An end-to-end safety design large model system and method for vehicle crashworthiness drive

By using an end-to-end safety design big model system and leveraging cross-modal interaction of a big language model and multimodal feature extraction and analysis module, the problem of low efficiency in traditional automotive crashworthiness design is solved, achieving efficient full-link mapping and automated design, and significantly shortening the design cycle.

CN122634753BActive Publication Date: 2026-11-10HUNAN UNIV
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202611124367.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-07-28
Publication Date
2026-11-10
Estimated Expiration
2046-07-28

AI Technical Summary

Technical Problem

Traditional automotive crashworthiness design processes are inefficient, requiring manual coordination at each stage, resulting in low iteration efficiency, long cycles, and high costs. Furthermore, manual interpretation can easily lead to deviations in requirements.

Method used

The end-to-end safety design large model system driven by vehicle crashworthiness achieves end-to-end mapping from design feature input to design scheme output that meets collision safety requirements through cross-modal interaction between the large language model module and the multimodal feature extraction and analysis module, replacing the traditional manual-led multi-round simulation iteration and physical test links.

Benefits of technology

It significantly improves the efficiency of the automotive crashworthiness design process, compressing the single-cycle design iteration from weeks to seconds, shortening the overall vehicle safety development cycle, and achieving full-process automation from natural language requirements to feasible structural solutions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122634753B_ABST
    Figure CN122634753B_ABST
Patent Text Reader

Abstract

This application belongs to the field of automotive passive safety design technology, specifically disclosing an end-to-end safety design large-scale model system and method driven by vehicle crashworthiness. The system includes a large language model module, and an input module, a multimodal feature extraction and analysis module, and an output module respectively connected to the large language model module. The input module is used to input crashworthiness design requirements information. The large language model module and the multimodal feature extraction and analysis module adopt a cross-modal interaction method to perform multimodal feature extraction and cross-modal fusion using the lexical sequences corresponding to the crashworthiness design requirements information, and to perform iterative analysis based on the fused features under preset crashworthiness constraints to obtain a target three-dimensional crashworthiness structure scheme. The output module is used to output the target three-dimensional crashworthiness structure scheme. This application can greatly improve the efficiency of the entire vehicle crashworthiness design process and shorten the vehicle safety development cycle.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application belongs to the field of automotive passive safety design technology, and more specifically, relates to an end-to-end safety design large model system and method for automotive crashworthiness driven by a specific approach. Background Technology

[0002] Automotive crashworthiness design is a core component of passive safety development, directly determining the vehicle's occupant protection capabilities and safety performance in collisions. Traditional automotive crashworthiness design typically employs a sequential process of "design-simulation-modification-testing." However, this traditional process suffers from low iteration efficiency due to the independent operation of each stage, resulting in a highly fragmented, lengthy, and costly workflow. Furthermore, during testing, the inability to automatically translate natural language requirements into engineering geometric parameters and physical performance indicators makes manual interpretation prone to errors, leading to extensive design rework and making the entire design process extremely inefficient.

[0003] In addition, there are currently design approaches that incorporate artificial intelligence (AI), but existing AI only serves partial aspects such as simulation acceleration, parameter optimization, or single part generation, and each aspect still requires manual coordination, making it impossible to streamline the entire process from requirement input to solution output, resulting in low overall design efficiency.

[0004] Therefore, how to achieve more efficient automotive crashworthiness design has become a pressing technical problem for the industry. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this application is to realize the crashworthiness design of automobiles more efficiently, and to solve the problem that the traditional automobile crashworthiness design process is inefficient because each link needs to be manually connected.

[0006] To achieve the above objectives, in a first aspect, this application provides an end-to-end safety design large-scale model system driven by vehicle crashworthiness, comprising: A large language model module, and an input module, a multimodal feature extraction and analysis module, and an output module respectively connected to the large language model module; The input module is used to input crashworthiness design requirements information; The large language model module and the multimodal feature extraction and analysis module adopt a cross-modal interaction method to extract multimodal features and perform cross-modal fusion using the word sequence corresponding to the crashworthiness design requirement information, and perform iterative analysis based on the fused features under preset crashworthiness constraints to obtain the target three-dimensional crashworthiness structure scheme. The output module is used to output the target three-dimensional impact-resistant structure scheme.

[0007] Secondly, this application provides a control method applied to any of the aforementioned end-to-end security design large model systems, comprising: The input crashworthiness design requirements are converted into corresponding word sequences. Multimodal feature extraction and cross-modal fusion are performed based on the word sequence corresponding to the crashworthiness design requirements information. Based on the fused features, iterative analysis is performed under preset crashworthiness constraints to obtain the target three-dimensional crashworthiness structure scheme. Output the target three-dimensional impact-resistant structure scheme.

[0008] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art: (1) The end-to-end safety design big model system and method driven by the crashworthiness of automobiles provided in this application, by introducing cross-modal interaction between the big language model module and the multimodal feature extraction and analysis module, constructs an end-to-end safety design big model system with crashworthiness as the core driving target. It realizes the end-to-end mapping of the entire link from design feature input to design scheme output that meets collision safety requirements. It can effectively replace the traditional manual-led multi-round, time-consuming simulation iteration and physical test links, and compress the single-round design iteration cycle from the traditional several weeks to the second level, which greatly improves the efficiency of the entire automobile crashworthiness design process and shortens the whole vehicle safety development cycle.

[0009] (2) The end-to-end safety design large model system and method for vehicle crashworthiness driven by this application reconstructs the traditional serial long-cycle trial and error process of "manual modeling-finite element simulation-manual modification" into an efficient parallel process of end-to-end forward calculation, embedded constraint guidance and automatic closed-loop fine adjustment in a unified feature space. It eliminates the overhead of manual connection and data conversion across links, which is superior to the traditional trial and error development mode driven by manual experience. It realizes the underlying paradigm of AI and physical dual-driven generative development of vehicle crashworthiness design.

[0010] (3) The end-to-end safety design big model system and method driven by vehicle crashworthiness provided in this application builds a big language model module by introducing an encoder group module, a cross-modal fusion module, an attention layer and a linear projection layer, and jointly introduces a constraint mapping module, a crashworthiness structure generation module, a collision simulation proxy prediction module, a compliance verification module and a closed-loop self-iterative optimization module, and constructs a four-modal unified representation closed-loop architecture driven by vehicle crashworthiness, which can effectively realize the full-process automation from natural language requirements to feasible structural solutions. Attached Figure Description

[0011] Figure 1 This is one of the structural schematic diagrams of the end-to-end safety design large model system for vehicle crashworthiness driven by the embodiments of this application; Figure 2 This is the second schematic diagram of the end-to-end safety design large model system for vehicle crashworthiness driven by the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the requirement parsing and constraint mapping module provided in an embodiment of this application; Figure 4 This is a schematic diagram of the impact-resistant structure generation module provided in the embodiments of this application; Figure 5 This is a schematic diagram of the collision simulation proxy prediction module provided in the embodiments of this application; Figure 6 This is a schematic diagram of the compliance verification module provided in the embodiments of this application; Figure 7 This is a schematic diagram of the closed-loop self-iterative optimization module provided in the embodiments of this application; Figure 8 This is a schematic diagram of the control method of the end-to-end security design large model system provided in the embodiments of this application. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0013] The terms "first" and "second," etc., used in the specification and claims of this application are used to distinguish different objects, not to describe a specific order of objects. For example, "first structural geometric feature" and "second structural geometric feature," etc., are used to distinguish different structural geometric features, not to describe a specific order of structural geometric features.

[0014] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0015] In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more, for example, multiple processing units means two or more processing units, multiple elements means two or more elements, etc.

[0016] The embodiments of this application are described below with reference to the accompanying drawings.

[0017] Figure 1This is a schematic diagram of the structure of the end-to-end safety design large model system for vehicle crashworthiness driven by the embodiments of this application, as shown below. Figure 1 As shown, it includes: The large language model module 1, and the input module 2, the multimodal feature extraction and analysis module 3, and the output module 4, which are respectively connected to the large language model module 1; Input module 2 is used to input crashworthiness design requirements information; The large language model module 1 and the multimodal feature extraction and analysis module 3 adopt a cross-modal interaction method to extract multimodal features and fuse them across modes using crashworthiness design requirements information. Based on the fused features, iterative analysis is performed under preset crashworthiness constraints to obtain the target three-dimensional crashworthiness structure scheme. Output module 4 is used to output the target three-dimensional impact-resistant structure scheme.

[0018] Specifically, the crashworthiness design requirements described in the embodiments of this application may include the type of vehicle to be designed, the crashworthiness standard of the front compartment, the installation space boundary of the front longitudinal beam, material and weight requirements, etc.

[0019] The multimodal feature extraction and analysis module described in this application refers to a comprehensive module that extracts multimodal features from crashworthiness design requirements and assists the large language model module in feature analysis and fusion processing. Specifically, it can be used to parse the natural language information of crashworthiness design requirements into engineering structural design boundary conditions, generate end-to-end three-dimensional geometry of the crashworthiness structure based on the engineering structural design boundary conditions, predict the nonlinear mechanical response of the generated crashworthiness structure, output compliance signals and reports through full-dimensional evaluation, provide clear optimization directions for closed-loop iteration, and realize end-to-end automatic optimization.

[0020] The preset crashworthiness constraints described in this application refer to all design constraint targets input by the user, covering original requirements such as regulatory conditions, performance indicators, installation boundaries, materials and processes, cost and weight, and are input-level targets of the system.

[0021] In the embodiments of this application, the large language model module serves as the foundation of the multimodal fusion large model and is the feature hub and cross-modal information interaction core of the safety design large model system. It is used to achieve unified representation and feature alignment among four different modal data: semantic constraint modality, three-dimensional (3D) geometric topology modality, collision physics field modality, and compliance verification modality.

[0022] In the embodiments of this application, after the user inputs the crashworthiness design requirements information through the input module, the input module converts the natural language requirements information into the corresponding token sequence and passes it into the large language model module.

[0023] Furthermore, in the embodiments of this application, the large language model module and the multimodal feature extraction and analysis module adopt a cross-modal interaction method. They use the token sequence corresponding to the crashworthiness design requirement information to perform multi-dimensional multimodal feature extraction and cross-modal fusion, including semantic constraints, geometric topology, collision physics field and compliance verification. Based on the fused features, iterative optimization analysis is performed under the preset crashworthiness constraints, and finally, the target three-dimensional crashworthiness structure scheme that meets the preset crashworthiness constraints can be obtained.

[0024] The end-to-end safety design large model system of this application, by introducing cross-modal interaction between the large language model module and the multimodal feature extraction and analysis module, constructs an end-to-end safety design large model system with crashworthiness as the core driving objective. It realizes the end-to-end mapping of the entire link from design feature input to design scheme output that meets collision safety requirements. It can effectively replace the traditional manual-led multi-round, time-consuming simulation iteration and physical test links, compressing the single-round design iteration cycle from the traditional several weeks to the second level, greatly improving the efficiency of the entire vehicle crashworthiness design process and shortening the vehicle safety development cycle.

[0025] Continue to refer to Figure 2 Based on the above embodiments, as an optional embodiment, the multimodal feature extraction and analysis module 3 includes a multimodal feature synthesis module 31 and a closed-loop self-iterative optimization module 32; the multimodal feature synthesis module 31 and the closed-loop self-iterative optimization module 32 are respectively connected to the large language model module 1; Large Language Model Module 1 is used to extract fusion features from the token sequence corresponding to the crashworthiness design requirements information and output a semantic fusion feature vector. The multimodal feature synthesis module 31 is used to generate an initial three-dimensional crash-resistant structure scheme based on semantic fusion feature vectors, and jointly with the large language model module to perform performance compliance verification on the initial three-dimensional crash-resistant structure scheme, and output the compliance fusion feature vector corresponding to the initial three-dimensional crash-resistant structure scheme. The closed-loop self-iterative optimization module 32 is used to determine the target structured engineering constraint set based on the compliance fusion feature vector, so that the large language model module can control the multimodal feature synthesis module to adjust the initial three-dimensional crash-resistant structure scheme according to the target structured engineering constraint set, until the target three-dimensional crash-resistant structure scheme that meets the preset crash-resistant constraint conditions is output.

[0026] Specifically, the structured engineering constraint set described in the embodiments of this application is used to characterize a data set that includes all constraints such as regulatory conditions, performance, installation boundaries, materials and processes, cost and weight.

[0027] In the embodiments of this application, the multimodal feature extraction and analysis module can specifically consist of a multimodal feature synthesis module and a closed-loop self-iterative optimization module. Through the multimodal feature synthesis module, the semantic fusion feature vector corresponding to the crashworthiness design requirements information is subjected to structured engineering constraint parsing to generate an initial three-dimensional crashworthiness structure scheme that meets the constraint conditions. This initial three-dimensional crashworthiness structure scheme is then jointly verified for performance compliance with the large language model module, outputting the initial three-dimensional crashworthiness structure scheme and its corresponding compliance fusion feature vector.

[0028] Continue to refer to Figure 2 Based on the above embodiments, as an optional embodiment, the multimodal feature synthesis module 31 includes a requirement analysis and constraint mapping module 311, a crash-resistant structure generation module 312, a collision simulation proxy prediction module 313, and a compliance verification module 314, which are respectively connected to the large language model module 1. The requirement parsing and constraint mapping module 311 is used to transform the semantic fusion feature vector into a structured engineering constraint set, so that the large language model module 1 can output the constraint semantic fusion feature vector based on the structured engineering constraint set and the semantic fusion feature vector; The crash-resistant structure generation module 312 is used to generate an initial three-dimensional crash-resistant structure scheme and the corresponding first structural geometric features based on the constraint semantic fusion feature vector, so that the large language model module 1 can output the constraint geometric fusion feature vector based on the structured engineering constraint set and the first structural geometric features. The collision simulation proxy prediction module 313 is used to output the performance feature vector corresponding to the initial three-dimensional crash-resistant structure scheme based on the constraint geometry fusion feature vector, so that the large language model module 1 can output the constraint geometry performance fusion feature vector according to the performance feature vector and the constraint geometry fusion feature vector; The compliance verification module 314 is used to output the compliance feature vector corresponding to the initial three-dimensional crash-resistant structure scheme based on the constraint geometric performance fusion feature vector, so that the large language model module 1 can output the compliance fusion feature vector based on the compliance feature vector.

[0029] Specifically, in the embodiments of this application, the multimodal feature synthesis module may include a requirement analysis and constraint mapping module, a crash-resistant structure generation module, a collision simulation proxy prediction module, and a compliance verification module, which are respectively connected to the large language model module.

[0030] The requirement parsing and constraint mapping module can also be described as the requirement parsing and multi-dimensional engineering constraint mapping module. It is used to parse the semantic fusion feature vector corresponding to the natural language requirement into the engineering structural design boundary conditions to obtain the structured engineering constraint set, so that the large language model module can output the constraint semantic fusion feature vector based on the structured engineering constraint set and the semantic fusion feature vector.

[0031] It should be noted that the structured engineering constraint set is a digital and structured execution carrier that completes entity standardization, constraint quantification, conflict detection and completion after the constraint data covered by the preset crashworthiness constraint conditions are processed by the requirement analysis and constraint mapping module. It serves as the direct basis for internal system iterative verification and structure generation.

[0032] The crashworthiness structure generation module, also known as the multi-constraint embedded crashworthiness structure generation module, is the core execution unit of the end-to-end safety design model for automotive crashworthiness. It generates end-to-end 3D geometry of the crashworthiness structure based on the boundary conditions of the engineered structural design. Its core function is to be fully compatible with the standardized engineering constraint set output by the upstream requirement analysis module, using user requirement constraints as the sole generation condition to generate a complete 3D structural model, from force transmission topology to mass production process details.

[0033] The collision simulation proxy prediction module can also be described as a physics-fidelity collision simulation proxy prediction module, which is used to achieve millisecond-level, high-fidelity prediction of the nonlinear mechanical response of the generated crashworthy structure.

[0034] The compliance verification module is used to evaluate the generated solution from all dimensions, output compliance signals and reports, and provide clear optimization directions for closed-loop iteration.

[0035] In a specific implementation, the large language model module processes the token sequence corresponding to the crashworthiness design requirement information to obtain a semantic fusion feature vector, and inputs it into the multimodal feature synthesis module. First, the requirement parsing and constraint mapping module transforms the semantic fusion feature vector into a structured engineering constraint set, and inputs it into the large language model module. The large language model module further performs feature fusion based on the structured engineering constraint set and the aforementioned semantic fusion feature vector, and outputs a constraint semantic fusion feature vector.

[0036] Furthermore, the collision-resistant structure generation module generates an initial three-dimensional collision-resistant structure scheme and the corresponding first structural geometric features by fusing the constraint semantic feature vector; the large language model module performs feature fusion on the structured engineering constraint set and the first structural geometric features, and outputs the constraint geometric fusion feature vector.

[0037] Here, the first structural geometric feature refers to the complete geometric feature vector corresponding to the initial three-dimensional crash-resistant structure output by the crash-resistant structure generation module after completing the generation of a single wheel. It includes all geometric information such as topological morphology, cross-sectional parameters, wall thickness distribution, and process features.

[0038] The collision simulation proxy prediction module performs performance prediction analysis on the constraint geometry fusion feature vector, outputting the performance feature vector corresponding to the initial 3D crash-resistant structure scheme. The large language model module further performs feature fusion processing based on the performance feature vector and the constraint geometry fusion feature vector, outputting the constraint geometry performance fusion feature vector. Then, the compliance verification module outputs the compliance feature vector corresponding to the initial 3D crash-resistant structure scheme based on the constraint geometry performance fusion feature vector, and inputs the compliance feature vector token sequence into the large language model module. The large language model module completes cross-modal alignment with the system's unified feature space, and finally outputs the compliance fusion feature vector.

[0039] like Figure 2 As shown, the large language model module 1 includes an encoder group module 11, a cross-modal fusion module 12, an attention layer 13, and a linear projection layer 14 connected in sequence. The encoder module 11 is used to encode any input token sequence and output the single-modal initial feature vector corresponding to any token sequence. The cross-modal fusion module 12 is used to perform cross-modal feature fusion between the initial feature vector of a single modality and the original semantic features of the demand by calling the expert unit, and outputs the preliminary fused feature vector corresponding to any token sequence; Attention layer 13 is used to constrain and enhance the initial fused feature vector through a preset attention mechanism, and output the enhanced multimodal feature sequence; Linear projection layer 14 is used to output the target fused feature vector based on the multimodal feature sequence; Each Token sequence includes a Token sequence corresponding to crashworthiness design requirements information, a Token sequence corresponding to structured engineering constraint set, a Token sequence corresponding to first structural geometric features, a Token sequence corresponding to performance feature vector, and a Token sequence corresponding to compliance feature vector. The target fusion feature vector includes semantic fusion feature vector, constraint semantic fusion feature vector, constraint geometric fusion feature vector, constraint geometric performance fusion feature vector, and compliance fusion feature vector.

[0040] Specifically, in the embodiments of this application, such as Figure 2As shown, the large language model module 1 can specifically include an encoder group module 11, a cross-modal fusion module 12, an attention layer 13, and a linear projection layer 14 connected in sequence. The encoder group module 11 is a parallel four-modal dedicated encoder, specifically including a semantic constraint encoder, a 3D geometric encoder, a physics encoder, and a compliance encoder. The output of the input module is converted into a standardized token sequence and then connected to the input of the semantic constraint encoder; the outputs of the requirement parsing and constraint mapping module and the closed-loop self-iterative optimization module are also converted into standardized token sequences and then connected to the input of the semantic constraint encoder; the output of the crash-resistant structure generation module is converted into a standardized token sequence and then connected to the input of the 3D geometric encoder; the output of the collision simulation proxy prediction module is converted into a standardized token sequence and then connected to the input of the physics encoder; the output of the compliance verification module is converted into a standardized token sequence and then connected to the input of the compliance encoder. The inputs of the requirement parsing and constraint mapping module, the crash-resistant structure generation module, the collision simulation proxy prediction module, the compliance verification module, and the closed-loop self-iterative optimization module are all connected to the output of the linear projection layer 14, receiving the unified multimodal fusion feature vector output by the multimodal large model base.

[0041] More specifically, in the embodiments of this application, the encoder group module can encode any input token sequence and output a single-modal initial feature vector corresponding to any token sequence. Here, for different types of input token sequences, the single-modal initial feature vectors point to the initial encoding results of four specific single modalities: semantic, geometric, physical field, and compliance.

[0042] For example, the token sequence corresponding to the crashworthiness design requirements and the token sequence corresponding to the structured engineering constraint set can be input into the semantic constraint encoder to generate their respective semantic modal initial feature vectors; the token sequence corresponding to the first structural geometric features can be input into the 3D geometry encoder to output its corresponding geometric modal initial feature vector; the token sequence corresponding to the performance feature vector can be input into the physics encoder to output its corresponding physics modal initial feature vector; and the token sequence corresponding to the compliance feature vector can be input into the compliance encoder to output its corresponding composite modal initial feature vector.

[0043] Furthermore, in the embodiments of this application, the cross-modal fusion module can call different expert units to perform feature fusion on the corresponding single-modal initial feature vector and the original requirement semantic features of the crashworthiness design requirement information, thereby outputting the corresponding preliminary fusion feature vector; the attention layer further enhances the preliminary fusion feature vector through a preset attention mechanism, outputting the enhanced multimodal feature sequence; finally, the linear projection layer projects the multimodal feature sequence to output the corresponding target fusion feature vector.

[0044] Regarding the information interaction process between the requirement parsing and constraint mapping module and the large language model module, any token sequence is the token sequence corresponding to the structured engineering constraint set, and the target fusion feature vector is the constraint semantic fusion feature vector. The detailed process is as follows: Input transformation: The structured engineering constraint set output by the requirement analysis and constraint mapping module is converted into a standardized token sequence through the constraint feedback interface and input into the semantic constraint encoder of the large language model module.

[0045] Single-modal encoding: The semantically constrained encoder performs self-attention encoding on the token sequence, mapping the discrete tokens to a 768-dimensional single-modal initial feature vector, i.e., the semantic modal initial feature vector.

[0046] Mixture of Experts (MoE) cross-modal fusion: The initial feature vector of the single modality enters the MoE cross-modal fusion architecture in the cross-modal fusion module. Through the Top-K gated routing with noise, the corresponding expert unit is called to complete the initial interactive fusion of the initial feature vector of the single modality with the original requirement semantic features (i.e., semantic fusion feature vector) of the crashworthiness design requirement information, and output the corresponding initial fusion feature vector.

[0047] Constraint-guided attention aggregation: The fused features enter the bidirectional cross-attention layer guided by global constraints. The initial feature vector of the single modality is used as a prefix token and concatenated into the key-value sequence of each layer's attention calculation, which strengthens the guiding role of constraints on the feature space and outputs a constraint-enhanced multimodal feature sequence.

[0048] Feature projection output: The multimodal feature sequence is mapped to a unified 1024-dimensional high-dimensional feature space through a linear projection layer, and finally outputs a constraint semantic fusion feature vector, which is sent to the collision-resistant structure generation module through the constraint feature adapter head as a constraint condition for generating collision-resistant structure schemes.

[0049] Here, the information interaction process between other modules and the large language model module strictly follows the fixed link of the four-layer network of the large language model module, which is processed sequentially. This belongs to the unified processing paradigm of the multimodal large model base for the five types of input words. For details, please refer to the processing method of each module / network layer in the large language model module. This application will not elaborate on this.

[0050] Furthermore, in the embodiments of this application, the closed-loop self-iterative optimization module adjusts and optimizes the original structured engineering constraint set according to the aforementioned compliance fusion feature vector to determine the target structured engineering constraint set. This allows the large language model module to control the multimodal feature synthesis module to adjust the initial three-dimensional crash-resistant structure scheme according to the target structured engineering constraint set, until the target three-dimensional crash-resistant structure scheme that meets the preset crash-resistant constraint conditions is output.

[0051] The system in this embodiment of the application builds a large language model module by introducing an encoder group module, a cross-modal fusion module, an attention layer and a linear projection layer, and jointly constructs a constraint mapping module, a crashworthiness structure generation module, a collision simulation proxy prediction module, a compliance verification module and a closed-loop self-iterative optimization module. This constructs a four-modal unified representation closed-loop architecture driven by the core of vehicle crashworthiness, which can effectively realize the full-process automation from natural language requirements to feasible structural solutions.

[0052] Traditional crashworthiness design relies heavily on engineers' experience, making it difficult to simultaneously address the safety requirements of various collision scenarios, including frontal, side, offset, and pole impacts. Furthermore, it fails to achieve global synergistic optimization of multiple dimensions such as crashworthiness, lightweighting, manufacturing cost, and manufacturability, easily falling into the limitations of local optima. The large-scale model system proposed in this application uses crashworthiness as the core driving factor. Based on massive amounts of collision data and multi-physics laws, it can perform global optimization across multiple scenarios and objectives, overcoming the limitations of human experience. While ensuring the overall vehicle crash safety performance meets global mainstream regulations and star rating requirements, it achieves an optimal balance between additional objectives such as lightweighting and cost control.

[0053] Traditional automotive crashworthiness design suffers from industry pain points such as fragmented data across design, simulation, and testing stages, lagging compliance control, and difficulty in accumulating engineering experience. This often leads to large-scale R&D rework due to non-compliance in later testing stages. The end-to-end large-scale model system proposed in this application connects the entire data chain of design parameters, simulation data, physical test results, and regulatory compliance requirements, enabling proactive crashworthiness compliance control from the design stage.

[0054] like Figure 3As shown, based on the above embodiments, as an optional embodiment, the requirement parsing and constraint mapping module 311 includes a feature encoding interface 3111, a vehicle crashworthiness vertical language model 3112, a constraint mapping structured layer 3113, and a constraint feedback interface 3114 connected in sequence; the feature encoding interface 3111 and the constraint feedback interface 3114 are also connected to the large language model module respectively. Feature encoding interface 3111 is used to convert semantic fusion feature vectors into corresponding first embedding vectors; The vertical language model 3112 for vehicle crashworthiness is used to output intermediate text results based on the first embedding vector; The constraint mapping structured layer 3113 is used for terminology standardization based on intermediate text results, and outputs a standardized set of structured engineering constraints. The constraint feedback interface 3114 is used to convert the structured engineering constraint set into the corresponding token sequence and feed it back to the large language model module.

[0055] Specifically, in the embodiments of this application, the requirement parsing and constraint mapping module may consist of a feature encoding interface, a vehicle crashworthiness vertical language model, a constraint mapping structured layer, and a constraint feedback interface.

[0056] In the embodiments of this application, the feature encoding interface receives the semantic fusion feature vector output by the large language model module, such as a 768-dimensional natural language semantic feature vector. This semantic fusion feature vector can be converted into an embedding vector (maintaining 768 dimensions) recognizable by the automotive crashworthiness vertical language model, i.e., the first embedding vector. After the automotive crashworthiness vertical language model is input with the aforementioned embedding vector, the output is an intermediate text result containing entity extraction results, preliminary constraint mapping results, conflict detection results, and requirement completion suggestions.

[0057] Furthermore, the constraint mapping structured layer takes the intermediate text output from the aforementioned vertical language model as input, completes terminology standardization through entity linking and normalization units, transforms qualitative descriptions into quantitative engineering indicators through constraint quantification units, and completes constraint conflict resolution and missing constraint completion through conflict detection and completion reasoning units. Finally, it can output a standardized structured engineering constraint set, which includes all constraints such as regulatory conditions, performance, boundaries, materials and processes, cost and weight.

[0058] Furthermore, the structured engineering constraint set is transformed into a unified-dimensional token embedding vector (768-dimensional) through the constraint feedback interface, and fed back to the large language model module, which serves as the foundation of the multimodal large model, to achieve cross-modal alignment of constraint features with geometric and physical features. The large language model module distributes the multimodal feature vector (i.e., constraint semantic fusion feature vector) that integrates constraint features to the collision-resistant structure generation module.

[0059] The system in this embodiment utilizes a four-level cascaded processing architecture to achieve automatic end-to-end transformation from natural language requirements to structured engineering constraints. Specifically, the feature encoding interface maps the semantic fusion feature vectors output from the upstream to standard embedding representations, providing standardized input for domain-specific models. The automotive crashworthiness vertical language model, leveraging its domain prior knowledge, parses general semantic features into intermediate text containing crashworthiness terminology, effectively compensating for the knowledge gaps of the general language model in automotive safety engineering. The constraint mapping structured layer performs terminology standardization on the intermediate text, eliminating synonym ambiguity and inconsistent expressions, and outputting a structured constraint set conforming to engineering specifications. The constraint feedback interface converts the structured constraint set into a token sequence and feeds it back to the large language model module, forming a closed-loop feedback mechanism that ensures the subsequent generation process of the large language model is always precisely guided by domain-specific constraints. Overall, this module achieves a deep integration of general language understanding capabilities and automotive crashworthiness expertise, significantly improving the accuracy and standardization of engineering constraint extraction, and ensuring inter-module consistency through a bidirectional interface design.

[0060] like Figure 4 As shown, based on the above embodiments, as an optional embodiment, the crash-resistant structure generation module 312 includes a constraint feature adapter head 3121, a macroscopic topology generation network 3122, a cross-sectional detail generation network 3123, a process feature generation network 3124, and a first performance prediction head 3125; the constraint feature adapter head 3121 and the first performance prediction head 3125 are also connected to the large language model module respectively; The constraint feature adapter 3121 is used to transform the constraint semantic fusion feature vector into the corresponding second embedding vector; The macroscopic topology generation network 3122 is used to output a voxelized macroscopic topology vector based on the second embedding vector; The section detail generation network 3123 is used to generate corresponding point cloud data based on the second embedding vector and the voxelized macroscopic topology vector; The process feature generation network 3124 is used to output the original three-dimensional impact-resistant structure scheme and the corresponding second structural geometric features based on point cloud data and the second embedding vector. The first performance prediction head 3125 is used to determine the performance deviation value based on the second structural geometric features during the training phase, so as to adjust the network parameters of the process feature generation network; during the inference phase, it generates an initial three-dimensional crash-resistant structural scheme with converged performance deviation value and its corresponding first structural geometric features, and converts the first structural geometric features into the corresponding token sequence and feeds it back to the large language model module.

[0061] Specifically, in the embodiments of this application, the macroscopic topology generation network, the cross-sectional detail generation network, and the process feature generation network adopt a strictly sequential cascaded structure, and the specific feature transfer method is as follows: The conditional embedding vector output by the macroscopic topology generator network input constraint feature adapter head, i.e. the second embedding vector, can output the corresponding voxelized macroscopic topology structure vector, such as a 64×64×64-dimensional voxelized macroscopic topology structure vector, which can define the overall force transmission path and main shape of the structure.

[0062] The cross-section detail generation network takes as input the voxelized macro-topology structure vector output by the macro-topology generation network, as well as the aforementioned constraint feature embedding vector (i.e., the second embedding vector), and outputs the corresponding point cloud data, such as 1024×3-dimensional high-precision point cloud data, which contains detailed features such as cross-sectional shape and wall thickness distribution.

[0063] The process feature generation network takes as input the point cloud data output by the cross-sectional detail generation network and the aforementioned second embedding vector, and outputs the original three-dimensional impact-resistant structure scheme, which is a complete 3D geometric model containing process features such as stamping fillets, welding edges, and draft angles. During the generation process, intermediate geometric features can be obtained from multiple denoising steps (e.g., every 100 denoising steps), and these intermediate geometric features constitute the second structural geometric features.

[0064] Among them, the macroscopic topology generation network can adapt to the input and output of design space voxels based on the 3D voxel diffusion model; the cross-sectional detail generation network can adapt to the detail generation of high-precision point clouds based on the point cloud Transformer diffusion model; and the process feature generation network can realize the generation of geometric details with fine process features based on the implicit diffusion model of the Signed Distance Function (SDF).

[0065] In the embodiments of this application, the crashworthy structure generation module may specifically consist of a constraint feature adapter head, a macroscopic topology generation network, a cross-sectional detail generation network, a process feature generation network, and a first performance prediction head. The constraint feature adapter head takes into account the constraint-semantic fusion feature vector output by the large language model module, such as a 1024-dimensional constraint-semantic fusion feature vector, and transforms this fusion feature vector into a conditional embedding vector (maintaining 1024 dimensions) that can be read by the diffusion model, i.e., the second embedding vector.

[0066] Furthermore, through the aforementioned three-level cascaded generative network—namely, the macroscopic topology generation network, the cross-sectional detail generation network, and the process feature generation network—the initial 3D crash-resistant structural scheme represented by SDF and the corresponding second structural geometric features in the generation process can be obtained. Then, a first performance prediction head is used during the training phase to determine the performance deviation value corresponding to the original 3D crash-resistant structural scheme based on the second structural geometric features, and to adjust the network parameters of the process feature generation network according to this deviation value. During the inference phase, an initial 3D crash-resistant structural scheme with converged performance deviation value and its corresponding first structural geometric features can be generated, and these first structural geometric features are converted into corresponding word sequences and fed back to the large language model module.

[0067] The system in this embodiment of the application uses a five-level progressive generation architecture to achieve end-to-end automatic generation from engineering constraint semantics to a manufacturable 3D crash-resistant structure. Specifically, the constraint feature adaptation head maps the constraint semantic fusion feature vector from upstream into a second embedding vector that adapts to the input format of the generation network, completing cross-modal feature alignment; the macro-topology generation network, driven by the second embedding vector, outputs a coarse-grained macro-topology structure vector in voxel space to determine the overall load-bearing skeleton layout of the crash-resistant structure; the cross-section detail generation network, guided by both macro-topology constraints and embedding vectors, completes the local cross-sectional geometric details in the form of point clouds, achieving a progressive resolution from coarse to fine; the process feature generation network integrates point cloud data and constraint embeddings to output the original 3D crash-resistant structure scheme and structural geometric features containing process feasibility features; the first performance prediction head, under the semantic supervision of the large language model module, predicts the performance deviation value based on the structural geometric features, and uses this deviation as a feedback signal to iteratively adjust the parameters of the process feature generation network, driving the generation process to evolve towards the performance deviation convergence direction. Overall, this module organically combines a three-level progressive generation process of "macro topology - cross-sectional details - process features" with a closed-loop performance feedback mechanism. While ensuring the manufacturability of the structure, it significantly improves the performance compliance and generation efficiency of the crash-resistant structure design, effectively avoiding the inefficiency of repeated manual iterations in traditional methods.

[0068] like Figure 5 As shown, based on the above embodiments, as an optional embodiment, the collision simulation proxy prediction module 313 includes a geometric performance feature adapter head 3131, a physical information neural fusion network 3132, a physical residual calculation unit 3133, and a second performance prediction head 3134; the geometric performance feature adapter head 3131 and the second performance prediction head 3134 are respectively connected to the large language model module. The geometric performance feature adapter 3131 is used to transform the constrained geometry fusion feature vector into a third embedding vector; The physical information neural fusion network 3132 is used to output preliminary collision physical field prediction results based on the third embedding vector; The physical residual calculation unit 3133 is used to determine the corresponding physical residual value based on the preliminary collision physical field prediction result, so as to optimize the network parameters of the physical information neural fusion network based on the physical residual value, so that the physical information neural fusion network outputs the full collision physical field prediction result. The second performance prediction head 3134 is used to extract the performance feature vector corresponding to the initial three-dimensional crash-resistant structure scheme from the full collision physics field prediction results, and feed back the token sequence corresponding to the performance feature vector to the large language model module.

[0069] Specifically, in the embodiments of this application, the collision simulation proxy prediction module may include: a geometric performance feature adaptation head, a physical information neural network, a physical residual calculation unit, and a second performance prediction head. The physical information neural network may specifically adopt a physical information neural network architecture that combines a graph neural network (GNN) and a physical information neural network (PINN).

[0070] In the embodiments of this application, the geometric performance feature adapter inputs the constraint geometry fusion feature vector output by the large language model module, such as a 1024-dimensional constraint-geometry fusion feature vector, and outputs a geometric structure embedding vector (512-dimensional), i.e., the third embedding vector, which can be directly read by the physical information neural fusion network. The physical information neural fusion network processes this geometric structure embedding vector, encodes geometric topological features through GNN, and introduces collision dynamics physical laws through PINN, outputting preliminary collision physical field prediction results, including stress field, strain field, acceleration field, etc.

[0071] Furthermore, after the initial collision physics prediction results are input into the physical residual calculation unit, the momentum conservation residual and energy conservation residual are calculated based on the collision dynamics control equations, and the physical residual values ​​are output. These residual values ​​can be used for backpropagation to optimize network parameters and output the full collision physics prediction results. After the final full collision physics prediction results are input into the second performance prediction head, the performance feature vector corresponding to the initial 3D crashworthy structure scheme can be extracted. Based on the performance feature vector, all core crashworthiness performance indicators (specific energy absorption, intrusion amount, maximum acceleration, etc.) can be extracted and output as performance indicator values. The performance feature vector is then fed back to the large language model module for use by the compliance verification module. Optionally, the performance evaluation results can also be fed back to the crashworthy structure generation module through the large language model module for real-time guidance during the generation process.

[0072] The system in this embodiment utilizes a physical information-driven four-level processing architecture to achieve high-precision proxy prediction and closed-loop feedback of the collision performance of initial three-dimensional crashworthy structural schemes. Overall, this module replaces traditional finite element full-scale simulation with a physical residual self-supervised mechanism, significantly reducing computational overhead while ensuring the accuracy of collision physics field prediction. Furthermore, its bidirectional interface design enables the large language model to continuously perceive the structural performance status, providing precise physical basis for subsequent scheme iteration and optimization.

[0073] like Figure 6 As shown, based on the above embodiments, as an optional embodiment, the compliance verification module 314 includes a multimodal feature interface port 3141, a multi-scale hierarchical neural network 3142, a multi-task compliance decoding layer 3143, a compliance report output port 3144, and a signal output port 3145; the multimodal feature interface port 3141 and the signal output port 3145 are respectively connected to the large language model module; The multimodal feature docking port 3141 is used to transform the constraint geometric performance fusion feature vector into a fourth embedding vector; The multi-scale hierarchical neural network 3142 is used to perform multi-scale feature encoding and fusion based on the fourth embedding vector, and output a full-dimensional fused feature vector. The multi-task compliance decoding layer 3143 is used to perform compliance verification analysis on the full-dimensional fused feature vector in different dimensions and output multi-dimensional compliance verification results. The compliance report output port 3144 is used to generate a compliance verification report based on the multi-dimensional compliance verification results. The signal output port 3145 is used to extract compliance feature vectors based on the multi-dimensional compliance verification results and output the token sequence corresponding to the compliance feature vectors to the large language model module.

[0074] Specifically, in the embodiments of this application, the compliance verification module may include a multimodal feature interface port, a multi-scale hierarchical neural network, a multi-task compliance decoding layer, a compliance report output port, and a signal output port. The multi-task compliance decoding layer may include three verification heads: a performance constraint compliance verification head, a manufacturability compliance verification head, and a boundary constraint compliance verification head.

[0075] More specifically, in the embodiments of this application, the constraint-geometry-performance fusion feature vector output by the large language model module base is input to the multimodal feature docking port. For example, a 1024-dimensional constraint-geometry-performance fusion feature vector can output an embedding vector recognizable by the GNN network (e.g., 768-dimensional), i.e., the fourth embedding vector. The multi-scale hierarchical neural network, which can also be described as the main body of a multi-scale hierarchical GNN network, takes the aforementioned embedding vector as input and, through multi-scale encoding and fusion of constraint, geometric, and performance features, can output the corresponding full-dimensional fusion feature vector.

[0076] Furthermore, the multi-task compliance decoding layer takes the aforementioned full-dimensional fused feature vector as input and performs compliance verification analysis in different dimensions through three parallel verification heads. It can output multi-dimensional compliance verification results, which can specifically include compliance / non-compliance signals, non-compliant area coordinates, and performance difference information output by the performance constraint compliance verification head; compliance / non-compliance signals, non-compliant area coordinates, and non-compliant process clause information output by the manufacturability compliance verification head; and compliance / non-compliance signals, out-of-bounds position coordinates, and boundary difference information output by the boundary constraint compliance verification head.

[0077] Furthermore, in this embodiment, by inputting the output results of the three verification heads through the compliance report output port and integrating all verification results, a visualized compliance verification report can be generated and output. This structured compliance verification report can include information such as the results of all verification items, problem descriptions, and optimization suggestions. Finally, by inputting the multi-dimensional compliance verification results of the three verification heads through the signal output port, and by comprehensively judging whether the overall solution is qualified, a compliance feature vector can be extracted based on the multi-dimensional compliance verification results. Specifically, this can include global compliance / non-compliance signals, and the token sequence corresponding to the compliance feature vector is output to the large language model module for the closed-loop self-iterative optimization module to perform iterative optimization operations.

[0078] The system in this embodiment of the application features a compliance verification module that, through a five-level collaborative processing architecture, achieves multi-dimensional automated compliance verification and closed-loop feedback for crash-resistant structural solutions. This module, through a collaborative design of multi-scale feature fusion and multi-task parallel decoding, achieves a leap from single-dimensional verification to comprehensive multi-dimensional compliance assessment, significantly improving the coverage and efficiency of compliance verification for crash-resistant structural solutions. Furthermore, a bidirectional interface ensures the continuous and effective constraint of compliance on the overall design process.

[0079] like Figure 7 As shown, based on the above embodiments, as an optional embodiment, the closed-loop self-iterative optimization module 32 includes a compliance signal receiving unit 321, an iterative judgment unit 322, and an optimization generation link 323 connected in sequence; the compliance signal receiving unit 321 and the optimization generation link 323 are respectively connected to the large language model module; The compliance signal receiving unit 321 is used to determine the non-compliance information of the initial three-dimensional crash-resistant structure scheme based on the compliance fusion feature vector; The iteration judgment unit 322 is used to send non-compliant item information to the generation optimization link when it is determined that the current iteration does not meet the preset iteration termination condition; The optimized link 323 is used to adjust and reconstruct the structured engineering constraint set based on the non-compliance information, determine the target structured engineering constraint set, and convert the target structured engineering constraint set into the corresponding Token sequence and output it to the large language model module.

[0080] Specifically, in the embodiments of this application, the closed-loop self-iterative optimization module may include a compliance signal receiving unit, an iteration judgment unit, and a generation optimization link. After the compliance verification module outputs a compliance feature vector including global compliance signals and detailed verification results, it is converted into a corresponding Token sequence and output to the large language model module. Then, the compliance signal receiving unit, by inputting the compliance fusion feature vector sent by the large language model module, can directly determine the non-compliance information of the initial three-dimensional crashworthiness structure scheme based on the compliance fusion feature vector, including information such as non-compliance type, severity, specific location, and difference. When the iteration judgment unit determines that the current iteration does not meet the preset iteration termination condition (all constraints of the preset crashworthiness constraint condition are met or the maximum number of iterations is reached), it will output a signal to continue iteration and forward the structured non-compliance information to the generation optimization link.

[0081] Furthermore, in this embodiment, non-compliant item information of the optimized link input structure is generated. By reconstructing and optimizing the aforementioned structured engineering constraint set, the weight coefficient of non-compliant items in the loss function is increased. Finally, the reconstructed constraint conditions can be output, i.e., the target structured engineering constraint set is obtained. The target structured engineering constraint set can be converted into a corresponding token sequence and output to the large language model module to guide the structure generation module for targeted optimization.

[0082] In embodiments of this application, the closed-loop self-iterative optimization module may further include a model self-optimization link and a requirement verification link, wherein the model self-optimization link is connected to the large language model module. The model self-optimization link is used to utilize the first structural geometric features, performance feature vectors, and compliance feature vectors in the current iteration process, using these data as online training samples to fine-tune the model parameters of the large language model module, update the model parameters, and improve the accuracy of subsequent generation and prediction.

[0083] In this embodiment, the requirement verification link is used to analyze the reasons for the constraint conflict when it is determined that the current iteration meets the preset iteration termination condition but does not meet the preset collision resistance constraint condition, evaluate the rationality of the requirement, and output requirement adjustment suggestion information to the requirement parsing module, which then sends it to the front end to prompt the user to adjust the requirement priority or target value.

[0084] The system in this embodiment utilizes a closed-loop self-iterative optimization module. This module employs a five-in-one closed-loop architecture encompassing compliance awareness, condition judgment, constraint reconstruction, parameter fine-tuning, and requirement feedback to achieve autonomous iterative convergence of crash-resistant structural designs and continuous system self-evolution. This module organically integrates compliance assessment, adaptive constraint reconstruction, online fine-tuning of model parameters, and requirement feasibility feedback, constructing a complete closed loop from single generation to continuous self-optimization. This significantly improves the iterative efficiency of crash-resistant structural design and the compliance convergence rate of the final design.

[0085] The following sections will describe the construction process and related implementation methods of each module in the end-to-end safety design large model system provided in this application embodiment, including the large language model module, the requirement parsing and constraint mapping module, the crash-resistant structure generation module, the collision simulation proxy prediction module, the compliance verification module, and the closed-loop self-iterative optimization module.

[0086] In the embodiments of this application, the specific construction process of the large language model module may include the following steps: S101, based on a dedicated knowledge base and database in the field of automotive crashworthiness, such as Figure 2 As shown, four-modal paired data were collected throughout the entire process of vehicle crashworthiness development, and data cleaning, deduplication, and outlier removal were completed.

[0087] Among them, the dedicated knowledge base and database in the field of automotive crashworthiness can provide underlying data support for the entire system, including: regulatory standards library, vehicle platform database, crashworthiness structure database, full-scale collision simulation database, materials and processes database, and manufacturing constraint database. The minimum training dataset size of the database is: ≥1000 sets of complete paired data for four vehicle modes. Specifically, the complete paired data for four modes includes requirement documents (semantic constraint mode), CAD models (3D geometric topology mode), simulation results (collision physics field mode), and compliance verification reports (compliance verification mode) that correspond one-to-one with the same vehicle model project, the same development stage, and the same working condition.

[0088] More specifically, in the embodiments of this application, the storage format of the regulatory standard library is structured XML, including regulatory number, applicable region, collision condition type, mandatory / recommended item identifier, performance index threshold, test conditions, effective version, and a matching "Regulatory Clause-Engineering Indicator Mapping Comparison Table" to realize the automatic mapping of regulatory clauses to quantitative engineering indicators.

[0089] The vehicle platform database includes vehicle class, platform architecture, hard point coordinates, installation space boundaries, vehicle weight target, and wheelbase / track parameters; the crashworthiness structure database includes part ID, CAD model, topology features, material grade, cross-sectional parameters, process features, and performance data for corresponding working conditions; the full collision simulation database includes vehicle ID, working condition type, mesh parameters, material cards, solver settings, time step, acceleration curves, passenger compartment intrusion, stress field distribution, and structural energy absorption data, with a mandatory unified timing alignment rule of 0-150ms collision duration and 1ms fixed time step; the materials and processes database includes material grade, mechanical property parameters, forming process, cost coefficient, welding performance, and stamping limits; and the manufacturing constraints database includes stamping process constraints, welding process constraints, assembly tolerance requirements, mold cost thresholds, and mass production feasibility judgment rules.

[0090] S102, standardize and preprocess the four-modal data. The four-modal data includes semantic constraint modal data, 3D geometric modal data, collision physics modal data, and regulatory compliance modal data.

[0091] Here, the semantic constraint modal data comes from the regulatory standard library (regulatory clause text) and the vehicle platform database (platform constraint text), supplemented by requirement documents from previous projects. For the semantic constraint modal data, structured encoding, priority labeling, constraint type normalization, and transformation into a standardized token sequence are completed.

[0092] The 3D geometric modal data is sourced from a crash-resistant structure database (part CAD models, topological features). The 3D geometric modal data undergoes format standardization, rigid body alignment, scale normalization, and topological feature annotation, and is then converted into 1024×3D standardized point cloud data.

[0093] The collision physics modal data comes from the full collision simulation database (acceleration curves, intrusion amount, stress field, energy absorption data) and the material process database (material mechanical property parameters). For this modal data, time alignment, dimension normalization, field data gridding and key performance index labeling are completed, and it is uniformly converted into 64×64×64×3-dimensional gridded field data.

[0094] The regulatory compliance modal data originates from the regulatory standard library (compliance rules), the manufacturing constraint database (manufacturing compliance rules), and compliance verification reports from historical projects. For this modal data, knowledge graph standardization, entity linking, compliance rule coding, and mandatory item annotation are performed, transforming it into standardized graph structure data.

[0095] The specific process for constructing the training dataset for subsequent model training can be as follows: using "same vehicle model project + same development stage + same collision condition" as the unique pairing identifier, extract four types of data under the corresponding identifier from the above six databases; process the four types of data according to the standardized preprocessing process to form a complete pairing sample containing "semantic constraint token sequence + 1024×3 point cloud + 64×64×64×3 physical field + compliance graph structure"; finally, after data cleaning, deduplication, and outlier removal, construct the final training dataset.

[0096] S103 is used to build four modality-specific encoders: a semantic constraint encoder, a 3D geometry encoder, a physics field encoder, and a compliance encoder.

[0097] The semantic constraint encoder can adopt the BERT-base architecture, which can include a 12-layer Transformer encoder and 12 attention heads, with a hidden layer dimension of 768. The input is a normalized token sequence with a maximum length of 512, and the output is a 768-dimensional semantic feature vector. The activation function is GELU.

[0098] The 3D geometric encoder can use the “PointNet++” hierarchical point cloud encoder, which can include 4 multi-scale SA layers, each SA layer has 3 sets of convolutional kernels of different scales, its input is a 1024×3-dimensional normalized point cloud, and its output is a 1024-dimensional geometric feature vector.

[0099] The physics encoder can adopt a hybrid architecture of 3D CNN and Transformer, specifically including 8 layers of 3D convolutional layers (3×3×3 kernels, stride 2), 4 layers of Transformer encoders and 8 attention heads. Its input is 64×64×64×3-dimensional gridded collision field data, and its output is a 1024-dimensional physics feature vector.

[0100] The compliance encoder can specifically include a 2-layer Graph Attention Network (GAT) and 8 attention heads, with a hidden layer dimension of 256. Its input is a compliance knowledge graph structure, and its output is a 768-dimensional initial compliance feature vector, i.e., a single-modal initial feature vector.

[0101] S104. A cross-modal fusion architecture for MoE is constructed, resulting in a cross-modal fusion module. Specifically, it can be configured with 16 expert units, including 4 modality-specific experts and 8 globally shared experts. Each expert unit adopts a standard Transformer FFN structure with a hidden layer dimension of 2048, a GELU activation function, and a dropout coefficient of 0.1. It is equipped with a noisy Top-K gated routing network, where K=2, and a load balancing loss function to prevent expert collapse.

[0102] S105 constructs a globally constrained bidirectional cross-attention layer, specifically a 6-layer network with 12 attention heads per layer and a hidden layer dimension of 768. Semantic constraint features are used as prefix tokens and concatenated into the key-value sequence of each layer's attention calculation. This enables bidirectional attention interaction between geometric / physical, semantic / compliant, and geometric / pluralistic scales, along with a masking mechanism and attention weight calculation formula.

[0103] S106 constructs a linear projection layer to map the fused multimodal token sequence to a unified 1024-dimensional high-dimensional feature space, adapting to the different task requirements of various downstream modules (including the requirement analysis and constraint mapping module, the crash-resistant structure generation module, the collision simulation proxy prediction module, and the compliance verification module).

[0104] S107 provides dedicated feature adapters for each downstream module, transforming the output vector of the unified feature space into feature inputs that can be directly read by the corresponding module, thus achieving seamless connection between the base and the downstream modules.

[0105] It should be noted that the large language model module, as the foundation of the multimodal fusion large model, is the sole feature hub and data transfer center of the entire system. All downstream modules do not directly interact with each other; all data flow is achieved through the unified feature space of the multimodal fusion large model foundation. The specific logic is as follows: The input data of all downstream modules comes from the unified multimodal feature vector output by the multimodal fusion large model base; the output data of all downstream modules is first converted into a token embedding vector of a unified dimension and fed back to the multimodal fusion large model base; the multimodal fusion large model base realizes the fusion and alignment of the output features of different modules through a bidirectional cross-attention mechanism; the fused features are then distributed by the multimodal fusion large model base to the corresponding downstream modules for further processing.

[0106] S108 can specifically employ a two-stage training mechanism to complete model training and resolve circular dependency issues.

[0107] The first stage (pre-training) involves using a contrastive learning loss to align the features of the four modalities based on a four-modal pairing dataset. The optimization goal is to reduce the feature distance between different modalities of the same paired sample and increase the feature distance between different samples, thereby completing the convergence of the backbone network.

[0108] Phase Two (Fine-tuning): Freeze the parameters of the main network of the base station, and fine-tune only the task networks of the multi-task feature adapter head and each downstream module. The weighted sum of the losses of all downstream tasks is used as the total loss. Total loss The calculation formula is: ; in To analyze the loss for the requirements, For structural generation loss, For performance prediction loss, Losses due to compliance verification.

[0109] Among them, the demand parsing loss is the weighted sum of the losses of multiple tasks, and its calculation formula is: ; In the formula, For entity extraction loss, the cross-entropy loss function can be used to measure the accuracy of entity recognition and classification; To constrain the mapping loss, the mean squared error (MSE) loss function can be used to measure the mapping accuracy from natural language descriptions to quantitative engineering indicators; For conflict detection loss, a binary cross-entropy loss function can be used to measure the accuracy of constraint conflict identification; To compensate for the loss in demand completion, the cross-entropy loss function can be used to measure the accuracy of missing constraint completion.

[0110] Among them, the compliance verification loss is also a weighted sum of the losses from multiple tasks, and the calculation formula is as follows: ; In the formula, To account for the performance constraint compliance verification loss, a binary cross-entropy loss function can be used to measure the accuracy of the judgment on whether the performance index meets the threshold requirements. For manufacturability compliance verification loss, a binary cross-entropy loss function can be used to measure the accuracy of the judgment on whether the manufacturing process constraints are met. To measure the accuracy of the judgment on whether the boundary constraints are satisfied, a binary cross-entropy loss function can be used to determine the loss of boundary constraints compliance.

[0111] In this embodiment, the training hyperparameters can be set as follows: optimizer AdamW (weight decay 1e-4), initial learning rate 5e-5 (cosine annealing decay), batch size 32, training epochs 100, and early stopping mechanism is to stop training if the validation set perplexity does not decrease for 10 consecutive epochs; the training termination condition can be set as geometric generation constraint satisfaction rate ≥95%, performance prediction accuracy ≥95%, and compliance verification accuracy ≥99%.

[0112] In some embodiments, the process of building the requirement analysis and constraint mapping module can specifically adopt the following steps: S201 selects a large language model base that has strong contextual understanding capabilities, long text support, excellent reasoning capabilities, and is open-source and commercially viable.

[0113] S202: Extract corpora from domain-specific knowledge bases and databases, process them according to the pre-training data format, and construct an incremental pre-training dataset.

[0114] The incremental pre-training dataset includes: regulatory conditions entities (collision condition type, regulatory version, rating requirements), performance indicators entities (core crashworthiness indicators, target values, limits), vehicle boundary entities (vehicle class, hard point boundaries, installation space, platform constraints), materials and processes entities (material type, strength requirements, forming process), manufacturing constraints entities (stamping, welding, assembly process requirements), and cost and weight entities (weight target, cost ceiling, lightweighting requirements).

[0115] S203, based on the incremental pre-trained dataset, extracts entity nodes and their relationships, constructs a knowledge graph, and defines constraint rules, coupling relationships, and conflict rules between entities, providing a reasoning basis for subsequent conflict detection and constraint completion.

[0116] S204, based on an incremental pre-training dataset, uses self-supervised language model loss as its core and convergence of validation set perplexity to a preset threshold as the training termination condition to incrementally train a large language model base and obtain a vertical language model for vehicle crashworthiness.

[0117] S205 constructs a dedicated instruction set for analyzing vehicle crashworthiness requirements. Each instruction contains a supervised sample of input requirements and labeled output.

[0118] Specifically, in the embodiments of this application, the instructions include entity extraction instructions, constraint mapping instructions, conflict detection instructions, and requirement completion instructions. Entity extraction instructions are used to accurately extract core entities and indicators for different types of requirement texts; constraint mapping instructions are used to transform natural language requirements into standardized and quantified sets of engineering constraints; conflict detection instructions are used to identify constraint conflicts in requirements and output conflict reasons and optimization suggestions; requirement completion instructions are used to complete the necessary constraints missing in the requirements and output the basis for completion and the adjustable range.

[0119] S206, combining the dedicated instruction set for analyzing vehicle crashworthiness requirements, and performing supervised fine-tuning based on the vehicle crashworthiness vertical language model obtained in step S204, to obtain an instruction-enhanced vehicle crashworthiness vertical language model.

[0120] S207 establishes a feature encoding interface to convert the semantic feature vectors output by the upstream unified multimodal base into embedding vectors that can be recognized by the automotive crashworthiness vertical language model.

[0121] S208 establishes a constraint mapping structured layer, including entity linking and normalization units, constraint quantization units, and conflict detection and completion reasoning units.

[0122] More specifically, the entity linking and normalization unit is used to link the extracted entities with the standard terms in the knowledge base, realize the standardization and normalization of colloquial expressions, and output a unified entity identifier; the constraint quantification unit is used to transform qualitative descriptions into quantitative engineering indicators and boundary conditions based on the knowledge base, and complete the mathematical encoding of constraints; the conflict detection and completion reasoning module is used to perform constraint conflict detection and missing constraint completion based on the knowledge graph rule engine and large model reasoning built in step S203.

[0123] S209 establishes a constraint feedback interface, transforming the structured constraint set into a unified-dimensional token embedding vector, which is then fed back to the multimodal large model base to achieve cross-modal alignment of constraint features with geometric and physical features.

[0124] In some embodiments, the construction process of the impact-resistant structure generation module may specifically include the following steps: S301, Select the base model. This base model supports 3D conditional generation, can flexibly embed custom conditional guidance and loss functions, and adapts to the geometric features of thin-walled vehicle body structures. The generation accuracy meets engineering design requirements. Specifically, open-source bases such as the 3D Diffusion model and the Shap-E model can be used.

[0125] S302, based on the above base model, the network architecture is modified to obtain a multi-constraint embedded crash-resistant structure generation model, specifically including three networks: macro topology, cross-sectional details, and process features. At the same time, feature transfer interfaces between the three networks are developed to ensure that the output features of the previous level can be seamlessly used as the generation conditions of the next level, while reserving feature docking interfaces with the unified multimodal base.

[0126] S303, set differentiated weight coefficients according to constraint priority, and construct the multi-constraint total loss function, as shown in the following formula: ; in, This represents the total loss due to the first-level constraint. This represents the total loss due to the second-level constraint. This represents the total loss due to Level 3 constraints. , , These are the weighting coefficients corresponding to different levels of loss.

[0127] Here, the default value for the weighting coefficient can be... , , When the first-level constraint is not satisfied, Increased to 0.8 to ensure that hard constraints are met first.

[0128] The first-level constraints include boundary constraint losses (penalizing geometric generation that exceeds the vehicle's hard points and installation space to ensure the structure is within the design space), topology force transmission constraint losses (penalizing topological structures with interrupted force transmission, stress concentration, and insufficient energy absorption efficiency), and manufacturing process constraint losses (penalizing geometric features that violate stamping, welding, and assembly hard constraints).

[0129] The total loss due to the first-level constraint can be calculated using the following formula: ; The specific formulas for each sub-loss are as follows: Boundary constraint loss: ; in, To generate a set of structural elements; The distance from the voxel to the outer boundary of the design space; As the tolerance threshold, it can be taken as... .

[0130] Topological force constraint loss: ; in, For maximum stress, For the allowable stress of the material, For the target energy absorption value, This represents the actual energy absorption value.

[0131] Manufacturing process constraint losses: ; in, To achieve the minimum stamping fillet, To design rounded corners, Minimum weld edge width, To design the width of the welding edge.

[0132] The second-level constraint includes a performance constraint loss (based on real-time performance prediction results, penalizing the geometric generation of performance deviations from the target value), which can be calculated using the following formula: ; in, To predict performance values ​​in real time, To constrain the target value, Let be the mean square error function.

[0133] The third-level constraint includes cost weight loss (penalizing the generated result for exceeding the weight cost limit), which can be calculated using the following formula: ; in, This is the actual weight. This is the upper limit of weight. For actual cost, This is the upper limit of the cost.

[0134] The standard denoising loss for diffusion models typically uses the mean squared error (MSE) loss, which measures the difference between the noise predicted by the model and the actual noise. Its calculation formula is as follows: ; in, For diffusion time step, For the original clean data, This is real noise. For the noise predicted by the model, This is the data at step t after adding noise.

[0135] S304 embeds the multi-constraint total loss function into each time step of the backpropagation denoising of the diffusion model, and weights it with the standard denoising loss of the diffusion model to obtain the total loss for each backpropagation step. This enables constraints to guide the entire generation process.

[0136] The weighted summation formula for the total multi-constraint loss and the standard denoising loss is as follows: ; Among them, it is possible to set S305, constructs the interface with the multimodal fusion large model base, including a constraint feature adapter head and a first performance prediction head. The constraint feature adapter head transforms the constraint semantic fusion feature vector output by the base into a conditional embedding vector readable by the diffusion model, achieving accurate transmission of constraints. The first performance prediction head, as a real-time performance prediction head, establishes a bidirectional interface with the base's physical field encoder. During the generation process, every 100 denoising steps, intermediate geometric features are transmitted to the base, and the performance prediction result is output in milliseconds. Performance deviations are fed back to the loss function, achieving real-time embedding of performance constraints.

[0137] In some embodiments, the process of building a collision simulation proxy prediction module may specifically include the following steps: S401, Select a base model to construct a performance prediction model, namely a physical information neural fusion network. This base model can take the geometric structure encoding of the crash-resistant structure as input and the corresponding performance prediction result as output.

[0138] S402, based on the control equations, constitutive relations and contact conditions of the collision scenario, constructs a modular physical residual calculation unit and designs a multi-component loss function with adaptive weights.

[0139] More specifically, the physical residual calculation unit is constructed based on the collision dynamics control equations, specifically including momentum conservation residual terms and energy conservation residual terms; a multi-component loss function with adaptive weights is designed. Specifically, it can be expressed as: ; in, For the predicted collision force, For realistic collision force; For the predicted stress tensor, The predicted acceleration; For the predicted internal energy, It is real internal energy; MSE Let be the mean square error function. MAE This represents the mean absolute error.

[0140] Here, weight Adaptive dynamic balancing is performed based on the gradient norm proportion of each component during training iterations, with the update rule as follows: ,in As the smoothing coefficient, it can be taken as... ; Indicates the total performance prediction loss The next j The gradient of each sub-loss component. This mechanism ensures the numerical stability of physical quantities with different dimensions during backpropagation.

[0141] S403 establishes the interface with the multimodal fusion large model base, including a geometric performance feature adapter head and a second performance prediction head. The geometric performance feature adapter head transforms the constraint geometric fusion feature vector output by the upstream unified multimodal fusion large model base into a geometric structure embedding vector that the collision simulation proxy prediction module can directly read. The second performance prediction head, as a real-time performance prediction feedback head, can be used to feed the performance evaluation results back to the multi-constraint embedded crashworthiness generation module in real time through the multimodal fusion large model base, guiding the generation of crashworthiness schemes and improving the quality of the initial structural scheme generated in a single round.

[0142] In some embodiments, the process of building a compliance verification module may specifically include the following steps: S501 can select a multi-scale hierarchical GNN network as the base to build an input layer and a feature encoding layer, i.e., a multi-scale hierarchical neural network. Among them, the input layer processes the constraint-geometric-performance feature fusion vector, and the feature encoding layer completes the full-dimensional feature encoding.

[0143] S502 establishes a multi-task compliance decoding layer, including three verification heads: a performance constraint compliance verification head, a manufacturability compliance verification head, and a boundary constraint compliance verification head. The performance constraint compliance verification head outputs a performance constraint threshold compliance / non-compliance signal, the coordinates and characteristics of the node / region causing performance failure, and the difference between the actual performance and the constraint threshold. The manufacturability compliance verification head outputs an overall manufacturability compliance / non-compliance signal, the coordinates and characteristics of the node / region causing manufacturability failure, and the corresponding process rule clause. The boundary constraint compliance verification head outputs a boundary constraint compliance / non-compliance signal, the specific location of interference / boundary violation, and the difference between the violation and the boundary threshold. The results from these three verification heads form a multi-dimensional compliance verification result.

[0144] S503 establishes a compliance report output port to integrate the output of the compliance code layer and generate a visual verification report; S504, constructs a docking port with the multimodal fusion large model base, used to convert the constraint-geometric-performance feature fusion vector output by the upstream unified multimodal fusion large model base into an embedding vector that the compliance verification module network can recognize; S505 is used to construct a signal output port for connecting a closed-loop self-iterative optimization module via a multimodal fusion large model base.

[0145] In some embodiments, the closed-loop self-iterative optimization module is used to achieve end-to-end automatic optimization. When an unacceptable signal exists, iteration is automatically triggered; when the solution is acceptable (i.e., the loss function value converges), iteration is automatically terminated; simultaneously, a maximum iteration threshold is set to avoid infinite iteration, and the entire process is recorded and traceable, meeting development requirements. Its setup process may specifically include the following steps: S601 constructs a generation optimization link, feeding back the reconstructed optimization objectives and constraints to the crash-resistant structure generation module through the multimodal fusion large model base. In the constraint function of the crash-resistant structure generation module, the weight coefficient of the non-conforming items is increased to guide the directional optimization of geometric features.

[0146] The S602 can also construct a model self-optimization link, using the geometric, simulation, and evaluation data from the current iteration as online training samples, feeding them back to the multimodal fusion large model base to complete online fine-tuning of the model and continuously improve model accuracy.

[0147] S603 can also construct a requirement verification chain. If multiple iterations still fail to meet the preset collision resistance constraints, the constraint conflict information is fed back to the requirement parsing and constraint mapping module, and requirement adjustment suggestions are output to the front end to prompt the user to adjust the requirement priority or target value, thus avoiding invalid iterations.

[0148] It should be noted that the implementation methods for the model training and inference stages in this application differ in the following aspects: First, the two methods use different types of loss. During the training phase, the multi-task joint training total loss, i.e., the total loss, is used. It integrates the task losses of four modules: requirement analysis, structure generation, performance prediction, and compliance verification. Its purpose is to jointly optimize the parameters of all networks in the entire system, representing a global training loss across modules. During the inference phase, only the total loss of multi-constraint satisfaction is used, i.e., the third-level constraint loss from the structure generation module. It is only used to evaluate the degree to which the current structural scheme meets the engineering constraints. It does not involve the task losses of the requirement analysis and constraint mapping module, the collision simulation proxy prediction module, and the compliance verification module. These three modules only perform forward inference output during inference and do not need to calculate their own losses.

[0149] The structure generation module is trained using a single-step loss function. As the smallest unit of execution, after aggregation, it becomes As a module-level loss component, it is included in the global total loss; during the inference phase, only the total loss of multi-constraint satisfaction is used, i.e., the total loss of the third-level constraints of the structure generation module. It is only used to evaluate the degree to which the current structural scheme meets the engineering constraints, and does not participate in backpropagation and parameter updates.

[0150] Second, the objects and mechanisms of iteration differ. During training iteration, the core function of loss-driven parameter updates is backpropagation of gradients, updating the network weight parameters of the large model base and the four downstream modules layer by layer. In each iteration, the same batch of training data is input, and the model learns patterns through backpropagation of loss, iterating on the model's own capabilities. During inference iteration, loss-guided scheme reconstruction uses loss only as an evaluation signal and optimization direction signal, without backpropagation or modification of any network parameters. In each iteration, the input constraint weights are dynamically adjusted, generating a completely new structural scheme, iterating on the structural scheme itself tailored to the current user needs.

[0151] Third, the termination objectives differ. Model training iterations terminate when the overall model performance converges (generation accuracy, prediction accuracy, and compliance accuracy all meet targets), and the trained model weights are output. Model inference iterations terminate when the total loss due to multiple constraints of the current generation scheme is required to be output. If the convergence threshold is lower than the preset threshold and all first-level hard constraint losses are 0, the final structural scheme that meets the user's requirements will be output.

[0152] In a specific embodiment of this application, a detailed description is provided, taking into account a specific implementation method of the front longitudinal beam crashworthiness structure design for a Class A SUV: Step 1: Input Requirements: Receive crashworthiness design requirements information through the input / output module: For this A-class SUV platform, develop a frontal crashworthiness structure that meets the C-NCAP frontal collision requirements. The front longitudinal beam installation space boundaries are 800mm in the X direction, 120mm in the Y direction, and 150mm in the Z direction. The material requirement is hot-formed steel, with the weight of the hot-formed steel front longitudinal beam ≤3.5kg and the manufacturing cost ≤120 yuan, meeting the requirements of stamping and welding mass production processes.

[0153] Step two, perform requirement analysis and constraint mapping: The requirement analysis and constraint mapping module completes entity extraction, constraint quantification, and conflict detection for the preset crashworthiness constraints, outputting a structured engineering constraint set, which may specifically include: Regulatory constraints: 100% frontal rigid barrier collision; Performance constraints: front longitudinal beam specific energy absorption ≥ 20 kJ / kg, crew compartment firewall intrusion ≤ 120 mm; Boundary constraints: Hard point boundaries of 800mm in the X direction, 120mm in the Y direction, and 150mm in the Z direction, with no interference; Material and process constraints: hot-formed steel, minimum stamping fillet radius R3mm, welded edge width ≥15mm, draft angle ≥3°; Cost and weight constraints: weight ≤ 3.5kg, unit manufacturing cost ≤ 120 yuan.

[0154] Step 3: Generate the crash-resistant structure scheme: The crash-resistant structure generation module uses the structured engineering constraint set as a condition and generates a three-dimensional structural scheme for the front longitudinal beam through a three-level cascaded generation network. During the generation process, a real-time performance prediction is completed every 100 denoising steps, with constraint guidance throughout the process. The total generation time is 28 seconds.

[0155] Step 4: Perform collision performance prediction: The collision simulation proxy prediction module performs dual-condition collision performance prediction on the generated structural scheme in Step 3. The output results are: front longitudinal beam energy absorption ratio 22.3kJ / kg, passenger compartment firewall intrusion 102mm, weight 3.28kg, unit cost 112 yuan, and all indicators meet the constraint requirements.

[0156] Step 5: Perform compliance verification: Complete full-dimensional verification through the compliance verification module, output compliance signals in the compliance feature vector, and achieve a 100% constraint satisfaction rate with no non-compliance items.

[0157] Step 6: Output the structural scheme: Output the 3D STEP format CAD model of the front longitudinal beam, performance prediction report, and compliance verification report. This scheme can be directly imported into the industry-leading analysis software LS-DYNA for simulation verification and mass production mold development.

[0158] The control method of the end-to-end safety design large model system provided in this application is described below. The control method of the end-to-end safety design large model system described below can be referred to in correspondence with the end-to-end safety design large model system of vehicle crashworthiness drive described above.

[0159] Figure 8 This is a schematic flowchart of the control method for an end-to-end safety design large model system provided in this application embodiment. It can be understood that it can be applied to any of the aforementioned end-to-end safety design large model systems driven by vehicle crashworthiness, such as... Figure 8 As shown, the method includes: S1, converts the input crashworthiness design requirements into a corresponding token sequence; S2, based on the token sequence corresponding to the crashworthiness design requirements, perform multimodal feature extraction and cross-modal fusion, and based on the fused features, perform iterative analysis under the preset crashworthiness constraints to obtain the target three-dimensional crashworthiness structure scheme; S3 outputs the target three-dimensional impact-resistant structural design.

[0160] It should be understood that the above method is applied to the system in the above embodiments. The implementation principle and technical effect of the corresponding steps in the method are similar to those of the relevant modules / units in the above system. The implementation process of the method can refer to the corresponding working process in the above system, which will not be repeated here.

[0161] Optionally, this application also provides a method for using an end-to-end safety design large model system driven by vehicle crashworthiness, the steps of which are as follows: Step 1: The input module receives the vehicle crashworthiness design requirements information in natural language form input by the user.

[0162] Step 2: The requirement parsing and constraint mapping module performs semantic understanding, entity extraction, and conflict detection on the natural language requirement, transforming it into a structured multi-dimensional engineering constraint set, thus obtaining the structured engineering constraint set.

[0163] Step 3: The multi-constraint embedded crash-resistant structure generation module generates an initial three-dimensional crash-resistant structure scheme based on the structured engineering constraint set.

[0164] Step 4: The physics-based collision simulation proxy prediction module performs physics field prediction for the initial three-dimensional crash-resistant structure scheme under full collision conditions, and outputs crashworthiness performance indicators and full simulation data.

[0165] Step 5: The compliance verification module evaluates and verifies the compliance of the initial three-dimensional impact-resistant structure scheme and outputs the corresponding compliance signal.

[0166] Step 6: The closed-loop self-iterative optimization module determines whether the initial three-dimensional impact-resistant structure scheme meets all constraints and qualification thresholds. If yes, proceed to step 7; otherwise, proceed to step 8.

[0167] Step 7: The output module outputs the initial three-dimensional impact-resistant structure design.

[0168] Step 8: The closed-loop self-iterative optimization module executes the reverse feedback link based on the evaluation results, returns to step 3, until the target three-dimensional crash-resistant structure scheme that meets the preset crashworthiness constraints is output.

[0169] The method in this application, by introducing cross-modal interaction between the large language model module and the multimodal feature extraction and analysis module, constructs an end-to-end safety design large model system with crashworthiness as the core driving objective. It realizes the end-to-end mapping of the entire link from design feature input to design scheme output that meets collision safety requirements. It can effectively replace the traditional manual-led multi-round, time-consuming simulation iteration and physical test links, compressing the single-round design iteration cycle from the traditional several weeks to the second level, greatly improving the efficiency of the entire vehicle crashworthiness design process and shortening the vehicle safety development cycle.

[0170] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0171] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0172] It is understood that the various numerical designations used in the embodiments of this application are merely for the convenience of description and are not intended to limit the scope of the embodiments of this application.

[0173] It should be understood that expressions such as “comprising” and “may include” used in this application indicate the existence of the disclosed functions, operations, or constituent elements, and do not limit one or more additional functions, operations, and constituent elements. In this application, terms such as “comprising” and / or “having” are to be interpreted as indicating a particular characteristic, number, operation, constituent element, component, or combination thereof, but not to exclude the existence or possibility of adding one or more other characteristics, numbers, operations, constituent elements, components, or combinations thereof.

[0174] Furthermore, in this application, the expression "and / or" includes any and all combinations of the associated listed words. For example, the expression "A and / or B" may include A, may include B, or may include both A and B.

[0175] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A large-scale model system for end-to-end safety design based on vehicle crashworthiness, characterized in that, include: A large language model module, and an input module, a multimodal feature extraction and analysis module, and an output module respectively connected to the large language model module; The input module is used to input crashworthiness design requirements information; The large language model module and the multimodal feature extraction and analysis module adopt a cross-modal interaction method to extract multimodal features and perform cross-modal fusion using the word sequence corresponding to the crashworthiness design requirement information, and perform iterative analysis based on the fused features under preset crashworthiness constraints to obtain the target three-dimensional crashworthiness structure scheme. The output module is used to output the target three-dimensional impact-resistant structure scheme; The multimodal feature extraction and analysis module includes a multimodal feature synthesis module and a closed-loop self-iterative optimization module; The multimodal feature synthesis module includes a requirement parsing and constraint mapping module, a crash-resistant structure generation module, a collision simulation proxy prediction module, and a compliance verification module, all of which are connected to the large language model module. The large language model module is used to extract fusion features from the word sequence corresponding to the crashworthiness design requirement information and output a semantic fusion feature vector. The requirement parsing and constraint mapping module is used to transform the semantic fusion feature vector into a structured engineering constraint set, so that the large language model module can output a constraint semantic fusion feature vector based on the structured engineering constraint set and the semantic fusion feature vector. The crash-resistant structure generation module is used to generate an initial three-dimensional crash-resistant structure scheme and a corresponding first structural geometric feature based on the constraint semantic fusion feature vector, so that the large language model module can output a constraint geometric fusion feature vector based on the structured engineering constraint set and the first structural geometric feature. The collision simulation proxy prediction module is used to output the performance feature vector corresponding to the initial three-dimensional crash-resistant structure scheme based on the constraint geometry fusion feature vector, so that the large language model module can output the constraint geometry performance fusion feature vector according to the performance feature vector and the constraint geometry fusion feature vector; The compliance verification module is used to output the compliance feature vector corresponding to the initial three-dimensional crash-resistant structure scheme based on the constraint geometric performance fusion feature vector, so that the large language model module can output the compliance fusion feature vector based on the compliance feature vector; The impact-resistant structure generation module includes a constraint feature adapter, a macroscopic topology generation network, a cross-sectional detail generation network, a process feature generation network, and a first performance prediction head; the constraint feature adapter and the first performance prediction head are also connected to the large language model module. The constraint feature adapter is used to transform the constraint semantic fusion feature vector into a corresponding second embedding vector. The macro-topology generation network is used to output a voxelized macro-topology structure vector based on the second embedding vector; The cross-sectional detail generation network is used to generate corresponding point cloud data based on the second embedding vector and the voxelized macroscopic topology vector. The process feature generation network is used to output the original three-dimensional impact-resistant structure scheme and the corresponding second structural geometric features based on the point cloud data and the second embedding vector. The first performance prediction head is used to determine the performance deviation value corresponding to the original three-dimensional crash-resistant structure scheme based on the second structural geometric features during the training phase, so as to adjust the network parameters of the process feature generation network; during the inference phase, it generates an initial three-dimensional crash-resistant structure scheme with converged performance deviation value and its corresponding first structural geometric features, and converts the first structural geometric features into corresponding word sequences and feeds them back to the large language model module.

2. The end-to-end security design large model system according to claim 1, characterized in that, The multimodal feature synthesis module and the closed-loop self-iterative optimization module are respectively connected to the large language model module; The multimodal feature synthesis module is used to generate an initial three-dimensional crash-resistant structure scheme based on the semantic fusion feature vector, and jointly with the large language model module to perform performance compliance verification on the initial three-dimensional crash-resistant structure scheme, and output the compliance fusion feature vector corresponding to the initial three-dimensional crash-resistant structure scheme. The closed-loop self-iterative optimization module is used to determine the target structured engineering constraint set based on the compliance fusion feature vector, so that the large language model module can control the multimodal feature synthesis module to adjust the initial three-dimensional crash-resistant structure scheme according to the target structured engineering constraint set, until the target three-dimensional crash-resistant structure scheme that meets the preset crash-resistant constraint conditions is output.

3. The end-to-end security design large model system according to claim 1, characterized in that, The closed-loop self-iterative optimization module includes a compliance signal receiving unit, an iterative judgment unit, and an optimization generation link connected in sequence; the compliance signal receiving unit and the optimization generation link are respectively connected to the large language model module. The compliance signal receiving unit is used to determine the non-compliance information of the initial three-dimensional impact-resistant structure scheme based on the compliance fusion feature vector; The iteration judgment unit is used to send the non-compliance information to the generation optimization link when it is determined that the current iteration does not meet the preset iteration termination condition; The generated optimization link is used to adjust and reconstruct the structured engineering constraint set according to the non-compliance information, determine the target structured engineering constraint set, and convert the target structured engineering constraint set into the corresponding lexical sequence and output it to the large language model module.

4. The end-to-end security design large model system according to claim 1, characterized in that, The requirement parsing and constraint mapping module includes a feature encoding interface, a vehicle crashworthiness vertical language model, a constraint mapping structured layer, and a constraint feedback interface connected in sequence; the feature encoding interface and the constraint feedback interface are also connected to the large language model module. The feature encoding interface is used to convert the semantic fusion feature vector into a corresponding first embedding vector; The vehicle crashworthiness vertical language model is used to output intermediate text results based on the first embedding vector; The constraint mapping structured layer is used to standardize terminology based on the intermediate text results and output a standardized set of structured engineering constraints. The constraint feedback interface is used to convert the structured engineering constraint set into a corresponding lexical sequence and feed it back to the large language model module.

5. The end-to-end security design large model system according to claim 1, characterized in that, The collision simulation proxy prediction module includes a geometric performance feature adapter head, a physical information neural fusion network, a physical residual calculation unit, and a second performance prediction head; the geometric performance feature adapter head and the second performance prediction head are respectively connected to the large language model module; The geometric performance feature adapter is used to transform the constrained geometric fusion feature vector into a third embedding vector. The physical information neural fusion network is used to output preliminary collision physical field prediction results based on the third embedding vector; The physical residual calculation unit is used to determine the corresponding physical residual value based on the preliminary collision physical field prediction result, so as to optimize the network parameters of the physical information neural fusion network based on the physical residual value, so that the physical information neural fusion network outputs the full collision physical field prediction result. The second performance prediction head is used to extract the performance feature vector corresponding to the initial three-dimensional crash-resistant structure scheme from the full collision physics field prediction results, and feed back the word sequence corresponding to the performance feature vector to the large language model module.

6. The end-to-end security design large model system according to claim 1, characterized in that, The compliance verification module includes a multimodal feature interface port, a multi-scale hierarchical neural network, a multi-task compliance decoding layer, a compliance report output port, and a signal output port; the multimodal feature interface port and the signal output port are respectively connected to the large language model module; The multimodal feature docking port is used to transform the constrained geometric performance fusion feature vector into a fourth embedding vector; The multi-scale hierarchical neural network is used to perform multi-scale feature encoding and fusion based on the fourth embedding vector, and output a full-dimensional fused feature vector. The multi-task compliance decoding layer is used to perform compliance verification analysis on the full-dimensional fused feature vector in different dimensions and output multi-dimensional compliance verification results. The compliance report output port is used to generate a compliance verification report based on the multi-dimensional compliance verification results. The signal output port is used to extract the compliance feature vector based on the multi-dimensional compliance verification result, and output the word sequence corresponding to the compliance feature vector to the large language model module.

7. The end-to-end security design large model system according to any one of claims 1-6, characterized in that, The large language model module includes an encoder group module, a cross-modal fusion module, an attention layer, and a linear projection layer connected in sequence. The encoder group module is used to encode any input lexical sequence and output the single-modal initial feature vector corresponding to the lexical sequence. The cross-modal fusion module is used to perform cross-modal feature fusion on the single-modal initial feature vector and the original demand semantic features by calling the expert unit, and output the preliminary fusion feature vector corresponding to any word sequence; The attention layer is used to constrain and enhance the preliminary fused feature vector through a preset attention mechanism, and output the enhanced multimodal feature sequence. The linear projection layer is used to output a target fusion feature vector based on the multimodal feature sequence; The lexical sequence includes the lexical sequence corresponding to the crashworthiness design requirements information, the lexical sequence corresponding to the structured engineering constraint set, the lexical sequence corresponding to the first structural geometric features, the lexical sequence corresponding to the performance feature vector, and the lexical sequence corresponding to the compliance feature vector. The target fusion feature vector includes the semantic fusion feature vector, the constraint semantic fusion feature vector, the constraint geometric fusion feature vector, the constraint geometric performance fusion feature vector, and the compliance fusion feature vector.

8. A control method applied to an end-to-end security design large model system as described in any one of claims 1-7, characterized in that, include: The input crashworthiness design requirements are converted into corresponding word sequences. Multimodal feature extraction and cross-modal fusion are performed based on the word sequence corresponding to the crashworthiness design requirements information. Based on the fused features, iterative analysis is performed under preset crashworthiness constraints to obtain the target three-dimensional crashworthiness structure scheme. Output the target three-dimensional impact-resistant structure scheme.

Citation Information

Patent Citations

  • Vehicle collision simulation optimization method based on machine learning

    CN114880934A

  • Automobile body innovative design system based on multi-modal knowledge

    CN121166938A

  • Automobile wire harness design intention analysis and CAD graph generation system based on large model

    CN122020856A