Multi-modal data fusion representation method and system for commercial vehicle frame performance prediction

Abstract

The application provides a kind of commercial vehicle frame performance prediction multi-modal data fusion representation method and system, applied to vehicle frame data modeling technical field, the method first carries out automatic mining to multi-source heterogeneous engineering data, extracts structure topological parameter, section configuration parameter, material attribute parameter, connection parameter, P-S-N curve data and field response data, and forms target domain data set and independent test set;Then establish the full parameterization finite element model of frame, combined with the virtual load sample generated by statistical derivation of measured load spectrum, extract two-dimensional load damage feature matrix, and construct source domain data set;Finally, the geometric topological feature, parameter feature, connection feature and load feature are mapped to a unified three-dimensional voxel space, generating a three-dimensional multi-channel engineering data tensor.The application can improve the efficiency of commercial vehicle frame sample construction, multi-modal data utilization efficiency and the generalization ability of performance prediction model.

Description

Technical Field

[0001] This invention belongs to the field of chassis data modeling technology, specifically relating to a multimodal data fusion characterization method and system for predicting the performance of commercial vehicle chassis. 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 pressures of increased power battery weight and range anxiety. This inherent conflict between structural load-bearing requirements and overall vehicle energy efficiency goals forces chassis design to achieve extreme lightweighting while ensuring high reliability. However, due to the strong nonlinearity of chassis fatigue damage evolution and the discreteness of manufacturing processes, traditional R&D models face significant bottlenecks: on the one hand, performance prediction methods based on finite element analysis (FEA) have long computation times and significant accuracy problems when solving nonlinear fatigue damage; on the other hand, iterative optimization methods based on surrogate models require a large number of sample points, have limited design space, slow convergence speed, and large optimization margins. This current state of technology, which struggles to balance prediction confidence and optimization convergence efficiency, is no longer sufficient to support the current R&D needs of commercial vehicle products, which require multiple configurations, extreme lightweighting, and a fast pace.

[0003] Given the inherent advantages of deep learning technology in capturing strong nonlinear mapping relationships and achieving second-level inference and prediction, constructing a high-precision prediction and intelligent design method for vehicle frame performance based on deep learning has become a key path for the industry to overcome the bottlenecks in R&D accuracy and efficiency. However, when applying deep learning technology to commercial vehicle frame R&D, it faces a dual technical bottleneck: "high-quality sample data acquisition" and "multi-modal feature fusion representation."

[0004] On the one hand, at the level of sample data acquisition and generation, existing technologies struggle to simultaneously address the mining of unstructured historical data and the efficient generation of globally topologically parameterized samples. The generalization accuracy of deep learning models is highly dependent on the scale, quality, and distribution diversity of training samples. Currently, although commercial vehicle companies and the industry have accumulated massive amounts of R&D data, this data is widely dispersed across public literature, industry databases, and internal company reports, manifesting in heterogeneous forms such as tables, experimental scalar results, characteristic curves, and fatigue life cloud maps. The lack of effective mining methods in existing technologies has resulted in the inability to fully utilize this highly valuable historical engineering experience data as effective training data samples.

[0005] Regarding simulation sample generation, existing technologies typically employ local parametric sampling or single load condition compilation, lacking comprehensive coverage. For example, Chinese patent application CN202511352601.3 discloses a machine learning-driven lightweight chassis design method. While this method constructs a sample space, it primarily samples dimensional variables such as plate thickness, failing to cover variations in topological features like beam layout and connection parameters, resulting in insufficient sample richness. Similarly, Chinese patent application CN202411577629.2 discloses an accelerated compilation method for multidimensional load spectra of chassis under WLTC conditions. This method focuses on extracting a single standard test spectrum from standard cyclic conditions, without involving statistical derivation techniques for load spectra based on measured road spectrum characteristics. This makes it impossible to generate virtual load samples in batches that combine physical realism with diverse operating conditions, limiting the generalization of data across varying operating conditions.

[0006] On the other hand, at the level of multimodal data fusion and representation, existing technologies, such as Chinese patent application number CN202511445349.0, disclose a multimodal data-driven method and system for predicting the performance of commercial vehicle frames. The graph representation method it adopts mainly focuses on the topological connections of discrete nodes. It has inherent limitations in expressing cloud map features such as stress and lifespan that are continuously distributed throughout the field, resulting in the loss of spatial information due to discretization. Moreover, when fusing material properties and two-dimensional time-varying load features, this method mainly relies on distribution alignment and weighted fusion, lacking a voxelized spatial semantic mapping mechanism based on a unified coordinate system, which makes it difficult to achieve deep coupling of multi-source features in the physical channel dimension.

[0007] In summary, there is an urgent need in the field of commercial vehicle chassis R&D for a data engineering method that can achieve automated mining of multi-source heterogeneous engineering data, full-domain enhancement of structural and working condition samples, and unified representation of physical space semantics. 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 multimodal data fusion and characterization method for predicting the performance of commercial vehicle frames. This method realizes the automated mining, global enhancement and unified physical semantic characterization of multi-source heterogeneous engineering data of commercial vehicle frames, thereby improving the scale, diversity and availability of training samples, and providing a standardized data foundation for intelligent prediction of frame performance and cross-topology generalization.

[0009] A multimodal data fusion characterization method for predicting the performance of commercial vehicle chassis includes the following steps:

[0010] S1. Automated mining of multi-source heterogeneous engineering data, identifying and extracting structural topology parameters, cross-sectional construction parameters, material property parameters, connection parameters, PSN curve data and field response data related to cross-topology frames, performing dimensional unification and coordinate space alignment on the extracted results, and encapsulating them into a target domain dataset for model virtual-real migration calibration and an independent test set for verifying the model's cross-topology generalization ability.

[0011] S2. Establish a fully parameterized finite element model of the chassis, and adjust the structural topology variables, cross-sectional construction variables, material property variables and connection parameter variables in a coordinated manner. Combine experimental design sampling to generate simulation samples with different parameter combinations, and generate virtual load samples through statistical analysis of measured load spectra. Extract a two-dimensional load damage feature matrix, and encapsulate the parameter combinations, the two-dimensional load damage feature matrix and the corresponding simulation truth labels to form a source domain dataset.

[0012] S3. Map the multimodal data obtained in steps S1 and S2 to a unified three-dimensional voxel space, construct geometric topology channels, attribute distribution channels, connection feature channels and load feature channels respectively, and stitch the channels together to generate a three-dimensional multi-channel engineering data tensor, which serves as the input representation for the commercial vehicle frame performance prediction model.

[0013] Preferably, in step S1, the automated mining of multi-source heterogeneous engineering data includes:

[0014] Document layout analysis was performed on the multi-source heterogeneous engineering data to segment it into a text area, a table area, and a field response contour map area; the field response contour map area includes stress contour maps, deformation contour maps, fatigue life contour maps, or damage contour maps.

[0015] Optical character recognition is performed on the text area and table area to extract structural topology parameters, cross-sectional construction parameters, material property parameters, and connection parameters;

[0016] The field response cloud map area is preprocessed to extract the field feature target area according to the preset saturation threshold and brightness threshold, and the field feature target area is digitally reverse-converted to extract PSN curve data and field response data.

[0017] Preferably, in step S1, performing a digital reverse transformation on the field response cloud map region includes:

[0018] Skeleton extraction is performed on the material property curve image to obtain the center contour of the PSN curve, and the numerical stress-life data sequence is reconstructed by combining the coordinate axis scale recognition results.

[0019] Pixel-level color inversion processing is performed on the field response cloud map. The pixel color vector of each pixel is matched with the standard color vector in the scale area to determine the corresponding physical quantity value, and a field response dataset with image coordinate index is generated.

[0020] Preferably, in step S1, after completing the unification of dimensions and alignment of coordinate space, the processed parameter data and PSN curve data are used as input features, the corresponding test performance scalar and the spatially aligned field response data are used as truth labels for sample encapsulation, and the topological similarity between the sample and the frame to be designed is calculated by extracting the topological feature vector of the structured sample. Samples with similarity higher than a preset threshold are divided into target domain datasets, and samples with similarity lower than the preset threshold are divided into independent test sets.

[0021] Preferably, in step S2, the design variables of the fully parametric finite element model include structural topology variables, cross-sectional construction variables, material property variables, and connection parameter variables;

[0022] The structural topology variables include the number of beams and their longitudinal distribution positions; the cross-sectional construction variables include the selection of cross-sectional shapes, cross-sectional dimensions, and plate thickness parameters for longitudinal and transverse beams; the material property variables include the material grade; and the connection parameter variables include the number, diameter, and spacing of bolts or rivets.

[0023] Preferably, in step S2, a multi-dimensional joint sampling is performed in the preset variable value space using an experimental design sampling algorithm to determine the parameter combination of each frame finite element model and to construct a simulation task flow;

[0024] The simulation task flow includes at least free modal analysis, typical working condition static analysis, bending stiffness analysis, torsional stiffness analysis, and fatigue life analysis; the typical working condition static analysis includes at least bending working condition, torsional working condition, braking working condition, and turning working condition.

[0025] Preferably, the fatigue life analysis includes:

[0026] Obtain the full-field unit stress response of each parameter combination under a unit load applied at a preset load application point;

[0027] The load time history can be obtained by solving the virtual prototype model of commercial vehicle and the road surface model, or the load spectrum data of the real vehicle collected in step S1 can be directly used as the load time history input.

[0028] Based on the principle of linear superposition, the full-field unit stress response is superimposed with the load time history to synthesize the actual stress history, and rainflow counting is performed on the actual stress history. The cumulative damage is calculated using the fatigue damage accumulation theory in combination with the material PSN curve.

[0029] The lifespan value is calculated based on the cumulative damage, and the minimum lifespan value across the entire field is taken as the true value of the frame fatigue life.

[0030] Preferably, in step S2, the statistical derivation of the measured load spectrum includes:

[0031] Rainflow counting was performed on the measured road spectrum to obtain the baseline rainflow matrix;

[0032] Statistical analysis was performed on the joint distribution characteristics of load amplitude and mean in the benchmark rainflow matrix, and a statistical model was established.

[0033] Based on the statistical model, random perturbation and probability sampling are performed to generate virtual rainflow matrix samples;

[0034] The virtual rainflow matrix sample is input into a two-dimensional residual convolutional neural network to extract a two-dimensional load damage feature matrix.

[0035] Preferably, step S3 includes:

[0036] The chassis mesh model is transformed into a three-dimensional geometric topology channel using voxel mapping.

[0037] The structural topology parameters, cross-sectional construction parameters, and material property parameters are mapped to the corresponding solid voxel space to construct property distribution channels.

[0038] The connection parameters are encoded and filled into the corresponding voxel units of the connection region to construct the connection feature channels;

[0039] The two-dimensional load damage feature matrix is ​​mapped to the voxel unit corresponding to the key stress point or sensor installation position, and local spatial diffusion is performed to construct the load feature channel.

[0040] The three-dimensional geometric topology channel, attribute distribution channel, connection feature channel, and load feature channel are spliced ​​together to generate a three-dimensional multi-channel engineering data tensor.

[0041] The second objective of this invention is to propose a multimodal data fusion characterization system for predicting the performance of commercial vehicle chassis, which executes the aforementioned multimodal data fusion characterization method for predicting the performance of commercial vehicle chassis. The system includes:

[0042] The multi-source data automated mining module is used to perform identification, extraction, digitization, unit unification, coordinate space alignment, and dataset partitioning on multi-source heterogeneous engineering data.

[0043] The multi-source data automated mining module includes a document layout analysis unit, an optical character recognition unit, a digital image processing unit, a coordinate space calibration unit, and a dataset management unit. The document layout analysis unit divides the input document into text areas, table areas, and field response cloud map areas. The optical character recognition unit extracts structural topology parameters, cross-sectional construction parameters, material property parameters, and connection parameters. The digital image processing unit extracts PSN curve data and field response data. The coordinate space calibration unit performs dimensional unification and coordinate space alignment on the extracted data. The dataset management unit encapsulates the data to form a target domain dataset and an independent test set.

[0044] The simulation data generation and enhancement module is used to perform parametric simulation sample generation of the chassis, statistical derivation of measured load spectrum, load feature extraction, and source domain dataset construction.

[0045] The simulation data generation and enhancement module includes a parametric modeling engine, a simulation task flow unit, a measured load spectrum statistical derivation unit, and a two-dimensional feature extraction unit. The parametric modeling engine is used to establish a fully parametric finite element model of the chassis. The simulation task flow unit is used to output simulation truth labels. The measured load spectrum statistical derivation unit is used to generate virtual rainflow matrix samples. The two-dimensional feature extraction unit is used to extract a two-dimensional load damage feature matrix. The simulation data generation and enhancement module is also used to encapsulate and form a source domain dataset.

[0046] The multimodal feature fusion and representation module is used to map the multimodal data in the target domain dataset, independent test set and source domain dataset to a unified three-dimensional voxel space and output a three-dimensional multi-channel engineering data tensor;

[0047] The multimodal feature fusion and characterization module includes a voxelization encoding engine, an attribute feature mapping unit, a load field spatial diffusion unit, and a tensor splicing unit. The voxelization encoding engine is used to construct geometric topology channels, the attribute feature mapping unit is used to construct attribute distribution channels and connection feature channels, the load field spatial diffusion unit is used to construct load feature channels, and the tensor splicing unit is used to splice the channels to form a three-dimensional multi-channel engineering data tensor.

[0048] The multi-source data automated mining module, the simulation data generation and enhancement module, and the multimodal feature fusion and characterization module are communicatively connected.

[0049] The beneficial effects of this invention are:

[0050] This invention establishes an automated mining mechanism for multi-source heterogeneous engineering data in commercial vehicle chassis R&D scenarios. It can uniformly identify, extract, and digitize text, tables, performance curves, and field response cloud maps from public literature, industry manuals, databases, and internal R&D and test reports. This transforms historical engineering data that was originally scattered, unstructured, and difficult to use directly into standardized sample data that can be used for model training, significantly improving the efficiency of acquiring, organizing, and reusing heterogeneous engineering data.

[0051] This invention constructs a global parametric simulation framework for the vehicle frame and combines it with statistical derivation techniques based on measured load spectra. This enables the joint expansion of training samples in both structural topology and load condition dimensions. It not only covers variations in multiple design variables such as structural topology, cross-sectional construction, material properties, and connection parameters, but also preserves the physical characteristics of measured loads and forms virtual load samples with statistical diversity. This effectively improves the scale, richness, and distribution coverage of the source domain dataset, providing more sufficient data support for the training and generalization of subsequent models under multiple configurations and load conditions.

[0052] This invention establishes a spatial semantic mapping mechanism for geometric topological features, attribute features, connectivity features, and load features in a unified three-dimensional voxel space. By constructing three-dimensional geometric topological channels, three-dimensional attribute distribution channels, three-dimensional connectivity feature channels, and three-dimensional load feature channels, and further forming a unified format of three-dimensional multi-channel engineering data tensors, it realizes the collaborative expression of multimodal engineering data in the same physical space coordinate system. This effectively avoids the problem of spatial correlation information loss in existing discretization splicing or simple graph structure representation methods, making the obtained data representation results more in line with the requirements of frame mechanical analysis characteristics and performance prediction.

[0053] This invention constructs a data architecture that combines a target domain dataset, an independent test set, and a source domain dataset. This architecture enables high-value samples mined from historical engineering data to work synergistically with large-scale samples derived from parametric simulation and load spectra. This provides a unified data foundation for subsequent virtual-to-real migration calibration, cross-topology generalization verification, and performance prediction optimization of the model. Compared to existing technologies that rely on numerous physical experiments or local sample modeling, this invention reduces the dependence on high-value test samples while improving data utilization efficiency and model training feasibility. This is beneficial for shortening the data preparation and solution iteration cycles in commercial vehicle chassis development. Attached Figure Description

[0054] 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:

[0055] Figure 1 This is a flowchart of the method of the present invention;

[0056] Figure 2 This is a flowchart of the method for automatic mining of multi-source heterogeneous engineering data according to the present invention;

[0057] Figure 3 This is a flowchart of the source domain sample numerical simulation generation and load spectrum statistical derivation process of the present invention;

[0058] Figure 4 This is a schematic diagram of the multimodal feature space semantic mapping mechanism of the present invention;

[0059] Figure 5 This is a schematic diagram of the three-dimensional multi-channel engineering voxelized data tensor structure of the present invention. Detailed Implementation

[0060] 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. Without departing from the concept of this invention, those skilled in the art can make equivalent substitutions or conventional modifications to the technical features therein, and such substitutions or modifications should all fall within the scope of protection of this invention.

[0061] It should be noted that the commercial vehicle frame described in this invention can be an aluminum alloy frame, a steel frame, a magnesium alloy frame, or a steel-aluminum hybrid frame. The following embodiments use an aluminum alloy commercial vehicle frame as an example for illustration, but this invention is not limited to frame structures made of specific materials.

[0062] Example 1

[0063] like Figure 1 As shown in the figure, this embodiment provides a multimodal data fusion characterization method for predicting the performance of commercial vehicle chassis. The method includes the following steps:

[0064] Step S1: Automated Mining of Multi-Source Heterogeneous Engineering Data

[0065] This step is used to extract structural information, material information, field response information, and experimental performance information from multi-source heterogeneous engineering data that can be used for model training and calibration, and to construct a target domain dataset for model virtual-to-real migration calibration and an independent test set for verifying cross-topology generalization ability.

[0066] like Figure 2 As shown, the specific process is as follows:

[0067] S11: Heterogeneous Document Layout Analysis and Region Segmentation

[0068] First, multi-source heterogeneous engineering data is acquired, including but not limited to publicly available literature, industry design manuals, standard databases, and internal R&D reports and test reports. Document layout analysis is then performed on the multi-source heterogeneous engineering data to automatically identify and segment unstructured text areas, structured table areas, and field response contour map areas. The field response contour map area may include stress contour maps, deformation contour maps, fatigue life contour maps, or damage contour maps. In this embodiment, the document layout analysis can preferably be implemented using a deep learning-based document layout analysis algorithm, such as the LayoutLMv3 model.

[0069] For the segmented field response cloud region, image preprocessing is performed to convert the image from RGB color space to HSV color space, and generate a mask according to preset saturation and brightness thresholds to isolate the background grid, coordinate axes, and text annotations, thereby extracting the pure field feature target region. In this embodiment, a saturation threshold can be set during image preprocessing. Brightness threshold However, the present invention is not limited thereto.

[0070] S12: Text and Tabular Data Recognition and Extraction

[0071] Optical character recognition (OCR) is performed on the text area and table area identified in step S11 to extract key physical parameters of the cross-topology frame, including structural topology parameters, cross-sectional construction parameters, material property parameters, and connection parameters.

[0072] Among them, the structural topology parameters are used to characterize the overall configuration of the frame and the arrangement of components, including at least the number of crossbeams, the coordinates of the distribution position of each crossbeam along the longitudinal direction of the frame, and the selection of the cross-sectional shape of the longitudinal beams and crossbeams.

[0073] Cross-sectional structural parameters are used to characterize the cross-sectional geometric features of vehicle frame components, including at least one or more of the following parameters: web height, flange width, cross-sectional opening size, closing size, and plate thickness.

[0074] Material property parameters are used to characterize the mechanical properties of the corresponding vehicle frame components. They include at least the material grade and the corresponding density, elastic modulus, Poisson's ratio, yield strength, tensile strength, static mechanical property parameters, dynamic mechanical property parameters, and fatigue characteristic curve parameters. After identifying the material grade, a preset material database can be called to automatically complete the missing attributes corresponding to the grade, so as to form a complete material property vector.

[0075] Connection parameters are used to characterize the connection structure features of each frame connection node, including at least the specifications, diameter, quantity, and spacing of bolts or rivets used at the connection between the crossbeam and the longitudinal beam; in some embodiments, the connection layout, edge distance, row spacing, and connection area location index information can be further extracted.

[0076] In one specific embodiment, using a heavy truck chassis test report as the input document, after performing OCR recognition on the text area and table area, the following key physical parameters are extracted:

[0077] Structural topology parameters: The frame contains 6 crossbeams, and the coordinates of each crossbeam along the longitudinal direction of the frame are analyzed and recorded; at the same time, the cross-sectional shape of the longitudinal beams is selected as a channel section, and the cross-sectional shape of the crossbeams is selected as a channel section or a box section.

[0078] Sectional structural parameters: The web height of the longitudinal beam is 200mm, the flange width is 90mm, and the plate thickness is 8mm; if other section description information exists, the corresponding opening or closing dimensions can be further extracted.

[0079] Material property parameters: The material grade used for the vehicle frame components is 6061-T6 or 7075-T6 aluminum alloy. After identifying the material grade, the preset material database is called to complete its detailed properties. Taking 6061-T6 as an example, the completed properties are: density 2.7 g / cm³, elastic modulus 68.9 GPa, Poisson's ratio 0.33, and yield strength 276 MPa.

[0080] For the connection parameters, at the connection node between the third crossbeam and the longitudinal beam, the connector consists of 8 M14 high-strength bolts with a connection spacing of 50mm; in the implementation scenario using riveting connection, the specifications, quantity, and spacing parameters of the corresponding rivets are extracted.

[0081] Through the above processing, the original unstructured text information and structured table information can be transformed into a standardized set of structural parameters that can be directly called by subsequent parametric modeling, simulation sample generation and voxelized feature mapping.

[0082] S13: Digital Reverse Conversion of Field Feature Images

[0083] The field feature target region extracted in step S11 is subjected to digital reverse transformation processing to restore the material property curve information and field response distribution information carried in image form into calculable and storable numerical physical data. This digital reverse transformation processing includes two parts: digital processing of material property curves and numerical restoration processing of field response contour maps.

[0084] One aspect involves extracting the curve skeleton from the material property curve image to obtain the curve's center contour. This skeleton is then combined with coordinate axis scale recognition results to reconstruct the corresponding data point sequence, thus transforming the image-based PSN curve (Probability-Stress-Life curve) into a numerical stress-life data sequence. The PSN curve characterizes the relationship between material stress levels and fatigue life under different failure probabilities. This embodiment employs a skeleton extraction algorithm (such as the Zhang-Suen thinning algorithm) for this purpose.

[0085] After obtaining the curve skeleton, the coordinates of the skeleton pixels can be mapped by combining the coordinate axis scale recognition results to generate a set of curve data points with physical dimensions.

[0086] On the other hand, pixel-level color inversion processing is performed on the field response contour map. Specifically, based on a preset color mapping relationship, each pixel in the field response contour map is traversed. The pixel color vector corresponding to each pixel is extracted and matched with the standard color vector in the scale area to determine the physical quantity value corresponding to the pixel. Then, the physical quantity value is associated with the original image coordinate index of the pixel to generate a field response dataset with image coordinate information. The physical quantity value includes at least one of the following: deformation value, stress value, fatigue life value, or damage value.

[0087] In one specific implementation, color matching is performed using the minimum Euclidean distance matching method. For any pixel, its pixel color vector is denoted as... The standard color vector in the scale is denoted as Calculate the Euclidean distance between the pixel color vector and each standard color vector. Based on the principle of minimum distance, the standard color closest to the pixel is determined, and the scale physical quantity value corresponding to the standard color is used as the target physical quantity value of the pixel. By performing the above processing on all pixels of the field response cloud map, an original field dataset with original image coordinate index can be constructed, providing basic data support for subsequent unit unification, coordinate alignment, and multimodal feature fusion.

[0088] S14: Dimensional unification and coordinate space alignment

[0089] A pre-defined dimensional conversion rule library is retrieved, and dimensional standardization processing is performed on data items with physical dimensions to uniformly convert heterogeneous parameters from different sources and with different unit representations to the same standard physical unit system, thereby eliminating differences in unit systems and dimensional expressions between different data sources. In this embodiment, the standard physical unit system is the N-mm-s unit system.

[0090] After unifying the dimensions, a global geometric reference coordinate system for the chassis is established. Based on the structural boundary features in the field response cloud map and the sensor installation location, the image coordinate system and the global geometric reference coordinate system are determined. The spatial transformation relationship between them is determined. The image coordinate index of each pixel in the original field dataset is mapped to the global geometric reference coordinate system to achieve the alignment transformation from image data to vehicle frame physical space data.

[0091] In one specific implementation, the spatial transformation relationship is represented by an affine transformation model as follows:

[0092] (1)

[0093] in, This represents the coordinates of a pixel in the image coordinate system. This represents the spatial coordinates in the global geometric reference coordinate system after mapping. This represents the affine transformation matrix from the image coordinate system to the global geometric reference coordinate system.

[0094] Through the above processing, data from heterogeneous sources can be unified into standardized physical data with consistent dimensions and spatial location semantics, providing a foundation for subsequent training sample encapsulation and multimodal feature space mapping.

[0095] S15: Sample Packaging and Dataset Partitioning

[0096] The parameter data and PSN curve data obtained after step S14 are used as sample input features, and the corresponding test performance scalars and spatially aligned field response data are used as sample ground truth labels. These are then associated and encapsulated to generate structured training samples. Among them, the test performance scalars include at least the frame mass, low-order modal frequencies, maximum stress and maximum deformation, bending stiffness, torsional stiffness, and fatigue life under bending, torsional, cornering, and braking conditions.

[0097] After completing the sample encapsulation, topological feature vectors of each structured sample are extracted to construct the target domain dataset for model virtual-to-real transfer calibration and the independent test set for verifying cross-topology generalization ability. The topological similarity between the mined samples and the chassis to be designed is calculated. The topological feature vector is used to characterize the differences in chassis configuration, reflecting at least the number of crossbeams, the distribution location of crossbeams, and the overall configuration characteristics.

[0098] In one specific implementation, cosine similarity is used to calculate the topological similarity between samples. Its expression is:

[0099] (2)

[0100] in, This represents the topological feature vector of the mined sample. This represents the topological feature vector of the chassis to be designed. Based on the similarity calculation results, samples with similarity higher than a preset threshold (e.g., 0.8) are divided into the target domain dataset for subsequent model transfer calibration; samples with similarity lower than the preset threshold are divided into an independent test set to evaluate the model's predictive performance under heterogeneous topology conditions.

[0101] Through the above processing, a standardized sample set containing both input features and ground truth labels can be formed, and a hierarchical dataset construction for virtual-to-real transfer and cross-topology generalization verification can be achieved.

[0102] Step S2: Global Parametric Simulation and Statistical Derivation of Measured Load Spectrum

[0103] This step is used to construct a source domain dataset that covers the design space and load condition distribution of the vehicle frame structure through fully parametric finite element modeling, simulation task flow solving, and statistical derivation of measured load spectra.

[0104] like Figure 3 As shown, a parametric simulation model that can be dynamically adjusted is first established based on the vehicle frame structure topology, material properties, cross-sectional structure, and connection parameters. Then, sample parameter combinations are determined through experimental design, and multi-task performance solutions are executed. Simultaneously, virtual load samples are derived based on measured road spectra, and load damage features are extracted. Finally, the source domain dataset is encapsulated. The specific steps include the following:

[0105] S21: Construction of a Fully Parametric Finite Element Model

[0106] A fully parameterized finite element model of the chassis is established based on the finite element preprocessing platform, and the design variables are adjusted in a linked manner using automated scripts to generate finite element simulation models of chassis with different structural configurations and process properties in batches.

[0107] The design variables include at least: structural topology variables characterizing the overall frame configuration, cross-sectional construction variables characterizing the geometry of component sections, material property variables characterizing material selection, and connection parameter variables characterizing connection construction. By parametrically defining the design variables, longitudinal beams, transverse beams, and their connection nodes can automatically complete geometric reconstruction, attribute assignment, and connection relationship updates as parameters change, thereby forming batch simulation models for different topological configurations.

[0108] In one specific implementation, automatic modeling can be achieved based on the Altair HyperMesh platform and using Tcl / Tk scripts. An example of design variable parameter settings is as follows:

[0109] The structural topology variables include the number of beams and their longitudinal distribution. The number of beams is set to vary between 4 and 8, and the longitudinal position of the beams can be adjusted by sliding within ±100mm of the reference position.

[0110] The cross-sectional construction variables include the selection of the cross-sectional shape, cross-sectional dimensions and plate thickness parameters of the longitudinal beams and transverse beams. The web height of the longitudinal beams is set to 200mm to 400mm, the flange width to 80mm to 120mm, and the plate thickness to 6mm to 12mm. The cross-section of the transverse beams can be switched between circular tube, channel, I-beam or box-shaped sections depending on the arrangement position.

[0111] The material property variables include the material grades used in the vehicle frame components. The material grades can be selected from a discrete aluminum alloy grade library or a high-strength steel grade library. The aluminum alloy grades include at least 6061 and 7075, and the high-strength steel grades include at least Q345 and 510L.

[0112] Connection parameter variables include the number, diameter, and spacing of bolts or rivets. The number of connectors can be adjusted between 4 and 12, the diameter can be switched between M12, M14, and M16, and the spacing is 40mm to 80mm.

[0113] S22: Experiment Design and Task Flow Construction

[0114] After constructing the fully parameterized finite element model, a design-of-experiments sampling algorithm is used to perform multi-dimensional joint sampling within a preset variable value space to determine the parameter combinations for each chassis finite element simulation model. Based on this, a unified simulation task flow is constructed. This simulation task flow includes at least free modal analysis, static analysis under typical operating conditions (bending, torsional, braking, and cornering), bending stiffness analysis, torsional stiffness analysis, and fatigue life analysis over the entire life cycle. By inputting different parameter combinations into the simulation task flow, performance samples for multiple topologies, materials, connection types, and operating conditions can be generated in batches.

[0115] In one specific implementation, the experimental design sampling algorithm can adopt the Latin hypercube sampling (LHS) algorithm to improve the uniformity of sample distribution in the high-dimensional design space; the number of sample parameter combinations can be set according to the design space range, for example, more than 5,000 parameter combinations can be generated.

[0116] S23: Performance Simulation Calculation

[0117] The finite element solver (such as nCode or OptiStruct) is driven to execute the simulation task flow, obtain the performance results corresponding to each parameter combination, and generate the response data required for subsequent truth labeling.

[0118] For free modal analysis, the low-order modal frequencies of the frame can be obtained; for static analysis under typical working conditions, the stress and deformation responses under bending, torsion, braking and turning conditions can be obtained; for stiffness analysis, bending stiffness and torsional stiffness can be obtained; for fatigue life analysis, fatigue damage and life distribution are further solved by combining the unit load response and load history superposition method.

[0119] Specifically, fatigue life analysis includes the following process:

[0120] (1) Under the condition of obtaining various combinations of structural parameters through finite element analysis, apply a unit load at each preset load application point ( The full-field unit stress response tensor at time ) The preset load application points are set at key locations on the vehicle frame structure, including at least one or more of the following: suspension mounting location, powertrain mounting location, steering system mounting location, superstructure connection location, and large mass component support location.

[0121] (2) Obtain the payload time history Specifically, Adams Car is used to establish a virtual prototype model of a commercial vehicle and a road surface model. The measured road surface roughness excitation file corresponding to the seed load spectrum in step S25 is imported to extract the load time history of the connection points between the vehicle frame and each assembly. The measured road surface roughness excitation file can correspond to at least one typical scenario among Class B road surface, cobblestone road or Belgian road. Furthermore, when the load spectrum data collected from the actual vehicle has been obtained in step S1, the load spectrum data collected from the actual vehicle can be directly used as the load time history input.

[0122] (3) Based on the principle of linear superposition, the unit stress response is superimposed with the load time history to synthesize the actual stress history. The rainflow is counted, and the cumulative damage is calculated using Miner's linear fatigue damage accumulation theory, combined with the material PSN curve extracted in step S1.

[0123] In one specific implementation, the cumulative damage can be expressed as:

[0124] (3)

[0125] in, For the first The number of cycles corresponding to the stress amplitude. The formula for calculating the corresponding fatigue life limit obtained from the material's PSN curve is as follows: ;in, Indicates fatigue life. Indicates the stress amplitude. These are constants related to the fatigue properties of materials; The stress amplitude level is extracted from the rainflow matrix.

[0126] (4) Calculate the lifespan value based on the cumulative damage, and take the minimum lifespan value across the entire field as the true value of the chassis fatigue life. Simultaneously, output the corresponding damage distribution or lifespan distribution cloud map across the entire field. Wherein, the true value of the chassis fatigue life... The calculation formula is: .

[0127] S24: Truth Label Extraction

[0128] The simulation result file (such as .op2 or .h3d) is automatically parsed, and the performance scalars and full-field response data of each simulation sample are extracted as truth labels. In this embodiment, the truth label Y includes the frame mass, the first six modal frequencies, the maximum von Mises stress under each typical working condition, the maximum deformation, bending stiffness, torsional stiffness, the full-field fatigue life distribution cloud map, and the fatigue life truth value determined by the minimum full-field life value.

[0129] S25: Statistical Derivation of Measured Load Spectrum

[0130] To improve the model's adaptability and generalization ability to varying load conditions, statistical derivation processing is performed on the measured road spectrum obtained from test site road surface or actual road driving data to construct a virtual load sample set that combines original physical damage characteristics with statistical diversity.

[0131] Specifically, firstly, the measured road spectrum is processed by rainflow counting to convert the time-domain load history into a benchmark rainflow matrix characterizing the load cycle characteristics, and the measured road spectrum is used as the seed load spectrum. Subsequently, statistical analysis is performed on the joint distribution characteristics of the load amplitude and mean in the benchmark rainflow matrix to obtain the corresponding two-dimensional probability distribution law. Based on this, a statistical model is established to describe the two-dimensional probability distribution law; the statistical model can be a Gaussian mixture model (GMM) or a kernel density estimation model. After obtaining the statistical model, the model parameters are randomly perturbed and probability sampled using the Monte Carlo sampling method to generate multiple virtual rainflow matrix samples that have the same damage evolution characteristics as the seed load spectrum but differ in amplitude distribution, mean distribution, and cycle combination.

[0132] Through the above processing, the load samples can be statistically expanded while retaining the main physical characteristics of the original load spectrum, thereby obtaining a virtual load sample set that covers different operating condition fluctuations, extreme load situations, and long-tail distribution characteristics.

[0133] S26: Load Feature Extraction

[0134] A two-dimensional residual convolutional neural network (2D-ResNet) is constructed to extract features from the virtual rainflow matrix samples generated in step S25, so as to obtain a two-dimensional load damage feature matrix that characterizes the temporal variation law of load and the distribution characteristics of fatigue damage.

[0135] Specifically, the virtual rainflow matrix sample is first converted into a two-dimensional image input form that can be processed by the network, and then the size is standardized to obtain a single-channel matrix sample with a preset resolution. Subsequently, the standardized matrix sample is input into the two-dimensional residual convolutional neural network. Through multi-layer convolution operations, residual connections and nonlinear mapping, the load amplitude distribution, mean distribution, cyclic clustering features and local texture change features in the rainflow matrix are extracted to generate a deep feature representation result that can reflect the time-frequency damage characteristics of the load.

[0136] Based on this, the deep feature representation results are subjected to dimensionality transformation or feature compression to output a standardized two-dimensional load damage feature matrix, which serves as the load feature input in subsequent multimodal fusion representation. This two-dimensional load damage feature matrix is ​​used to characterize the statistical distribution characteristics and local sensitivity characteristics of virtual load samples during fatigue damage evolution, and serves as one of the input channels for subsequently constructing a three-dimensional multi-channel engineering data tensor.

[0137] In one specific implementation, the virtual rainflow matrix sample can be processed as a 64×64 or 128×128 grayscale image input, and the output two-dimensional load damage feature matrix can be a 64×64-dimensional feature matrix, but is not limited to the above size.

[0138] S27: Source Domain Dataset Encapsulation

[0139] The variable parameters (including topology, material, cross-section, and connection features) determined in steps S21 and S22, along with the two-dimensional load-damage feature matrix extracted in step S26, are used as the input feature set. The truth labels extracted in step S24 are used as the output set. A one-to-one mapping is performed to construct the source domain dataset. The resulting source domain dataset can contain tens of thousands of samples.

[0140] In step S2, steps S23 to S24 constitute the simulation label generation branch, which is used to extract the truth labels corresponding to each parameter combination; steps S25 to S26 constitute the load feature generation branch, which is used to generate load feature input; the output results of the above two branches are correlated and uniformly encapsulated in step S27 to form the source domain dataset.

[0141] Step S3: Multimodal feature fusion and voxelization representation

[0142] This step maps the multimodal data obtained in steps S1 and S2 into a unified three-dimensional voxel space, forming a standardized three-dimensional multichannel engineering data tensor that can be directly read by subsequent models. For example... Figure 4 , Figure 5 As shown, the specific operation is as follows:

[0143] S31: Constructing the geometric topology channel (Channel1)

[0144] The unstructured mesh model is transformed using voxel mapping. The regular three-dimensional density matrix is ​​obtained. Regular three-dimensional mesh cells are divided within the circumscribed cuboid space containing the overall chassis configuration. It is determined whether a chassis solid structure exists within each mesh cell; if it does, it is defined as a solid voxel and assigned a value of 1; otherwise, it is defined as a background voxel and assigned a value of 0, thereby constructing a three-dimensional geometric topology channel.

[0145] In this embodiment, the dimensions of the circumscribed cuboid space can be 5500mm in length, 948mm in width, and 243mm in height, and the voxel mesh resolution can be 512×128×32, corresponding to a single voxel being approximately a 10mm cube. The above parameters are for illustrative purposes only.

[0146] S32: Construct the attribute distribution channel (Channel2)

[0147] A voxel-based attribute mapping mechanism based on component topology is established. The structural attributes and material parameters determined in steps S1 and S2 are mapped to a three-dimensional voxel space. First, discrete material grades and cross-sectional shape selections are encoded, and continuous cross-sectional dimensions and plate thickness parameters are encoded, generating attribute feature vectors. Then, based on the position indices of each longitudinal and transverse beam, the attribute feature vectors are filled into the corresponding entity voxel space, constructing a three-dimensional attribute distribution channel reflecting material distribution and cross-sectional structural features. In this embodiment, an embedding layer can be used to encode discrete features, and a multilayer perceptron (MLP) can be used to encode continuous features.

[0148] S33: Constructing the Connectivity Feature Channel (Channel 3)

[0149] Spatial semantic mapping of connection parameters is performed. For each crossbeam and longitudinal beam connection of the frame, its local coordinate region in three-dimensional voxel space is identified. A feature encoder is used to encode the connection parameters of the corresponding connection nodes into a unified-dimensional connection feature vector. Based on the local coordinate region, the connection feature vector is filled into the corresponding voxel unit to construct a three-dimensional connection feature channel. The connection parameters include at least the diameter, number, and spacing of bolts or rivets. Through the above processing, the connection construction information is explicitly expressed in a unified three-dimensional voxel space.

[0150] S34: Construct the load feature channel (Channel 4)

[0151] Establish spatial mapping rules for the load characteristic field. Obtain the coordinates of each key stress point or sensor installation location (such as leaf spring support, cab suspension point) in three-dimensional voxel space, and map the two-dimensional load damage characteristic matrix obtained in step S26 to the corresponding voxel element. To simulate the load transmission effect in the structure, the load characteristics can be locally diffused to construct a three-dimensional load characteristic channel.

[0152] In this embodiment, a Gaussian kernel function can be used for spatial diffusion, and its expression can be written as:

[0153] (4)

[0154] in, For the load eigenvector, ( , , () represents the center coordinates of the key stress point or sensor installation location. The spatial coordinates of the voxel unit to be assigned values. This is the diffusion radius parameter.

[0155] S35: Generate 3D multichannel engineering data tensors

[0156] In terms of channel dimension, the three-dimensional geometric topology channel constructed in step S31, the three-dimensional attribute distribution channel constructed in step S32, the three-dimensional connection feature channel constructed in step S33, and the three-dimensional load feature channel constructed in step S34 are spliced ​​together to generate a three-dimensional multi-channel engineering data tensor in a unified format.

[0157] In this embodiment, the shape of the three-dimensional multi-channel engineering data tensor can be 4×512×128×32. This tensor can be used as input to a subsequent three-dimensional convolutional neural network model. The tensor dimensions described above are merely an example, and the present invention is not limited thereto.

[0158] Example 2

[0159] This embodiment provides a multimodal data fusion characterization system for predicting the performance of commercial vehicle chassis. The system can be built on a computer hardware platform, and the functional modules are connected and interact through a data bus or application programming interface to execute the method described in Embodiment 1.

[0160] The system includes a multi-source data automated mining module, a simulation data generation and enhancement module, and a multi-modal feature fusion and characterization module.

[0161] The multi-source data automated mining module is used to execute step S1 to identify, extract, quantify, spatially align, and partition multi-source heterogeneous engineering data. This module includes a document layout analysis unit, an optical character recognition unit, a digital image processing unit, a coordinate space calibration unit, and a dataset management unit.

[0162] The document layout analysis unit is used to identify the page layout structure of the input document and divide the document content into unstructured text areas, structured table areas, and field response cloud areas to form data inputs for different processing units.

[0163] The optical character recognition unit is connected to the document layout analysis unit and is used to perform character recognition and field parsing on the unstructured text area and the structured table area, extract frame-related parameter information and convert it into processable digital features.

[0164] The digital image processing unit is connected to the document layout analysis unit and is used to perform digital reverse processing on the field response cloud map region and characteristic curve image. The digital image processing unit includes at least a skeleton extraction subunit and a color mapping conversion subunit. The skeleton extraction subunit is used to extract the contour features of the material performance curve and generate the corresponding data sequence. The color mapping conversion subunit is used to restore the color gradient information in the field response cloud map to the corresponding physical quantity value.

[0165] The coordinate space calibration unit is connected to the optical character recognition unit and the digital image processing unit respectively. It is used to retrieve the preset dimension transformation rule library, perform dimension unification processing on the extracted heterogeneous data, and establish a global geometric reference coordinate system. The coordinate space calibration unit is also used to determine the spatial transformation relationship based on the structural boundary features and reference position features in the image, and map data from different sources to a unified physical coordinate space to achieve spatial alignment.

[0166] The dataset management unit is connected to the coordinate space calibration unit to encapsulate the sample data after dimensional unification and spatial alignment, and to perform dataset partitioning based on the similarity of sample topological features, so as to construct the target domain dataset for model virtual-real transfer calibration and an independent test set for verifying cross-topology generalization ability.

[0167] With the above settings, the multi-source data automated mining module can automatically convert heterogeneous engineering documents into standardized structural samples, providing basic data support for subsequent simulation data enhancement, multimodal feature fusion, and three-dimensional voxelization representation.

[0168] The simulation data generation and enhancement module is used to execute step S2 to achieve the generation of parametric simulation samples for the chassis, statistical derivation of measured load spectra, load feature extraction, and source domain dataset construction. This simulation data generation and enhancement module includes a parametric modeling engine, a simulation task flow unit, a measured load spectrum statistical derivation unit, and a two-dimensional feature extraction unit.

[0169] The parametric modeling engine is used to build a parametric finite element model of the chassis based on preset design variables, and drives the linkage adjustment of structural topology, cross-sectional construction, material properties and connection parameters through automated scripts to generate chassis finite element simulation models with different parameter combinations in batches. In one specific implementation, the parametric modeling engine has built-in Tcl / Tk scripts based on the HyperMesh platform to realize the automatic construction and updating of chassis finite element models.

[0170] The simulation task flow unit is connected to the parametric modeling engine to receive the finite element simulation model of the chassis and perform free modal analysis, typical working condition static analysis, bending stiffness analysis, torsional stiffness analysis, and fatigue life analysis according to a preset task flow to generate performance response results corresponding to each parameter combination. The simulation task flow unit is also used to perform fatigue life calculation in conjunction with multibody dynamic load input and output the true fatigue life value and the corresponding field response results. In one specific embodiment, the simulation task flow unit is configured with an Adams-nCode co-simulation interface to import excitation data corresponding to the measured seed road spectrum and complete fatigue analysis calculations.

[0171] The measured load spectrum statistical derivation unit is used to perform statistical modeling and random derivation processing on the measured load spectrum to generate virtual rainflow matrix samples with statistical diversity. Specifically, the measured load spectrum statistical derivation unit establishes a statistical model based on the amplitude and mean distribution patterns of the seed road spectrum, and generates multiple virtual rainflow matrix samples through random sampling to expand the load sample space.

[0172] The two-dimensional feature extraction unit is connected to the measured load spectrum statistical derivation unit and is used to perform feature extraction on the virtual rainflow matrix sample and output a two-dimensional load damage feature matrix characterizing the time-frequency damage characteristics of the load. In a specific embodiment, the two-dimensional feature extraction unit is configured with a two-dimensional residual convolutional neural network to perform deep feature learning and feature encoding on the virtual rainflow matrix sample.

[0173] The simulation data generation and enhancement module is also used to correlate and encapsulate the parametric modeling results, load feature extraction results, and ground truth labels output by the simulation task flow, forming a source domain dataset that covers the chassis design space and load condition distribution patterns. Through these settings, the simulation data generation and enhancement module can achieve a complete processing flow from generating chassis parametric simulation samples and statistically amplifying the load spectrum to expressing load features and encapsulating source domain samples, providing a data foundation for subsequent multimodal feature fusion characterization and model training.

[0174] The multimodal feature fusion and characterization module is used to execute step S3 to achieve the fusion and characterization of geometric topological features, attribute parameter features, connectivity parameter features, and load features in a unified three-dimensional voxel space, and outputs a standardized three-dimensional multi-channel engineering data tensor. This module includes a voxelization encoding engine, an attribute feature mapping unit, a load field spatial diffusion unit, and a tensor stitching unit.

[0175] Among them, the voxelization coding engine is used to construct a regular three-dimensional voxel mesh in the outer space of the overall frame configuration, and generate geometric topological channels according to the occupancy of the frame entity structure in the voxel space, so as to realize the voxelization expression of the frame geometry and topological connection relationship.

[0176] The attribute feature mapping unit is connected to the multi-source data automated mining module and the simulation data generation and enhancement module. It receives material properties, cross-sectional construction parameters, and connection parameters, and maps these parameters to corresponding voxel space regions to construct attribute feature channels and connection feature channels. Specifically, the attribute feature mapping unit includes an embedding layer and a feature encoder. The embedding layer is used to vectorize discrete parameters, and the feature encoder is used to encode continuous parameters, filling the encoded results into the corresponding voxel units based on the component location index or connection region location index.

[0177] The load field spatial diffusion unit is connected to the simulation data generation and enhancement module. It receives the two-dimensional load damage feature matrix and establishes a spatial mapping relationship of the load feature field based on the coordinates of key stress points or load input positions in three-dimensional voxel space. The load field spatial diffusion unit also performs local spatial diffusion processing on the mapped load features to construct a load feature channel reflecting the spatial distribution law of time-varying loads. In one specific embodiment, the local spatial diffusion processing is implemented using a Gaussian kernel function.

[0178] The tensor splicing unit is connected to the voxelization encoding engine, the attribute feature mapping unit, and the load field spatial diffusion unit, respectively. It is used to perform spatial alignment and channel dimension splicing on the geometric topology channel, attribute feature channel, connection feature channel, and load feature channel to construct and output a standard three-dimensional multi-channel engineering data tensor containing physical semantics.

[0179] Through the above settings, the multimodal feature fusion and characterization module can uniformly map data from different sources, dimensions, and physical meanings to the same three-dimensional voxel coordinate system, thereby achieving deep fusion and characterization of multimodal engineering information and providing standardized input data for subsequent performance prediction models.

[0180] 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 multimodal data fusion and characterization method for predicting the chassis performance of commercial vehicles, characterized in that, Includes the following steps: S1. Automated mining of multi-source heterogeneous engineering data, identifying and extracting structural topology parameters, cross-sectional construction parameters, material property parameters, connection parameters, PSN curve data and field response data related to cross-topology frames, performing dimensional unification and coordinate space alignment on the extracted results, and encapsulating them into a target domain dataset for model virtual-real migration calibration and an independent test set for verifying the model's cross-topology generalization ability. S2. Establish a fully parameterized finite element model of the chassis, and adjust the structural topology variables, cross-sectional construction variables, material property variables and connection parameter variables in a coordinated manner. Combine experimental design sampling to generate simulation samples with different parameter combinations, and generate virtual load samples through statistical analysis of measured load spectra. Extract a two-dimensional load damage feature matrix, and encapsulate the parameter combinations, the two-dimensional load damage feature matrix and the corresponding simulation truth labels to form a source domain dataset. S3. Map the multimodal data obtained in steps S1 and S2 to a unified three-dimensional voxel space, construct geometric topology channels, attribute distribution channels, connection feature channels and load feature channels respectively, and stitch the channels together to generate a three-dimensional multi-channel engineering data tensor, which serves as the input representation for the commercial vehicle frame performance prediction model.

2. The multimodal data fusion characterization method for predicting commercial vehicle frame performance according to claim 1, characterized in that, Step S1, the automated mining of multi-source heterogeneous engineering data includes: Document layout analysis is performed on the multi-source heterogeneous engineering data to segment it into a text area, a table area, and a field response cloud map area. The field response cloud map area includes a stress cloud map, a deformation cloud map, a fatigue life cloud map, or a damage cloud map. Optical character recognition is performed on the text area and table area to extract structural topology parameters, cross-sectional construction parameters, material property parameters, and connection parameters; The field response cloud map area is preprocessed to extract the field feature target area according to the preset saturation threshold and brightness threshold, and the field feature target area is digitally reverse-converted to extract PSN curve data and field response data.

3. The multimodal data fusion characterization method for predicting commercial vehicle frame performance according to claim 2, characterized in that, In step S1, performing a digital reverse transformation on the field response cloud map region includes: Skeleton extraction is performed on the material property curve image to obtain the center contour of the PSN curve, and the numerical stress-life data sequence is reconstructed by combining the coordinate axis scale recognition results. Pixel-level color inversion processing is performed on the field response cloud map. The pixel color vector of each pixel is matched with the standard color vector in the scale area to determine the corresponding physical quantity value, and a field response dataset with image coordinate index is generated.

4. The multimodal data fusion characterization method for predicting commercial vehicle chassis performance according to claim 1, characterized in that, In step S1, after completing the unification of dimensions and alignment of coordinate space, the processed parameter data and PSN curve data are used as input features, and the corresponding test performance scalar and the spatially aligned field response data are used as truth labels for sample encapsulation. The topological similarity between the sample and the frame to be designed is calculated by extracting the topological feature vector of the structured sample. Samples with similarity higher than a preset threshold are divided into target domain datasets, and samples with similarity lower than the preset threshold are divided into independent test sets.

5. The multimodal data fusion characterization method for predicting the performance of commercial vehicle frames according to claim 1, characterized in that, In step S2, the design variables of the fully parametric finite element model include structural topology variables, cross-sectional construction variables, material property variables, and connection parameter variables; The structural topology variables include the number of beams and their longitudinal distribution positions; the cross-sectional construction variables include the selection of cross-sectional shapes, cross-sectional dimensions, and plate thickness parameters for longitudinal and transverse beams; the material property variables include the material grade; and the connection parameter variables include the number, diameter, and spacing of bolts or rivets.

6. The multimodal data fusion characterization method for predicting commercial vehicle chassis performance according to claim 5, characterized in that, In step S2, the experimental design sampling algorithm is used to perform multidimensional joint sampling in the preset variable value space to determine the parameter combination of each frame finite element model and to construct the simulation task flow; The simulation task flow includes at least free modal analysis, typical working condition static analysis, bending stiffness analysis, torsional stiffness analysis, and fatigue life analysis; the typical working condition static analysis includes at least bending working condition, torsional working condition, braking working condition, and turning working condition.

7. The multimodal data fusion characterization method for predicting the performance of commercial vehicle chassis according to claim 6, characterized in that, The fatigue life analysis includes: Obtain the full-field unit stress response of each parameter combination under a unit load applied at a preset load application point; The load time history can be obtained by solving the virtual prototype model of commercial vehicle and the road surface model, or the load spectrum data of the real vehicle collected in step S1 can be directly used as the load time history input. Based on the principle of linear superposition, the full-field unit stress response is superimposed with the load time history to synthesize the actual stress history, and rainflow counting is performed on the actual stress history. The cumulative damage is calculated using the fatigue damage accumulation theory in combination with the material PSN curve. The lifespan value is calculated based on the cumulative damage, and the minimum lifespan value across the entire field is taken as the true value of the frame fatigue life.

8. The multimodal data fusion characterization method for predicting the performance of commercial vehicle frames according to claim 1, characterized in that, In step S2, the statistical derivation of the measured load spectrum includes: Rainflow counting was performed on the measured road spectrum to obtain the baseline rainflow matrix; Statistical analysis was performed on the joint distribution characteristics of load amplitude and mean in the benchmark rainflow matrix, and a statistical model was established. Based on the statistical model, random perturbation and probability sampling are performed to generate virtual rainflow matrix samples; The virtual rainflow matrix sample is input into a two-dimensional residual convolutional neural network to extract a two-dimensional load damage feature matrix.

9. The multimodal data fusion characterization method for predicting the performance of commercial vehicle chassis according to claim 1, characterized in that, Step S3 includes: The chassis mesh model is transformed into a three-dimensional geometric topology channel using voxel mapping. The structural topology parameters, cross-sectional construction parameters, and material property parameters are mapped to the corresponding solid voxel space to construct property distribution channels. The connection parameters are encoded and filled into the corresponding voxel units of the connection region to construct the connection feature channels; The two-dimensional load damage feature matrix is ​​mapped to the voxel unit corresponding to the key stress point or sensor installation position, and local spatial diffusion is performed to construct the load feature channel. The three-dimensional geometric topology channel, attribute distribution channel, connection feature channel, and load feature channel are spliced ​​together to generate a three-dimensional multi-channel engineering data tensor.

10. A multimodal data fusion characterization system for predicting the performance of a commercial vehicle chassis, characterized in that, The system for performing the multimodal data fusion characterization method for predicting the chassis performance of commercial vehicles as described in any one of claims 1 to 9 includes: The multi-source data automated mining module is used to perform identification, extraction, digitization, unit unification, coordinate space alignment, and dataset partitioning on multi-source heterogeneous engineering data. The multi-source data automated mining module includes a document layout analysis unit, an optical character recognition unit, a digital image processing unit, a coordinate space calibration unit, and a dataset management unit. The document layout analysis unit divides the input document into text areas, table areas, and field response cloud map areas. The optical character recognition unit extracts structural topology parameters, cross-sectional construction parameters, material property parameters, and connection parameters. The digital image processing unit extracts PSN curve data and field response data. The coordinate space calibration unit performs dimensional unification and coordinate space alignment on the extracted data. The dataset management unit encapsulates the data to form a target domain dataset and an independent test set. The simulation data generation and enhancement module is used to perform parametric simulation sample generation of the chassis, statistical derivation of measured load spectrum, load feature extraction, and source domain dataset construction. The simulation data generation and enhancement module includes a parametric modeling engine, a simulation task flow unit, a measured load spectrum statistical derivation unit, and a two-dimensional feature extraction unit. The parametric modeling engine is used to establish a fully parametric finite element model of the chassis. The simulation task flow unit is used to output simulation truth labels. The measured load spectrum statistical derivation unit is used to generate virtual rainflow matrix samples. The two-dimensional feature extraction unit is used to extract a two-dimensional load damage feature matrix. The simulation data generation and enhancement module is also used to encapsulate and form a source domain dataset. The multimodal feature fusion and representation module is used to map the multimodal data in the target domain dataset, independent test set and source domain dataset to a unified three-dimensional voxel space and output a three-dimensional multi-channel engineering data tensor; The multimodal feature fusion and characterization module includes a voxelization encoding engine, an attribute feature mapping unit, a load field spatial diffusion unit, and a tensor splicing unit. The voxelization encoding engine is used to construct geometric topology channels, the attribute feature mapping unit is used to construct attribute distribution channels and connection feature channels, the load field spatial diffusion unit is used to construct load feature channels, and the tensor splicing unit is used to splice the channels to form a three-dimensional multi-channel engineering data tensor. The multi-source data automated mining module, the simulation data generation and enhancement module, and the multimodal feature fusion and characterization module are communicatively connected.

Images