number of connected longitudinal or lateral beams) can be taken as the specific value of A ij . In order to facilitate subsequent graph convolution, self-loop normalization processing is required for the adjacency matrix. The self-loop means that the node is considered to have an edge connection relationship with itself, which is reflected in the adjacency matrix as A ii taking a non-zero value (usually 1), and normalization is to normalize all element values of the entire adjacency matrix.

[0029] Through the above method, this step converts complex frame design parameters into a standardized, readable, and unified model input form, which is the basis for the subsequent multi-task performance prediction process. Through standardization processing and encoding methods, not only the physical meaning and engineering semantics of the design variables are preserved, but also the modeling efficiency and expression ability of the intelligent model are significantly improved, making it suitable for multi-type, multi-material, and multi-structure commercial vehicle frame performance prediction tasks.

[0030] Step 2: Construction of multi-modal performance response dataset After the encoding of the frame design parameters is completed, the next step is to establish a comprehensive multi-modal performance response dataset, which is a very critical step in this program. The multi-modal refers to the diversity of data sources and data types, including both data generated by numerical simulation and data collected from real bench or real vehicle tests. The purpose of constructing this dataset is to enable the subsequent performance prediction model not only to understand the design parameters of the frame, but also to learn how these parameters affect the performance of the frame.

[0031] This embodiment is based on the existing multiple production commercial vehicle frames of the enterprise, and a large amount of finite element simulation and bench / real vehicle test data completed in the product development process. A multi-modal performance response dataset of the frame under typical working conditions is constructed. The collected data includes but is not limited to: maximum stress and maximum deformation under full load bending, full load torsion, emergency braking and emergency turning, natural frequency (especially low order modal frequency, such as the first six orders), and fatigue life under bending load and torsion load, etc. Key performance indicators. Since these data are mainly derived from the historical design and verification process, they have high authenticity and coverage, and do not need to redevelop simulation and test, which is beneficial to quickly build the dataset and reduce the modeling cost.

[0032] To improve the data quality and the stability of subsequent fusion processing, the simulation and test results need to be systematically sorted out. In the initial stage of data processing, first, all the original data are subjected to anomaly detection and cleaning. Outlier data exceeding 3 times the standard deviation are removed to reduce the interference of extreme values on data distribution. For example, if the deformation output of a simulation is much larger than that of other similar structures, it may be caused by modeling or boundary condition setting errors, and such abnormal values need to be removed. In order to ensure a more balanced distribution of data, a small number of class data in the sample (such as responses under special structural morphology and extreme working conditions) also need to be oversampled and expanded. In this embodiment, the SMOTE-NC (Synthetic Minority Over-sampling Technique for Nominal and Continuous variables) method is used to alleviate the class imbalance problem and enhance the representativeness of boundary samples, which can prevent the model from only leaning towards the majority samples and ignoring the important but rare structural configurations during training.

[0033] All the processed data are further subjected to standardized conversion, and finally form a consistent and statistically stable basic dataset, providing good support for the alignment and fusion of subsequent simulation-test data.

[0034] Step three: alignment and fusion of simulation data and test data To further improve the quality and stability of the multimodal dataset and reduce the deviation between simulation data and experimental data, this embodiment introduces a simulation-experimental data alignment and fusion processing step after the multimodal performance response dataset is constructed.

[0035] First, it is necessary to unify the dimensions and coordinate systems of simulation and experimental data from a physical perspective. For example, in finite element simulation, the stress of a vehicle frame structure may be expressed in MPa (megapascals), while the stress measured in an experiment may be in psi (pounds-forces per square inch). Furthermore, the three-dimensional coordinate system used in the simulation may have its origin at the vehicle's centerline, while the actual coordinates of the sensors installed in the experiment may be defined relative to the equipment's coordinate system. Without unifying these physical foundations, the model will generate incorrect analogies during the learning phase, conflating fundamentally different structural representations. Therefore, this embodiment establishes a unified dimension conversion module and coordinate system mapping mechanism in the early stages of data processing. Through a pre-defined physical mapping table, all simulation and experimental samples are transformed to the same physical reference system, ensuring comparability and fusion between various performance indicators.

[0036] After physical alignment, it is also necessary to perform distribution alignment on the same performance indicators in the two types of data sources. Since simulation data is usually more regular and continuous, while experimental data is limited by testing conditions and may have more fluctuations, missing values, or measurement errors, these differences can cause the model to tend to rely on the more smoothly distributed simulation data during training, neglecting the more representative experimental data in reality. To address this issue, this embodiment uses the Maximum Mean Difference (MMD) algorithm to perform distribution alignment on the same performance indicators in the two types of data sources, minimizing the distribution differences caused by different data sources and reducing the impact of "domain offset" on subsequent analysis and modeling. Simultaneously, the Correlation Alignment (CORAL) algorithm further optimizes the statistical consistency between the data sources, ensuring alignment of simulation data and experimental data in higher-order statistical distributions, thus improving the quality and stability of data fusion. The MMD algorithm compares whether there is a shift in the "mean distribution" of the two data sources under the same performance indicator and attempts to bring them closer through algorithmic means; while the CORAL algorithm further examines the differences in the "covariance matrix" (i.e., the statistical correlation between indicators) of the two data sources, adjusting the statistical structure of the two sources to be consistent through mathematical transformation. This is similar to transforming two engineering drawings with "different ways of speaking" into a language with "the same grammar" through a certain conversion method.

[0037] After the above alignment, the simulation data and the test data are mapped to a unified data representation space, and then fused according to the weight coefficients determined automatically according to the Bayesian optimization algorithm. After all, the number of simulation data and test data may be very different in different enterprises or projects, and cannot be simply mixed according to the number of samples, otherwise it will lead to data imbalance. The embodiment adopts the Bayesian optimization algorithm to automatically find the optimal fusion weight. The algorithm gradually converges to a data weight ratio most suitable for the current task by continuously evaluating the fusion effect in multiple attempts, ensuring that the fused data not only retains the regularity of simulation, but also reflects the authenticity of the test.

[0038] In order to further improve the representativeness of the data and avoid over-expansion, the embodiment also adopts recalibration and sample weighting methods to optimize the balance of the data set samples. Recalibration is used to adjust the abnormal values or rare configurations on the sample distribution boundary, so that these samples are more easily identified by the model; while sample weighting gives each sample a representativeness score after fusion, so that the model pays more attention to samples with high information value in the training process, and reduces the interference of samples with low value. These processes together improve the usability and learning value of the fused data. Finally, the consistency of data fusion is evaluated by calculating the distribution overlap and Kullback-Leibler divergence of the fused samples, and the alignment strategy is automatically adjusted according to the evaluation results until the predetermined quality requirements are met.

[0039] Through the above process, a structure-performance control data set with high consistency and representativeness is finally formed. In this data set, each sample includes clear input (i.e. parameterized frame design information) and corresponding output (i.e. performance response under multiple typical working conditions), and these data come from different sources and cover a wide range of working conditions and structure types. This data set not only meets the basic needs of model training, but also significantly enhances the generalization ability of the prediction model when facing new structure designs. Most of the samples in the data set are used for training the frame performance prediction model, and a small number of samples are left for verification and testing to evaluate the calibration model performance.

[0040] Step four: construction of the graph-aware multi-task prediction model After the parametric coding of the frame design variables and the fusion of the simulation test data are completed, an intelligent model that can truly understand the data and make performance predictions is needed next. Traditional prediction models are often single-task regressors that can only learn for a certain performance indicator, such as predicting stress or predicting deformation. However, the performance evaluation of a commercial vehicle frame involves not only a single indicator, but also multiple interrelated characteristics such as stress, deformation, modal frequency, fatigue life, etc. If independent models are established for each indicator, not only is the calculation efficiency low, but there is also a lack of information sharing between the models, which can easily lead to one-sided or even unstable prediction results. Therefore, a multi-task intelligent prediction model is constructed in this embodiment, which can simultaneously predict multiple performance indicators in a unified framework and achieve information complementation between them, thereby more accurately reflecting the overall performance of the frame.

[0041] As shown in Figure 1 The intelligent prediction model adopts a Transformer architecture based on the fusion of physical structure and graph relationship, and is composed of an input embedding module, a graph perception encoding module and a multi-task decoding module. The input embedding module is used to input a feature matrix composed of feature vectors of each node in the frame structure design graph, and a self-loop normalized adjacency matrix reflecting the connection relationship of each node. The graph perception encoding module includes a graph convolution encoder and a Transformer encoder connected in front and back. The graph convolution encoder is used to extract the graph convolution features of each node in the frame structure design graph, and then send these graph convolution features to the Transformer encoder for deep modeling. The multi-task decoding module includes multiple prediction branches, which are used to predict the performance of the frame according to the deep features output by the Transformer encoder.

[0042] The graph convolution encoder includes multiple graph convolution layers connected in front and back, which can adopt traditional GCN (Graph Convolutional Network) based on mean aggregation or GAT (Graph Attention Network) based on attention mechanism, etc. Through comparative experiments, the graph convolution layer of this embodiment finally adopts GIN (Graph Isomorphism Network).

[0043] GIN is a graph neural network architecture with strong structural discriminability. Its core structure design is different from traditional GCN or GAT, mainly reflected in the expression ability of its aggregation method and update mechanism. In GIN, each layer of the network sums the features of the target node and its neighbor nodes through an additive aggregation operation, and then inputs this sum value into a multi-layer perceptron (MLP) for nonlinear mapping. Specifically, for any node in the graph, GIN does not simply weight and average the neighbor features, but combines the sum of its own node features and neighbor features with a weighting combination, and adjusts the degree of maintaining its own information through learnable parameters, and then uses MLP for deep feature extraction. This design makes GIN theoretically equivalent to the Weisfeiler-Lehman graph isomorphism test capability, meaning it can effectively distinguish subgraphs or node neighborhoods with different structures in the graph, even if the differences are very subtle.

[0044] This structure of GIN is particularly suitable for task scenarios that are extremely sensitive to structure. In the structural performance prediction of commercial vehicles, the connection method between the nodes of the frame, and the slight differences in the topology structure often directly affect the mechanical performance, durability or vibration response of the whole vehicle. Traditional graph convolution methods such as GCN may cause different topology structure of frame subgraph to be mapped to similar features due to the use of average aggregation, thus failing to capture key structural differences. GIN, on the other hand, combines the more structure-discriminative additive aggregation with MLP, can more accurately perceive complex structural information such as connection angle, branch number, closed path, etc. Such a mechanism enables the model to retain differentiated features when faced with slight structural changes, thus providing more discriminative input representations for subsequent performance prediction modules (such as Transformer, regression predictor). Therefore, applying GIN as an encoder in the frame structural performance prediction task not only can fully exploit the deep relationship between structural information and physical performance, but also effectively improve the model's sensitivity to structural changes and generalization ability, making it more suitable for high-precision, high-reliability engineering structure modeling and performance evaluation.

[0045] The Transformer encoder is a deep modeling architecture centered on multi-head self-attention mechanism, which has strong ability to model global dependencies. Transformer is used to receive the node embedding representation extracted by the graph convolutional encoder, further capturing the deep semantic relationship between non-adjacent nodes in the graph. When the graph convolutional encoder completes the structure-aware feature extraction, it outputs a node graph convolutional feature matrix, each row representing the representation of a node after fusing its own and neighborhood information. The Transformer encoder takes this node sequence as input, assigns attention weights to each node with all other nodes through self-attention mechanism, and realizes information interaction and fusion in a global range. This mechanism breaks through the limitation of traditional graph neural networks that can only propagate information locally, enabling the model to perceive potential long-distance dependencies between any two nodes in the graph, especially suitable for graph structure modeling tasks with complex structure and significant long-range interaction.

[0046] The standard Transformer encoder is stacked by multiple encoder layers with the same structure, each of which includes two sub-layers: multi-head self-attention mechanism (MSA) and feedforward network (FFN). MSA realizes global interaction and fusion of features by taking each input node representation as query (Q), key (K), and value (V) vectors, and calculating its attention weights with all other nodes. After passing through the attention layer, each node is independently enhanced by an FFN (usually two linear layers + ReLU) to further enhance the non-linear modeling capability. In addition, each sub-layer has a residual connection (Residual Connection) and layer normalization (LayerNorm) to ensure gradient stability and training efficiency.

[0047] The standard Transformer is designed for sequences, originally used to process natural language sequences, and its attention mechanism is highly sensitive to the order of positions, but when processing graph data, the nodes of the graph are unordered and the connection relationship is irregular, and the standard Transformer will lose the topological constraints when modeling the graph structure. To this end, the embodiment adopts a Transformer variant for graph data, Graph Transformer. Graph Transformer has been adaptively adjusted in structure. It not only retains the multi-head self-attention mechanism to capture the global interaction between nodes, but also introduces a guide mechanism related to the graph structure, which injects the adjacency matrix as a bias term into the attention calculation, so that the model can not only consider the feature similarity between each pair of nodes when calculating the attention weight, but also refer to their relative position and connectivity in the graph structure. In the task of predicting the performance of the whole vehicle based on the frame structure, Graph Transformer shows high applicability. The frame structure can be abstracted as a topological graph in essence, and the dynamics performance, structural strength and durability of the whole vehicle are not only affected by the local node attributes, but also by the overall structure configuration, force conduction path, loop structure and cross-block connection method. Therefore, although the graph convolutional network can extract the structure perception features of the local neighborhood, it is difficult to effectively perceive the complex relationship between distant nodes. Graph Transformer can realize the direct interaction of information between different regions of the frame through the global attention mechanism, so that the coupling effect of the chassis diagonal brace on the middle load-bearing beam, the adjustment effect of the edge support structure on the stiffness response of the whole vehicle and other long-range structure correlations can be modeled, thereby greatly improving the accuracy of structure performance prediction.

[0048] In the embodiment, the multi-task decoding module includes three prediction branches: static performance branch, modal frequency branch and fatigue life branch. Each branch adopts a fully connected network, including an input layer, an output layer and a plurality of hidden layers between the input and output layers. The input layer is used to receive the deep features output by the graph perception encoding module. The hidden layers can be set according to the specific branch task, such as Figure 1As shown, in this embodiment, the static performance branch contains two hidden layers, i.e., a fully connected layer FC64 with 64 nodes and a fully connected layer FC32 with 32 nodes. The hidden layers of the modal frequency branch are set the same as those of the static performance branch, and the hidden layers of the fatigue life branch are set to a fully connected layer FC128 with 128 nodes and a fully connected layer FC64. The static performance branch is used to predict the static performance indicators under typical working conditions, for example, the maximum stress and maximum deformation under four typical working conditions of bending, torsion, braking, and turning, a total of 8 indicators, and the output layer includes multiple (8) output nodes for outputting the corresponding performance indicators. The modal frequency branch is used to predict the low-order modal frequencies (the first 6 orders) of the frame, and the output layer contains output nodes for each order of modal frequency. The fatigue life branch is used to predict the fatigue life of the frame, and the output layer only contains one output node.

[0049] Step five: training of the image-aware multi-task prediction model In this embodiment, the image-aware multi-task prediction model is trained end-to-end using a weighted multi-task loss function. The loss terms include the prediction errors of the maximum stress, the maximum deformation, the modal frequency, and the fatigue life, each of which is measured by the mean square error (MSE). The overall loss function is written as:

[0050] wherein, , , , represent the stress, displacement, frequency, and fatigue life values predicted by the model, , , , are task weighting coefficients. The weights can be manually set according to the importance of the tasks, for example, set to a ratio of 1:1:0.5:2. The network parameter update uses the AdamW optimizer, and the initial learning rate is , combined with a linear warm-up of 5% training steps and a cosine annealing schedule; the weight decay coefficient is set to . To ensure the stability of the Transformer training, a Dropout rate of 0.1 (acting on the self-attention output and the feedforward network) is introduced, and a gradient clipping threshold of 1.0 is applied in the backpropagation stage. The early stopping mechanism monitors the comprehensive loss of the validation set, and if there is no significant improvement within 10 epochs, the training is terminated to prevent overfitting.

[0051] Step six: model deployment and inference The trained deep learning model is exported in ONNX format and deployed to a server or cloud platform through the TensorRT inference engine to provide real-time online prediction services. When inferring, the frame structure design drawing is constructed according to the method described above and input, and the maximum stress, maximum deformation, modal frequency and fatigue life prediction results under each typical working condition can be quickly obtained. The time consumption of a single inference is less than 20 ms, which meets the real-time requirements of actual engineering.

[0052] Figure 2 The R 2 The fitting comparison chart, (a) is the comparison chart of the maximum stress under the bending condition, (b) is the comparison chart of the maximum displacement under the torsion condition, (c) is the comparison chart of the first-order modal frequency, and (d) is the comparison chart of the fatigue life. As can be seen from the chart, the prediction output value of the trained image perception multi-task prediction model is basically consistent with the simulation analysis value, and the prediction error is within the allowable range.

[0053] Step seven: model incremental update mechanism This embodiment ensures that the model can adapt to new structure data and continuously optimize its prediction ability by establishing a periodic incremental update mechanism. Whenever simulation data or test data of new vehicle models are available, first standardize and process the data for outliers to ensure consistency between the new data and the original data.

[0054] During the incremental training process, the model is updated with a lower learning rate to ensure that only the parameters related to the new data are updated, avoiding the need to retrain the entire model, thereby improving computational efficiency. To prevent catastrophic forgetting, the elastic weight fixing (EWC) technique is used to protect key parameters, ensuring that the model does not lose old knowledge when learning new data.

[0055] The incremental update period is adjusted according to the frequency of new data arrival, and the update is usually started when new data is available. After the updated model is evaluated by the validation set and its prediction ability is confirmed, it is immediately put into use for real-time prediction. Through this incremental update mechanism, the model can continuously improve prediction accuracy and adapt to new designs and working conditions in long-term application.

[0056] Embodiment two Based on the above method, this embodiment provides a multi-modal data-driven commercial vehicle frame performance prediction system. The system is mainly composed of a frame structure coding module and a multi-task intelligent prediction module, and is equipped with a visual interactive interface and a performance display module to realize the automatic and intelligent processing flow of frame structure performance from modeling to prediction.

[0057] The system as a whole takes the structural diagram modeling and the deep graph neural network as the core, supports users to quickly evaluate the influence of the structural scheme on the overall vehicle performance in the early design stage, and intuitively presents the prediction results through the interactive interface to assist engineers in scheme screening and optimal design.

[0058] When the user inputs the design scheme of the frame to be predicted through the graphical interface, whether through parameter form filling, CAD model import or graphical depiction, the system will automatically convert the frame structure into a standardized structural design diagram. In this process, the system first performs grid division on the external boundary of the frame, dividing the frame into regular two-dimensional or three-dimensional grid cells, each cell corresponding to a graph node, and generating edges between nodes according to the connection relationship between regions in the frame design, and taking the number or complexity of connecting beams as the weight of the edge. Subsequently, the system extracts the structural parameter information in each grid cell, including the number of longitudinal and transverse beams, cross-section type, material properties, geometric dimensions, etc., and after standardization and one-hot encoding processing, a unified dimension node feature vector is formed. The node feature matrix and the normalized adjacency matrix together constitute the frame structural design diagram, which is the standard input representation of the structure to be predicted in the system.

[0059] After completing the structural modeling, the system sends the structural design diagram to the deployed multi-task intelligent prediction module. The module internally integrates a pre-trained graph perception multi-task intelligent prediction network, which is composed of a GIN graph convolutional encoder in the front end and a Graph Transformer encoder in the back end, and can extract the graph structure perception features of each node from the structural design diagram and realize global semantic interaction, thereby deeply understanding the influence of the overall configuration of the frame on performance. Finally, the output of the network is sent to a multi-task decoder, which respectively predicts the maximum stress and maximum deformation under typical working conditions such as bending, twisting, braking, turning, etc., the first six order modal frequencies, and the structural fatigue life and other key performance indicators.

[0060] To improve user experience and engineering decision-making efficiency, the system is also equipped with a performance display and human-computer interaction module. Users can view the structural modeling process, graphical layout of design variables and connection paths in real time in the visualization interface, and after the model inference is completed, the system automatically generates visualization charts of performance prediction results. Including stress distribution heat map, deformation trend curve, modal response spectrum and fatigue life estimate, etc. can be displayed in graphical form to help designers quickly identify structural weak points or performance bottlenecks. The system also supports comparative analysis of different frame design schemes, users can import multiple schemes, the system will complete structural coding and performance prediction in parallel, and output horizontal comparison chart and performance score report to assist structural optimization decision-making.

[0061] In addition, the system supports the prediction result export function, and users can export the structural diagram, prediction value and chart result in the form of a report as a PDF or Excel file for project archiving or cross-team communication. To adapt to the frequent changes in the design phase, the system also has a quick update mechanism. Once the design parameters are modified, the structural diagram will be automatically updated and re-predicted to ensure the real-time and consistency of the results.

[0062] The above system can perform the commercial vehicle frame performance prediction method described in embodiment one, has the corresponding functional modules and beneficial effects of the method, and the technical details not described in detail in this embodiment can be referred to the commercial vehicle frame performance prediction method provided in embodiment one of the present application.

[0063] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions or the part that contributes to the related art can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes a plurality of instructions to make a computer device (which can be a personal computer, server, or network device, etc.) execute the method described in each embodiment or some part of the embodiment.

[0064] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; under the idea of the present application, the technical features in the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other changes of the different aspects of the present application as described above. In order to be brief, they are not provided in detail; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A multi-modal data-driven commercial vehicle frame performance prediction method, characterized in that, The method comprises the steps of: constructing a frame structure design graph for reflecting a frame topology structure; inputting the frame structure design graph into a graph-aware multi-task prediction model to obtain predicted values of various performance indicators of the frame; the frame structure design graph comprises a plurality of nodes with feature vectors and a node adjacency matrix for representing connection relationships of the nodes, and the feature vectors of the nodes collectively constitute a node feature matrix; the graph-aware multi-task prediction model is trained by a multi-modal performance response dataset, and the prediction model comprises an input embedding module, a graph-aware encoding module, and a multi-task decoding module; the input embedding module is used to input the node feature matrix of the frame structure design graph and the node adjacency matrix after self-loop normalization processing; the graph-aware encoding module comprises a graph convolution encoder and a Transformer encoder, the graph convolution encoder is used to extract graph convolution features of the nodes in the frame structure design graph, and the Transformer encoder is used to perform deep modeling on the graph convolution features; the multi-task decoding module comprises a plurality of prediction branches, which are used to respectively predict various performance indicators of the frame according to deep features output by the Transformer encoder.

2. The commercial vehicle frame performance prediction method of claim 1, wherein, The construction of the frame structure design graph comprises: performing grid processing on an outer rectangle of the entire frame structure to uniformly divide the outer rectangle into a plurality of grids of the same size, each grid representing a node, and if there is a mutual connection or an integral longitudinal or transverse beam between the frame regions corresponding to adjacent grids, the corresponding two nodes are regarded as having a connection edge, and the number or order of the longitudinal or transverse beam connected to the edge is taken as the weight of the edge; quantitative coding is respectively performed on the frame design variables in the frame region corresponding to each grid, and the frame design variables are spliced into the feature vector of the corresponding node; the frame design variables include the number, geometric size, cross-section type and material type of the longitudinal or transverse beam, for the numerical continuous variable in the design variable, a standardization processing mode is used for normalization processing, and for the non-numerical discrete variable in the design variable, a one-hot encoding is used for numerical coding; The node adjacency matrix A is used to represent the connection relationship of each node, and the matrix element A ij is nonzero, which means that there is a connection edge between node i and node j. The weight of the edge is taken as the specific value of A ij , and the self-loop normalization is performed on the node adjacency matrix A.

3. The commercial vehicle frame performance prediction method of claim 1, wherein, the multi-modal performance response dataset comprises a plurality of training samples, each training sample comprises a frame structure design graph and corresponding performance indicator values, the performance indicators include maximum stress and maximum deformation under various typical working conditions, low-order modal frequency and fatigue life; the multi-modal refers to the diversity of performance data sources and data types, which includes both data generated by numerical simulation and data collected in real bench or real vehicle tests, for unified data representation, the multi-modal data needs to be preprocessed; the preprocessing comprises: firstly, performing anomaly detection and cleaning on all original data, eliminating outlier data, and performing oversampling expansion processing on a small number of class data, then unifying the dimensions and coordinate systems of the simulation data and the test data, and aligning the distributions of the same performance indicators in the two types of data sources, then performing standardization processing, and finally fusing the simulation data and the test data of the same frame according to the weight coefficients determined by the Bayesian optimization algorithm.

4. The commercial vehicle frame performance prediction method of claim 3, wherein, The SMOTE-NC method is used to alleviate the class imbalance problem and enhance the representativeness of boundary samples during the oversampling expansion process, and the rescaling and sample weighting methods are used to optimize the balance of the dataset samples. The MMD algorithm is used to align the same performance indicators in the two data sources during distribution alignment, minimizing the distribution difference caused by different data sources, and the CORAL algorithm is used to optimize the statistical consistency between the data sources, ensuring the alignment of the simulation data and the test data in the high-order statistical distribution, and improving the quality and stability of the data fusion.

5. The commercial vehicle frame performance prediction method of claim 1, wherein, The graph convolutional encoder comprises a plurality of front and rear connected graph convolutional layers, which adopt GCN based on mean aggregation or GAT or GIN based on attention mechanism.

6. The commercial vehicle frame performance prediction method of claim 1, wherein, The Transformer encoder comprises a plurality of stacked encoder layers, each of which comprises a multi-head self-attention mechanism and a feed-forward neural network, the multi-head self-attention mechanism calculates the attention weight of each node with other nodes by taking the input of each node as a query, key and value vector, thereby realizing global interaction and fusion of features, and each node is independently further enhanced by a feed-forward neural network after the attention layer, thereby improving the nonlinear modeling capability, and each sub-layer has a residual connection and layer normalization to ensure gradient stability and training efficiency.

7. The commercial vehicle frame performance prediction method of claim 1, wherein, The Transformer encoder adopts Graph Transformer, and the node adjacency matrix is injected as a bias term into the attention calculation.

8. The commercial vehicle frame performance prediction method of claim 1, wherein, The multi-task decoding module comprises three prediction branches: a static performance branch, a modal frequency branch and a fatigue life branch, each branch adopts a fully connected network comprising an input layer, an output layer and a plurality of hidden layers between the input and output layers, and the input layer is used to receive the deep features output by the graph perception coding module; the static performance branch is used to predict the static performance indicators under typical working conditions, including a total of 8 indicators of maximum stress and maximum deformation under four typical working conditions of bending, torsion, braking and turning, and the modal frequency branch is used to predict the low-order modal frequencies of the vehicle frame, including the first six-order modal frequencies.

9. A commercial vehicle frame performance prediction system based on the method of any one of claims 1 to 8, characterized by The system comprises a vehicle frame structure coding module and a multi-task intelligent prediction module, the vehicle frame structure coding module is used to construct a vehicle frame structure design diagram of a vehicle frame to be predicted, and the multi-task intelligent prediction module is provided with a trained graph perception multi-task prediction network, which is used to output corresponding performance prediction values according to the input vehicle frame structure design diagram of the vehicle frame to be predicted.

10. The commercial vehicle frame performance prediction system of claim 9, wherein, The system is also provided with a performance display and human-computer interaction module, and the user can view the structure modeling process, the graphical layout and connection path of the design variables in real time in the visual interface, and the system automatically generates a visual chart of the performance prediction results after the model inference is completed. The system also supports comparative analysis of different vehicle frame design schemes, and can perform performance prediction of multiple vehicle frame design schemes input by the user in parallel and output a horizontal comparison chart and a performance score report. The system also supports a prediction result export function, which exports the prediction values and chart results in the form of a report as a PDF or Excel file. The system also supports a prediction result export function, which exports the prediction values and chart results in the form of a report as a PDF or Excel file.

Images