CN-122021309-A - Whole vehicle performance integrated balance prediction method, model training method and system

Abstract

The invention discloses a method, a system, equipment and a medium for predicting integrated balance of whole vehicle performance, and relates to the technical field of automobile engineering and artificial intelligence. The method comprises the steps of obtaining a vehicle design parameter set containing structural feature data, inputting the vehicle design parameter set into a hybrid integrated AI proxy model in parallel, extracting decision tree path information by using a first feature learning branch to serve as a first intermediate feature vector representing a structural rule, extracting a full-connection layer high-order semantic vector by using a second feature learning branch to serve as a second intermediate feature vector representing a physical coupling relation, carrying out weighted fusion on the two feature vectors and carrying out multi-task prediction by using a feature fusion network, and finally outputting a balance scheme containing performance indexes and cost estimation. According to the invention, the advantages of rule learning and nonlinear fitting are combined through the hybrid model architecture, the traditional time-consuming simulation iteration is replaced, and the rapid and accurate collaborative optimization of the whole vehicle performance and cost is realized.

Inventors

• LI YICHEN
• WANG XIN
• TUO TE
• TAN HU
• SHI XINQI

Assignees

• 东风汽车集团股份有限公司

Dates

Publication Date

20260512

Application Date

20260129

Claims (20)

1. 1. The integrated balance prediction method for the whole vehicle performance is characterized by comprising the following steps of: The method comprises the steps of acquiring a design parameter set of a vehicle to be developed, wherein the design parameter set comprises structural feature data defining physical properties and configuration of the vehicle; The characteristic extraction step is that the design parameter set is input into a first characteristic learning branch and a second characteristic learning branch of the hybrid integrated AI proxy model; discretizing space division is carried out on the design parameter set by utilizing the first feature learning branch, and path position information of a sample in a decision tree is extracted to be used as a first intermediate feature vector so as to represent a structural rule relation between design parameters and performance indexes; Carrying out multi-layer nonlinear mapping on the design parameter set by utilizing the second feature learning branch, and extracting a high-order abstract semantic vector output by a full-connection layer as a second intermediate feature vector so as to represent an implicit physical field coupling relation between design parameters; A mixed reasoning step, namely fusing the first intermediate feature vector and the second intermediate feature vector in feature dimension by utilizing a feature fusion network of the mixed integrated AI proxy model, and performing multi-task prediction based on the fused features; And the decision output step is to acquire a multi-dimensional whole vehicle performance index predicted value and a cost estimated value which are output by the hybrid integrated AI proxy model, and generate a whole vehicle performance integrated balance scheme based on the trade-off relation between the predicted value and the estimated value.
2. 2. The method of claim 1, wherein the first feature learning branch employs a tree structure-based ensemble learning architecture, and wherein extracting path location information of the sample in the decision tree as the first intermediate feature vector specifically comprises: Traversing a plurality of decision trees in the integrated learning architecture, and acquiring leaf node indexes of the input design parameter set in each decision tree; And converting the leaf node index into a single-hot coding or embedding vector, and splicing the coding vectors of all trees to form a meta-feature comprising expert rules and data distribution boundaries as the first intermediate feature vector.
3. 3. The method according to claim 2, wherein the tree structure-based ensemble learning architecture specifically employs a gradient-lifting decision tree model and is configured to extract the index path of leaf nodes and convert it into a vector form for output as the first intermediate feature vector, so as to preserve the original rule decision boundary information.
4. 4. The method of claim 1, wherein the second feature learning branch employs a deep neural network architecture, and wherein the extracting the high-order abstract semantic vector of the full connection layer output as the second intermediate feature vector specifically comprises: The design parameter set is input into a multi-layer perceptron after standardized treatment; And intercepting the hidden layer output of the multi-layer perceptron before the final output layer as the second intermediate feature vector so as to retain nonlinear physical coupling information among design parameters and inhibit the feature compression loss of the final regression layer.
5. 5. The method of claim 1, wherein the feature fusion network comprises an adaptive weighting layer, and wherein the fusing in the feature dimension specifically comprises: dynamically calculating contribution weights of the first intermediate feature vector and the second intermediate feature vector for different prediction tasks through an attention mechanism; And carrying out weighted splicing on the first intermediate feature vector and the second intermediate feature vector based on the contribution weight, wherein the attention mechanism is configured to automatically adjust the weight parameter of each intermediate feature vector according to the loss gradient information of each prediction task in the training process so that a model can identify and depend on feature branches with higher contribution degree to a specific task.
6. 6. The method of claim 1, further comprising a feature construction step prior to the feature extraction step: And searching marginal change values of the manufacturing cost or the material cost of the parts corresponding to the unit change amount of the key physical parameters based on a preset engineering database, generating cost gradient characteristics, and inputting the cost gradient characteristics into the first characteristic learning branch as explicit characteristics in the design parameter set.
7. 7. The method according to claim 1, wherein the overall vehicle performance integrated balancing scheme generated in the decision output step specifically includes: Constructing an optimization model taking the total cost minimization as an objective function and taking the predicted value of the performance index of the multi-dimensional whole vehicle as a constraint condition, wherein the predicted value meets a preset threshold; And calculating gradient information based on the micro-characteristics of the hybrid integrated AI proxy model, and reversely solving the optimal design parameter combination meeting the constraint condition.
8. 8. A method of training a hybrid integrated AI proxy model for constructing the hybrid integrated AI proxy model of claim 1, comprising the steps of: Constructing a multi-dimensional engineering data system, wherein the multi-dimensional engineering data system comprises attribute-parameter data, system associated data and cost attribute data; The characteristic engineering step is that the data from the multidimensional engineering data system is cleaned and normalized, and the key characteristic vector is screened by utilizing a global sensitivity analysis algorithm to construct a training data set; The model construction step comprises the steps of constructing an initial model framework comprising the first characteristic learning branch, the second characteristic learning branch and the characteristic fusion network; And in the training process, updating the weights of the second feature learning branch and the feature fusion network through a back propagation algorithm, and optimizing model parameters through minimizing a multi-task loss function until the evaluation index of the model on the verification set meets the preset convergence condition.
9. 9. The method according to claim 8, wherein the screening of the key feature vectors in the feature engineering step using a global sensitivity analysis algorithm specifically comprises: calculating a main effect index and an interactive effect index of each design parameter to model output based on a variance decomposition principle by adopting a Sobol sequence method or a Morris screening method; preserving parameters with the total effect index higher than a preset threshold value, and identifying parameter pairs with high interaction effect indexes; the parameter pairs are combined feature constructed as part of the input feature.
10. 10. The method of claim 8, wherein the data construction step further comprises a physical consistency constraint cleaning: Presetting physical limit boundary conditions and physical conservation law constraints of each performance index; Abnormal sample points in the training data set, which violate the physical limit boundary conditions or physical conservation law constraints, are identified and removed from the training data set to prevent pollution of the model weights by the wrong physical laws.
11. 11. The method of claim 8, wherein in the iterative training step, a dynamic regularization strategy is performed: the method comprises the steps that an evaluation index difference value of a real-time monitoring model on a training set and a verification set is defined as a generalization error; And when the generalization error exceeds a preset warning threshold, triggering a structure adjustment mechanism, namely dynamically increasing the Dropout ratio in the second characteristic learning branch or dynamically limiting the maximum depth of the growth of the decision tree in the first characteristic learning branch until the generalization error falls back into a preset convergence threshold range.
12. 12. The method of claim 8, wherein the multitasking loss function is constructed as a weighted composite loss function: Respectively calculating prediction errors for dynamic performance, economy, NVH performance and cost indexes, and defining task weights according to priorities of the indexes in the whole vehicle development project; and taking the sum of the products of the prediction errors of the indexes and the corresponding task weights as the total loss, and introducing an L2 norm regularization term of the model weights.
13. 13. The method according to claim 8, wherein the multi-dimensional engineering data hierarchy comprises in particular: the attribute-parameter database is used for storing the original value, the performance target value and the historical vehicle model reference data of the key design parameters; The system association database is used for storing response surface data, transfer functions and coupling weight data for representing the interface performance of the subsystem; And the cost attribute database is used for storing the material price of the parts, the manufacturing cost data of the process and the weight engineering detail data, and establishing the vehicle configuration association among the data.
14. 14. The utility model provides a whole car performance integrated balance and prediction device which characterized in that includes: The system comprises a parameter acquisition module, a development module and a development module, wherein the parameter acquisition module is used for acquiring a design parameter set of a vehicle to be developed, and the design parameter set comprises structural characteristic data defining physical properties and configuration of the vehicle; The mixed integrated AI proxy model is configured to extract structural features representing rule relations by utilizing integrated learning branches based on a tree structure, extract nonlinear features representing physical coupling relations by utilizing deep neural network branches in parallel, and perform multitask reasoning calculation after fusing the structural features and the nonlinear features; And the decision analysis module is used for acquiring the multidimensional whole vehicle performance index predicted value and the cost estimated value output by the hybrid reasoning module, displaying the trade-off relation between the performance index and the cost estimated value in a visual form and assisting in generating an integrated balance scheme.
15. 15. The apparatus of claim 14, further comprising a data governance module and a model update module; the data management module is used for executing an incremental learning mechanism, and automatically triggering global sensitivity analysis and updating a key feature list when new actual measurement data of the vehicle model are generated; And the model updating module is used for calling a training algorithm to carry out fine adjustment updating on the hybrid integrated AI proxy model.
16. 16. A computer device comprising a memory, a processor and a computer program stored on the memory, characterized in that the processor executes the computer program to carry out the steps of the method of any one of claims 1 to 7.
17. 17. A computer device comprising a memory, a processor and a computer program stored on the memory, characterized in that the processor executes the computer program to carry out the steps of the method of any one of claims 8 to 13.
18. 18. The computer device of claim 16 or 17, wherein the processor comprises a graphics processor, tensor processor, or neural network processor capable of performing parallel computations configured to perform matrix operations in the hybrid integrated AI proxy model in parallel.
19. 19. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method according to any one of claims 1 to 7.
20. 20. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method of any of claims 8 to 13.

Description

Whole vehicle performance integrated balance prediction method, model training method and system Technical Field The invention relates to the technical field of intersection of artificial intelligence and vehicle engineering, in particular to a prediction and training method of integrated balance of multi-objective performance of a whole vehicle based on a hybrid artificial intelligence model, and a device, a system, computer equipment, a computer readable storage medium and a computer program product for realizing the method. Background In the V-type development process of modern automobiles, the integration and balance of the performance of the whole automobile is a core challenge throughout. With the development of automobiles to electric, intelligent and networking, the whole automobile is used as a complex engineering system, the number of subsystems is increased, and the physical field coupling relationship is increasingly complex. There are strong nonlinear coupling and constraint relations among multidimensional targets such as dynamic property, economical efficiency, NVH (noise, vibration and harshness), control stability, safety and cost. Traditional performance development methods rely heavily on the experience of domain experts and separate Computer Aided Engineering (CAE) simulations. The team of engineers in different performance fields independently perform simulation analysis to form a 'data island'. When different performance simulation results conflict, a balance point is often required to be sought through a plurality of manual iteration loops of parameter adjustment, re-simulation and meeting review. The process is low in efficiency, the development period is prolonged, the research and development cost is increased, and the balance result is often local and suboptimal due to lack of global insight on the coupling relation among multiple performances, so that the influence of design change on the overall vehicle-level comprehensive performance is difficult to accurately predict in early stage of projects. Although artificial intelligence technology has been applied in the automotive field, most of the technology is focused on downstream applications such as autopilot and intelligent cabins, and the upstream performance development and integration balance links, the application is still obviously insufficient, and in particular, a set of schemes capable of systematically solving the following technical problems is lacking: 1. The heterogeneous feature fusion problem is that the whole vehicle engineering data not only contains structural parameters with clear physical meaning and regularity boundary, but also contains hidden physical field coupling relation which is difficult to describe by a formula. It is difficult for a single type of AI model to efficiently handle both heterogeneous features simultaneously. 2. The difficult problem of fitting of small samples is that the whole vehicle development, in particular the development of a brand new platform, has limited initial high-quality simulation or actual measurement data sample size, and the traditional deep learning model is easy to be fit, so that generalization capability is poor. 3. The multi-objective optimization balancing problem is that the dimension difference of a plurality of performance indexes is huge, the importance of the performance indexes in projects is different, and how to balance the optimization direction of the multi-tasks in model training is a key challenge. 4. The difficulty in data quality and knowledge utilization is that "dirty data" caused by calculation divergence can exist in simulation data, and how to effectively structure and use massive historical data and engineering experience for guiding model training is also a bottleneck currently faced. Therefore, developing an intelligent prediction and decision method and system that can deeply integrate domain knowledge, efficiently process heterogeneous engineering data, remain robust under small sample conditions, and realize multi-objective collaborative optimization has become an urgent need in the industry. Disclosure of Invention Aiming at the defects in the background art, the invention provides a method for predicting the integrated balance of the performance of a whole vehicle, a method for training a model and a system thereof, and aims to solve the core technical problems that a single AI model in the prior art cannot simultaneously and effectively process heterogeneous data (regular data and physical field data) in the whole vehicle engineering, multi-physical field coupling analysis is difficult, performance and cost are synergistically optimized and disjointed, and model generalization capability under a small sample is poor. In order to solve the technical problems, in a first aspect, the present invention provides the following technical solutions: A method for predicting the integrated balance of the performance of a whole vehicle comprises