A method and system for lightweight optimization design of a vehicle frame considering fatigue strength constraints

Abstract

The application discloses a kind of lightweight optimization design method and system of frame considering fatigue strong constraint, by modeling the frame design task as constrained Markov decision process, using Hybrid-PPO algorithm to construct the hybrid strategy agent capable of simultaneously processing discrete selection and continuous size variable, and calling high-precision multi-performance prediction model to realize sub-second level fast evaluation of scheme performance;In optimization iteration, introduce hierarchical constraint processing mechanism, through bottom layer action shielding, middle layer safety layer projection and upper layer dynamic penalty, forced to maintain manufacturing process limit, space interference and fatigue life and other hard constraints, so as to generate non-inferior solution set covering Pareto front under the premise of guaranteeing engineering feasibility;Finally, based on combination weighting and TOPSIS method to establish decision-making closed loop to select the final scheme.The application realizes the autonomous intelligent design of frame structure, significantly improves the optimization efficiency and lightweight level of frame design.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of vehicle frame structure design and artificial intelligence optimization decision-making, and particularly relates to a lightweight optimization design method and system for vehicle frames that considers strong fatigue constraints. Background Technology

[0002] With the development trends of intelligentization, electrification, and decarbonization in commercial vehicles, the chassis, as the fundamental load-bearing component of the entire vehicle, faces the dual technical challenges of increased battery weight and range anxiety. This inherent conflict between structural load-bearing requirements and overall vehicle energy efficiency goals necessitates that chassis design achieve extreme lightweighting while ensuring high reliability. Currently, with the continuously shortening product update cycle in commercial vehicles, more stringent engineering requirements are being placed on the precision and efficiency of chassis structural design. Efficient and intelligent optimization decision-making methods and automated design systems are the core support for achieving forward design for structural lightweighting.

[0003] In the current structural design and optimization decision-making process, existing technical solutions still face the following bottlenecks:

[0004] First, traditional structural optimization strategies struggle to balance optimization efficiency, convergence accuracy, and global search capability when dealing with high-dimensional, highly nonlinear chassis optimization problems. Currently, the industry primarily employs direct search strategies or surrogate model-based optimization methods. Direct search methods (such as CN202511490809.1 and CN202511148781.3) require frequent calls to high-fidelity CAE models for nonlinear fatigue analysis, resulting in excessively long optimization cycles and poor engineering practicality. While surrogate model-based methods (such as CN202410887126.9) reduce computational load through mathematical fitting, they are limited by the traditional surrogate model's ability to represent highly nonlinear response surfaces, easily leading to prediction biases in local extrema regions. More importantly, when facing highly non-convex design spaces involving discrete selection and continuous dimension coupling, these methods are prone to getting trapped in local optima, making it difficult to uncover the ultimate lightweight potential of the structure within complex parameter combinations.

[0005] Secondly, existing deep reinforcement learning methods lack effective maintenance mechanisms for strong constraints when applied to the design of complex engineering assemblies. While some technologies have attempted to apply reinforcement learning to the design optimization of biomechanical structures such as implantable devices (e.g., CN202511510033.5), automating the process by embedding deep learning models into reinforcement learning frameworks, these methods primarily focus on the geometric adaptation of structures and have not yet established hard-processing logic for commercial vehicle frames, which involve strong constraints across multiple disciplines. Frame design involves a hybrid action space deeply coupled with discrete variables (material / bolt selection, etc.) and continuous variables (plate thickness / section dimensions, etc.), making it highly susceptible to generating illegal solutions that violate manufacturing process limits (e.g., bolt edge distance exceeding limits) or geometric interference during the optimization process. The lack of a hierarchical control mechanism for the engineering feasibility of complex assemblies leads to frequent idle cycles within invalid spaces, making it difficult for the optimization process to converge to a feasible solution that meets actual engineering requirements.

[0006] Finally, existing optimization processes lack a closed-loop system that coordinates subjective and objective factors in multi-objective decision-making. Currently, after obtaining a Pareto non-dominated solution set covering the design space through optimization algorithms, the final selection of the solution often relies excessively on human experience or a single performance weight allocation. Such methods (e.g., CN202210419640.0) are either too susceptible to subjective biases or easily affected by random data fluctuations, making it difficult to achieve a scientific balance between lightweight benefits and performance risks. The lack of a comprehensive evaluation mechanism that deeply integrates expert engineering intuition and statistical data characteristics leads to a decision-making gap between algorithm optimization and engineering implementation, which not only limits design efficiency but also affects the robustness and reliability of the final optimal solution. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a lightweight chassis optimization design method and system that considers strong fatigue constraints. This method can efficiently handle high-dimensional hybrid design spaces, strictly maintain strong performance constraints, and possess a scientific decision-making closed loop.

[0008] Specifically, the technical solution provided by this invention is as follows:

[0009] A lightweight optimization design method for vehicle frames considering strong fatigue constraints includes the following steps:

[0010] S1. Model the frame structure optimization task as a constrained Markov decision process, define a composite state space including geometric configuration features, design parameter attributes, load condition features and performance response, construct a hybrid action space that couples discrete and continuous variables, and establish a guided reward function with lightweight benefits as positive feedback and engineering constraint violation as negative feedback.

[0011] S2. Construct a hybrid policy agent based on a parameterized action space network. The composite state space is used as input. The probability distribution of discrete selection actions and the distribution parameters of continuous adjustment actions are output synchronously using the hybrid action space. The proximal policy optimization algorithm is used with a pre-trained high-precision multi-performance prediction model as the interactive environment to perform policy iterative training on the hybrid policy agent.

[0012] S3. During the strategy iteration process, the bottom-level action masking for discrete selection decision, the middle-level safety layer projection for continuous variable adjustment, and the upper-level Lagrange dynamic penalty for performance index constraints are executed in sequence to constrain the actions output by the agent to the engineering feasible domain and generate a Pareto non-dominated solution set that meets the engineering feasibility requirements.

[0013] S4. Based on the combination weighting and approximation ideal solution sorting method, perform multi-criteria decision optimization on the Pareto non-dominated solution set, and output the optimal engineering lightweight solution.

[0014] Furthermore, the composite state space mentioned in step S1 includes:

[0015] Geometric configuration features are the three-dimensional binary density matrix obtained by mapping the chassis model;

[0016] The design parameter attribute vector contains the real-time updated plate thickness values, cross-sectional dimensions, connection parameters, and material grade indexes for each component;

[0017] Load condition feature matrix, extract the load distribution characteristics of the frame under bending, torsion, braking and steering conditions;

[0018] Performance response characteristics include frame mass, first-order modal frequency, bending stiffness, torsional stiffness, maximum stress, maximum deformation, and minimum fatigue life.

[0019] Furthermore, the hybrid action space mentioned in step S1 includes a discrete action subspace and a continuous action subspace. The discrete action subspace is used to make decisions on material selection, cross-sectional shape selection of beams and crossbeams, and connection parameter selection. The continuous action subspace is used to make decisions on plate thickness adjustment, cross-sectional size adjustment, beam layout coordinate adjustment, and connection spacing adjustment.

[0020] Furthermore, the guided reward function mentioned in step S1 is:

[0021]

[0022] in, As a reward value, The percentage reduction in frame weight. This indicates the degree of violation of the i-th performance metric. For the corresponding to the firsti The Lagrange multiplier of each performance indicator; when all performance indicators meet the constraints, the reward value is the percentage reduction in frame mass; when any performance indicator violates the constraints, a negative penalty is imposed according to the degree of violation.

[0023] Furthermore, the parameterized action space network described in step S2 adopts a layered architecture with a shared feature extraction layer and a dual-head output structure:

[0024] The shared feature extraction layer extracts the spatial topological relationships of geometric configuration features in the composite state space through a three-dimensional residual convolutional network, and fuses them with design parameter attribute vectors and performance response features to output high-dimensional semantic features. The dual-head output structure includes a discrete output head and a continuous output head. The discrete output head outputs the probability distribution of discrete actions based on the high-dimensional semantic features. The continuous output head outputs the mean and logarithmic standard deviation of continuous actions based on the high-dimensional semantic features, and generates continuous action values ​​through reparameterization techniques. a : Where μ and logσ are the mean and log standard deviation of the output from the continuous output head, respectively. This is sampling noise.

[0025] Furthermore, in step S3,

[0026] The underlying action masking is as follows: a mask vector is constructed based on the current geometric topology of the chassis and a preset process rule library; a mask operation is performed on the probability distribution of discrete actions; and the selection probability of illegal actions is set to zero.

[0027] The projection of the middle security layer is as follows: when the continuous action exceeds the preset feasible region, the continuous action is projected to the nearest point on the boundary of the feasible region by solving a quadratic programming problem, and the projection process is embedded in the neural network to realize gradient backpropagation.

[0028] The upper-level Lagrange dynamic penalty is as follows: the Lagrange multipliers corresponding to each performance index in the reward function are dynamically updated using the dual gradient descent method; when a certain performance index fails to meet the standard, its corresponding Lagrange multiplier value is automatically increased, forcing the agent to abandon excessive weight reduction and prioritize meeting the reliability requirements; conversely, when the constraint is met, the Lagrange multiplier decays to zero, allowing the agent to freely explore the lightweight boundary.

[0029] Furthermore, the process of generating the Pareto non-dominated solution set in step S3 is as follows:

[0030] An external archive is established during the strategy iteration process to store non-dominated solutions generated during the optimization process; each time a new effective solution is generated, the dominance relationship of the new solution is compared with the existing solutions in the external archive.

[0031] If the new scheme is dominated by any existing scheme in the external archive set, then the new scheme is rejected from being added to the external archive set.

[0032] If the new scheme dominates several existing schemes in the external archive set, then the new scheme is added to the external archive set, and the existing schemes dominated by it are removed.

[0033] If the new scheme is not mutually exclusive with any existing schemes in the external archive set, then the new scheme will be directly added to the external archive set.

[0034] After the iteration is completed, all non-dominant solution points retained in the external archive set are output as Pareto non-dominant solution set, which constitutes the optimal trade-off front of the chassis lightweight design space.

[0035] Further, step S4 includes:

[0036] From the Pareto non-dominated solution set, m candidate solutions are extracted, and n evaluation indicators are extracted for each solution to construct an original decision matrix. The original decision matrix is ​​then dimensionless using the range transformation method, converting all indicators into benefit-type indicators to obtain a standardized matrix. ;

[0037] The subjective weight vector is calculated by scoring the relative importance of each indicator based on the analytic hierarchy process (AHP) and after consistency testing. The information entropy and difference coefficient of each indicator in the standardized matrix are calculated based on the entropy weight method, and the objective weight vector is obtained after normalization. ;

[0038] , ,

[0039] in, For the first j The coefficient of difference of the indicators For the first j Information entropy of the evaluation indicators is the constant coefficient in the calculation of information entropy, and , For the first j The first item under the indicator i The weight of the indicator values ​​of each candidate solution, and have , For the standardized matrix elements, represent the first... i The candidate solution is in the... j Standardized values ​​for each evaluation indicator;

[0040] The comprehensive weight vector is calculated using a linear weighted model. α and β are preset preference coefficients;

[0041] The standardized matrix is ​​weighted using the comprehensive weight vector to construct a weighted normalized decision matrix. ,in ;

[0042] Define the positive ideal solution in the weighted normalized decision matrix. and negative ideal solution Calculate the Euclidean distance from each candidate solution to the positive ideal solution. Euclidean distance to the negative ideal solution According to the formula Calculate the relative schedule of each candidate scheme, and determine the scheme with the highest relative schedule as the optimal lightweight engineering scheme.

[0043] Furthermore, in step S2, the hybrid policy agent uses a pre-trained high-precision multi-performance prediction model as the interaction environment for policy iterative training.

[0044] The high-precision multi-performance prediction model is constructed as follows: Sample points are generated within the design space of the chassis using an experimental design method, forming a training sample set covering the range of design variable values; high-fidelity finite element simulation software is used to perform performance simulation calculations on each sample point in the training sample set, obtaining the performance response data corresponding to each sample point. The performance response data includes at least chassis mass, fatigue life, first-order modal frequency, bending stiffness, torsional stiffness, maximum stress, and maximum deformation; the sample points in the training sample set are used as input features, and the corresponding performance response data are used as output labels to construct a training dataset; a machine learning algorithm is selected to fit the training dataset, establishing a nonlinear mapping relationship between design parameters and performance response; and the model accuracy is verified through cross-validation to obtain the high-precision multi-performance prediction model.

[0045] During the policy iteration training process, the hybrid policy agent inputs the action vector output by the current decision step into the high-precision multi-performance prediction model. The high-precision multi-performance prediction model returns the corresponding performance evaluation results in real time, which serve as the basis for agent state updates and reward function calculations.

[0046] A lightweight chassis optimization design system based on the above method, the system includes the following modules:

[0047] The chassis design decision task modeling module is equipped with a state and action space construction unit, which is used to encapsulate the three-dimensional voxelized configuration of the chassis, design parameter attributes, load condition characteristics and current performance response into a composite state vector, and map discrete decisions and continuous decisions into a hybrid action space.

[0048] The hybrid policy agent construction and optimization module has a core component of a parameterized action space network, which integrates a shared feature perception layer, a discrete output head for outputting discrete action probability distributions, and a continuous output head for outputting continuously adjusted action distribution parameters, thereby achieving collaborative decision-making on hybrid design variables. The hybrid policy agent construction and optimization module is also equipped with a policy iteration optimization unit, which uses a near-end policy optimization algorithm combined with a trust region truncation mechanism to update network weights.

[0049] The engineering feasibility maintenance and non-dominated solution generation module is equipped with a hierarchical constraint processing engine, which contains three cascaded constraint operators: a bottom-level action masking operator to filter illegal discrete decisions that violate manufacturing process limits; a middle-level safety layer projection operator to correct continuous actions that violate spatial interference constraints to the feasible region boundary; and an upper-level dynamic penalty correction unit to handle strongly nonlinear performance constraints based on the Lagrange multiplier method. The module also integrates a performance evaluation proxy interface, which calls a pre-trained high-precision multi-performance prediction model to evaluate the performance of the corrected design scheme and iteratively generates a multi-objective non-dominated solution set covering the Pareto front of the design space.

[0050] The multi-criteria comprehensive decision-making optimization module integrates a subjective and objective combination weighting component, and is configured with a hierarchical analysis operator and an information entropy calculation operator, which are used to quantify subjective preferences and mine objective statistical characteristics of solution set data to generate a comprehensive weight vector; and a comprehensive evaluation unit for approximating the ideal solution ranking method, which is used to construct a weighted normalized decision matrix, calculate the relative distance between each candidate scheme and the ideal solution and the negative ideal solution, and output the optimal lightweight engineering scheme with the highest relative progress.

[0051] Compared with the prior art, the present invention has at least the following beneficial effects:

[0052] First, this invention achieves efficient autonomous optimization in a high-dimensional hybrid space covering multiple design variables such as frame structure topology, material grade, cross-sectional configuration, and geometric dimensions by establishing a constrained Markov decision process model and the Hybrid-PPO algorithm. Compared with traditional heuristic optimization algorithms or direct search optimization methods based on surrogate models, this invention utilizes a hybrid policy agent to achieve parallel collaborative decision-making of multi-dimensional design variables, significantly improving global search efficiency. It can deeply explore the lightweight potential of the frame from the perspective of interdisciplinary coupling and effectively solve the technical problem of easily getting trapped in local optima under complex parameter combinations.

[0053] Secondly, this invention establishes a hierarchical, strongly constrained processing mechanism to ensure the engineering feasibility and overall performance compliance of the optimized output solution. Through gradient-based control of bottom-level action shielding, mid-level safety layer projection, and upper-level dynamic penalties, this invention effectively filters out illegal actions that violate manufacturing processes, spatial interference, and overall performance indicators during the agent exploration phase. Compared to the "engineering distortion" problem often encountered in structural design in existing deep reinforcement learning, this invention forcibly maintains the physical boundaries and performance baselines of the chassis as a complex assembly, ensuring that the generated Pareto non-dominated solution set is manufacturable.

[0054] Furthermore, this invention introduces a pre-trained high-precision multi-performance prediction model as the performance evaluation environment, achieving high-fidelity performance evaluation feedback in the optimization loop. Combined with the trust region truncation mechanism of the PPO algorithm, this system effectively suppresses the impact of local fluctuations in the evaluation environment on decision-making, achieving a closed loop of efficient exploration and robust optimization. Compared to the traditional optimization mode of directly calling the finite element solver, this invention, while ensuring performance evaluation accuracy, significantly shortens the convergence cycle and design iteration time of lightweight solutions, providing an efficient tool for the rapid forward development of commercial vehicle chassis.

[0055] Finally, this invention significantly improves the scientific rigor and robustness of lightweight solution selection by introducing a combination of subjective and objective weighting and a TOPSIS comprehensive decision-making mechanism. This mechanism quantifies expert experience using the AHP method and mines the statistical characteristics of Pareto solution data using the entropy weighting method, achieving a deep trade-off between lightweight level, manufacturing cost, and all performance criteria, including fatigue life margin, static strength safety margin, and low-order modal frequencies. Compared to single-weighting selection methods that rely solely on subjective experience or are based solely on data distribution, this invention eliminates preference biases under a single decision-making logic, ensuring that the selected optimal solution possesses greater overall advantages during engineering implementation. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0057] Figure 1 This is a flowchart illustrating the overall process framework of the lightweight chassis optimization design method provided in this embodiment of the invention.

[0058] Figure 2 This is a schematic diagram of the initial commercial vehicle chassis model to be optimized, provided in an embodiment of the present invention.

[0059] Figure 3 This is a schematic diagram of the logical flow of the hierarchical strong constraint processing mechanism provided in the embodiments of the present invention;

[0060] Figure 4 This is a schematic diagram of the final solution optimization process provided by the embodiments of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort are all within the scope of protection of the present invention.

[0062] Example 1

[0063] like Figure 1 As shown, this embodiment provides a lightweight chassis optimization design method considering strong fatigue constraints. To achieve extreme lightweighting while meeting full-life-cycle fatigue reliability and various performance constraints, this embodiment models the chassis structural optimization task as a constrained Markov decision process (CMDP). An agent autonomously fine-tunes the chassis's topological variables, material properties, and cross-sectional dimensions at each decision step, and calls a pre-trained high-precision multi-performance prediction model to perform sub-second performance evaluation of the generated solutions. Through tens of thousands of policy interaction iterations between the agent and the evaluation environment, a lightweight design scheme that meets all engineering feasibility requirements is finally selected from the generated solution set.

[0064] like Figure 2 As shown, this embodiment uses a commercial vehicle side-beam steel frame as the target for lightweight optimization. The frame initially consists of 2 longitudinal beams and 7 transverse beams. The longitudinal beams adopt a channel-shaped cross-section structure, and multiple weight-reduction holes are distributed in the web. For the transverse beams, the first 4 transverse beams adopt a cylindrical tubular structure, while the last 3 transverse beams adopt an I-beam, box-shaped, or multi-cavity irregular cross-section structure. The intersections of the transverse beams and longitudinal beams are connected by bolt groups of preset specifications.

[0065] Specifically, the execution steps of this embodiment are detailed below:

[0066] I. Constrained Markov Decision Process Modeling for Chassis Design Decisions

[0067] This step aims to transform the complex chassis engineering design problem into a mathematical model that a reinforcement learning agent can understand and interact with. In this embodiment, the time step of the decision-making process is defined as... The specific modeling process is as follows:

[0068] 1. Construct a composite state space that includes a three-dimensional voxelized configuration and preset engineering semantic information.

[0069] intelligent agents in The state vector observed at each time step It is composed of features from the following four dimensions:

[0070] (1) Geometric configuration features: Map the current chassis model to a resolution of The three-dimensional binary density matrix is ​​used to characterize the spatial topological occupancy of the two longitudinal beams and seven transverse beams (including the first four circular tube beams and the last three functional transverse beams).

[0071] (2) Design parameter attribute vector: contains the material and size information of each component of the current frame. In the initial state of this embodiment, according to the original design scheme (see Table 1), the material of the left / right longitudinal beam and the reinforcing beam is set to ZQS500L, the material of the first 4 round tube beams is set to Q345B, the material of the body of the last 3 crossbeams is set to 20D2, and the material of the connecting plate is set to QSTE420. The state vector records the current plate thickness value (continuous variable), cross-sectional dimensions, connection parameters and material grade index (discrete variable) of the above components in real time. The cross-sectional dimensions refer to the shape and size parameters of the cross-section of the cross and longitudinal beams. The most typical example is the width of the upper and lower flanges and the height of the web of the channel longitudinal beam. The connection parameters include the selection (quantity, diameter) and spacing of the bolts or rivets and other connecting parts at each connection.

[0072] (3) Load condition feature matrix: Extract the load distribution characteristics of the frame under four typical working conditions of bending, torsion, braking and steering to form a damage feature matrix, which serves as the third dimension of the state input;

[0073] (4) Performance response characteristics: including the current frame mass, first-order modal frequency, bending stiffness, torsional stiffness, maximum stress, maximum deformation, and minimum life value in the full-field fatigue life cloud map. In this embodiment, the initial state... The performance benchmarks are set as follows: frame mass 655 kg, first-order modal frequency 7.28 Hz, and bending stiffness 5.2 × 10⁻⁶. 3 N / mm, torsional stiffness 2×10 6 N·mm / °, minimum fatigue life 2.794×10 6 The loop continues.

[0074] Table 1. Original frame materials and mechanical properties

[0075]

[0076] 2. Construct a hybrid action space that couples discrete and continuous variables.

[0077] An action vector is established for the agent's output at each decision step. This action vector is formed by concatenating a discrete action subspace and a continuous action subspace. In this embodiment, the discrete action subspace is used for decision-making variables, specifically including:

[0078] Material selection: Switch between ZQS500L, 6061-T6, and 7075-T6 materials for each component;

[0079] Selection of cross-sectional shape for horizontal and vertical beams: switch between circular tube, channel, I-beam, box and multi-cavity cross-sections; Selection of connection parameters: select the bolt diameter at each connection of horizontal and vertical beams between M8, M10, M12 and M16, and adjust the number of bolts between 2, 4 and 6.

[0080] The continuous action subspace is used to determine numerical adjustments, specifically including:

[0081] Plate thickness adjustment: The plate thickness of each longitudinal / transverse beam is continuously fine-tuned within the range of 4mm to 14mm; Cross-sectional dimension adjustment: Continuous fine-tuning is performed within the preset minimum and maximum dimension range;

[0082] Crossbeam layout adjustment: The layout coordinates of the crossbeam along the longitudinal direction of the frame are offset within ±150mm.

[0083] Connection spacing adjustment: The bolt spacing at the connection of each horizontal and vertical beam is finely corrected.

[0084] 3. Design a guided reward function based on lightweight benefits and engineering constraint penalties.

[0085] Construct a guided reward function that is primarily driven by the percentage reduction in chassis mass and provides negative feedback on the degree of performance constraint violation. If the fatigue life under the current state is less than 10... 6 If, during the second cycle, the first-order frequency exceeds 10 Hz, or the maximum stress under any operating condition exceeds the yield strength of the material at the corresponding location, or the maximum deformation under any operating condition exceeds 12 mm, it is considered a violation of engineering constraints, and the reward function outputs a large negative weight penalty value R. If all performance parameters meet the constraints, the reward value is determined by the decrease in frame mass. Through this weighted feedback mechanism, the agent is guided to find the optimal solution in a complex hybrid decision space that satisfies both engineering feasibility and achieves the highest level of lightweighting.

[0086] II. Construction and Training of Hybrid-PPO Algorithm-Based Hybrid Policy Agent

[0087] This step aims to construct an intelligent decision-making "brain" capable of simultaneously handling discrete selection and continuous size adjustment, namely a hybrid policy agent. In this embodiment, a parametric action space network (PAS-Net) combined with a hybrid action proximal policy optimization (Hybrid-PPO) algorithm is used to achieve this goal. The specific execution logic is as follows:

[0088] 1. Construct a hybrid policy network (PAS-Net) based on a parameterized action space.

[0089] The core design idea of ​​this network is that the optimization design of the chassis involves two fundamentally different types of decision variables: discrete variables (such as material selection and cross-sectional shape) and continuous variables (such as plate thickness and layout coordinates). Traditional reinforcement learning networks can only handle discrete actions (such as DQN) or only continuous actions (such as DDPG), making it difficult to achieve collaborative decision-making for both types of variables within the same framework. Therefore, this embodiment designs an end-to-end hybrid policy network architecture that can simultaneously output the probability distributions of discrete actions and continuous actions, and optimizes both within the same policy gradient update framework.

[0090] (1) Overall network architecture

[0091] The hybrid strategy network constructed in this embodiment adopts a hierarchical architecture of "shared feature extraction - dual-head output". The network input receives the composite state vector defined in the preceding steps. The state vector contains three-dimensional voxelized geometric configuration features, design parameter attribute vectors, and performance response features. The network first extracts and fuses features from the multi-source heterogeneous input states through a shared bottom layer network, extracting high-dimensional semantic features that can characterize the current design state of the chassis. Subsequently, the network branches into two parallel output heads above the shared feature layer, namely, a discrete output head and a continuous output head, which are responsible for outputting the selection probability of discrete actions and the distribution parameters of continuous actions, respectively.

[0092] This architecture design with a shared underlying layer and branched outputs has the following technical advantages: On the one hand, the shared underlying network enables discrete and continuous decisions to reason based on the same state representation, ensuring the synergy between material selection and size adjustment; on the other hand, the branched output heads ensure that the decision logic of the two types of actions maintains a certain degree of independence, avoiding training instability caused by network parameter coupling.

[0093] (2) Design of shared feature extraction layer

[0094] The shared feature extraction layer employs a cascaded 3D residual convolutional network as its backbone structure, specifically designed to process the 3D voxelized geometric features in the input state. The reason for using 3D convolution instead of 2D convolution is that the chassis geometry is essentially a 3D spatial structure. The 3D convolution kernel can simultaneously capture the spatial topological relationships between the crossbeams and longitudinal beams in the X (longitudinal), Y (lateral), and Z (vertical) directions, including the connection positions of the crossbeams and longitudinal beams, the spatial extension of the cross-sectional shape, and the distribution pattern of the weight-reducing holes.

[0095] The specific network configuration is as follows:

[0096] The first layer is a 3D convolutional layer with 1 input channel (corresponding to the binary density matrix) and 16 output channels. The kernel size is 3×3×3, the stride is 2, and the padding is set to "same" to maintain the spatial size of the feature map. This layer is responsible for extracting local geometric features, such as the boundary between the web and flange of the longitudinal beam, and the overlapping area between the end of the crossbeam and the longitudinal beam.

[0097] Four 3D residual blocks are then stacked, each containing two 3D convolutional layers, with skip connections introduced to alleviate the vanishing gradient problem in deep networks. The number of output channels for each residual block is 16, 32, 64, and 64, respectively. Each convolutional layer employs batch normalization and the ReLU activation function. A convolutional layer with a stride of 2 doubles the number of channels while halving the spatial size of the feature map, thereby expanding the receptive field while reducing computational cost, enabling the network to gradually perceive a wider range of geometric structural information. After the above convolutional layer processing, the spatial size of the 3D voxel feature map is compressed to 8×8×8, with 64 channels.

[0098] Following the 3D convolutional layer, the network connects to a flattening layer, flattening the 3D feature map into a 1D vector with dimensions of 8×8×8×64 = 32768. This vector then passes through two fully connected layers, each containing 512 neurons, both using the ReLU activation function and incorporating a Dropout mechanism (dropout rate set to 0.2) to prevent overfitting. The purpose of these two fully connected layers is to deeply fuse the local geometric features extracted by the convolutional layers with the design parameter attribute vector and performance response features from the input state, forming a high-dimensional semantic representation of the overall chassis design state.

[0099] It should be noted that the parameter configuration of the above-mentioned three-dimensional convolutional network (kernel size, stride, number of channels, number of residual blocks) was determined through experimental optimization based on the characteristics (size and geometric complexity) of the vehicle frame model in this embodiment. For different sizes or types of vehicle frame structures, those skilled in the art can make appropriate adjustments to the above parameters according to actual needs, while keeping the network architecture principle unchanged. For example, they can increase the input voxel resolution, increase or decrease the convolutional layer depth, adjust the number of channels, etc., and these adjustments will not depart from the protection scope of this invention.

[0100] (3) Discrete output head design

[0101] The discrete output head is responsible for outputting the selection probability of each action in the discrete action subspace. The discrete output head takes the high-dimensional feature vector output by the shared feature extraction layer as input, maps the feature vector to a logical vector through an independent single-layer fully connected network (the number of neurons equals the total number of discrete actions), and then converts it into a probability distribution through the Softmax function.

[0102] As mentioned above, in this embodiment, the discrete action space includes the following three types of decisions:

[0103] Material selection: This involves the selection of material grades for components such as longitudinal beams, transverse beams, and connecting plates. Taking the longitudinal beam as an example, the available discrete actions include keeping ZQS500L high-strength steel, switching to 6061-T6 aluminum alloy, and switching to 7075-T6 aluminum alloy, for a total of three options.

[0104] Crossbeam cross-section shape selection: For the 7 crossbeams, each crossbeam can be selected from 5 cross-section shapes: circular tube, channel, I-beam, box, and multi-cavity. Considering the different stress characteristics and spatial layout of the different crossbeams, the cross-section selection of each crossbeam is output as an independent discrete action.

[0105] Connection parameter selection: The combination of bolt specifications (M8, M10, M12, M16, a total of 4 types) and bolt quantity (2, 4, 6, a total of 3 types) at the connection of each horizontal and vertical beam.

[0106] The total number of logical output channels for the aforementioned discrete actions is equal to the sum of the number of all discrete action options. For each discrete action dimension (e.g., longitudinal beam material selection), the network outputs a corresponding logical vector, which, after Softmax normalization, yields the probability distribution for that dimension. During the sampling phase, the agent performs random sampling based on this probability distribution to determine the specific discrete decision value.

[0107] (4) Continuous output head design

[0108] The continuous output head is responsible for outputting the adjustment amounts of each action in the continuous action subspace. Unlike the discrete output head, the values ​​of continuous actions are continuous numerical values, so they need to be modeled using a probability distribution to achieve a balance between exploration and exploitation. This embodiment uses a Gaussian distribution (normal distribution) to model each continuous action, that is, it is assumed that each continuous action follows a Gaussian distribution with a mean of μ and a standard deviation of σ.

[0109] The continuous output head takes the high-dimensional feature vector output from the shared feature extraction layer as input and outputs the mean μ and log standard deviation log σ for each continuous action in parallel through an independent single-layer fully connected network. The reason for outputting the log standard deviation instead of the standard deviation directly is to avoid negative values ​​of the standard deviation, and at the same time, the value of log σ is unrestricted, which is beneficial to the training stability of the neural network.

[0110] As mentioned earlier, in this embodiment, the continuous action space includes plate thickness adjustment, beam layout coordinate adjustment, and bolt spacing adjustment. For each of the above continuous action dimensions, the continuous output head outputs a pair of parameters: the mean μ and the logarithmic standard deviation log σ. During the sampling phase, the agent first samples a noise term from a standard normal distribution. Then, actual motion values ​​are generated using reparameterization techniques. a :

[0111]

[0112] The reparameterization technique makes the sampling process differentiable, thus allowing the gradient to propagate back from the action values ​​to the network parameters. This is a key technique for the effective training of policy gradient algorithms.

[0113] 2. Performing collaborative sampling decisions for discrete and continuous design variables.

[0114] During the interaction between the agent and the prediction model environment, action sampling is performed synchronously from the probability distribution generated by the policy network. For discrete actions, sampling is performed directly according to the probability distribution output by Softmax to determine the values ​​of discrete variables such as material grade, cross-sectional shape, and bolt specifications. For continuous variables, the policy network outputs standardized action values ​​ranging from [-1, 1]. a It is then transformed into a physical value using a linear mapping operator:

[0115]

[0116] Standardized actions a The mapping is to the plate thickness value of each longitudinal / transverse beam in the range of 4~14mm, or to the coordinate offset of the transverse beam in the range of ±150mm, or the bolt spacing value.

[0117] 3. Utilize the proximal policy optimization algorithm to perform robust policy updates.

[0118] The Hybrid-PPO algorithm is used to perform multiple rounds of weight iteration on the agent policy network. In this embodiment, a discount factor is set. Generalized dominance estimation coefficient The training batch size for a single policy update is set to 256. To improve optimization stability, a trust region truncation mechanism is introduced, with the truncation threshold set to... The value is set to 0.2 to forcibly limit the range of change between the old and new policies, preventing the agent from making decision jumps when dealing with complex surfaces containing strong nonlinear fatigue constraints. At the same time, an entropy regularization term (with a weight of 0.01) is introduced into the loss function to encourage the agent to explore fully in the early stages of training, ultimately guiding the policy network to converge to the globally optimal solution that meets all performance compliance requirements and has the highest level of lightweightness.

[0119] III. Hierarchical Strong Constraint Processing and Pareto Nondominated Solution Set Generation

[0120] This step aims to establish a rigorous engineering constraint control mechanism to ensure that when the agent explores in a broad design space, the resulting design solutions always meet the requirements of manufacturing process, assembly space, and physical performance.

[0121] like Figure 3 As shown, the specific execution logic and algorithm details are as follows:

[0122] 1. Execute low-level action masking for discrete selection decisions

[0123] Before generating the probability distribution of the discrete output head of the agent policy network, a binary mask vector is first constructed based on the current geometric topology of the chassis. Specifically, the system uses a built-in process rule library to predict each discrete action. The rule library covers manufacturing specifications such as bolt edge distance standards, minimum welding space, and stamping process limitations.

[0124] For example, when a narrower I-shaped cross section is selected for the crossbeam (such as a flange width of less than 60mm), for the option of M16 bolt specification at the connection between the crossbeam and the longitudinal beam, because it violates the minimum edge distance installation standard, the system sets the mask value corresponding to this action to 0 (or negative infinity); then, the original value output by the policy network is subjected to Hadamard product operation with the mask vector, forcing the selection probability of illegal actions to be reduced to zero, and the probability distribution of the remaining legal actions is renormalized.

[0125] This underlying shielding mechanism ensures that all sampled discrete selection schemes are absolutely compliant at the manufacturing process level.

[0126] 2. Perform the middle-level safety layer projection for continuous variable adjustment.

[0127] For continuous movements such as plate thickness, cross-sectional dimensions, and beam coordinates, a safety layer correction mechanism based on a differentiable projection operator is introduced. In this embodiment, the assembly space of the vehicle chassis is set as a hard constraint boundary, for example, the longitudinal movement of the beam must not interfere with the battery pack mounting point, and the plate thickness must not be less than the minimum rollable thickness of the material. The system calculates the continuous movement points output by the intelligent agent in real time. With feasible region boundary The Euclidean distance between them, when an action is detected. Out of the feasible region (e.g., beam position coordinates) When intruding into the rear axle space, the actions are solved by solving a quadratic programming problem (QP). Project onto the nearest point on the feasible region boundary ,Right now:

[0128]

[0129] This process, acting as a differential layer embedding within a neural network, not only corrects for physical interference but also allows gradient backpropagation to guide the policy network in learning boundary constraints. This projection correction ensures that the continuous design parameters of the final output always fall within a fabrication- and assembly-feasible geometric space.

[0130] 3. Implement upper-level Lagrange dynamic penalties based on performance index constraints.

[0131] To address strongly nonlinear performance constraints such as fatigue life, static strength, and stiffness, a dynamic penalty-reward reshaping mechanism based on the Lagrange multiplier method is adopted. The augmented reward function constructed in this embodiment is...

[0132]

[0133] in, For the first The degree of violation of a performance metric (e.g.: ), These are the corresponding Lagrange multipliers (penalty weights). During training, they are dynamically updated using dual gradient descent. When a certain performance metric (such as fatigue life) consistently fails to meet the standard, the system automatically increases its corresponding value. A higher value significantly increases the penalty, forcing the agent to abandon excessive weight reduction and prioritize reliability requirements; conversely, when the constraint is satisfied, The decay rate is zero, allowing agents to freely explore lightweight boundaries.

[0134] 4. Perform large-scale policy iteration and extract multi-objective Pareto non-dominated solution sets.

[0135] Under the strict control of the aforementioned three-layer constraint mechanism, the driving agent invokes a pre-trained high-precision multi-performance prediction model environment to perform, for example, 20,000 steps of interactive optimization. During the iteration process, an external archive set is established to record all generated effective solutions in real time. For each newly generated solution, it is checked whether it is dominated by existing solutions in the archive set in terms of target dimensions such as frame mass and fatigue life; if the solution reduces weight without significantly reducing lifespan (or is lighter in mass for the same lifespan), it is added to the archive set and inferior solutions dominated by it are removed. Finally, as... Figure 4 As shown in (a), the series of non-dominant solution points retained in the archive set constitute a Pareto non-dominated solution set covering the frontier of the design space, providing a rich candidate library containing different performance preferences for subsequent multi-criteria decision-making.

[0136] It should be noted that Pareto optimality is a classic theory in the field of multi-objective optimization, and its core concept is "dominance relationship." In this embodiment, frame optimization involves conflicting objectives: lightweighting (minimizing mass) and performance preservation (maximizing fatigue life, stiffness, etc.). Because there are trade-offs between these objectives, a single solution is usually not optimal for all objectives simultaneously. Therefore, the Pareto non-dominated solution set is defined as: in a set of solutions, no single solution can further improve any of the objectives without compromising at least one other objective. These solutions constitute the "optimal trade-off front" of multi-objective optimization problems, providing designers with a basis for flexibly selecting solutions based on engineering requirements.

[0137] The pre-trained high-precision multi-performance prediction model used in this step serves as the "environment" for agent interaction. It receives the action vectors output by the agent (i.e., the modified design scheme) and returns multiple performance indicators (mass, fatigue life, modal frequency, stiffness, stress, etc.) as evaluation feedback. This prediction model belongs to the mature surrogate model technology in the field of structural optimization, and its construction and application are well-known to those skilled in the art. Specifically, the standard construction process of the surrogate model includes: generating sample points in the design space using experimental design methods (such as Latin hypercube sampling); calling high-precision finite element simulation software (such as Abaqus, ANSYS, etc.) to calculate the performance response of each sample point to construct a training dataset; then selecting appropriate machine learning algorithms (such as Kriging model, radial basis function network, graph neural network, or three-dimensional convolutional neural network, etc.) to fit the dataset, establishing a nonlinear mapping relationship between design parameters and performance response; and finally verifying the model accuracy through methods such as cross-validation. The aforementioned technical solutions have been documented in numerous publicly available documents. For example, Chinese patent application CN202410887126.9 discloses a method for predicting vehicle frame performance using a Kriging proxy model, and CN202311485672.1 discloses a technical solution for predicting vehicle body structure performance using a graph neural network. Those skilled in the art, combined with existing technologies, are fully capable of constructing a performance prediction model that meets the accuracy requirements and using it as the interactive environment for the intelligent agent in this embodiment.

[0138] IV. Comprehensive Decision-Making Optimization Based on Combinatorial Weighting and TOPSIS

[0139] This step aims to establish a scientific evaluation system that integrates expert experience and statistical data patterns, and to select the unique engineering solution with the best overall performance from the Pareto non-dominated solution set generated in the previous steps.

[0140] like Figure 4 As shown, the specific calculation logic and execution steps are as follows:

[0141] 1. Construct a multi-criteria design scheme evaluation matrix

[0142] Extract from Pareto non-dominated solution set Candidate solutions (e.g., in this embodiment) ), and extract for each solution One evaluation index (in this embodiment) Construct the original decision matrix The evaluation indicators include: frame mass reduction (the larger the better), fatigue life safety margin (the larger the better), first-order modal frequency (the larger the better), manufacturing cost (the smaller the better), and static strength safety margin and static stiffness safety margin (both the larger the better). For indicators with different dimensions, the range transformation method is used to perform dimensionless processing, converting all indicators into "benefit-oriented" values ​​to obtain a standardized matrix. .

[0143] 2. Determining the subjective weight vector based on the AHP (Analytic Hierarchy Process)

[0144] Senior chassis design experts were organized to construct pairwise judgment matrices, and the relative importance of each indicator was scored using the Saaty scale (1-9 scale). In one application scenario of this embodiment, considering fatigue reliability as the primary objective, followed by lightweighting, and then cost, the subjective weight vector calculated after consistency verification is shown in the example below. These correspond to mass, lifespan, frequency, cost, strength, and stiffness, respectively.

[0145] 3. Determining the objective weight vector based on the entropy weight method

[0146] Mining standardized matrices using information entropy theory The degree of data variation. First, calculate the... Information entropy of the indicator:

[0147]

[0148] in

[0149]

[0150] Then, the coefficient of variation for each indicator was calculated:

[0151]

[0152] And normalize to obtain objective weights:

[0153]

[0154] This step ensures that indicators with high data dispersion (i.e., indicators that can better distinguish the merits of different solutions) receive greater weight, avoiding blind spots caused by human bias.

[0155] 4. Execution of subjective and objective combination weighting and construction of weighted decision matrix

[0156] To balance expert experience with data patterns, a linear weighted model is used to calculate the comprehensive weight vector:

[0157]

[0158] In this embodiment, a preference coefficient is set to highlight the data-driven decision-making characteristics. Using the calculated comprehensive weights For the standardized matrix Weighting is performed to construct a weighted normalized decision matrix. ,in .

[0159] 5. Comprehensive ranking and scheme determination based on TOPSIS

[0160] Define a weighted normalized decision matrix The positive ideal solution (Composed of the maximum values ​​of each column of indicators) and negative ideal solution (Constituted by the minimum value of each column of indicators). Calculate each candidate solution. Euclidean distance to the ideal solution and the Euclidean distance to the negative ideal solution According to the formula:

[0161]

[0162] Calculate the relative tracking schedule for each option. (Value range 0~1). Ultimately, based on... The numerical values ​​are sorted in descending order of all candidate solutions, and the solution with the highest progress tracking is selected as the final optimal lightweight engineering solution. For example... Figure 4 As shown in (d), the optimal solution achieves the best balance between quality and performance while satisfying all strong constraints.

[0163] In this embodiment, the chassis achieves a generational optimization from materials to structure in the selected optimal lightweight engineering scheme. In terms of parameter configuration, the entire chassis is optimized from the original steel to an aluminum alloy structure, with the longitudinal beam material changed from ZQS500L to 7075-T6, and the web thickness reduced from 8.0mm to 6.2mm; all 7 crossbeams are changed from steel to 6061-T6 material, and the intermediate functional crossbeams are uniformly optimized from the initial round tube or channel structure to a multi-cavity extruded cross-section structure, and the longitudinal layout coordinates are corrected by an average position offset of about 55mm based on the load characteristics.

[0164] In terms of final performance, compared to the initial state, this optimal solution achieved a comprehensive performance improvement while significantly reducing weight. The total frame mass was reduced from 655kg to 386.5kg, a weight reduction of 41%; dynamic characteristics were significantly enhanced, with the first-order modal frequency increasing from 7.28Hz to 10.42Hz; static stiffness was steadily improved, with torsional stiffness optimized to 2.25×10⁻⁶. 6 The maximum stress and maximum deformation are within the acceptable range. In terms of reliability, thanks to the optimization of local stress by the multi-cavity cross section, the minimum fatigue life of the whole field has been increased from 2.794×10^6 cycles to 3.58×10^6 cycles (equivalent to about 1.16 million kilometers), achieving the ultimate balance between quality and performance while meeting all engineering constraints.

[0165] Example 2

[0166] Based on the above method, this embodiment provides a lightweight chassis optimization design system considering strong fatigue constraints. This system is built on a computer hardware platform including a high-performance processor (CPU), a graphics acceleration unit (GPU), and large-capacity memory. The functional modules are connected and interact via a data bus or API interface to execute the method described in Embodiment 1. Specifically, the system includes:

[0167] Chassis Design Decision Task Modeling Module: This module is equipped with state and action space construction units, which are responsible for encapsulating the three-dimensional voxelized configuration of the chassis, design parameter attributes, load condition characteristics and current performance response into a composite state vector, and mapping discrete decisions such as material selection and cross-section switching with continuous decisions such as size fine-tuning and layout offset into a hybrid action space.

[0168] Hybrid Policy Agent Construction and Optimization Module: The core component of this module is the Parametric Action Space Network (PAS-Net), which integrates a shared feature perception layer, a Softmax output head for outputting discrete action probability distributions, and a linear output head for outputting continuously adjusted Gaussian distribution parameters of actions, enabling collaborative decision-making on hybrid design variables. In addition, this module is also equipped with a policy iteration optimization unit, which uses the proximal policy optimization algorithm (PPO) combined with a trust region truncation mechanism to update network weights, ensuring the agent's exploration stability and convergence robustness when facing high-dimensional non-convex design spaces.

[0169] The engineering feasibility maintenance and non-dominated solution generation module is equipped with a hierarchical constraint processing engine, which contains three cascaded constraint operators: a bottom-level action masking operator for real-time filtering of illegal discrete decisions that violate manufacturing process limits; a middle-level safety layer projection operator forcibly correcting continuous actions that violate spatial interference constraints to the feasible region boundary through quadratic programming; and an upper-level dynamic penalty correction unit for handling strongly nonlinear performance constraints such as fatigue life based on the Lagrange multiplier method. Furthermore, this module integrates a performance evaluation proxy interface, which calls a pre-trained high-precision multi-performance prediction model to perform sub-second performance evaluation of the corrected design scheme and generates a multi-objective non-dominated solution set covering the Pareto front of the design space through large-scale iterations.

[0170] Multi-criteria integrated decision optimization module: This module integrates a subjective and objective weighting component, and is equipped with an AHP hierarchical analysis operator and an information entropy calculation operator, which are used to quantify experts' subjective preferences for lightweighting and reliability and to mine the objective statistical characteristics of the solution set data to generate a comprehensive weight vector; and a TOPSIS comprehensive evaluation unit, which is used to construct a weighted normalized decision matrix, calculate the Euclidean distance between each Pareto candidate solution and the ideal solution and the negative ideal solution, and output the unique optimal lightweighting solution with the highest relative progress.

[0171] The above system can execute the lightweight frame optimization design method described in Embodiment 1, and has the corresponding functional modules and beneficial effects of the method. For technical details not described in detail in this embodiment, please refer to the lightweight frame optimization design method provided in Embodiment 1 of this invention.

[0172] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0173] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; under the concept of the present invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the present invention as described above, which are not provided in detail for the sake of brevity; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A lightweight chassis optimization design method considering strong fatigue constraints, characterized in that, Including the following steps: S1. Model the frame structure optimization task as a constrained Markov decision process, define a composite state space including geometric configuration features, design parameter attributes, load condition features and performance response, construct a hybrid action space that couples discrete and continuous variables, and establish a guided reward function with lightweight benefits as positive feedback and engineering constraint violation as negative feedback. S2. Construct a hybrid policy agent based on a parameterized action space network. The composite state space is used as input. The probability distribution of discrete selection actions and the distribution parameters of continuous adjustment actions are output synchronously using the hybrid action space. The proximal policy optimization algorithm is used with a pre-trained high-precision multi-performance prediction model as the interactive environment to perform policy iterative training on the hybrid policy agent. S3. During the strategy iteration process, the following steps are executed sequentially: the bottom-level action masking for discrete selection decision, the middle-level safety layer projection for continuous variable adjustment, and the upper-level Lagrange dynamic penalty for performance index constraints. The action output by the agent is constrained to the engineering feasible domain to generate a Pareto non-dominated solution set that meets the engineering feasibility requirements. The performance index constraints are strong nonlinear performance constraints, including fatigue life, static strength, and stiffness. S4. Based on the combination weighting and approximation ideal solution sorting method, perform multi-criteria decision optimization on the Pareto non-dominated solution set, and output the optimal engineering lightweight solution.

2. The lightweight frame optimization design method as described in claim 1, characterized in that, The composite state space mentioned in step S1 includes: Geometric configuration features are the three-dimensional binary density matrix obtained by mapping the chassis model; The design parameter attribute vector contains the real-time updated plate thickness values, cross-sectional dimensions, connection parameters, and material grade indexes for each component; Load condition feature matrix, extract the load distribution characteristics of the frame under bending, torsion, braking and steering conditions; Performance response characteristics include frame mass, first-order modal frequency, bending stiffness, torsional stiffness, maximum stress, maximum deformation, and minimum fatigue life.

3. The lightweight frame optimization design method as described in claim 1, characterized in that, The hybrid action space mentioned in step S1 includes a discrete action subspace and a continuous action subspace. The discrete action subspace is used to make decisions on material selection, cross-sectional shape selection of beams and crossbeams, and connection parameter selection. The continuous action subspace is used to make decisions on plate thickness adjustment, cross-sectional size adjustment, beam layout coordinate adjustment, and connection spacing adjustment.

4. The lightweight frame optimization design method as described in claim 1, characterized in that, The guided reward function mentioned in step S1 is: in, As a reward value, The percentage reduction in frame weight. This indicates the degree of violation of the i-th performance metric. For the corresponding to the first i The Lagrange multiplier of each performance indicator; when all performance indicators meet the constraints, the reward value is the percentage reduction in frame mass; when any performance indicator violates the constraints, a negative penalty is imposed according to the degree of violation.

5. The lightweight frame optimization design method as described in claim 1, characterized in that, The parameterized action space network described in step S2 adopts a layered architecture with a shared feature extraction layer and a dual-head output structure: The shared feature extraction layer extracts the spatial topological relationships of geometric configuration features in the composite state space through a three-dimensional residual convolutional network, and fuses them with design parameter attribute vectors and performance response features to output high-dimensional semantic features. The dual-head output structure includes a discrete output head and a continuous output head. The discrete output head outputs the probability distribution of discrete actions based on the high-dimensional semantic features. The continuous output head outputs the mean and logarithmic standard deviation of continuous actions based on the high-dimensional semantic features, and generates continuous action values ​​through reparameterization techniques. : Where μ and log σ are the mean and log standard deviation of the output from the continuous output head, respectively. This is sampling noise.

6. The lightweight frame optimization design method as described in claim 4, characterized in that, In step S3, The underlying action masking is as follows: a mask vector is constructed based on the current geometric topology of the chassis and a preset process rule library; a mask operation is performed on the probability distribution of discrete actions; and the selection probability of illegal actions is set to zero. The projection of the middle security layer is as follows: when the continuous action exceeds the preset feasible region, the continuous action is projected to the nearest point on the boundary of the feasible region by solving a quadratic programming problem, and the projection process is embedded in the neural network to realize gradient backpropagation. The upper-level Lagrange dynamic penalty is as follows: the Lagrange multipliers corresponding to each performance index in the reward function are dynamically updated using the dual gradient descent method; when a certain performance index fails to meet the standard, its corresponding Lagrange multiplier value is automatically increased, forcing the agent to abandon excessive weight reduction and prioritize meeting the reliability requirements; conversely, when the constraint is met, the Lagrange multiplier decays to zero, allowing the agent to freely explore the lightweight boundary.

7. The lightweight frame optimization design method as described in claim 1, characterized in that, The process of generating the Pareto non-dominated solution set in step S3 is as follows: An external archive set is established during the strategy iteration process to store non-dominated solutions generated during the optimization process; Each time a new valid solution is generated, the new valid solution is compared with the existing solutions in the external archive set in terms of dominance; If the new valid solution is dominated by any existing solution in the external archive set, then the new valid solution is rejected from being added to the external archive set. If the new effective scheme dominates several existing schemes in the external archive set, then the new effective scheme is added to the external archive set, and the existing schemes dominated by it are removed. If the new effective solution is not mutually exclusive with any existing solutions in the external archive set, then the new effective solution is directly added to the external archive set. After the iteration is completed, all non-dominant solution points retained in the external archive set are output as Pareto non-dominant solution set, which constitutes the optimal trade-off front of the chassis lightweight design space.

8. The lightweight frame optimization design method as described in claim 1, characterized in that, Step S4 includes: From the Pareto non-dominated solution set, m candidate solutions are extracted, and n evaluation indicators are extracted for each solution to construct an original decision matrix. The original decision matrix is ​​then dimensionless using the range transformation method, converting all indicators into benefit-type indicators to obtain a standardized matrix. ; The subjective weight vector is calculated by scoring the relative importance of each indicator based on the analytic hierarchy process (AHP) and after consistency testing. The information entropy and difference coefficient of each indicator in the standardized matrix are calculated based on the entropy weight method, and the objective weight vector is obtained after normalization. ; ， ， in, For the first j The coefficient of difference of the indicators For the first j Information entropy of the evaluation indicators is the constant coefficient in the calculation of information entropy, and , For the first j The first item under the indicator i The weight of the indicator values ​​of each candidate solution, and have , For the standardized matrix elements, represent the first... i The candidate solution is in the... j Standardized values ​​for each evaluation indicator; The comprehensive weight vector is calculated using a linear weighted model. α and β are preset preference coefficients; The standardized matrix is ​​weighted using the comprehensive weight vector to construct a weighted normalized decision matrix. ,in ; Define the positive ideal solution in the weighted normalized decision matrix. and negative ideal solution Calculate the Euclidean distance from each candidate solution to the positive ideal solution. Euclidean distance to the negative ideal solution According to the formula Calculate the relative schedule of each candidate scheme, and determine the scheme with the highest relative schedule as the optimal lightweight engineering scheme.

9. The lightweight frame optimization design method as described in claim 1, characterized in that, In step S2, the hybrid policy agent uses a pre-trained high-precision multi-performance prediction model as the interaction environment for policy iterative training. The high-precision multi-performance prediction model is constructed by using experimental design methods to generate sample points in the design space of the chassis, forming a training sample set that covers the range of design variable values. High-fidelity finite element simulation software is used to perform performance simulation calculations on each sample point in the training sample set to obtain the performance response data corresponding to each sample point. The performance response data includes at least the frame mass, fatigue life, first-order modal frequency, bending stiffness, torsional stiffness, maximum stress, and maximum deformation. A training dataset is constructed by using sample points in the training sample set as input features and the corresponding performance response data as output labels. A machine learning algorithm is selected to fit the training dataset, a nonlinear mapping relationship between design parameters and performance response is established, and the model accuracy is verified by cross-validation to obtain the high-precision multi-performance prediction model. During the policy iteration training process, the hybrid policy agent inputs the action vector output by the current decision step into the high-precision multi-performance prediction model. The high-precision multi-performance prediction model returns the corresponding performance evaluation results in real time, which serve as the basis for agent state updates and reward function calculations.

10. A chassis lightweight optimization design system based on the method of any one of claims 1 to 9, characterized in that, The system includes the following modules: The chassis design decision task modeling module is equipped with a state and action space construction unit, which is used to encapsulate the three-dimensional voxelized configuration of the chassis, design parameter attributes, load condition characteristics and current performance response into a composite state vector, and map discrete decisions and continuous decisions into a hybrid action space. The hybrid policy agent construction and optimization module has a core component of a parameterized action space network, which integrates a shared feature perception layer, a discrete output head for outputting discrete action probability distributions, and a continuous output head for outputting continuously adjusted action distribution parameters, thereby achieving collaborative decision-making on hybrid design variables. The hybrid policy agent construction and optimization module is also equipped with a policy iteration optimization unit, which uses a near-end policy optimization algorithm combined with a trust region truncation mechanism to update network weights. The engineering feasibility maintenance and non-dominated solution generation module is equipped with a hierarchical constraint processing engine, which contains three cascaded constraint operators: a bottom-level action masking operator to filter illegal discrete decisions that violate manufacturing process limits; a middle-level safety layer projection operator to correct continuous actions that violate spatial interference constraints to the feasible region boundary; and an upper-level dynamic penalty correction unit to handle strongly nonlinear performance constraints based on the Lagrange multiplier method. The module also integrates a performance evaluation proxy interface, which calls a pre-trained high-precision multi-performance prediction model to evaluate the performance of the corrected design scheme and iteratively generates a multi-objective non-dominated solution set covering the Pareto front of the design space. The multi-criteria comprehensive decision-making optimization module integrates a subjective and objective combination weighting component, and is configured with a hierarchical analysis operator and an information entropy calculation operator, which are used to quantify subjective preferences and mine objective statistical characteristics of solution set data to generate a comprehensive weight vector; and a comprehensive evaluation unit for approximating the ideal solution ranking method, which is used to construct a weighted normalized decision matrix, calculate the relative distance between each candidate scheme and the ideal solution and the negative ideal solution, and output the optimal lightweight engineering scheme with the highest relative progress.

Images