A vehicle crashworthiness multi-objective optimization method based on hyper-computing assistance

Abstract

The application discloses a kind of based on super-dimensional calculation auxiliary vehicle crashworthiness multi-objective optimization method, first all the vehicle structures that have been evaluated are evenly divided into the same scale " high-quality structure set " and " poor structure set ", provide balanced training data for constructing classification model.Using a kind of approximate structure thickness value coding mode, all structures are encoded into corresponding hyper vector. These hyper vectors are added according to the category of the bit to construct classification model, obtain the category hyper vector of " high-quality structure " and " poor structure ".Finally, the candidate structure newly generated using genetic operator is encoded into hyper vector in the same way, and the similarity with category hyper vector is calculated using cosine similarity to predict the category of candidate structure. The candidate structure whose predicted category is " high-quality structure " is screened out and evaluated by real objective function. The method has better effect when solving standard test problem set and vehicle crashworthiness optimization problem.

Description

Technical Field

[0001] This invention relates to the field of evolutionary algorithms, and proposes an evolutionary optimization algorithm based on hyperdimensional computational assistance for multi-objective optimization problems such as vehicle crashworthiness optimization, which have high computational costs. Background Technology

[0002] Crashworthiness is a critical yet demanding design requirement when developing high-quality, low-cost industrial products that involve potential impacts. In the automotive industry, advancements in high-performance computing and advanced numerical algorithms have made it possible to simulate comprehensive laboratory crash events. To date, nonlinear explicit finite element analysis has been widely used in vehicle design to meet various safety guidelines.

[0003] However, in the context of material and geometric nonlinearity and frictional contact dynamics, the implicit relationships and enormous complexity of sensitivity analysis can significantly impair the feasibility of practical applications of general mathematical programming techniques. Therefore, surrogate modeling methods have been widely applied in automotive design, attracting considerable interest from the industry. For multi-objective optimization problems modeled from complex practical applications such as vehicle structural design, function evaluation requires expensive computational simulations, consuming significant time and material resources. Such problems are often referred to as expensive multi-objective optimization problems.

[0004] Surrogate-assisted evolutionary algorithms are widely used to solve such expensive multi-objective optimization problems. They employ computationally efficient surrogate models to replace the truly expensive functions. In the last decade, many surrogate-assisted evolutionary algorithms have been proposed, such as CSEA, K-RVEA, and EDN-ARMOEA. These algorithms utilize different machine learning techniques to construct surrogate models, pre-select promising solutions, and then evaluate them using the real, expensive functions, thereby saving computational costs.

[0005] Some existing surrogate-assisted evolutionary algorithms use regression models as surrogate models to directly approximate the target value and find promising candidate solutions. Gaussian processes and neural networks are the most commonly used regression models. However, in expensive optimization processes with only a few iterations, limited training samples can hinder regression models from achieving sufficient accuracy. In recent years, methods using classification models to directly predict whether candidate solutions can be evaluated by the true function have gained increasing attention. These methods train surrogate models for simple classification tasks, enabling them to achieve high prediction accuracy with limited training samples. Neural networks and support vector machines are frequently used classification models.

[0006] Traditional machine learning models often have high computational complexity during training, leading to inefficient problem-solving. For example, the computational complexity of a Gaussian process is the cube of the number of training data points. Furthermore, the Gaussian model must be rebuilt after new training data is generated, requiring a very complex model-building computation process in each iteration. The computational complexity of training a neural network is the number of iterations required multiplied by the batch size, then multiplied by the power of the number of hidden neurons in each layer and the number of layers. Although the model can be updated by continuing to train the neural network after new data is generated, increasing the number of layers leads to an exponential increase in the computational complexity of the neural network model. The computational complexity of training a support vector machine (SVM) classification model is between the square and the cube of the number of training data points. Although it has lower computational complexity compared to Gaussian processes and neural networks, it still relies on complex kernel functions for computation during training. Therefore, to improve the efficiency of the algorithm, it is necessary to use simpler and faster computational processes to build classification proxy models.

[0007] Hyperdimensional computing is a lightweight, neuromorphic computing technique with highly scalable learning mechanisms. This technique can achieve higher efficiency and even better performance than traditional machine learning models in classification tasks across various domains, and can solve classification tasks using only simple computational methods. To solve expensive multi-objective optimization problems more efficiently and accurately, this invention, for the first time, introduces hyperdimensional computing as an auxiliary classification surrogate model into solving this costly multi-objective optimization problem related to vehicle crashworthiness. Summary of the Invention

[0008] The purpose of this invention is to provide a new method for solving the expensive multi-objective optimization problem of vehicle crashworthiness. The key to solving this type of multi-objective optimization problem is that the algorithm needs to construct an efficient and accurate surrogate model and use the surrogate model to pre-select high-quality vehicle structures for evaluation of the real function.

[0009] To achieve the above objectives, this invention proposes a multi-objective evolutionary optimization method for vehicle crashworthiness based on hyperdimensional computational assistance, characterized in that:

[0010] 1. A costly multi-objective evolutionary optimization method for vehicle crashworthiness based on hyperdimensional computational assistance, characterized by comprising the following steps:

[0011] 1) Set the initial number of vehicle structures N, and the number of evaluation iterations (FES) for the vehicle damage optimization objective function;

[0012] 2) Randomly generate an initial vehicle structure set P = {x1, x2, ..., x...} NEach structure consists of five decision components, each representing the thickness of the five reinforcing members surrounding the frontal structure that significantly impacts collision safety. The entire vehicle structure set is then evaluated using a realistic objective function. Considering lightweighting, the overall vehicle mass is chosen as the first objective; considering the worst-case scenario of occupant biomechanical injury caused by acceleration, the integral of the frontal full-scale collision acceleration is selected as the second objective; finally, considering the most severe mechanical damage, toe plate intrusion in an offset-frontal collision is chosen as the third objective. The initial vehicle structure set is then assigned to structure archive set A.

[0013] 3) Divide all structures in the structure archive A into a "high-quality structure set" and a "low-quality structure set": First, apply Pareto dominance to the three objective values ​​mentioned above to divide all structures in the structure archive into several clusters, where structures in each cluster are non-dominated. Then, calculate the crowding distance for each structure in each cluster. Rank all structures in the structure archive according to Pareto rank and crowding distance. Assign the top half of the ranked structures to the "high-quality structure set" and the bottom half to the "low-quality structure set".

[0014] 4) Perform regularization on all component thickness values; for a structure x = {x1, x2, ..., x5}, regularize the thickness of each component to between 0 and 1 using formula (1).

[0015]

[0016] x max and x min These are the upper and lower limits of the component thickness domain.

[0017] 5) Initialization of hypervectors representing vehicle reinforcement components: At the start of each hyperdimensional computational model construction, five h=10000-dimensional hypervectors B={B1,B2,…,B5} are randomly initialized to represent the five reinforcement components of the vehicle's frontal structure, respectively. Each component is B i ={b1,b2,…,b h}, b i ∈{0,1}.

[0018] 6) Initialization of the hypervector representing the thickness value of the vehicle reinforcement component: First, initialize the hypervector L1 = {l1, l2, ..., l...} representing the minimum thickness value. h Random initialization, l i ∈{0,1}. Then, h / 10 random positions are selected each time for inversion to generate a larger-order thickness value hypervector L. i+1 This generation process ensures that the supervector L representing the maximum thickness value is... 10The hypervector representing the minimum thickness value, L1, is the least similar (approximately orthogonal), while the hypervectors representing other thickness values ​​are relatively similar.

[0019] 7) Generate category hypervectors representing “high-quality structure” and “low-quality structure”: For each vehicle reinforcement component in a structure, select the corresponding thickness value hypervector with the closest thickness value. The encoding of each structure is achieved by binding each hypervector representing the vehicle reinforcement component with the corresponding thickness value hypervector bitwise AND-OR and adding them, as shown in Equation (2).

[0020]

[0021] This represents an AND / OR operation. Then, all encoded structures belonging to the "high-quality structure" category are summed bitwise to obtain the hypervector H representing the corresponding "high-quality structure" category. n The hypervector H for the "inferior structure" category is obtained. d The process is the same as the one above.

[0022] 8) Candidate structure generation: After obtaining the hypervectors representing the categories, new candidate structures are generated using genetic operators, and these offspring candidate structures are encoded into hypervectors in the same way.

[0023] 9) Pre-selection based on a hyperdimensional computational classification model: The generated candidate structures are compared with two hypervectors representing categories using a cosine similarity test to obtain the category labels of the candidate structures. High-quality candidate structures (predicted as "high-quality structures") are retained, while low-quality candidate structures (predicted as "low-quality structures") are removed.

[0024] 10) Structure archive set update: Finally, a set of child structures that are predicted to be excellent are obtained. They are evaluated using the true objective function and put into the structure archive set A.

[0025] 11) Determine if the termination condition is met, i.e., whether the number of times the structure is evaluated reaches FES. If it is met, output all non-dominated structures in the structure archive set A; otherwise, return to step 3.

[0026] In step 3), when partitioning all structures in the structure archive, the three optimization objective values ​​of each structure are uniformly divided using Pareto dominance sorting. This is because Pareto dominance is a suitable choice for partitioning training set labels for multi-objective optimization problems. Since the number of evaluations is small, each evaluated structure should be used to train the classification model. Furthermore, to improve the model's classification performance, the partitioning operation ensures the balance of training data labels.

[0027] In step 6), the initialization of the hypervector operation representing the thickness value of the vehicle reinforcement component is to ensure that the encoded hypervector retains the similarity of the original data. For the encoding of thickness values, since the thickness values ​​of the candidate structures are continuous, it's not feasible to directly generate hypervectors randomly for an infinite number of thickness values. Therefore, we consider generating hypervectors randomly only for a finite number of thickness values. Here, we consider setting ten hypervectors representing thickness values ​​within the domain. The actual thickness value is represented by the hypervector corresponding to the predefined thickness value. Furthermore, thickness values ​​differ from vehicle component indices; their magnitudes are correlated, so the similarity of their corresponding hypervectors should also differ.

[0028] Compared with existing methods, the present invention has the following significant advantages and innovations:

[0029] This invention proposes a multi-objective evolutionary optimization method for vehicle crashworthiness based on hyperdimensional computational assistance. The method has the following characteristics: 1) This paper proposes an expensive multi-objective evolutionary optimization method based on hyperdimensional computational assistance, in which a surrogate model based on hyperdimensional computation is carefully designed for the first time, enabling efficient and rapid screening of promising candidate structures. 2) A hypervector encoding method that approximates the component thickness value is specifically adopted, which maintains the similarity between the original structures after encoding the structures into hypervectors. 3) The proposed algorithm is empirically compared with five state-of-the-art surrogate-assisted evolutionary algorithms on 17 benchmark problems in the DTLZ and WFG test problem sets. The comparison results show that the proposed algorithm has faster speed and competitive performance in solving the expensive multi-objective optimization problem of vehicle crashworthiness. Attached Figure Description

[0030] Figure 1 Schematic diagram of a multi-objective optimization method for vehicle crashworthiness based on hyperdimensional computational assistance

[0031] Figure 2 Schematic diagram of approximate variable value encoding method Detailed Implementation

[0032] To further explain the technical solution of the present invention, the present invention will be described in detail below through specific embodiments.

[0033] like Figure 1 As shown in the diagram, this invention provides a multi-objective evolutionary optimization method for expensive vehicle structures based on hyperdimensional computation assistance. The algorithm HDCMOEA of this invention uses hyperdimensional computation as a surrogate model to solve the multi-objective optimization of vehicle structure crashworthiness. The optimization process is carried out using three main operations: training set partitioning, hyperdimensional computation model construction, and pre-screening of high-quality candidate structures.

[0034] The performance of the HDCMOEA algorithm was tested using 17 test functions from DTLZ and WFG, and compared with five of the most representative algorithms, including EDNARMOEA, KRVEA, CSEA, MOEADEGO, and HeEMOEA. Each algorithm was run independently 21 times in each test function, and its average IGD and single-run time were calculated.

[0035] This invention relates to a multi-objective optimization algorithm for vehicle crashworthiness based on hyperdimensional computation, as detailed below:

[0036] To achieve the above objectives, this invention proposes a multi-objective evolutionary optimization method for vehicle crashworthiness based on hyperdimensional computational assistance, characterized in that:

[0037] 2. A costly multi-objective evolutionary optimization method for vehicle crashworthiness based on hyperdimensional computational assistance, characterized by comprising the following steps:

[0038] 1) Set the initial number of vehicle structures N, and the number of evaluation iterations (FES) for the vehicle damage optimization objective function;

[0039] 2) Randomly generate an initial vehicle structure set P = {x1, x2, ..., x...} N Each structure consists of five decision components, each representing the thickness of the five reinforcing members surrounding the frontal structure that significantly impacts collision safety. The entire vehicle structure set is then evaluated using a realistic objective function. Considering lightweighting, the overall vehicle mass is chosen as the first objective; considering the worst-case scenario of occupant biomechanical injury caused by acceleration, the integral of the frontal full-scale collision acceleration is selected as the second objective; finally, considering the most severe mechanical damage, toe plate intrusion in an offset-frontal collision is chosen as the third objective. The initial vehicle structure set is then assigned to structure archive set A.

[0040] 3) Divide all structures in the structure archive A into a "high-quality structure set" and a "low-quality structure set": First, apply Pareto dominance to the three objective values ​​mentioned above to divide all structures in the structure archive into several clusters, where structures in each cluster are non-dominated. Then, calculate the crowding distance for each structure in each cluster. Rank all structures in the structure archive according to Pareto rank and crowding distance. Assign the top half of the ranked structures to the "high-quality structure set" and the bottom half to the "low-quality structure set".

[0041] 4) Perform regularization on the thickness values ​​of all components; for a structure x = {x1, x2, ..., x5}, regularize the thickness of each component to between 0 and 1 using formula (1).

[0042]

[0043] x max and x min These are the upper and lower limits of the component thickness domain.

[0044] 5) Initialization of hypervectors representing vehicle reinforcement components: At the start of each hyperdimensional computational model construction, five h=10000-dimensional hypervectors B={B1,B2,…,B5} are randomly initialized to represent the five reinforcement components of the vehicle's frontal structure, respectively. Each component is B i ={b1,b2,…,b h}, b i ∈{0,1}.

[0045] 6) Initialization of the hypervector representing the thickness value of the vehicle reinforcement component: First, initialize the hypervector L1 = {l1, l2, ..., l...} representing the minimum thickness value. h Random initialization, l i ∈{0,1}. Then, h / 10 random positions are selected each time for inversion to generate a larger-order thickness value hypervector L. i+1 This generation process ensures that the supervector L representing the maximum thickness value is... 10 The hypervector representing the minimum thickness value, L1, is the least similar (approximately orthogonal), while the hypervectors representing other thickness values ​​are relatively similar.

[0046] 7) Generate category hypervectors representing “high-quality structure” and “low-quality structure”: For each vehicle reinforcement component in a structure, select the corresponding thickness value hypervector with the closest thickness value. The encoding of each structure is achieved by binding each hypervector representing the vehicle reinforcement component with the corresponding thickness value hypervector bitwise AND-OR and adding them, as shown in Equation (2).

[0047]

[0048] This represents an AND / OR operation. Then, all encoded structures belonging to the "high-quality structure" category are summed bitwise to obtain the hypervector H representing the corresponding "high-quality structure" category. n The hypervector H for the "inferior structure" category is obtained. d The process is the same as the one above.

[0049] 8) Candidate structure generation: After obtaining the hypervectors representing the categories, new candidate structures are generated using genetic operators, and these offspring candidate structures are encoded into hypervectors in the same way.

[0050] 9) Pre-selection based on a hyperdimensional computational classification model: The generated candidate structures are compared with two hypervectors representing categories using a cosine similarity test to obtain the category labels of the candidate structures. High-quality candidate structures (predicted as "high-quality structures") are retained, while low-quality candidate structures (predicted as "low-quality structures") are removed.

[0051] 10) Structure archive set update: Finally, a set of child structures that are predicted to be excellent are obtained. They are evaluated using the true objective function and put into the structure archive set A.

[0052] 11) Determine if the termination condition is met, i.e., whether the number of times the structure is evaluated reaches FES. If it is met, output all non-dominated structures in the structure archive set A; otherwise, return to step 3.

[0053] Tables 1 and 2 compare the performance of the algorithm of this invention with that of five other algorithms in the target space on the DTLZ and WFG test functions, as well as the time required for one run (in seconds). It can be seen that HDCMOEA outperforms the compared surrogate-assisted evolutionary algorithms in both performance and runtime.

[0054] Table 3 compares the HV (Hardness Value) index of the proposed algorithm with that of six classical evolutionary algorithms in the target space for solving practical vehicle crashworthiness problems. This multi-objective optimization problem for vehicle crashworthiness aims to optimize the crashworthiness of the front structure of a vehicle, involving five decision variables and three objectives. The decision variables include the thickness of five reinforcing members surrounding the front structure. Simultaneously, the vehicle mass, deceleration in a full-frontal collision (proportional to the biomechanical damage to the occupants), and toe plate intrusion in a non-full-frontal collision (considering the vehicle's structural integrity) are taken as objectives and minimized. It can be seen that the proposed algorithm has the best performance in solving practical vehicle crashworthiness problems.

[0055] Table 1 compares the IGD index of the proposed method in the target space with other algorithms.

[0056]

[0057] Table 2 compares the time required for one run of the method proposed in this invention with other algorithms.

[0058]

[0059]

[0060] Table 3. HV index of the proposed method and other algorithms in the vehicle crashworthiness optimization problem.

[0061]

Claims

1. A costly multi-objective evolutionary optimization method for vehicle crashworthiness based on hyperdimensional computational assistance, characterized in that, Includes the following steps: 1) Set the initial number of vehicle structures N, and the number of evaluation iterations (FES) for the vehicle damage optimization objective function; 2) Randomly generate an initial vehicle structure set P={x1,x2,…,x N Each structure consists of five decision components, each representing the thickness of the five reinforcing members surrounding the frontal structure that significantly impacts collision safety. The entire vehicle structure set is then evaluated using a realistic objective function. Considering lightweighting, the overall vehicle mass is chosen as the primary objective. Taking into account the worst-case scenario of occupant biomechanical injury caused by acceleration, the integral of the frontal full-scale collision acceleration is selected as the second objective. Finally, considering the most severe mechanical damage, toe plate intrusion in an offset-frontal collision is chosen as the third objective. The initial vehicle structure set is then assigned to structure archive set A. 3) Divide all structures in the structure archive set A into "high-quality structure set" and "low-quality structure set": First, apply Pareto dominance to the values ​​of the three objectives mentioned above to divide all structures in the structure archive set into several clusters, where structures in each cluster are non-dominated to each other; then, calculate the crowding distance for each structure in each cluster; rank all structures in the structure archive set according to Pareto rank and crowding distance; divide the top half of the structures into the "high-quality structure set" and the bottom half into the "low-quality structure set"; 4) Perform regularization on the thickness values ​​of all components; for a structure The thickness of each component is regularized to between 0 and 1 using formula (1); (1) and These are the upper and lower limits of the component thickness domain, respectively; 5) Initialization of hypervectors representing vehicle reinforcement components: At the start of each hyperdimensional computational model construction, five h=10000-dimensional hypervectors representing the five reinforcement components of the vehicle's frontal structure are randomly initialized. Each of the components is , {0,1}; 6) Initialization of the hypervector representing the thickness value of the vehicle reinforcement component: First, initialize the hypervector representing the minimum thickness value. Random initialization, 𝜖{0,1}; then each time, h / 10 random positions are selected for inversion to generate a hypervector with a larger thickness value. ; This generation process ensures that the supervector representing the maximum thickness value is... hypervector with minimum thickness value The most dissimilar hypervectors are those representing thickness values, while those representing other thickness values ​​are relatively similar. 7) Generate category hypervectors representing "high-quality structure" and "low-quality structure": For each vehicle reinforcement component in a structure, select the corresponding thickness value hypervector with the closest thickness value; the encoding of each structure is achieved by binding each hypervector representing the vehicle reinforcement component with the corresponding thickness value hypervector bitwise AND-OR and adding them, as shown in formula (2). (2) This represents an AND / OR operation; then, all encoded structures belonging to the "high-quality structure" category are added bitwise to obtain a hypervector representing the corresponding "high-quality structure" category. ; Obtain the hypervector of the "inferior structure" category. The process is the same as the one above; 8) Candidate structure generation: After obtaining the hypervectors representing the categories, new candidate structures are generated using genetic operators, and these offspring candidate structures are encoded into hypervectors in the same way. 9) Pre-selection based on hyperdimensional computation classification model: The generated candidate structure is compared with two hypervectors representing categories using cosine similarity test to obtain the category label of the candidate structure; high-quality candidate structures, which are predicted as "high-quality structures", are retained, and low-quality candidate structures, which are predicted as "low-quality structures", are removed. 10) Structure archive set update: Finally, a set of predicted excellent child structures is obtained, evaluated using the true objective function, and put into the structure archive set A; 11) Determine if the termination condition is met, i.e., whether the number of times the structure is evaluated reaches FES. If it is met, output all non-dominated structures in the structure archive set A; otherwise, return to step 3.

Images