Integrated intelligent design method for electric truck aluminum alloy frame

Abstract

The application provides a kind of integrated intelligent design method of electric truck aluminum alloy frame, belongs to the technical field of automobile structure design.The method comprises the following steps: establishing a collaborative parametric finite element model of the frame, unifying the structural size, section shape and bolt connection parameters as design variables, and obtaining a performance sample set; based on MLP, an agent model is constructed to predict the frame mass, fatigue life, modal, displacement and stress response; a PPO-enhanced NSGA-III algorithm is used for multi-objective optimization, and the PPO dynamically adjusts the crossover and mutation parameters to obtain a Pareto candidate solution; and then the CRITIC-TOPSIS algorithm is used to select the best design scheme. The application can reduce the amount of repeated finite element calculation, improve the efficiency of high-dimensional multi-objective optimization, and achieve the collaborative balance between frame lightweight, stiffness, strength, modal performance and fatigue life.

Description

Technical Field

[0001] This invention belongs to the field of automotive design technology, specifically relating to an integrated intelligent design method for an aluminum alloy chassis of an electric cargo truck. Background Technology

[0002] With the increasing demand for lightweight electric trucks, aluminum alloys, with their low density, high strength-to-weight ratio, and excellent corrosion resistance, have become the preferred material to replace traditional steel frames and achieve structural lightweighting. However, aluminum alloys and traditional steel have significant differences in mechanical properties, and simple weight reduction designs often lead to a decline in the overall performance of the structure. In practical engineering, the frame design process is essentially a highly nonlinear optimization problem involving multiple conflicting objectives such as mass, stiffness, strength, and fatigue life. Traditional empirical design methods are no longer sufficient to meet the precision design requirements of modern frames.

[0003] Currently, the main technical bottlenecks in this field are as follows: 1. Low computational efficiency: Direct optimization using the finite element model requires a large amount of computational resources, has a long cycle, and is difficult to handle complex design spaces.

[0004] 2. Limited algorithm optimization capability: The traditional NSGA-III optimization algorithm uses fixed values ​​for crossover and mutation parameters set by experience. When dealing with highly nonlinear chassis design problems, the convergence speed is slow and it is very easy to get trapped in local optima.

[0005] 3. Strong subjectivity in decision-making: Multi-objective optimization usually produces a Pareto solution set containing a large number of alternative solutions. Existing decision-making methods mostly adopt subjective weighting methods. Due to the influence of the decision-maker's subjective will and experience level, they often ignore the inherent conflicts and correlations between objective data indicators, resulting in the final selected solution not necessarily being the global optimal solution. Summary of the Invention

[0006] In view of the above-mentioned problems in the prior art, the purpose of this invention is to provide an integrated intelligent design method for an aluminum alloy frame of an electric cargo truck.

[0007] An integrated intelligent design method for an aluminum alloy chassis of an electric cargo truck includes the following steps: S1: Establish a collaborative parameterized finite element model of the aluminum alloy frame of the electric truck, and define the frame structure size variables, cross-sectional shape variables and bolt connection parameters as collaborative design variables. Then, obtain an initial performance sample set composed of the collaborative design variables and the corresponding frame performance response quantities through sampling simulation. S2: Using the collaborative design variables as input and the chassis performance response as output, construct an MLP integrated proxy model. The MLP integrated proxy model is used to predict the chassis performance response corresponding to the candidate chassis design scheme. S3: Perform multi-objective optimization based on the PPO-enhanced NSGA-III algorithm. Utilize the MLP ensemble surrogate model to predict the performance of candidate individuals in the NSGA-III population and extract the population evolutionary state, including evolutionary progress, population distribution diversity, hypervolume index, hypervolume increment, and the proportion of non-dominated solutions. Input the population evolutionary state into the PPO agent. The PPO agent outputs action vectors, which are then mapped to genetic operator parameters of the NSGA-III algorithm. Use the hypervolume increment as a reward signal and apply an evolutionary stagnation penalty when the hypervolume does not increase to update the PPO agent and obtain the Pareto candidate design scheme set. S4: The Pareto candidate design scheme set is comprehensively evaluated and ranked based on the CRITIC-TOPSIS algorithm, and the candidate design scheme with the highest relative similarity is selected as the final recommended design scheme for the aluminum alloy frame.

[0008] Preferably, in step S1, SFE software is used to parametrically model the shape variables and structural dimension variables of the frame, and Python parametric scripts are used to automatically define bolt connection parameters, including bolt specifications, bolt spacing, bolt quantity, and bolt hole positions, so that the bolt connection parameters are updated synchronously with the changes in the frame structural dimension variables and cross-sectional shape variables.

[0009] Preferably, the collaborative design variables constitute a multi-dimensional design variable vector, including frame structure dimension variables, cross-sectional shape variables, and bolt connection parameters; The frame structure dimensional variables include the thickness, length, width, height, or installation position parameters of the longitudinal beams, crossbeams, connecting plates, or reinforcing plates; the cross-sectional shape variables include the cross-sectional height, cross-sectional width, fillet radius, or reinforcing rib parameters of the longitudinal beams or crossbeams; the bolt connection parameters include the bolt specifications, bolt quantity, bolt spacing, bolt edge distance, or bolt arrangement position parameters for different connection areas.

[0010] Preferably, in step S1, the Hamsley sampling method is used to uniformly sample within the variable space formed by the collaborative design variables, and batch simulation is performed by driving the Optistruct solver and nCode fatigue analysis software through an automated script to obtain 10 frame performance response quantities corresponding to each group of collaborative design variables. The 10 vehicle frame performance response parameters include vehicle frame mass, fatigue life, first-order modal frequency, maximum displacement under torsional conditions, maximum displacement under braking conditions, maximum displacement under bending conditions, maximum stress under torsional conditions, maximum stress under braking conditions, maximum stress under bending conditions, and maximum stress under steering conditions.

[0011] Preferably, in step S2, constructing the MLP integrated agent model includes: Logarithmic transformation was performed on the fatigue life data, and the input variables and output response were standardized. The sample set is divided into a training set and a test set, and an MLP network topology containing an input layer, several hidden layers and an output layer is constructed. A two-stage grid search strategy based on five-fold cross-validation is used to determine the hyperparameters of the MLP model. The hyperparameters include the L2 regularization penalty coefficient, batch size, number of hidden layer neurons, and initial learning rate. Under optimal hyperparameter configuration, the Adam optimization algorithm and mean squared error loss function are used to train the MLP model, and an early stopping mechanism based on validation set error is introduced. For each frame performance response quantity, multiple MLP base models are trained independently, and the average of the prediction results of the multiple MLP base models is used as the integrated prediction output of the corresponding frame performance response quantity.

[0012] Preferably, in step S3, the frame mass, maximum displacement under torsional conditions, and fatigue life are taken as the main optimization objectives, and corresponding engineering constraints are set to establish a multi-objective optimization mathematical model. Among them, the frame mass and maximum displacement under torsional conditions are indicators that are better the smaller they are, and the fatigue life is an indicator that is better the larger it is.

[0013] Preferably, in step S3, the state vector constructed by the PPO agent based on the population evolution state is:

[0014] in, Represents the state vector of generation t; Indicates the progress of evolution. This indicates the diversity of population distribution in the target space; The overvolume index represents the current population size. Indicates the volume increment; Indicates the proportion of non-dominated solutions; in, t represents the current generation. Indicates the maximum number of generations; , This indicates the overvolume index of the previous generation population.

[0015] Preferably, in step S3, the Actor policy network of the PPO agent is based on the state vector. Output a four-dimensional continuous action vector: ,in, All are continuous action values ​​located in the interval [0,1]. The four-dimensional continuous action vector is converted into genetic operator parameters for the NSGA-III algorithm through the following mapping relationship: in, It is the cross-distribution index; It is the distribution index of variation; The crossover probability; Probability of mutation; The number of decision variables.

[0016] Preferably, in step S3, the hypervolume increment of the (t+1)th generation population relative to the tth generation population is used. As a reward signal, and with the introduction of an evolutionary stagnation penalty, the t-th generation receives an immediate reward. for:

[0017] After each generation of evolution, the interaction trajectory will be... , , , The data is stored in the trajectory buffer, and the Actor network and Critic network of the PPO agent are periodically updated based on the interaction trajectory.

[0018] Preferably, in step S4, the selection of the optimal design scheme based on the CRITIC-TOPSIS algorithm includes: An initial decision matrix was constructed using 10 chassis performance response parameters corresponding to the Pareto candidate design schemes; For fatigue life and first-order modal frequency, positive index standardization is used; for frame mass, maximum displacement under each working condition, and maximum stress under each working condition, inverse index standardization is used. Based on the CRITIC algorithm, objective weights are calculated according to the standard deviation of each evaluation indicator and the correlation coefficient between indicators. A weighted standardized decision matrix is ​​constructed based on the objective weights to determine the positive and negative ideal solutions, and the relative closeness of each Pareto candidate design is calculated. The Pareto candidate designs were sorted from highest to lowest relative similarity, and the candidate design with the highest relative similarity was selected as the final recommended design.

[0019] The beneficial effects of this invention are: the integrated intelligent design method for the aluminum alloy frame of the electric cargo truck, This invention enables the coordinated optimization of structural and connection parameters of aluminum alloy vehicle frames: It defines frame structural dimensions, cross-sectional shape, and bolt connection parameters as coordinated design variables, and uses a coordinated parametric finite element model to synchronously update the main structural components and bolt connection schemes. Compared to the traditional approach of designing and optimizing beam structures and connection structures separately and in stages, this invention can simultaneously consider the impact of longitudinal beams, cross beams, connecting plates, reinforcing plates, and bolt connection areas on the overall frame performance within the same design space, reducing the overall performance imbalance caused by local optimization.

[0020] This invention significantly reduces the amount of repetitive finite element calculations in high-dimensional chassis optimization: Based on finite element simulation samples, it constructs an MLP integrated surrogate model, establishing a mapping relationship between multi-dimensional collaborative design variables and performance responses such as chassis mass, fatigue life, first-order modal frequency, maximum displacement under each working condition, and maximum stress under each working condition. During multi-objective optimization, the MLP integrated surrogate model can be used to quickly predict the performance response of candidate design schemes, thereby reducing repeated calls to simulation software such as Optistruct and nCode and improving optimization computation efficiency.

[0021] This invention enhances the adaptive search capability of high-dimensional multi-objective optimization processes by introducing a PPO agent into the NSGA-III algorithm. The population evolutionary state is used as the input to the PPO agent, which dynamically outputs the crossover distribution index, mutation distribution index, crossover probability, and mutation probability based on this state. This allows the NSGA-III algorithm to automatically adjust its search strategy according to different evolutionary stages. Compared to the traditional NSGA-III algorithm with fixed genetic operator parameters, this invention better balances global exploration capability and local convergence capability, reducing the risk of getting trapped in local optima or search stagnation.

[0022] This invention enables a direct correlation between reinforcement learning training objectives and Pareto front improvement directions: It uses the hypervolume increment as the reward signal for the PPO agent and imposes an evolutionary stagnation penalty when the hypervolume is not improved. Therefore, the learning direction of the PPO agent directly corresponds to the improvement of Pareto solution set quality. This encourages the algorithm to prioritize learning genetic operator parameter adjustment strategies that promote Pareto front expansion and improve solution set distribution, thereby improving the convergence accuracy and distribution uniformity of the candidate solution set.

[0023] This invention improves the objectivity of the final frame selection: After obtaining Pareto candidate designs, it employs the CRITIC-TOPSIS algorithm for comprehensive evaluation. The CRITIC algorithm determines objective weights based on the variability of each performance indicator and the correlation between indicators, reducing subjective bias caused by manual weighting. The TOPSIS algorithm calculates the relative proximity of candidate designs to positive and negative ideal solutions, enabling unified ranking of multiple candidate designs. Therefore, this invention can objectively select the aluminum alloy frame design with superior overall performance from multiple Pareto compromise solutions.

[0024] This invention achieves a comprehensive balance between lightweight design and structural performance: It prioritizes frame mass, maximum displacement under torsional conditions, and fatigue life as primary optimization objectives, while using bending displacement, braking displacement, maximum stress under each condition, and first-order modal frequency as constraints or evaluation indicators. This ensures that the final recommended design not only focuses on weight reduction but also considers stiffness, strength, modal characteristics, and fatigue durability. Therefore, this invention provides an aluminum alloy frame design more suitable for the engineering applications of electric freight vehicles. Attached Figure Description

[0025] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the MLP of the present invention; Figure 3 This is a schematic diagram illustrating the prediction accuracy of the MLP integrated agent model of the present invention; Figure 4 This is a schematic diagram of the PPO-NSGA-III process of the present invention; Figure 5 This is a schematic diagram of the 3D distribution of the Pareto front of the present invention. Detailed Implementation

[0026] Example 1 like Figure 1 As shown, an integrated intelligent design method for an aluminum alloy frame of an electric cargo truck includes the following steps: S1: Establish the finite element model of the chassis and obtain the initial performance sample set. In this embodiment, a parametric finite element model of the aluminum alloy frame of the electric truck is first established based on the finite element analysis software (Optistruct). The aluminum alloy frame includes longitudinal beams, crossbeams, connecting plates, reinforcing plates, and bolted connection structures for connecting the longitudinal beams and crossbeams, the longitudinal beams and connecting plates, or the crossbeams and connecting plates.

[0027] To address the challenge of achieving global coordination in traditional chassis optimization design where connection parameters, shape parameters, and structural dimension parameters are optimized sequentially in stages, this embodiment employs a multi-software collaborative parametric automatic modeling method. Specifically, SFE software is used to parametrically model the shape and structural dimension variables of the chassis, enabling rapid and flexible reconstruction of these parameters. Simultaneously, parametric scripts written in Python are used to automatically define and batch-arrange connection parameters such as bolt specifications, bolt spacing, bolt quantity, and bolt hole positions. The generated connection scheme adheres to the relevant specifications for aluminum alloy structure design and fastener mechanical properties, ensuring the manufacturability of the design.

[0028] In the specific implementation process, taking the bolt design at the connection position between the first crossbeam connecting plate and the longitudinal beam of the aluminum alloy frame as an example, firstly, the mid-surface of the thickness of each part of the initial aluminum alloy frame is extracted in the SFE software, the preset bolt connection holes are deleted, and the frame and connection structure are meshed using finite element methods and the mesh quality is checked. Subsequently, the position of the bolt center hole is determined according to the preliminary connection scheme, and temporary nodes are arranged at the corresponding positions for coordinate positioning. Finally, a Python parametric script is called to automatically generate the bolt holes and the corresponding bolt connection models based on the coordinates of the temporary nodes. By setting loop logic in the Python script, different bolt arrangement schemes can be automatically updated as structural size variables and cross-sectional shape variables change, thereby realizing the collaborative parametric design of the frame main structure and the connection structure.

[0029] Furthermore, through the collaborative driving of SFE software and Python scripts, the frame structure dimensions, cross-sectional shape, and bolt connection parameters are uniformly incorporated into the same design variable space. This embodiment extracts and defines N collaborative design variables, which cover the structural dimensions, cross-sectional shape, and bolt connection parameters of the aluminum alloy frame.

[0030] The N collaborative design variables constitute the design variable vector:

[0031] In the formula, Represents the N-dimensional collaborative design variable vector of the aluminum alloy frame; Represents the nth collaborative design variable. .

[0032] Among them, structural dimension variables include the thickness, length, width, height or installation position parameters of longitudinal beams, transverse beams, connecting plates and reinforcing plates; cross-sectional shape variables include the cross-sectional height, cross-sectional width, fillet radius or reinforcing rib parameters of longitudinal beams and transverse beams; bolt connection parameters include bolt specifications, bolt quantity, bolt spacing, bolt edge distance and bolt arrangement position parameters for different connection areas.

[0033] For continuous variables such as plate thickness, beam height, beam width, and connecting plate dimensions, values ​​are taken continuously within the preset engineering range; for bolt specifications, bolt quantity, and some bolt placement parameters, integer encoding is used. Specifically, the discrete variables related to bolt connections satisfy the following:

[0034] In the formula, This indicates the bolt specification or arrangement type code for different connection areas. Codes indicating the number or arrangement of bolts in different connection areas; Represents a set of integers.

[0035] To ensure the initial samples have good representativeness and uniformity of space filling in the N-dimensional design space, this embodiment uses the Hammsley sampling method to uniformly sample within the N-dimensional variable space, generating 500 sets of design sample points. Subsequently, batch finite element simulation calculations are performed using an automated script-driven Optistruct solver and nCode fatigue analysis software to extract 10 key performance response quantities corresponding to each sample point.

[0036] The 10 key performance response quantities include frame mass, fatigue life, first-order modal frequency, maximum displacement under torsional, braking, and bending conditions, and maximum stress under torsional, braking, bending, and steering conditions. This forms the basis of an initial performance sample set with "N-dimensional input – 10-dimensional output".

[0037] For ease of subsequent description, the performance response vector corresponding to the i-th group of design variables is represented as:

[0038] In the formula, This represents the frame mass corresponding to the i-th sample group. This represents the fatigue life corresponding to the i-th sample group; This represents the first-order mode frequency corresponding to the i-th sample group; These represent the maximum displacements of the i-th sample under torsional, braking, and bending conditions, respectively. These represent the maximum stresses of the i-th sample under torsional, braking, bending, and steering conditions, respectively.

[0039] The initial performance sample set is represented as follows: In the formula, This represents the initial performance sample set; Represents the i-th group of collaborative design variables; This represents the performance response vector corresponding to the i-th group of collaborative design variables; n represents the number of samples, in this embodiment n=500.

[0040] S2: Building a high-precision proxy model based on MLP To replace the time-consuming finite element simulation calculations, this embodiment constructs a multilayer perceptron (MLP) integrated surrogate model based on the initial performance sample set obtained in step S1. This model is used to establish a high-dimensional nonlinear mapping relationship between N collaborative design variables and 10 performance response quantities.

[0041] Specifically, such as Figure 2 As shown, the N collaborative design variables obtained in step S1 are used as input features, and the 10 performance response quantities are used as output targets. Considering that fatigue life data usually spans multiple orders of magnitude, the fatigue life data is first logarithmically transformed to improve the data distribution; then, the Z-score normalization method is used to normalize all input variables and output response quantities to eliminate the influence of dimensional differences on model training.

[0042] After data preprocessing, the processed sample set is randomly divided into training and test sets at a ratio of 80% and 20%. An MLP network topology is constructed, consisting of an input layer, several hidden layers, and an output layer. The input layer corresponds to N collaborative design variables, the output layer corresponds to the performance response to be predicted, and the hidden layers use the ReLU activation function to enhance the model's ability to express high-dimensional nonlinear mapping relationships.

[0043] Before model training, this embodiment employs a two-stage grid search strategy based on five-fold cross-validation to automatically optimize the hyperparameter space. First, a coarse global search is performed over a broad range to quickly identify the range of dominant parameters. Then, a fine-grained local search with small steps is performed within the range of dominant parameters to balance computational overhead and optimization accuracy. Finally, the L2 regularization penalty coefficient, batch size, number of hidden layer neurons, and initial learning rate corresponding to each performance response are determined.

[0044] Under optimal hyperparameter configuration, the Adam optimization algorithm and mean squared error loss function are used to iteratively update the network parameters. During five-fold cross-validation training, the mean squared error of the validation set is monitored in real time, and an early stopping mechanism with a tolerance of 30 generations is introduced. Through the above rolling training mechanism, five MLP base models are independently trained for each performance response. When faced with unknown samples, the average of the prediction results of the five MLP base models is taken as the final prediction output for that performance response.

[0045] In the formula, This represents the integrated predicted value of the performance response of the k-th item; This represents the k-th performance response corresponding to the th term. Predictions from a base MLP model; Represents the collaborative design variables to be predicted; These correspond to the frame mass, fatigue life, first-order modal frequency, maximum displacement under each working condition, and maximum stress under each working condition, respectively.

[0046] To verify the predictive performance of the constructed MLP ensemble agent model, test set data was input into the final ensemble model, and the determination coefficients of each performance response were calculated. .like Figure 3 As shown in this embodiment, the prediction of 10 performance indicators... All values ​​exceeded 0.90, indicating that the MLP integrated proxy model can meet the accuracy requirements for rapid prediction of performance response in subsequent multi-objective optimization processes, and can significantly reduce the number of repeated calls to finite element simulation during the optimization process.

[0047] S3: Multi-objective optimization based on PPO-enhanced NSGA-III algorithm like Figure 4 As shown, after the MLP integrated agent model is constructed in step S2, the aluminum alloy frame is then subjected to multi-objective collaborative optimization using an enhanced adaptive multi-objective optimization algorithm (NSGA-III algorithm) driven by the proximal policy optimization algorithm (PPO). This algorithm framework maps the adjustment process of the NSGA-III genetic operation parameters to a Markov decision process, utilizing the PPO agent to perceive the real-time evolutionary state of the population and dynamically outputting the genetic operator parameters. This overcomes the problem of the traditional NSGA-III algorithm easily getting trapped in local optima or experiencing slow convergence in complex nonlinear design spaces due to static hyperparameter configuration.

[0048] Specifically, based on the N collaborative design variables extracted in step S1, this embodiment selects fatigue life, frame mass, and maximum displacement under torsional conditions as the three main optimization objectives. Among these, frame mass and maximum displacement under torsional conditions are preferred when they are smaller, while fatigue life is preferred when it is larger. Therefore, fatigue life is converted into... This is then incorporated into the minimization objective function. Simultaneously, nine engineering constraints are set, and the optimization mathematical model is expressed as:

[0049] In the formula, Represents a multi-objective optimization function; This represents the frame mass corresponding to the design variable X; This represents the fatigue life corresponding to the design variable X; This represents the maximum displacement under torsional conditions corresponding to the design variable X; Indicates the maximum displacement under braking conditions; () indicates the maximum displacement under bending conditions; This indicates the maximum stress under torsional conditions; Indicates the maximum stress under braking conditions; Indicates the maximum stress under bending conditions; This indicates the maximum stress under steering conditions; This represents the first-order modal frequency. The unit for the displacement threshold is mm, the unit for the stress threshold is MPa, the unit for the fatigue life threshold is the number of cycles, and the unit for the first-order modal frequency threshold is Hz.

[0050] In each generation of the NSGA-III algorithm, for each candidate individual in the population, the MLP ensemble agent model trained in step S2 is first called to predict its corresponding 10 performance responses, and then the objective function value and constraint violation are calculated according to the optimization mathematical model formula.

[0051] For candidate individuals that meet the constraints, selection is made according to the non-dominated sorting and reference direction association mechanism; for candidate individuals that do not meet the constraints, the feasibility priority rule is adopted, that is, feasible solutions take precedence over infeasible solutions. When all candidate individuals are infeasible, the one with the smaller degree of constraint violation takes precedence.

[0052] To enable the PPO agent to accurately perceive the population's evolutionary state, this embodiment extracts five key features to construct a state vector:

[0053] In the formula, Represents the state vector of generation t; Indicates the progress of evolution, and Where t represents the current generation number, Indicates the maximum number of generations; The diversity of population distribution in the target space is represented by the mean of the standard deviation of the normalized objective function value. The hypervolume index represents the current population and is used to measure the overall quality of the solution set; This indicates the increase in volume, and In the formula, Indicator of the supervolume of the previous generation population; This represents the proportion of non-dominated solutions, that is, the proportion of the number of solutions at the first non-dominated frontier to the total population.

[0054] The state vector is received in the Actor policy network of PPO. Then, output a four-dimensional continuous action vector: ,in .

[0055] To apply the aforementioned dimensionless actions to the genetic operations of NSGA-III, this embodiment designs a linear mapping function to convert the actions output by the PPO agent into genetic operator parameters of NSGA-III:

[0056] in, It is the cross-distribution index; It is the distribution index of variation; The crossover probability; The mutation probability; The number of decision variables.

[0057] Through the above equation, the continuous actions output by the PPO agent are transformed into four core hyperparameters of the NSGA-III genetic operation, enabling the NSGA-III algorithm to dynamically adjust its search strategy based on the current population state. When population diversity is insufficient or the hypervolume increment slows down, the PPO agent can enhance the search perturbation by adjusting the mutation probability and mutation distribution index; when the quality of the solution set continues to improve, the PPO agent can strengthen the information recombination among superior individuals by adjusting the crossover probability and crossover distribution index.

[0058] Furthermore, the reward function is the core driving force guiding the PPO to learn adaptive strategies. To encourage the agent to continuously seek parameter combinations that can propel the Pareto front forward, this embodiment uses hypervolume increments. As a reward signal, a penalty mechanism is introduced for performance stagnation. The immediate reward for generation t is... Designed as follows: ; when When this occurs, it indicates an improvement in population quality, and the agent receives a positive incentive; when... When this occurs, it indicates that evolution has stalled or degenerated. At this point, an additional penalty of -0.1 is applied to force the policy network to adjust its actions, thereby triggering the population to escape the local trap.

[0059] After each generation of evolution, NSGA-III adjusts its actions accordingly. Complete state transition to and the interaction trajectory , , , These trajectories are temporarily stored in a trajectory buffer. Subsequently, the PPO algorithm uses these trajectories to calculate the advantage function and, by maximizing the composite objective function of PPO, periodically updates the weights of the Actor and Critic networks, ultimately achieving adaptive optimal control of the multi-objective evolutionary path.

[0060] The formula for calculating the composite objective function is:

[0061] in For a composite objective function, To replace the objective function with a pruning tool, The value function loss is used to train the Critic network; It is a strategy In state Entropy measures the randomness or uncertainty of a strategy; and It is a hyperparameter used to balance the importance of various loss terms.

[0062] To verify the effectiveness and robustness of the PPO-NSGA-III algorithm proposed in this embodiment in solving complex multi-objective optimization problems, benchmark tests were conducted using standard multi-objective test functions from DTLZ1 to DTLZ7 before applying it to chassis optimization.

[0063] All test problems were set to have three objective dimensions, and the population size was determined to be 190 based on the number of reference directions. To ensure fairness in the comparison, all parameters of PPO-NSGA-III, except for the PPO reinforcement learning control module, were kept consistent with those of the traditional NSGA-III. To eliminate random bias from a single run, each algorithm was independently run 10 times on each test function.

[0064] This embodiment uses three evaluation metrics—GD, IGD, and HV—to quantitatively evaluate the Pareto solution set.

[0065] The GD metric represents the average distance from the solution set to the true Pareto front obtained by the algorithm, and its calculation formula is as follows:

[0066] The IGD metric represents the average distance from the true Pareto front to the solution set obtained by the algorithm, and its calculation formula is as follows:

[0067] The HV index represents the hypervolume of the non-dominated solution set relative to the reference point coverage, and its calculation formula is as follows:

[0068] In the formula, P represents the solution set obtained by the algorithm; Represents the true Pareto frontier; This represents the Euclidean distance between point x and point y. This indicates the number of solutions in the solution set P; S represents the number of reference points in the true Pareto front; S represents the non-dominated solution set. Indicates the number of solutions in the non-dominated solution set; This represents the hypervolume region formed by the reference point and the i-th solution in the non-dominated solution set; This represents the Lebesgue measure.

[0069] Based on the above evaluation metrics, the quantitative comparison results of the two algorithms on the DTLZ test set are shown in Table 1.

[0070] Table 1. Comparison of algorithm performance on the DTLZ problem.

[0071] As shown in Table 1, the PPO-NSGA-III algorithm in this embodiment outperforms the traditional NSGA-III algorithm in terms of GD, IGD, and HV, indicating that the solution set obtained by this algorithm has better convergence accuracy and distribution uniformity. Subsequently, the PPO-NSGA-III algorithm was applied to the multi-objective optimization of the aluminum alloy frame of an electric truck. In this embodiment, 253 Pareto candidate design schemes were finally obtained, and their 3D distribution diagram is shown below. Figure 5 As shown.

[0072] S4: Select the best design scheme based on the CRITIC-TOPSIS algorithm.

[0073] The Pareto solution set obtained in step S3 is used to perform a comprehensive multi-criteria decision-making process, which is as follows: S41: Construct the initial decision matrix and standardize the data.

[0074] Suppose there are n candidate solutions in the Pareto solution set, and each candidate solution contains m evaluation indicators. Construct the initial decision matrix: Where X represents the initial decision matrix, Let represent the value of the i-th scheme on the j-th evaluation index, n represent the number of candidate schemes, and m represent the number of evaluation indexes. In this embodiment, the evaluation indexes include frame mass, fatigue life, first-order modal frequency, maximum displacement under three working conditions, and maximum stress under four working conditions.

[0075] Since different indicators have different dimensions and orders of magnitude, to eliminate the impact of scale differences on decision-making results, the extreme value normalization method is used to perform forward normalization and standardization on the original data, resulting in a standardized decision matrix. For indicators where larger values ​​are considered better, the standardization formula is:

[0076] in The standardized index value, and These are the maximum and minimum values ​​of the j-th index among all possible schemes, respectively.

[0077] S42: Extract objective weights based on the CRITIC algorithm.

[0078] The CRITIC algorithm is used to eliminate subjective bias in manual weighting. First, the variability within each indicator is calculated; a larger standard deviation indicates a stronger ability to distinguish between different schemes. The standard deviation of the j-th indicator is expressed as:

[0079] in This represents the standard deviation of the j-th indicator. Let be the average value of the j-th indicator.

[0080] Subsequently, the conflict between the indicators was calculated using the Pearson correlation coefficient. This reflects the degree of information redundancy between indicators. If two indicators are highly correlated, it indicates that they provide duplicate information. The total information content of the j-th indicator... The calculation is as follows:

[0081] In the formula, This represents the total information content of the j-th indicator; This represents the correlation coefficient between the j-th indicator and the k-th indicator.

[0082] The information content of each indicator is normalized to obtain the final objective weight of the j-th indicator:

[0083] In the formula, This represents the objective weight of the j-th indicator. This weight is determined by both indicator variability and indicator conflict, which can reduce subjective bias caused by manual weighting.

[0084] S43: Calculation of relative proximity and optimization of scheme based on TOPSIS.

[0085] By combining the obtained weights, a weighted standardized decision matrix is ​​constructed: In the formula, It represents the weighted standardized value of the i-th candidate solution on the j-th metric.

[0086] Determine the ideal solution within the weighted space. (Hypothetical solution composed of the optimal values ​​of each indicator) and negative ideal solution (Hypothetical scenario based on the worst values ​​of each indicator): ; ; Calculate the Euclidean distance from the i-th candidate solution to both the positive and negative ideal solutions. and : ; ; In the formula, This represents the distance from the i-th candidate solution to the positive ideal solution; This represents the distance from the i-th candidate solution to the negative ideal solution; This represents the value of the j-th index in the positive ideal solution; This represents the value of the j-th index in the negative ideal solution.

[0087] Finally, the relative similarity of each scheme is calculated. :

[0088] in . The larger the value, the closer the scheme is to the positive ideal solution and the farther it is from the negative ideal solution, and the better its overall performance.

[0089] Based on relative closeness Sort all candidate solutions in descending order and select... The largest candidate solution is selected as the final recommended design. By combining the CRITIC objective weighting method with the TOPSIS ranking method, this embodiment can achieve an objective comprehensive evaluation among conflicting engineering objectives such as frame quality, fatigue life, modal frequency, stiffness, and strength, thereby extracting the best compromise solution with engineering application value from the Pareto solution set.

[0090] After obtaining the final recommended design scheme, the corresponding N-dimensional collaborative design variables are fed back to the SFE parametric model and the Python bolt parametric script to automatically generate the final aluminum alloy frame design model. The Optistruct solver and nCode fatigue analysis software can be called again for verification to confirm that the final design scheme meets the preset engineering constraints in terms of performance indicators such as quality, fatigue life, first-order modal frequency, maximum displacement, and maximum stress.

[0091] Through the above implementation methods, this invention integrates the structural dimensions, cross-sectional shape, and bolt connection parameters of the aluminum alloy frame into a collaborative parametric design space. It utilizes an MLP integrated surrogate model to establish a high-dimensional nonlinear mapping relationship between design variables and structural performance response. Furthermore, a PPO agent dynamically adjusts the crossover distribution index, mutation distribution index, crossover probability, and mutation probability of the NSGA-III algorithm based on the population evolution state. Finally, the CRITIC-TOPSIS joint decision-making method is employed to select the optimal compromise solution from Pareto candidate solutions. This method reduces the computational burden of repetitive finite element simulations, improves the search efficiency and solution quality of multi-objective optimization in a high-dimensional complex frame design space, and achieves collaborative optimization of the aluminum alloy frame for electric trucks in terms of lightweighting, stiffness, strength, fatigue life, and modal performance.

[0092] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An integrated intelligent design method for an aluminum alloy frame of an electric cargo truck, characterized in that, Includes the following steps: S1: Establish a collaborative parameterized finite element model of the aluminum alloy frame of the electric truck, and define the frame structure size variables, cross-sectional shape variables and bolt connection parameters as collaborative design variables. Then, obtain an initial performance sample set composed of the collaborative design variables and the corresponding frame performance response quantities through sampling simulation. Specifically, SFE software is used to parametrically model the shape and structural dimensions of the vehicle frame, and a Python parametric script is used to automatically define bolt connection parameters, including bolt specifications, bolt spacing, bolt quantity, and bolt hole positions. The position of the bolt center hole is determined according to the bolt connection scheme, and temporary nodes are arranged at the corresponding positions for coordinate positioning. The Python parametric script is called to automatically generate bolt holes and corresponding bolt connection models based on the coordinates of the temporary nodes. By setting loop logic in the Python parametric script, the bolt connection parameters are updated synchronously with the changes in the structural dimensions and cross-sectional shape of the vehicle frame. S2: Using the collaborative design variables as input and the chassis performance response as output, construct an MLP integrated proxy model. The MLP integrated proxy model is used to predict the chassis performance response corresponding to the candidate chassis design scheme. S3: Perform multi-objective optimization based on the PPO-enhanced NSGA-III algorithm. Utilize the MLP ensemble surrogate model to predict the performance of candidate individuals in the NSGA-III population and extract the population evolutionary state, including evolutionary progress, population distribution diversity, hypervolume index, hypervolume increment, and the proportion of non-dominated solutions. Input the population evolutionary state into the PPO agent. The PPO agent outputs action vectors, which are then mapped to genetic operator parameters of the NSGA-III algorithm. Use the hypervolume increment as a reward signal and apply an evolutionary stagnation penalty when the hypervolume does not increase to update the PPO agent and obtain the Pareto candidate design scheme set. In step S3, the state vector constructed by the PPO agent based on the population evolution state is: in, Represents the state vector of generation t; Indicates the progress of evolution. This indicates the diversity of population distribution in the target space; The overvolume index represents the current population size. Indicates the volume increment; Indicates the proportion of non-dominated solutions; in, t represents the current generation. Indicates the maximum number of generations; , Indicator of the supervolume of the previous generation population; The Actor policy network of the PPO agent is based on the state vector Output a four-dimensional continuous action vector: ,in, All are continuous action values ​​located in the interval [0,1]. The four-dimensional continuous action vector is converted into genetic operator parameters for the NSGA-III algorithm through the following mapping relationship: in, It is the cross-distribution index; It is the distribution index of variation; The crossover probability; Probability of mutation; The number of decision variables; The supervolume increment of the (t+1)th generation population relative to the tth generation population As a reward signal, and with the introduction of an evolutionary stagnation penalty, the t-th generation receives an immediate reward. for: After each generation of evolution, the interaction trajectory will be... , , , The data is stored in the trajectory buffer, and the Actor network and Critic network of the PPO agent are periodically updated based on the interaction trajectory. S4: The Pareto candidate design scheme set is comprehensively evaluated and ranked based on the CRITIC-TOPSIS algorithm, and the candidate design scheme with the highest relative similarity is selected as the final recommended design scheme for the aluminum alloy frame.

2. The integrated intelligent design method for the aluminum alloy frame of the electric cargo truck according to claim 1, characterized in that, The collaborative design variables constitute a multidimensional design variable vector, including frame structure size variables, cross-sectional shape variables, and bolt connection parameters; The frame structure dimensional variables include the thickness, length, width, height, or installation position parameters of the longitudinal beams, crossbeams, connecting plates, or reinforcing plates; the cross-sectional shape variables include the cross-sectional height, cross-sectional width, fillet radius, or reinforcing rib parameters of the longitudinal beams or crossbeams; the bolt connection parameters include the bolt specifications, bolt quantity, bolt spacing, bolt edge distance, or bolt arrangement position parameters for different connection areas.

3. The integrated intelligent design method for the aluminum alloy frame of the electric cargo truck according to claim 1, characterized in that, In step S1, the Hamsley sampling method is used to uniformly sample within the variable space formed by the collaborative design variables, and batch simulation is performed by driving the Optistruct solver and nCode fatigue analysis software through an automated script to obtain 10 frame performance response quantities corresponding to each group of collaborative design variables. The 10 vehicle frame performance response parameters include vehicle frame mass, fatigue life, first-order modal frequency, maximum displacement under torsional conditions, maximum displacement under braking conditions, maximum displacement under bending conditions, maximum stress under torsional conditions, maximum stress under braking conditions, maximum stress under bending conditions, and maximum stress under steering conditions.

4. The integrated intelligent design method for the aluminum alloy frame of the electric cargo truck according to claim 3, characterized in that, In step S2, constructing the MLP integrated agent model includes: Logarithmic transformation was performed on the fatigue life data, and the input variables and output response were standardized. The sample set is divided into a training set and a test set, and an MLP network topology containing an input layer, several hidden layers and an output layer is constructed. A two-stage grid search strategy based on five-fold cross-validation is used to determine the hyperparameters of the MLP model. The hyperparameters include the L2 regularization penalty coefficient, batch size, number of hidden layer neurons, and initial learning rate. Under optimal hyperparameter configuration, the Adam optimization algorithm and mean squared error loss function are used to train the MLP model, and an early stopping mechanism based on validation set error is introduced. For each frame performance response quantity, multiple MLP base models are trained independently, and the average of the prediction results of the multiple MLP base models is used as the integrated prediction output of the corresponding frame performance response quantity.

5. The integrated intelligent design method for the aluminum alloy frame of the electric cargo truck according to claim 3, characterized in that, In step S3, the frame mass, maximum displacement under torsional conditions, and fatigue life are taken as the main optimization objectives, and corresponding engineering constraints are set to establish a multi-objective optimization mathematical model. Among them, the frame mass and maximum displacement under torsional conditions are indicators that are better the smaller they are, while the fatigue life is an indicator that is better the larger it is.

6. The integrated intelligent design method for the aluminum alloy frame of the electric cargo truck according to claim 1, characterized in that, In step S4, the selection of the optimal design scheme based on the CRITIC-TOPSIS algorithm includes: An initial decision matrix was constructed using 10 chassis performance response parameters corresponding to the Pareto candidate design schemes; For fatigue life and first-order modal frequency, positive index standardization is used; for frame mass, maximum displacement under each working condition, and maximum stress under each working condition, inverse index standardization is used. Based on the CRITIC algorithm, objective weights are calculated according to the standard deviation of each evaluation indicator and the correlation coefficient between indicators. A weighted standardized decision matrix is ​​constructed based on the objective weights to determine the positive and negative ideal solutions, and the relative closeness of each Pareto candidate design is calculated. The Pareto candidate designs were sorted from highest to lowest relative similarity, and the candidate design with the highest relative similarity was selected as the final recommended design.

Images