Search

CN-121997466-A - Frame lightweight optimization design method and system considering fatigue strong constraint

CN121997466ACN 121997466 ACN121997466 ACN 121997466ACN-121997466-A

Abstract

The invention discloses a frame lightweight optimization design method and system considering fatigue strong constraint, which are characterized in that a frame design task is modeled into a constrained Markov decision process, a Hybrid-PPO algorithm is utilized to construct a mixed strategy intelligent body capable of simultaneously processing discrete type selection and continuous size variable, a sub-second level rapid evaluation of the performance of a scheme is realized by calling a high-precision multi-performance prediction model, a layered constraint processing mechanism is introduced in optimizing iteration, and the strong constraint such as manufacturing process limit, spatial interference, fatigue life and the like is forcefully maintained through bottom action shielding, middle safety layer projection and upper dynamic punishment, so that a non-inferior solution set covering the Pareto front edge is generated on the premise of ensuring engineering feasibility, and finally, a decision closed loop is established based on a combined empowerment and TOPSIS method to screen a final scheme. The invention realizes the autonomous intelligent design of the frame structure, and remarkably improves the optimizing efficiency and the light weight level of the frame design.

Inventors

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

Assignees

  • 吉林大学

Dates

Publication Date
20260508
Application Date
20260410

Claims (10)

  1. 1. A frame lightweight optimization design method considering fatigue strong constraint is characterized by comprising the following steps: S1, modeling a frame structure optimization task as a constrained Markov decision process, defining a composite state space containing geometric configuration characteristics, design parameter attributes, load working condition characteristics and performance response, constructing a mixed action space of coupling discrete variables and continuous variables, and establishing a guide type reward function taking light-weight benefit as positive feedback and engineering constraint violation as negative feedback; S2, constructing a mixed strategy intelligent agent based on a parameterized action space network, taking the composite state space as input, synchronously outputting probability distribution of discrete type selection actions and distribution parameters of continuous adjustment actions by utilizing the mixed action space, and carrying out strategy iterative training on the mixed strategy intelligent agent by adopting a near-end strategy optimization algorithm and taking a pre-trained high-precision multi-performance prediction model as an interaction environment; S3, in the strategy iteration process, sequentially executing bottom-layer action shielding facing discrete selection decision, middle-layer security layer projection facing continuous variable adjustment and upper-layer Lagrange dynamic punishment facing performance index constraint, and constraining actions output by an agent to an engineering feasible domain to generate a Pareto non-inferior solution set meeting engineering feasibility requirements; and S4, carrying out multi-criterion decision optimization on the Pareto non-inferior solution set based on a combined weighting and approximation ideal solution ordering method, and outputting an optimal engineering lightweight scheme.
  2. 2. The method for optimizing the weight of a vehicle frame according to claim 1, wherein the composite state space in step S1 includes: the geometric configuration features are three-dimensional binarization density matrixes obtained by mapping the frame models; Designing parameter attribute vectors, wherein the parameter attribute vectors comprise plate thickness values, section sizes, connection parameters and material grade indexes updated in real time by all parts; the load condition characteristic matrix is used for extracting load distribution characteristics of the frame under bending, torsion, braking and steering conditions; Performance response characteristics, including frame mass, first order modal frequencies, bending stiffness, torsional stiffness, maximum stress, maximum deformation, and minimum fatigue life values.
  3. 3. The method for optimizing the design of the light weight of the vehicle frame according to claim 1, wherein the mixed motion space in the step S1 comprises a discrete motion subspace and a continuous motion subspace, the discrete motion subspace is used for deciding material selection, cross-beam cross-sectional shape selection and connection parameter selection, and the continuous motion subspace is used for deciding plate thickness adjustment, cross-sectional size adjustment, cross-beam layout coordinate adjustment and connection interval adjustment.
  4. 4. The method for optimizing the weight of a vehicle frame according to claim 1, wherein the guided type reward function in step S1 is: Wherein, the In order to be a prize value, For the mass reduction ratio of the frame, Indicating the degree of violation of the ith performance index, The method comprises the steps of obtaining a Lagrangian multiplier corresponding to an ith performance index, obtaining a reward value which is a frame quality reduction proportion when all performances meet constraint, and giving negative punishment according to the degree of violation when any performance index violates constraint.
  5. 5. The method for optimizing the weight of a vehicle frame according to claim 1, wherein in the step S2, the parameterized action space network adopts a layered architecture of a shared feature extraction layer and a dual-head output structure: the shared feature extraction layer extracts the space topological relation of geometric configuration features in the composite state space through a three-dimensional residual convolution network, fuses the space topological relation with design parameter attribute vectors and performance response features, and outputs high-dimensional semantic features, the double-end output structure comprises a discrete output head and a continuous output head, the discrete output head outputs probability distribution of discrete actions based on the high-dimensional semantic features, the continuous output head outputs average and logarithmic standard deviation of continuous actions based on the high-dimensional semantic features, and continuous action values a are generated through a re-parameterization skill: wherein mu and log sigma are the mean and logarithmic standard deviation of the continuous output head output respectively, Is sampling noise.
  6. 6. The method for optimizing the weight of a vehicle frame according to claim 4, wherein in step S3, The bottom layer action shielding is that a mask vector is constructed according to the current geometric topological state of the frame and a preset process rule base, masking operation is carried out on probability distribution of discrete actions, and the selection probability of illegal actions is set to zero; When the continuous action exceeds a preset feasible region, the continuous action is projected to a point closest to the boundary of the feasible region by solving a quadratic programming problem, and the projection process is embedded into a neural network to realize gradient back propagation; The upper layer Lagrange dynamic penalty is that Lagrange multipliers corresponding to all performance indexes in the reward function are dynamically updated by a dual gradient descent method, when certain performance is continuously not up to standard, corresponding Lagrange multiplier values are automatically increased, an agent is forced to give up excessive weight reduction to preferentially meet the reliability requirement, otherwise, when constraint is met, lagrange multipliers are attenuated to zero, and the agent is allowed to freely explore the lightweight boundary.
  7. 7. The method for optimizing the design of the light frame according to claim 1, wherein the generating process of the Pareto non-inferior solution set in the step S3 is as follows: establishing an external archiving set in a strategy iteration process, wherein the external archiving set is used for storing non-inferior solutions generated in an optimizing process; If the new scheme is governed by any existing scheme in the external archive set, refusing to add the new scheme to the external archive set; If the new scheme dominates several existing schemes in the external archive set, adding the new scheme to the external archive set and rejecting the existing schemes dominated by it; if the new scheme is not mutually exclusive with all existing schemes in the external archive set, directly adding the new scheme into the external archive set; and after iteration is completed, outputting all mutually non-dominant solution points reserved in the external archiving set as a Pareto non-inferior solution set, wherein the Pareto non-inferior solution set forms an optimal balance front edge of the lightweight design space of the frame.
  8. 8. The frame lightweight optimization design method according to claim 1, wherein step S4 includes: extracting m candidate schemes from the Pareto non-inferior solution set, extracting n evaluation indexes for each scheme, constructing an original decision matrix, performing dimensionless treatment on the original decision matrix by adopting a range transformation method, and converting all indexes into benefit indexes to obtain a standardized matrix ; Scoring the relative importance of each index based on analytic hierarchy process, and calculating to obtain subjective weight vector after consistency test Calculating information entropy and difference coefficient of each index in the standardized matrix based on entropy weight method, and normalizing to obtain objective weight vector ; , , Wherein, the Is the difference coefficient of the j-th index, The information entropy of the j-th evaluation index, Is a constant coefficient in information entropy calculation, and , Is the index value proportion of the ith candidate scheme under the jth index, and has , Representing the standardized value of the ith candidate solution on the jth evaluation index for the standardized matrix element; calculating a composite weight vector using a linear weighting model Alpha and beta are preset preference coefficients; weighting the standardized matrix by using the comprehensive weight vector to construct a weighted standardized decision matrix Wherein ; Defining a positive ideal solution in the weighted normalized decision matrix And negative ideal solution Calculating Euclidean distance from each candidate scheme to positive ideal solution Euclidean distance from sum to negative ideal solution According to the formula And calculating the relative paste progress of each candidate scheme, and determining the scheme with the highest relative paste progress as the optimal project lightweight scheme.
  9. 9. The method for optimizing the design of the light frame according to claim 1, wherein the mixed strategy agent in the step S2 performs strategy iterative training by taking a pre-trained high-precision multi-performance prediction model as an interaction environment, The high-precision multi-performance prediction model is constructed by adopting a test design method to generate sample points in a design space of a frame to form a training sample set covering a design variable value range, invoking high-fidelity finite element simulation software to perform performance simulation calculation on each sample point in the training sample set to obtain performance response data corresponding to each sample point, wherein the performance response data at least comprises frame quality, fatigue life, first-order modal frequency, bending stiffness, torsional stiffness, maximum stress and maximum deformation; In the strategy iterative training process, the mixed strategy agent inputs the motion vector output by the current decision step into the high-precision multi-performance prediction model, and the high-precision multi-performance prediction model returns the corresponding performance evaluation result in real time to be used as the basis for agent state updating and rewarding function calculation.
  10. 10. A frame lightweight optimization design system based on the method of any one of claims 1-9, characterized in that the system comprises the following modules: The vehicle frame design decision task modeling module is configured with a state and action space construction unit and is used for packaging a three-dimensional voxelized configuration, design parameter attributes, load working condition characteristics and current performance response of a vehicle frame into a composite state vector and mapping discrete decisions and continuous decisions into a mixed action space; The system comprises a mixed strategy intelligent body construction and optimizing module, a strategy iteration optimizing unit, a near-end strategy optimizing algorithm and a trust domain cutting mechanism, wherein a core component of the mixed strategy intelligent body construction and optimizing module is a parameterized action space network, a shared feature sensing layer, a discrete output head for outputting discrete selection action probability distribution and a continuous output head for outputting continuous adjustment action distribution parameters are integrated in the mixed strategy intelligent body construction and optimizing module, so that collaborative decision on mixed design variables is realized; The engineering feasibility maintenance and non-inferior solution generation module is configured with a layered constraint processing engine, and three cascaded constraint operators are contained in the engineering feasibility maintenance and non-inferior solution generation module, wherein a bottom-layer action shielding operator is used for filtering illegal discrete decisions violating manufacturing process limits, a middle-layer safety-layer projection operator is used for correcting continuous actions violating space interference constraints to a feasible domain boundary, and an upper-layer dynamic penalty correction unit is used for processing strong non-linear performance constraints based on a Lagrangian multiplier method; the system comprises a multi-criterion comprehensive decision optimization module, an approximation ideal solution sequencing method comprehensive evaluation unit and an optimization method optimization module, wherein the multi-criterion comprehensive decision optimization module is integrated with a subjective and objective combination weighting component, is provided with an analytic hierarchy operator and an information entropy calculation operator and is respectively used for quantifying subjective preference and mining objective statistical characteristics of solution set data to generate comprehensive weight vectors, and the approximation ideal solution sequencing method comprehensive evaluation unit is used for constructing a weighted normalized decision matrix, calculating relative distances between candidate schemes and ideal solutions and negative ideal solutions and outputting an optimal engineering light-weight scheme with highest relative fitting progress.

Description

Frame lightweight optimization design method and system considering fatigue strong constraint Technical Field The invention belongs to the technical field of frame structural design and artificial intelligent optimizing decision crossing, and particularly relates to a frame lightweight optimizing design method and system considering fatigue strong constraint. Background Under the development trend of intellectualization, dynamism and low carbonization of commercial vehicles, the frame is used as a basic bearing part of the whole vehicle and faces the double technical challenges brought by weight increment and mileage anxiety of a power battery. The built-in conflict between the structural bearing requirement and the whole vehicle energy efficiency target makes the design of the vehicle frame extremely light on the premise of ensuring high reliability. Currently, with the continuous shortening of the updating iteration period of commercial vehicle products, more stringent engineering requirements are put forward on the precision and efficiency of the frame structural design. The efficient and intelligent optimizing decision method and the automatic design system are core supports for realizing the lightweight forward design of the structure. In the current structural design and optimizing decision links, the existing technical scheme still faces the following bottlenecks: Firstly, when the traditional structure optimizing strategy is used for solving the problem of high-dimensional and strong nonlinear frame optimization, the optimizing efficiency, the convergence accuracy and the global searching capability are difficult to be considered. At present, the industry mainly adopts a direct search strategy or an optimization method based on a proxy model. The direct search method (such as CN202511490809.1 and CN 202511148781.3) needs to frequently call the high-fidelity CAE model to execute nonlinear fatigue analysis, so that the single optimization period is overlong, and the engineering practicability is poor. While the computation amount of the agent model-based method (such as CN 202410887126.9) is reduced by mathematical fitting, the method is limited by the characterization capability of the traditional agent model on the strong nonlinear response curved surface, the prediction deviation is easy to generate in a local extremum region, and more importantly, when the method faces to a highly non-convex design space comprising discrete selection and continuous size coupling, the method is extremely easy to fall into a local optimal solution, and the extremely light weight potential of the structure is difficult to excavate in a complex parameter combination. Secondly, the existing deep reinforcement learning method lacks an effective maintenance mechanism aiming at strong and hard constraint when being applied to complex engineering assembly design. While the prior art attempts to apply reinforcement learning to design optimization of biomechanical structures such as implant interventional instruments (e.g., CN 202511510033.5), process automation is achieved by embedding a deep learning model into the reinforcement learning framework. However, such methods focus on geometric adaptation of the structure, and hard processing logic involving multidisciplinary strong constraints for commercial vehicle frames has not been established. The frame design involves the mixed action space of depth coupling of discrete variables (materials/bolt selection, etc.) and continuous variables (plate thickness/section size, etc.), and illegal schemes of violating manufacturing process limits (such as bolt margin overrun) or geometric interference are extremely easy to generate in the optimizing and exploring process. Due to the lack of a layered control mechanism aiming at the engineering feasibility of the complex assembly, the intelligent agent frequently idles in a dead space, and the optimizing process is difficult to converge to a feasible solution meeting the engineering practical requirements. Finally, the existing optimization flow lacks a subjective and objective cooperative closed-loop system in a multi-objective decision link. At present, after a Pareto non-inferior solution set covering a design space is obtained through an optimizing algorithm, the screening of a final scheme often depends on manual experience or single performance weight distribution excessively. Such methods (e.g., CN 202210419640.0) are either too much disturbed by the subjective preferences or are susceptible to random fluctuations in data, making it difficult to achieve a scientific balance between light weight gains and performance risks. Due to the lack of a comprehensive evaluation mechanism capable of deeply fusing expert engineering intuition and data statistics characteristics, decision faults exist between algorithm optimization and engineering landing, so that design efficiency is limited, and the robustness and reliability of a final p