Search

CN-121997468-A - Multi-mode data fusion characterization method and system for commercial vehicle frame performance prediction

CN121997468ACN 121997468 ACN121997468 ACN 121997468ACN-121997468-A

Abstract

The invention provides a multi-mode data fusion characterization method and a system for predicting the frame performance of a commercial vehicle, which are applied to the technical field of frame data modeling, and the method comprises the steps of firstly, automatically mining multi-source heterogeneous engineering data, extracting structural topology parameters, section construction parameters, material attribute parameters, connection parameters, P-S-N curve data and field response data, and forming a target domain data set and an independent test set; and finally mapping the geometric topological features, the parameter features, the connection features and the load features to a unified three-dimensional voxel space to generate a three-dimensional multi-channel engineering data tensor. The method can improve the construction efficiency of the commercial vehicle frame sample, the utilization efficiency of multi-mode data and the generalization capability of the performance prediction model.

Inventors

  • WANG DENGFENG
  • Ni Yenan
  • MENG ZIHAO
  • Lian Fengmin

Assignees

  • 吉林大学

Dates

Publication Date
20260508
Application Date
20260410

Claims (10)

  1. 1. A multi-mode data fusion characterization method for predicting the performance of a commercial vehicle frame is characterized by comprising the following steps: S1, automatically mining multi-source heterogeneous engineering data, identifying and extracting structural topology parameters, section construction parameters, material attribute parameters, connection parameters, P-S-N curve data and field response data related to a cross-topology frame, executing dimension unification and coordinate space alignment on the extraction result, and packaging to form a target domain data set for model virtual-real migration calibration and an independent test set for verifying cross-topology generalization capability of a model; S2, establishing a full-parameterized finite element model of the vehicle frame, carrying out linkage adjustment on structural topological variables, section construction variables, material attribute variables and connection parameter variables, generating simulation samples of different parameter combinations by combining test design sampling, generating virtual load samples by actual measurement load spectrum statistics derivation, extracting a two-dimensional load damage characteristic matrix, and packaging the parameter combinations, the two-dimensional load damage characteristic matrix and corresponding simulation truth labels to form a source domain data set; And S3, mapping the multi-mode data obtained in the step S1 and the step S2 to a unified three-dimensional voxel space, respectively constructing a geometric topological channel, an attribute distribution channel, a connection characteristic channel and a load characteristic channel, and splicing the channels to generate a three-dimensional multi-channel engineering data tensor serving as an input representation of a commercial vehicle frame performance prediction model.
  2. 2. The method for multi-modal data fusion characterization of commercial vehicle frame performance prediction according to claim 1, wherein in step S1, performing automatic mining on multi-source heterogeneous engineering data comprises: Performing document layout analysis on the multi-source heterogeneous engineering data, and dividing the multi-source heterogeneous engineering data to obtain a text region, a table region and a field response cloud picture region, wherein the field response cloud picture region comprises a stress cloud picture, a deformation cloud picture, a fatigue life cloud picture or a damage cloud picture; Performing optical character recognition on the text region and the table region to extract structural topology parameters, section construction parameters, material attribute parameters and connection parameters; and carrying out image preprocessing on the field response cloud region, extracting a field characteristic target region according to a preset saturation threshold value and a preset brightness threshold value, and carrying out digital reverse conversion on the field characteristic target region so as to extract P-S-N curve data and field response data.
  3. 3. The method for multi-modal data fusion characterization of commercial vehicle frame performance prediction according to claim 2, wherein in step S1, performing a digitized reverse transformation on the field response cloud image region comprises: performing skeleton extraction on the material performance curve image to obtain the central outline of the P-S-N curve, and reconstructing by combining the coordinate axis scale recognition result to obtain a numeric stress-life data sequence; And performing pixel-level color inversion processing on the field response cloud image, matching pixel color vectors of all pixel points with standard color vectors in the scale area to determine corresponding physical quantity values, and generating a field response data set with image coordinate indexes.
  4. 4. The multi-mode data fusion characterization method of the commercial vehicle frame performance prediction according to claim 1, wherein in step S1, after dimension unification and coordinate space alignment are completed, the processed parameter data and P-S-N curve data are used as input features, corresponding test performance scalar and field response data with completed space alignment are used as truth labels to perform sample packaging, the topological similarity between the mining sample and the frame to be designed is calculated by extracting topological feature vectors of the structured sample, samples with similarity higher than a preset threshold value are divided into target domain data sets, and samples with similarity lower than the preset threshold value are divided into independent test sets.
  5. 5. The multi-modal data fusion characterization method for commercial vehicle frame performance prediction according to claim 1, wherein in step S2, the design variables of the full parameterized finite element model include structural topology variables, cross-sectional construction variables, material property variables, and connection parameter variables; The structural topological variables comprise the number of the cross beams and the longitudinal distribution positions of the cross beams, the section construction variables comprise the section shape selection, the section size and the plate thickness parameters of the longitudinal beams and the cross beams, the material property variables comprise the material marks, and the connection parameter variables comprise the number, the diameter and the distance of bolts or rivets.
  6. 6. The multi-mode data fusion characterization method for predicting the performance of the commercial vehicle frame according to claim 5, wherein in the step S2, multi-dimensional joint sampling is performed in a preset variable value taking space by using a test design sampling algorithm so as to determine the parameter combination of each frame finite element model and construct a simulation task flow; The simulation task flow at least comprises free mode analysis, typical working condition statics analysis, bending stiffness analysis, torsional stiffness analysis and fatigue life analysis, wherein the typical working condition statics analysis at least comprises bending working conditions, torsional working conditions, braking working conditions and turning working conditions.
  7. 7. The method for multi-modal data fusion characterization of commercial vehicle frame performance prediction as defined in claim 6, wherein the fatigue life analysis comprises: acquiring full-field unit stress response of unit load applied by each parameter combination at a preset load application point; Solving and obtaining load time histories based on a commercial vehicle virtual prototype model and a road surface model, or directly adopting the real vehicle acquired load spectrum data obtained by excavation in the step S1 as load time histories for input; According to a linear superposition principle, the full-field unit stress response and the load time history are superposed to synthesize an actual stress history, the actual stress history is subjected to rain flow counting, and the accumulated damage is calculated by adopting a fatigue damage accumulation theory in combination with a material P-S-N curve; And obtaining a life value according to the accumulated damage, and taking the full-field minimum life value as a fatigue life true value of the frame.
  8. 8. The method for multi-modal data fusion characterization of commercial vehicle frame performance prediction according to claim 1, wherein in step S2, performing statistical derivation on the measured load spectrum comprises: Performing rain flow counting processing on the actually measured road spectrum to obtain a reference rain flow matrix; carrying out statistical analysis on the combined distribution characteristics of the load amplitude value and the mean value in the reference rain flow matrix, and establishing a statistical model; Random disturbance and probability sampling are carried out based on the statistical model, and a virtual rain flow matrix sample is generated; and inputting the virtual rain flow matrix sample into a two-dimensional residual convolution neural network, and extracting to obtain a two-dimensional load damage characteristic matrix.
  9. 9. The method for multi-modal data fusion characterization of commercial vehicle frame performance prediction according to claim 1, wherein step S3 comprises: converting the frame grid model into a three-dimensional geometric topological channel by adopting a voxelized mapping mode; Mapping structural topology parameters, section construction parameters and material attribute parameters to corresponding entity voxel spaces, and constructing attribute distribution channels; filling the connection parameters into voxel units corresponding to the connection areas after coding, and constructing connection characteristic channels; mapping the two-dimensional load damage characteristic matrix to voxel units corresponding to key stress points or sensor mounting positions, and performing local spatial diffusion to construct a load characteristic channel; and splicing the three-dimensional geometric topological channel, the attribute distribution channel, the connection characteristic channel and the load characteristic channel to generate a three-dimensional multi-channel engineering data tensor.
  10. 10. A multi-modal data fusion characterization system for commercial vehicle frame performance prediction, characterized in that a multi-modal data fusion characterization method for commercial vehicle frame performance prediction as defined in any one of claims 1 to 9 is performed, the system comprising: The multi-source data automatic mining module is used for executing identification, extraction, digital processing, dimension unification, coordinate space alignment and data set division on multi-source heterogeneous engineering data; The multi-source data automatic mining module comprises a document layout analysis unit, an optical character recognition unit, a digital imaging processing unit, a coordinate space calibration unit and a data set management unit, wherein the document layout analysis unit is used for dividing an input document into a text region, a table region and a field response cloud region, the optical character recognition unit is used for extracting structural topological parameters, section construction parameters, material attribute parameters and connection parameters, the digital imaging processing unit is used for extracting P-S-N curve data and field response data, the coordinate space calibration unit is used for executing dimension unification and coordinate space alignment on the extracted data, and the data set management unit is used for packaging and forming a target domain data set and an independent test set; The simulation data generation and enhancement module is used for executing frame parameterization simulation sample generation, actual measurement load spectrum statistical derivation, load characteristic extraction and source domain data set construction; The simulation data generation and enhancement module comprises a parameterized modeling engine, a simulation task flow unit, an actual measurement load spectrum statistics derivation unit and a two-dimensional feature extraction unit, wherein the parameterized modeling engine is used for building a frame full parameterization finite element model, the simulation task flow unit is used for outputting a simulation truth value label, the actual measurement load spectrum statistics derivation unit is used for generating a virtual rain flow matrix sample, the two-dimensional feature extraction unit is used for extracting a two-dimensional load damage feature matrix, and the simulation data generation and enhancement module is also used for packaging to form a source domain data set; The multi-modal feature fusion characterization module is used for mapping multi-modal data in the target domain data set, the independent test set and the source domain data set to a unified three-dimensional voxel space and outputting a three-dimensional multi-channel engineering data tensor; The multi-modal feature fusion characterization module comprises a voxelized coding engine, an attribute feature mapping unit, a load field space diffusion unit and a tensor splicing unit, wherein the voxelized coding engine is used for constructing a geometric topological channel, the attribute feature mapping unit is used for constructing an attribute distribution channel and a connection feature channel, the load field space diffusion unit is used for constructing a load feature channel, and the tensor splicing unit is used for splicing all channels to form a three-dimensional multi-channel engineering data tensor; the multi-source data automatic mining module, the simulation data generation and enhancement module and the multi-mode feature fusion characterization module are in communication connection.

Description

Multi-mode data fusion characterization method and system for commercial vehicle frame performance prediction Technical Field The invention belongs to the technical field of frame data modeling, and particularly relates to a multi-mode data fusion characterization method and system for predicting the performance of a commercial vehicle frame. Background Under the development trend of intellectualization, dynamism and low carbonization of commercial vehicles, the frame is used as a basic bearing part of the whole vehicle and faces the double pressure caused by weight increment of a power battery and anxiety of endurance mileage. The built-in conflict between the structural bearing requirement and the whole vehicle energy efficiency target forces the design of the vehicle frame to realize extremely light weight on the premise of ensuring high reliability. However, the traditional research and development mode faces significant bottlenecks due to strong nonlinearity of frame fatigue damage evolution and manufacturing process discreteness, namely, on one hand, a performance prediction method based on Finite Element Analysis (FEA) is long in calculation time and has significant poor precision when solving nonlinear fatigue damage, and on the other hand, an iterative optimization method based on a proxy model is required to collect more sample points, is limited in design space, is slow in convergence speed and is large in optimization margin. The technical situation that prediction confidence and optimizing convergence efficiency are difficult to achieve is that the research and development requirements of multiple configurations, extremely light weight and fast pace of the current commercial vehicle products are difficult to support. In view of the fact that the deep learning technology has natural advantages in capturing strong nonlinear mapping relation and realizing second-level reasoning prediction, the construction of the vehicle frame performance high-precision prediction and intelligent design method based on deep learning has become a key path for industry to break through the bottleneck of research and development precision and efficiency. However, when the deep learning technology is specifically applied to the development of a commercial vehicle frame, the dual technical bottlenecks of 'high-quality sample data acquisition' and 'multi-mode feature fusion characterization' are faced: On the one hand, in the aspect of obtaining and generating sample data, the prior art is difficult to combine the mining of unstructured historical data and the efficient generation of global topology parameterized samples. The generalization accuracy of the deep learning model is highly dependent on the scale, quality and distribution diversity of the training samples. At present, commercial vehicle enterprises and industries accumulate massive research and development data, but the data are widely dispersed in public documents, industry databases and enterprise internal reports and are in heterogeneous forms such as tables, test scalar results, characteristic graphs, fatigue life cloud charts and the like. The lack of effective mining means in the prior art results in the failure of the highly valuable historical engineering experience data to be converted into effective training data samples to be fully utilized. In the aspect of generating a simulation sample, the prior art generally adopts local parametric sampling or single load working condition compiling, and lacks global coverage capability. For example, chinese patent application No. CN202511352601.3 discloses a lightweight design method for a machine learning driven vehicle frame, which mainly samples dimensional variables such as plate thickness, and the like, while constructing a sample space, and is difficult to cover the change of topological features such as beam layout, connection parameters, and the like, resulting in insufficient sample richness. For another example, chinese patent application No. CN202411577629.2 discloses a method for accelerating the compilation of a multi-dimensional load spectrum of a vehicle frame under WLTC working conditions, which focuses on extracting a single standard test spectrum from standard cycle working conditions, does not involve a load spectrum statistical derivation technology based on actually measured road spectrum characteristics, cannot generate virtual load samples with physical reality and working condition diversity in batch, and limits generalization of data in variable working condition dimensions. On the other hand, in the aspect of fusion characterization of multi-mode data, the prior art, such as China patent application No. CN202511445349.0, discloses a multi-mode data-driven commercial vehicle frame performance prediction method and system, wherein the adopted Graph structure (Graph) characterization mode mainly focuses on topological connection of discrete nodes, natural limitation exists when cloud G