A machine learning-based intelligent design method and system for a vehicle connecting plate

Abstract

The application relates to the technical field of new energy automobile structure design, and discloses a vehicle connecting plate intelligent design method and system based on machine learning, wherein the vehicle connecting plate intelligent design method based on machine learning comprises the following steps: collecting vehicle connecting plate historical design data, performing data preprocessing and multi-source heterogeneous feature fusion; obtaining a design demand condition vector of a user, combining a multi-source heterogeneous data unified representation vector, and performing initial configuration generation and topology optimization coupling; extracting design parameters, constructing an adaptive multi-objective reward function to optimize the design parameters; constructing a proxy model, performing proxy model training and online updating; performing multi-working-condition performance prediction and evaluation on the service process of the vehicle connecting plate; performing integrated development and visual implementation to obtain a vehicle connecting plate intelligent design scheme; and the application performs integrated prediction on multiple working conditions and quantitative analysis on the coupling relationship between working conditions, and predicts the comprehensive performance of the connecting plate in a complex service environment.

Description

Technical Field

[0001] This invention relates to the field of new energy vehicle structural design technology, and more specifically, to a machine learning-based intelligent design method and system for automotive connecting plates. Background Technology

[0002] With the rapid development of the new energy vehicle industry, the chassis structure design of electric vehicles is facing unprecedented challenges. Traditional fuel vehicle chassis mainly support the powertrain, such as the engine and transmission, while electric vehicles need to support the heavier and more complex power battery pack. This places higher demands on the load-bearing capacity, stiffness distribution, and fatigue life of the chassis subframe connecting plates. Existing connecting plate designs mainly rely on engineers' experience and repeated finite element simulation verification, resulting in long design cycles, high costs, and difficulty in finding the optimal balance between multiple conflicting objectives such as lightweighting, strength, stiffness, and manufacturing cost.

[0003] The core technical problem with traditional design methods lies in the lack of an effective multi-source heterogeneous data fusion mechanism and intelligent design decision-making capabilities. Connector plate design involves multi-dimensional information such as geometric topology, material properties, multi-condition loads, and manufacturing processes. This heterogeneous data is difficult to represent and utilize in a unified manner, leading to severe information silos during the design process and hindering the full inheritance and reuse of historical design experience. Connector plates need to simultaneously meet multiple operating condition requirements, including static load-bearing, frontal collision, side collision, and fatigue vibration. Complex coupling relationships exist between different operating conditions, and traditional single-condition optimization methods are prone to overlooking some aspects, failing to achieve truly globally optimal design.

[0004] Furthermore, existing design processes heavily rely on finite element simulation for performance evaluation, with each simulation taking several hours, resulting in extremely low efficiency in design space exploration and making it difficult to support large-scale parameter optimization and configuration generation. At the same time, traditional methods lack effective consideration of manufacturing process constraints, often leading to design solutions encountering process feasibility issues in subsequent manufacturing stages, requiring repeated design modifications and further extending the product development cycle.

[0005] Therefore, there is an urgent need for an intelligent design method for automotive connection plates that can integrate multi-source heterogeneous data, achieve intelligent configuration generation, support multi-objective collaborative optimization, and embed manufacturing process constraints. Summary of the Invention

[0006] This invention provides a machine learning-based intelligent design method and system for automotive connecting plates, solving the technical problems of relying on experience, long design cycles, difficulty in multi-objective optimization, and insufficient consideration of manufacturing constraints in related technologies.

[0007] This invention provides a machine learning-based intelligent design method for automotive connection plates, comprising the following steps: Historical design data of automotive connection plates are collected, and data preprocessing and multi-source heterogeneous feature fusion are performed to obtain a unified representation vector of multi-source heterogeneous data. Obtain the user's design requirement condition vector, combine it with the unified representation vector of multi-source heterogeneous data, perform initial configuration generation and topology optimization coupling, and obtain initial geometric configuration candidate schemes; Design parameters are extracted based on the initial geometric configuration candidate schemes, and an adaptive multi-objective reward function is constructed to optimize the design parameters, thereby obtaining the Pareto front solution set; A proxy model is constructed based on a unified representation vector of multi-source heterogeneous data. The proxy model is trained and updated online to obtain an adaptively optimized proxy model. Based on an adaptive optimization surrogate model, the performance of automotive connecting plates under multiple operating conditions is predicted and evaluated, and quantitative analysis results of the coupling relationship between operating conditions are obtained. Based on the unified representation vector of multi-source heterogeneous data, candidate schemes of initial geometric configuration, Pareto front solution set, surrogate model of adaptive optimization, and quantitative analysis results of coupling relationship between working conditions, an integrated development and visualization implementation is carried out to obtain the intelligent design scheme of automotive connection plate.

[0008] In a preferred embodiment, the data preprocessing includes: The original dataset is constructed by collecting historical design data of automotive connection plates. A data integrity detection algorithm is used to scan each data field of each sample in the original dataset one by one, and the proportion of missing values, data type consistency and value range rationality are statistically analyzed to obtain a data quality assessment report. For missing fatigue test data, a collaborative filtering method based on similar samples is used for intelligent filling. By calculating the cosine similarity between the sample to be filled and other complete samples in terms of geometric features and material properties, the neighboring sample with the highest similarity is selected, and the weighted average of its fatigue test data is taken as the filling value. The isolated forest algorithm is used for outlier detection. The performance index of each sample is used to form a feature vector. Multiple isolated trees are constructed by randomly dividing the feature space. The average path length required for each sample to be isolated is calculated, and samples that are incorrectly identified are removed.

[0009] In a preferred embodiment, the multi-source heterogeneous feature fusion includes: A mesh simplification algorithm is used to convert the 3D geometric model into a graph structure representation. Feature points of the geometric model are extracted as graph nodes. Edges are established between connected feature points according to geometric topology. Each node is assigned spatial coordinates, local curvature and the number of adjacent edges. Graph convolutional neural networks are used for node feature learning. Multi-layer graph convolutional layers are constructed so that the final embedding of each node contains topological features within a multi-order neighborhood range. Global average pooling is used to average the embeddings of all nodes to obtain a global topological feature vector. A multimodal feature fusion network is used for high-dimensional unified representation learning. The topological feature vector, mechanical response feature vector and structured feature vector are nonlinearly transformed by independent fully connected layers. The correlation weights between modes are calculated by a self-attention layer. The weighted modal features are then concatenated to obtain a unified representation vector of multi-source heterogeneous data.

[0010] In a preferred embodiment, the initial configuration generation and topology optimization coupling includes: The user's design requirement condition vector is obtained, and a generative adversarial network is constructed to generate the initial configuration. It includes two sub-networks: a generator and a discriminator. The input of the generator is a concatenation of a unified representation vector of multi-source heterogeneous data, the user's input design requirement condition vector, and a random noise vector. A multi-layer deconvolutional neural network is used to map low-dimensional vectors to high-dimensional topological density fields. The output value at each position is mapped to a preset range to represent the material density through multiple deconvolutional blocks and an activation function. Based on the thickness range of the connecting plate specified in the user's design requirements, the density map is stretched vertically to form a three-dimensional voxel model. The moving cube algorithm is used to extract the isosurface of the density threshold to generate a preliminary three-dimensional mesh model.

[0011] In a preferred embodiment, the generative adversarial network further includes: A discriminator network is constructed to evaluate the authenticity and physical rationality of the initial configuration. The discriminator adopts a dual-branch structure, with the first branch being the authenticity discrimination branch and the second branch being the physical constraint discrimination branch. The two branches share the first few layers of convolutional feature extraction layers. In the physical constraint discrimination branch, the compliance minimization criterion of topology optimization is introduced as a physical constraint to perform fast finite element analysis on the input three-dimensional mesh model. Under given load and boundary conditions, the displacement field is solved, the total compliance value of the structure is calculated, and the stiffness score is obtained after normalizing the compliance value. An alternating training strategy is adopted to optimize the generator and discriminator. In each training iteration, the generator parameters are fixed and the discriminator is trained for several steps, while the discriminator parameters are fixed and the generator is trained, so that the generator and discriminator reach Nash equilibrium.

[0012] In a preferred embodiment, the step of constructing an adaptive multi-objective reward function to optimize the design parameters includes: The design parameter optimization problem of the connecting plate is modeled as a Markov decision process, and the state space is defined to include two parts: the parameter state and the performance state of the current design. A dual-deep Q-network is constructed as a value function approximator, consisting of two neural networks with identical structures: a current network and a target network. The current network is used to select actions, and the target network is used to calculate the target Q-value. An adaptive multi-objective reward function is designed, which comprehensively considers the lightweight objective, strength constraints, stiffness constraints, cost constraints, and manufacturing process constraints, and dynamically adjusts the weights of each objective based on the position of the current solution in the objective space.

[0013] In a preferred embodiment, the adaptive multi-objective reward function includes: Define the normalized reward components for each single objective: the lightweight reward is defined as the reduction ratio of the current mass to the initial mass, the strength reward is defined as a normalized function of the safety factor, the stiffness reward is defined as a normalized function of the maximum displacement, and the cost reward is defined as the percentage of manufacturing cost savings relative to the budget. Increase the reward weight of manufacturing process constraints, conduct manufacturing feasibility assessment of the current design scheme, stamping constraints are achieved by detecting the draft angle of the connecting plate surface, and welding assembly constraints are achieved by detecting the interface matching degree between the connecting plate and adjacent parts; Design adaptive weighting coefficients to dynamically adjust the weight of each objective based on the current solution's position in the objective space.

[0014] In a preferred embodiment, the construction of the proxy model, the training of the proxy model, and the online updating include: The Latin hypercube sampling method was used to select representative sample points. For each sample point, a three-dimensional model of the connecting plate was constructed based on the parameter values ​​of each sample point. The model was then imported into finite element analysis software for static structural analysis and modal analysis. A deep feedforward neural network is constructed as a surrogate model, with design parameters as input and performance index vector as output; An uncertainty estimation mechanism is introduced, and the standard deviation of the prediction results is calculated as an uncertainty index. When the uncertainty exceeds a preset threshold, finite element simulation is triggered for evaluation.

[0015] In a preferred embodiment, the process of training and updating the proxy model online further includes: When a new sample is obtained by triggering a finite element simulation, the new sample is added to the training set of the surrogate model, and the surrogate model is retrained periodically. The update process adopts an incremental learning strategy, which fine-tunes the original model parameters using new samples. The learning rate is set to a preset proportion of the initial learning rate, and the number of training cycles is set to a preset value. Through incremental learning, the surrogate model absorbs information from new samples and gradually expands the range of the parameter space.

[0016] In a preferred embodiment, a machine learning-based intelligent design system for automotive connector plates is used to execute the aforementioned machine learning-based intelligent design method for automotive connector plates, comprising: The multi-source heterogeneous data fusion module is used to collect historical design data of automotive connection boards, perform data preprocessing and multi-source heterogeneous feature fusion, and obtain a unified representation vector of multi-source heterogeneous data. The intelligent configuration generation module is used to obtain the user's design requirement condition vector, combine it with the unified representation vector of multi-source heterogeneous data, perform initial configuration generation and topology optimization coupling, and obtain initial geometric configuration candidate schemes. The multi-objective collaborative optimization module extracts design parameters based on the initial geometric configuration candidate schemes, constructs an adaptive multi-objective reward function to optimize the design parameters, and obtains the Pareto front solution set; The adaptive agent modeling module constructs an agent model based on a unified representation vector of multi-source heterogeneous data, performs agent model training and online updates, and obtains an adaptively optimized agent model. The multi-condition performance prediction module, based on an adaptive optimization surrogate model, performs multi-condition performance prediction and evaluation on the service process of automotive connecting plates, and obtains quantitative analysis results of the coupling relationship between conditions. An integrated visualization implementation module is developed and visualized based on a unified representation vector of multi-source heterogeneous data, candidate schemes for initial geometric configurations, Pareto front solution set, surrogate model for adaptive optimization, and quantitative analysis results of coupling relationships between working conditions, resulting in an intelligent design scheme for automotive connection plates.

[0017] The beneficial effects of this invention are as follows: This invention establishes a unified representation learning framework for multi-source heterogeneous data. It extracts geometric topological features through graph neural networks, extracts mechanical response features through convolutional neural networks, constructs structured features through feature engineering, and employs an attention mechanism for multimodal feature fusion, achieving a unified representation of heterogeneous information such as geometry, mechanics, materials, and processes. This method can fully mine and utilize the implicit knowledge in historical design data, improving the efficiency of design experience inheritance and reuse, and providing a solid data foundation for intelligent design decisions. Simultaneously, by combining generative adversarial networks with configuration generation methods based on topology optimization physical constraints, it can automatically generate diverse initial design schemes that satisfy mechanical optimization criteria, enhancing the breadth and depth of design space exploration.

[0018] This invention proposes an adaptive multi-objective reinforcement learning optimization strategy. Through Markov decision process modeling, dual-deep Q-network value function approximation, and adaptive reward function design, it achieves collaborative optimization of multiple conflicting objectives such as lightweighting, strength, stiffness, cost, and manufacturing process. It can automatically balance the trade-offs between different objectives, generating a uniformly distributed Pareto front solution set, providing engineers with a wealth of design options. Combined with the rapid performance prediction capabilities of the surrogate model, the computational efficiency of the optimization process is improved by several orders of magnitude compared to traditional methods, enabling intelligent exploration of large-scale design spaces. Through multi-condition integrated prediction and quantitative analysis of coupling relationships between conditions, it can accurately predict the comprehensive performance of the connecting plate in complex service environments, ensuring the robustness and reliability of the design scheme. Attached Figure Description

[0019] Figure 1 This is a flowchart of the main process of a machine learning-based intelligent design method for automotive connection plates in this invention. Figure 2 This is a detailed flowchart of a machine learning-based intelligent design method for automotive connection plates in this invention. Figure 3 This is a block diagram of an intelligent design system for automotive connection plates based on machine learning, as described in this invention. Detailed Implementation

[0020] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0021] At least one embodiment of the present invention discloses a machine learning-based intelligent design method for automotive connection plates, such as... Figures 1 to 2 As shown, it includes the following steps: Step 1: Collect historical design data of automotive connection plates, perform data preprocessing and multi-source heterogeneous feature fusion to obtain a unified representation vector of multi-source heterogeneous data; Step 1.1: Historical design data collection and quality assessment; We obtained historical design data of vehicle connection plates for 20 models of a new energy vehicle company over the past 5 years, including 3D geometric model files, material property tables, multi-condition load records, finite element simulation results, and actual test reports. We used data crawling tools to extract these data in batches from the product data management system, resulting in a raw dataset containing 512 design samples.

[0022] Based on the original dataset, a data integrity detection algorithm was used to scan each data field of each sample one by one, and the proportion of missing values, data type consistency, and reasonableness of value range were statistically analyzed to obtain a data quality assessment report. The report showed that 8.7% of the samples had some missing fields, mainly concentrated in fatigue test data and manufacturing cost information. At the same time, it was found that the material properties of 23 samples did not match the geometric model, indicating data labeling errors.

[0023] Step 1.2, Intelligent Data Repair and Anomaly Detection; Based on the original dataset and data quality assessment report output in step 1.1, intelligent filling is performed on the missing fatigue test data using a collaborative filtering method based on similar samples. By calculating the cosine similarity between the sample to be filled and other complete samples in terms of geometric features and material properties, the five neighboring samples with the highest similarity are selected, and the weighted average of their fatigue test data is taken as the filling value. The weight coefficient is proportional to the similarity, thus obtaining the completed fatigue data.

[0024] For samples where material properties do not match the geometric model, a multi-source data cross-validation mechanism is used for consistency correction. At the same time, the material labels in the 3D model file, the material definitions in the finite element model, and the material numbers in the bill of materials are queried. The material type that appears most frequently is selected as the correct label through a voting mechanism. When the material types of the three data sources are different, the sample is marked as pending manual review. The data after consistency correction is obtained to ensure the accurate correspondence between the geometric model and the material properties.

[0025] The isolated forest algorithm was used for outlier detection. The performance indicators of each sample, such as quality, strength, and stiffness, were used to construct a feature vector. Multiple isolated trees were constructed by randomly dividing the feature space. The average path length required for each sample to be isolated was calculated. Finally, 17 samples with incorrect confirmation were removed, and 15 samples with innovative value were retained, resulting in a cleaned dataset containing 495 high-quality samples.

[0026] Step 1.3, Geometric topological feature extraction and graph neural network learning; Based on the 495 high-quality samples cleaned from step 1.2, the 3D geometric model of the connecting plate is concentrated. A mesh simplification algorithm is used to convert the model into a graph structure representation. Feature points of the geometric model are extracted as graph nodes, including the vertices of the plate, the inflection points of the stiffeners, and the center points of the bolt holes. Edges are established between connected feature points according to the geometric topology. Each node is assigned attributes such as spatial coordinates, local curvature, and the number of adjacent edges. Each edge is assigned attributes such as connection length and connection angle. The edge weight is defined as the reciprocal of the Euclidean distance between two connected nodes. The closer the distance, the greater the weight. When the distance between two nodes exceeds 10% of the maximum size of the connecting plate, no edge connection is established to ensure the sparsity and local connectivity of the graph structure.

[0027] Based on graph structure representation, a graph convolutional neural network is used for node feature learning. A three-layer graph convolutional layer is constructed. The first layer aggregates the initial attributes of each node with the attributes of its first-order neighbor nodes in a weighted manner. The second layer aggregates the second-order neighbor information, and the third layer aggregates the third-order neighbor information, so that the final embedding of each node contains the topological structure features within its third-order neighborhood. A global average pooling operation is used to average the embeddings of all nodes to obtain a 128-dimensional global topological feature vector.

[0028] Step 1.4, Mechanical response feature extraction and structured feature construction; Based on the finite element simulation results of the connection plate in the cleaning dataset output in step 1.2, stress distribution cloud maps and strain distribution cloud maps are extracted as two-dimensional image data. A lightweight convolutional neural network is used to extract mechanical response features. A network structure containing four convolutional blocks is constructed. Each convolutional block consists of a depthwise separable convolutional layer, a batch normalization layer, and a ReLU activation layer. A 128-dimensional mechanical response feature vector is obtained through layer-by-layer feature extraction. At the same time, based on the material property parameters of the connection plate and the multi-condition load conditions, a structured feature vector is constructed using feature engineering methods. The physical parameters of the material, such as yield strength, elastic modulus, density, and Poisson's ratio, are numerically normalized. The load magnitude and direction of static load conditions, frontal collision conditions, side collision conditions, and fatigue cycle conditions are vectorized and encoded. Boundary conditions such as battery pack mass, suspension connection point position, and body interface size are numerically standardized. The resulting 64-dimensional structured feature vector is formed by concatenating these features.

[0029] Step 1.5: Multimodal feature fusion and unified representation learning; Based on the global topological feature vector output in step 1.3, the mechanical response feature vector output in step 1.4, and the structured feature vector, a multimodal feature fusion network is used for high-dimensional unified representation learning. A fusion network with an attention mechanism is constructed. The three feature vectors are nonlinearly transformed by three independent fully connected layers to obtain intermediate representations of the same dimension. The correlation weights between the three modes are calculated by a self-attention layer, enabling the network to adaptively emphasize the modal features that contribute significantly to the current design task. Finally, the weighted three modal features are concatenated and dimensional compression and nonlinear mapping are performed by two fully connected layers to obtain a unified representation vector of multi-source heterogeneous data. This vector is 256-dimensional and integrates comprehensive information such as geometric topology, mechanical response, material properties, and boundary conditions, providing sufficient design space representation for subsequent configuration generation and parameter optimization.

[0030] In some embodiments, since the design data of connecting plates for different vehicle models differ in geometric scale and performance index range, direct fusion may lead to the suppression of features of certain modalities. In the multimodal feature fusion process of step 1.5, an adaptive feature standardization method can be used to dynamically scale the features of each modality. The purpose is to ensure that the contribution of different modal features to the fusion result is balanced. Specifically, the mean and variance of each modal feature vector in the current batch of samples are calculated, and zero-mean unit variance standardization is performed on each modal feature to obtain the standardized feature vector. Then, the standardized features are subjected to affine transformation through learnable scaling factors and offset factors. The scaling factors and offset factors are jointly optimized with other parameters of the fusion network through backpropagation, so that the network can automatically adjust the relative importance of each modality according to the task requirements. Finally, the adjusted features are input into the attention fusion layer to obtain a balanced multimodal fusion representation. This method can adapt to the distribution differences of data of different vehicle models and improve the representation quality of fused features.

[0031] Step 2: Obtain the user's design requirement condition vector, combine it with the unified representation vector of multi-source heterogeneous data, perform initial configuration generation and topology optimization coupling, and obtain initial geometric configuration candidate schemes; Step 2.1, Generate the adversarial network construction; The user inputs a vector of design requirements, including parameters such as target mass range, strength requirements, stiffness requirements, battery pack size, and suspension connection point coordinates. These parameters are then standardized and converted into a 64-dimensional numerical vector.

[0032] A generative adversarial network (GAN) is constructed, comprising two sub-networks: a generator and a discriminator. Based on the unified representation vector of multi-source heterogeneous data output in step 1, the GAN is constructed to generate the initial configuration. The unified representation vector of multi-source heterogeneous data is used as the conditional input of the GAN, which is fused with the user design requirement condition vector to construct the GAN for initial configuration generation. The generator's input is a concatenation of the unified representation vector of multi-source heterogeneous data, the user-input design requirement condition vector, and a random noise vector. The random noise vector is sampled from a 128-dimensional Gaussian distribution. After concatenation, it forms the generator's input vector. The generator's input vector contains both design constraint information and randomness, ensuring the diversity of generated configurations.

[0033] Step 2.2, multi-layer deconvolution topological density field mapping; Based on the generator input vector output in step 2.1, a multi-layer deconvolutional neural network is used to map the low-dimensional vector to a high-dimensional topological density field. Specifically, the generator input vector is expanded to a 4096-dimensional feature vector through a two-layer fully connected network and reshaped into an 8×8×64 three-dimensional tensor as the initial feature map. Upsampling is then performed through five deconvolutional blocks. Each deconvolutional block contains a transposed convolutional layer, a batch normalization layer, and a LeakyReLU activation layer. The first deconvolutional block expands the feature map size from 8×8 to 16×16 and reduces the number of channels from 64 to 32. The second... The first deconvolutional block is expanded to 32×32 while maintaining 32 channels. The third deconvolutional block is expanded to 64×64 and reduced to 16 channels. The fourth deconvolutional block is expanded to 128×128 while maintaining 16 channels. The fifth deconvolutional block is expanded to 256×256 and outputs a single channel. The output value at each position is mapped to between 0 and 1 by the Sigmoid activation function, representing the material density at that position. Finally, a two-dimensional topological density distribution map with a resolution of 256×256 is obtained. In this density map, areas close to 1 indicate that material needs to be retained, and areas close to 0 indicate that material needs to be removed.

[0034] Step 2.3: Generate the three-dimensional geometric configuration by stretching; Based on the two-dimensional topological density distribution map output in step 2.2, a preliminary three-dimensional geometric configuration of the connecting plate is generated by a three-dimensional stretching operation. Specifically, according to the thickness range of the connecting plate specified in the design requirements, the density map is stretched along the vertical direction to form a three-dimensional voxel model. The material density of each voxel inherits the density value of the corresponding two-dimensional position. The moving cube algorithm is used to extract the isosurface with a density threshold of 0.5 to generate a preliminary three-dimensional mesh model. Then, the three-dimensional mesh model is smoothed to eliminate the stepped edges caused by voxelization, resulting in a continuous and smooth initial geometric configuration. This configuration retains the main material distribution pattern in the density distribution map.

[0035] Step 2.4, Construction of the dual-branch discriminator network; Based on the initial geometric configuration output in step 2.3, a discriminator network is constructed to evaluate the authenticity and physical rationality of the configuration. The discriminator adopts a dual-branch structure: the first branch is the authenticity discrimination branch, and the second branch is the physical constraint discrimination branch. The two branches share the first four convolutional feature extraction layers and are each connected to an independent fully connected layer to output the discrimination result. The authenticity discrimination branch outputs a probability value between 0 and 1, indicating whether the input configuration is a real design sample or a generated sample. The physical constraint discrimination branch outputs a physical rationality score, indicating whether the configuration meets the mechanical optimization criteria.

[0036] Step 2.5, embedding physical constraints for topology optimization; In the physical constraint discrimination branch constructed in step 2.4, the compliance minimization criterion of topology optimization is introduced as a physical constraint. Specifically, a fast finite element analysis is performed on the input three-dimensional mesh model. A simplified linear elastic model and coarse mesh generation are used to reduce the computational load. The displacement field is solved under given load and boundary conditions, and the total compliance value of the structure is calculated. Compliance is defined as the inner product of the displacement at the load application point and the load. The smaller the compliance, the greater the stiffness of the structure. The compliance value is normalized to the range of 0 to 1. The stiffness score is obtained by subtracting the normalized compliance from 1. At the same time, it is checked whether there are suspended material blocks in the configuration. Suspended material blocks refer to isolated regions that are not connected to the main structure. The volume ratio of suspended material blocks is counted. The connectivity score is obtained by subtracting the ratio from 1. The stiffness score and the connectivity score are weighted and summed to obtain the topological rationality score. This score can quantify the mechanical performance and topological rationality of the configuration.

[0037] Step 2.6, Joint Loss Function Design; Based on the authenticity discrimination branch constructed in step 2.4 and the topological rationality score output in step 2.5, a joint loss function for the generative adversarial network is designed. The generator's loss function consists of two parts: adversarial loss and physical constraint loss. The adversarial loss uses the Wasserstein distance to measure the difference between the generated distribution and the real distribution. By minimizing this distance, the generated configuration is made closer to the real design sample in the feature space. The physical constraint loss uses the negative value of the physical rationality score. By maximizing the physical rationality score, the generated configuration satisfies the mechanical optimization criterion. The two loss parts are linearly combined through weight coefficients, which are determined through performance tuning on the validation set. The discriminator's loss function is a weighted sum of the authenticity discrimination loss and the physical constraint discrimination loss. The authenticity discrimination loss measures the discriminator's ability to distinguish between real samples and generated samples, while the physical constraint discrimination loss measures the discriminator's accuracy in evaluating physical rationality. Thus, the complete adversarial training objective function is obtained.

[0038] Step 2.7, Optimize the alternating training strategy; Based on the joint loss function designed in step 2.6, an alternating training strategy is adopted to optimize the generator and discriminator. Specifically, in each training iteration, the generator parameters are fixed, and the discriminator is trained for 5 steps using real design samples and generated samples to improve the discriminator's discrimination and physical evaluation capabilities. Then, the discriminator parameters are fixed, and the generator is trained for 1 step to produce a more realistic and physically reasonable configuration. The Adam optimizer is used to update the network parameters, with the generator learning rate set to 0.0001 and the discriminator learning rate set to 0.0004. To maintain training stability, a gradient penalty term is added to the discriminator's loss function to penalize the degree to which the gradient norm of the discriminator deviates from 1 at the interpolation point between real and generated samples. After 500 training epochs, the generator and discriminator reach Nash equilibrium, resulting in a converged conditional generative adversarial network model.

[0039] Step 2.8: Generation of diverse initial configuration candidates; Based on the convergence condition generative adversarial network model output in step 2.7, multiple initial configuration candidate schemes are generated for the new design task. Specifically, the user inputs a design requirement condition vector, including parameters such as a target mass of 3.5 kg, a yield strength requirement of greater than 500 MPa, a first-order modal frequency requirement of greater than 80 Hz, a battery pack installation area size of 600 mm × 400 mm, and the coordinates of the front suspension connection point and the rear suspension connection point. These parameters are converted into a 64-dimensional design requirement condition vector through standardization. Ten different sets of 128-dimensional noise vectors are randomly sampled. The design requirement vector is concatenated with the 10 sets of noise vectors and then input into the generator, outputting 10 different topology density distribution maps. Each density map is subjected to 3D stretching and isosurface extraction to obtain initial geometric configuration candidate schemes. These schemes present different stiffener layouts and hole distribution forms under the premise of satisfying basic design constraints. The topology rationality scoring algorithm in step 2.5 is called to evaluate each candidate scheme, and the weighted sum of stiffness score and connectivity score is calculated and sorted from high to low.

[0040] In some embodiments, since the generated initial configuration may contain overly complex organic surface morphology that is difficult to manufacture using traditional stamping processes, a manufacturing constraint embedding strategy can be adopted in the generator loss function in step 2.6 to improve the process feasibility of the generated configuration. Specifically, a manufacturing constraint loss term is added to the generator loss function. This loss term assesses the manufacturing difficulty of the configuration. The Gaussian curvature and average curvature at various locations on the configuration surface are calculated using a curvature analysis algorithm to identify regions with drastic curvature changes. These regions are prone to springback and wrinkling defects during stamping. The area ratio of high curvature regions is statistically analyzed as an indicator of surface complexity. The draft angle detection algorithm is then used to determine the surface complexity along a preset curve. The stamping direction is analyzed on each surface of the configuration to check for undercut structures. Undercut structures can prevent demolding. The percentage of area with undercuts is used as a demolding difficulty index. The manufacturing constraint loss is obtained by weighted summation of the surface complexity index and the demolding difficulty index. This loss is minimized during generator training to make the generated configuration tend to have a smooth surface and a good draft angle. At the same time, in order to maintain design freedom, the weight coefficient of the manufacturing constraint loss is set relatively small, allowing the generator to generate complex shapes when necessary to meet performance requirements. The initial configuration generated by this method achieves a preliminary balance between mechanical properties and manufacturing process. Fine adjustments can be made on this basis in the subsequent optimization stage.

[0041] Step 3: Extract design parameters based on the initial geometric configuration candidate schemes, construct an adaptive multi-objective reward function to optimize the design parameters, and obtain the Pareto front solution set; Step 3.1, Markov decision process modeling; Based on the topological rationality score ranking results of the initial geometric configuration candidate schemes output in step 2.8, the three initial configuration schemes with the highest topological rationality scores are selected to enter the parameter optimization stage. Parametric modeling is performed on each initial geometric configuration candidate scheme, and key design parameters are automatically extracted through geometric feature recognition algorithms. Specifically, a voxel scanning algorithm is used to measure the thickness distribution at various positions of the connecting plate along the Z-axis, and the thickness variation area is divided into 8 control regions. The average thickness of each region is used as the plate thickness parameter. The center line of the reinforcing rib is identified through edge detection and skeleton extraction algorithms, and 12 control points are set at equal intervals along the center line. The height and thickness of each control point are used as the reinforcing rib parameters. Bolt holes are detected through circular Hough transform, and the diameter and center coordinates of 6 main bolt holes are extracted as hole position parameters, for a total of 50 parameters. To maintain optimization efficiency, the 32 parameters with the highest change sensitivity are selected as optimization variables to establish a parametric design space.

[0042] The design parameter optimization problem of the connecting plate is modeled as a Markov decision process. The state space is defined to include two parts: the parameter state and the performance state of the current design. The parameter state includes 32 continuous parameters such as the thickness of the connecting plate body, the height of the stiffener, the thickness of the stiffener, the spacing of the stiffener, the diameter of the bolt hole, and the position of the bolt hole. The performance state includes 12 performance indicators of the current configuration, specifically: mass, maximum stress, maximum displacement, first-order modal frequency, second-order modal frequency, third-order modal frequency, static load safety factor, frontal collision energy absorption, frontal collision intrusion, side collision displacement, fatigue life, and manufacturing cost estimation. The parameter state and the performance state are concatenated to form a 44-dimensional state vector, which comprehensively describes the characteristics of the current design scheme.

[0043] Step 3.2, Definition and mask of discrete action space; The action space is defined as the adjustment operations on each design parameter. For each continuous parameter, three discrete actions are defined: increase, decrease, and hold. The 32 parameters form a 96-dimensional discrete action space. To improve optimization efficiency, an action masking mechanism is introduced. For parameters that have reached the upper limit of their value, the increase action is prohibited. For parameters that have reached the lower limit of their value, the decrease action is prohibited. For parameters that have remained unchanged for multiple consecutive steps, the hold action is temporarily prohibited to encourage exploration. Through the action masking mechanism, the number of effective actions per step is reduced from 96 to about 20, thereby reducing the complexity of the strategy search space.

[0044] Step 3.3, Construction of the value function for the dual-depth Q-network; A Double Deep Q-Network (DDQN) is constructed as a value function approximator. This network consists of two identical neural networks: a current network and a target network. Each network contains an input layer, three hidden layers, and an output layer. The input layer receives a 44-dimensional state vector. The first hidden layer is a fully connected layer with 256 neurons, using the ReLU activation function. The second hidden layer is a fully connected layer with 256 neurons, using the ReLU activation function. The third hidden layer is a fully connected layer with 128 neurons, using the ReLU activation function. The output layer is a linear layer with 96 neurons. Each neuron corresponds to the Q-value of an action. During action selection, the action masking mechanism from step 3.2 is used to make decisions based only on the Q-values ​​corresponding to valid actions. The Q-value represents the expected long-term cumulative reward of taking that action in the current state. The current network is used to select actions, and the target network is used to calculate the target Q-value. The parameters of the target network are copied from the current network every 50 training steps. By decoupling action selection and Q-value evaluation, the Double Deep Q-Network can reduce the problem of Q-value overestimation and improve training stability.

[0045] Step 3.4, Design of adaptive multi-objective reward function; An adaptive multi-objective reward function is designed, which comprehensively considers lightweighting objectives, strength constraints, stiffness constraints, cost constraints, and manufacturing process constraints. Specifically, normalized reward components for each individual objective are defined: the lightweighting reward is defined as the reduction ratio of the current mass to the initial mass, with higher rewards for greater mass reduction; the strength reward is defined as a normalized function of the safety factor, which is the ratio of the material's yield strength to the maximum stress, with positive rewards when the safety factor is greater than the design threshold and negative rewards when it is less than the threshold; the stiffness reward is defined as a normalized function of the maximum displacement, with positive rewards when the maximum displacement is less than the design threshold and negative rewards when it is greater than the threshold; and the cost reward is defined as the percentage reduction in manufacturing cost relative to the budget, with higher rewards for greater cost savings. Secondly, a manufacturing process constraint reward component is added to assess the manufacturing feasibility of the current design scheme. Stamping constraints are achieved by detecting the draft angles on the connecting plate surfaces, with a manufacturing feasibility score awarded when the draft angles on all surfaces are greater than 3 degrees. The manufacturing feasibility score is 1 if the interface gap is between 0.5 and 2.0 mm, and decreases linearly according to the proportion of unmet area. The welding assembly constraint is achieved by detecting the interface matching degree between the connecting plate and the adjacent parts. When the interface gap is within the range of 0.5 to 2.0 mm, the assembly feasibility score is 1. If it exceeds the range, it decreases linearly according to the degree of deviation. The weighted average of the manufacturing feasibility score and the assembly feasibility score is used as the manufacturing constraint reward component. An adaptive weight coefficient is designed to dynamically adjust the weight of each objective according to the position of the current solution in the objective space. When an objective seriously fails to meet the constraint, the weight of that objective is increased. When all constraints are met, the weights are evenly distributed. The weight coefficient of the manufacturing constraint reward component is set to 0.1. The weight coefficients of the four objectives of lightweighting, strength, stiffness and cost are dynamically adjusted in the range of 0.1-0.4 according to the constraint satisfaction and the sum is 0.9. Finally, each reward component is multiplied by its corresponding weight coefficient and summed to obtain the comprehensive reward value. This reward function can adaptively guide the optimization process to balance multiple conflicting objectives and ensure the manufacturing feasibility of the design scheme.

[0046] Step 3.5, Training the experience replay mechanism; Based on the definition of Markov decision processes and the structure of dual-depth Q-networks, an experience replay mechanism is used for training. Specifically, the experience replay buffer is initialized as a circular queue with a capacity of 10,000, and the parameters of the current network and the target network are initialized to random values. For each initial configuration scheme, optimization iterations are performed starting from the initial state. At each time step, the current state is input into the current network to obtain the Q-value of each action. An epsilon greedy strategy is used to select actions, randomly selecting an effective action for exploration with probability epsilon and selecting the effective action with the largest Q-value with probability 1 minus epsilon. The initial value of epsilon is set to 1.0, which linearly decays to 0.1 with the number of training steps. After executing the selected action, the design parameters are updated according to the parameter adjustment rules, and the performance index of the new state is calculated by calling finite element simulation. The simulation settings are consistent with the parameters in step 4.1, using tetrahedral mesh elements with a unit size of 5 milliseconds. The training model is constructed using high-strength steel. Rewards are calculated based on performance changes. A five-tuple (current state, selected action, reward, new state, and termination status) is stored in an experience replay buffer. When the number of samples in the buffer exceeds 256, 64 experience samples are randomly sampled to form a training batch. For each sample in the batch, the target network is used to calculate the maximum Q-value of the new state. The target Q-value is calculated using the Bellman equation: the immediate reward plus a discount factor multiplied by the maximum Q-value of the new state (discount factor set to 0.95). The current network is used to calculate the predicted Q-value of the action selected in the current state. The mean squared error between the predicted Q-value and the target Q-value is calculated as the loss. The Adam optimizer minimizes the loss and updates the current network parameters with a learning rate of 0.001. Every 50 training steps, the current network parameters are copied to the target network. Optimization terminates when there is no improvement in performance metrics for 20 consecutive steps or when the 500-step iteration limit is reached, resulting in a well-trained deep Q-network model.

[0047] Step 3.6, Execution of deterministic optimization strategy; Based on the trained deep Q-network model output in step 3.5, deterministic optimization strategies are executed on the three initial configuration schemes selected in step 3.1. Specifically, starting from the initial state of each initial configuration, instead of using the epsilon greedy strategy for random exploration, the action with the largest Q value is selected at each step to ensure that the optimization process proceeds along the learned optimal strategy trajectory. All states visited during the optimization process and their performance indicators are recorded, and the optimization trajectory is plotted in the three-dimensional target space of mass-strength-stiffness. When the optimization converges, each initial configuration generates an optimized design scheme, resulting in three optimized schemes. The three optimized schemes are non-dominated and sorted to select Pareto front solutions. If there is a dominance relationship between the three schemes, the non-dominated solutions are retained; if the three schemes are not mutually dominant, all are retained, resulting in a Pareto front solution set. The schemes in this solution set have different focuses in multi-objective trade-offs, providing engineers with differentiated choices.

[0048] Step 3.7, Diversification and Expansion of Weight Vector Decomposition; To further enrich the diversity of the Pareto front solution set output in step 3.6, a weight vector decomposition strategy is adopted to generate more candidate solutions. Specifically, 10 sets of weight vectors are uniformly sampled in the target space. Each set of weight vectors is a linear combination coefficient of the three objectives of lightweighting, strength, and stiffness. The weight coefficients range from 0 to 1 and sum to 1. A uniform design method is used to ensure that the weight vectors are evenly distributed in the target space. For each set of weight vectors, the weight coefficients in the adaptive multi-objective reward function are modified to the corresponding fixed weight values, and the dual-depth Q-network optimization process is re-executed. Starting from the three initial configurations, optimization is performed separately to obtain 30 candidate solutions. All candidate solutions are non-dominated and sorted to select the solutions located at the Pareto front, resulting in a Pareto front solution set containing 15 non-dominated solutions. These solutions are evenly distributed in the target space, covering different performance orientations from extreme lightweighting to high stiffness and high strength, meeting the needs of different application scenarios.

[0049] In some embodiments, since the exploration efficiency of deep Q-networks in continuous high-dimensional action spaces is limited, a deep deterministic policy gradient method can be used in step 3 to replace the deep Q-network for parameter optimization. The aim is to improve optimization efficiency and convergence speed. Specifically, two sub-networks, an Actor network and a Critic network, are constructed. The Actor network receives the state vector as input and outputs a 32-dimensional continuous action vector through a 3-layer fully connected network. Each dimension corresponds to the adjustment amount of a design parameter. The output layer uses the Tanh activation function to limit the action range to between -1 and 1. The actual parameter adjustment amount is obtained by scaling according to the value range of each parameter. The Critic network receives the concatenation of the state vector and the action vector as input and outputs a scalar Q-value through a 3-layer fully connected network. The value of taking this action in this state is represented by the value of the action. During training, the Actor network learns the optimal policy by maximizing the Q-value output of the Critic network. The Critic network learns the value function by minimizing the mean square error between the predicted Q-value and the target Q-value. Similarly, the experience replay and target network mechanism are used to improve training stability. When selecting an action, Gaussian noise is added to the output of the Actor network for exploration. The standard deviation of the noise gradually decreases with the number of training steps. Through this method, the policy network can directly output continuous parameter adjustment, avoiding the resolution limitation caused by the discrete action space. The optimization process is more refined, and the convergence speed is improved by about 30%. The final Pareto front solution set is more continuously and densely distributed in the target space, providing engineers with richer design options.

[0050] Step 4: Construct a proxy model based on the unified representation vector of multi-source heterogeneous data, train and update the proxy model online to obtain an adaptively optimized proxy model; Step 4.1, Training the Latin hypercube sampling surrogate model; To accelerate performance evaluation during the deep reinforcement learning optimization process, a surrogate model was constructed to replace the time-consuming finite element simulation. Based on the unified representation vector of the multi-source heterogeneous data output in step 1, 200 representative sample points were selected in the design parameter space using the Latin hypercube sampling method. Latin hypercube sampling ensures that the samples are evenly distributed across all parameter dimensions through hierarchical random sampling, which can cover the high-dimensional parameter space with fewer samples. For each sample point, a three-dimensional model of the connecting plate was constructed according to its parameter values ​​and imported into the finite element analysis software for static structural analysis and modal analysis. The static analysis applied the battery pack gravity load and the road surface excitation load, and the modal analysis extracted the first three natural frequencies and mode shapes. The simulation used tetrahedral mesh elements with a unit size of 5 mm. The material used was high-strength steel with a yield strength of 500 MPa, an elastic modulus of 210 GPa, and a Poisson's ratio of 0.3. The simulation time for each sample was approximately 15 minutes, and the total simulation time for 200 samples was 50 hours, resulting in a training sample set containing design parameters and corresponding performance indicators.

[0051] Step 4.2, Construction of deep feedforward neural network agent; Based on the training sample set containing design parameters and corresponding performance indicators output in step 4.1, a deep feedforward neural network is constructed as a surrogate model. The network input is the design parameters, and the output is a 12-dimensional performance indicator vector. The network structure includes an input layer, four hidden layers, and an output layer. The first hidden layer has 512 neurons, the second hidden layer has 512 neurons, the third hidden layer has 256 neurons, and the fourth hidden layer has 256 neurons. Each hidden layer uses the ReLU activation function. A Dropout layer is inserted after the second and fourth hidden layers with a deactivation probability of 0.2 to prevent overfitting. The output layer is a linear layer with 12 neurons, outputting the quality, performance, and performance indicators respectively. The predicted values ​​of performance indicators such as maximum stress, maximum displacement, and first-order modal frequency were obtained. The training sample set was divided into a training set and a validation set in an 8:2 ratio. The training set was used for model parameter learning, and the validation set was used for hyperparameter tuning and early stopping. The mean squared error was used as the loss function, and the loss was minimized by the Adam optimizer. The learning rate was set to 0.001, the batch size was set to 32, and the training was carried out for 200 epochs. The training was stopped early when the loss of the validation set did not decrease for 20 consecutive epochs, resulting in a converged surrogate model. The average relative error of the model in predicting the performance of the test samples was 4.2%, and the time taken for a single prediction was only 0.8 milliseconds, which is more than 1 million times faster than finite element simulation.

[0052] Step 4.3, embedding the uncertainty estimation mechanism; In the deep reinforcement learning optimization process, the surrogate model output from step 4.2 is embedded into the state transition stage of a Markov decision process. Specifically, after executing an action to update the design parameters, the new parameter vector is input into the surrogate model. The predicted value of the performance index is obtained through forward propagation. Based on the predicted value, the reward function and the new state vector are calculated, achieving rapid environmental feedback. This reduces each optimization iteration from 15 minutes to 1 millisecond, and the total optimization time from several days to several hours, improving design efficiency. At the same time, to ensure prediction accuracy, an uncertainty estimation mechanism is introduced by training 10 surrogates with the same structure but different initial parameters. The model is composed of an integrated model. For a given input parameter, the prediction results of the 10 models differ. The standard deviation of the prediction results is calculated as an uncertainty index. The standard deviation of the prediction results of the 10 models is divided by the prediction mean to obtain the coefficient of variation as a normalized uncertainty index. When the coefficient of variation exceeds 0.15, the uncertainty is considered too high. This threshold is determined by analyzing the correlation between prediction error and coefficient of variation on the validation set. When the coefficient of variation is greater than 0.15, the prediction error usually exceeds 10%. At this time, finite element simulation is triggered for accurate evaluation. The prediction reliability is low, and the true performance index is obtained to ensure the reliability of the optimization process.

[0053] Step 4.4, Adaptive online incremental learning and updating; Based on the uncertainty estimation mechanism established in step 4.3, the adaptive online update of the surrogate model is realized. Specifically, during the optimization process, whenever a new sample is obtained by triggering finite element simulation, the sample is added to the training set of the surrogate model, and the surrogate model is retrained periodically. The update process adopts an incremental learning strategy, which does not start training from scratch, but fine-tunes the original model parameters with new samples. The learning rate is set to one-tenth of the initial learning rate, and the number of training cycles is set to 20. Through incremental learning, the surrogate model can quickly absorb the information of new samples and gradually expand the range of its accurate prediction parameter space. As the optimization iteration proceeds, the coverage of the surrogate model becomes larger and larger, and the frequency of triggering simulation becomes lower and lower. In the later stage of optimization, the performance evaluation can be basically completely relied upon by the surrogate model, resulting in an adaptively optimized surrogate model. The average relative error of this model in the early and late stages of optimization is 4.2% and 2.1%, respectively, and the prediction accuracy continues to improve.

[0054] In some embodiments, since the prediction accuracy of the neural network surrogate model may drop sharply in the extrapolation region, a Gaussian process regression model can be used in step 4 to replace the neural network in constructing the surrogate model. The aim is to obtain more reliable uncertainty estimation and better small-sample generalization ability. Specifically, the radial basis function is selected as the kernel function of the Gaussian process. The length scaling parameter and variance parameter of the kernel function are determined by maximizing the marginal likelihood function. Gaussian process regression can output not only the predicted mean of the performance index, but also the predicted variance. The predicted variance directly reflects the model's confidence in the prediction results. The predicted variance automatically increases in regions far from the training samples, providing self-confidence. However, uncertainty is quantified. When the prediction variance exceeds a threshold, finite element simulation is triggered. After adding new samples to the training set, the posterior distribution of the Gaussian process can be efficiently calculated through Bayesian updates without retraining the entire model. The update speed is about 10 times faster than that of neural networks. At the same time, the generalization ability of the Gaussian process is better than that of neural networks under small sample conditions. When the number of training samples is only 50, the prediction error of the Gaussian process is 6.8%, while the prediction error of the neural network is 12.3%. Through this method, the surrogate model maintains stable prediction accuracy throughout the optimization process, and the total number of simulations triggered is reduced from 80 times for the neural network method to 45 times, further reducing the optimization cost.

[0055] Step 5: Based on the adaptive optimization surrogate model, perform multi-condition performance prediction and evaluation on the service process of the vehicle connecting plate to obtain quantitative analysis results of the coupling relationship between the conditions. Step 5.1, Construction of a multi-condition robust performance evaluation framework; To meet the training requirements of the model under multiple working conditions, after the surrogate model is trained in step 4.1, the surrogate model is used to perform additional sampling of the parameter space to generate virtual samples to supplement the training set. Combined with the real samples in step 1.2, an expanded dataset is formed, which is then classified according to working conditions and used for training of each dedicated model.

[0056] To address the various operating conditions faced by automotive connecting plates during service, an integrated learning framework is constructed based on the adaptively optimized surrogate model output in step 4.4 to predict robust performance. Four typical operating conditions for evaluating the connecting plate are defined: Condition 1 is the static load condition, simulating the static pressure on the connecting plate from the battery pack weight and the vehicle body weight, with key performance indicators being maximum stress and maximum displacement; Condition 2 is the frontal collision condition, simulating the impact load borne by the connecting plate when the vehicle collides head-on with a rigid barrier at a speed of 50 km / h, with key performance indicators being energy absorption and passenger compartment intrusion; Condition 3 is the side collision condition, simulating the local deformation of the connecting plate during a side pole impact, with key performance indicators being the collision point displacement and B-pillar acceleration; and Condition 4 is the fatigue vibration condition, simulating the cyclic load experienced by the connecting plate when the vehicle is driving on uneven roads, with key performance indicators being stress amplitude and fatigue life. This results in a multi-condition performance evaluation framework.

[0057] Step 5.2, Static load random forest regression modeling; For each working condition defined in step 5.1, a dedicated performance prediction model is constructed. For static load-bearing working conditions, based on the cleaned dataset output in step 1, the data is classified according to the working condition type, including 200 static load-bearing working condition samples, 150 frontal collision working condition samples, 120 side collision working condition samples, and 180 fatigue vibration working condition samples. Some samples contain multiple working condition data. A random forest regression model is adopted, with the number of trees set to 100 and the maximum depth of each tree set to 15, to obtain the static working condition random forest prediction model. The prediction error of this model for the maximum stress is 3.1%, and the prediction error for the maximum displacement is 2.8%.

[0058] Step 5.3, frontal collision gradient boosting tree modeling; For the frontal collision scenario, based on the cleaned dataset output in step 1.2 containing 150 samples of frontal collision simulation data, a gradient boosting tree regression model was adopted. The collision simulation used an explicit dynamics solver. The first 100 milliseconds of the collision process were simulated. The number of trees was set to 200, the learning rate was set to 0.1, and the maximum depth of each tree was set to 10. The resulting frontal collision gradient boosting tree prediction model had a prediction error of 4.5% for energy absorption and a prediction error of 5.2% for the intrusion of the occupant compartment.

[0059] Step 5.4, Dedicated construction of side impact fatigue multi-model; For side-impact and fatigue vibration conditions, based on the cleaned dataset output in step 1.2, support vector regression and deep neural network models are used respectively. Support vector regression maps the data to a high-dimensional space through a kernel function for linear regression, which is suitable for handling nonlinear problems with small to medium samples. During training, a side-impact simulation dataset containing 120 samples is used, and the radial basis function is selected as the kernel function. The penalty parameter and kernel parameter are optimized through grid search to obtain the side-impact support vector regression prediction model. The deep neural network model adopts a 4-layer fully connected network. During training, a fatigue simulation dataset containing 180 samples is used. Fatigue life is calculated using the rainflow counting method and Miner's linear cumulative damage criterion to obtain the fatigue vibration neural network prediction model. The dedicated models for the four conditions are optimized according to the characteristics of their respective conditions to achieve the best prediction performance in their respective fields.

[0060] Step 5.5, multi-condition weighted fusion integrated prediction; A multi-condition integrated prediction framework is constructed. The four condition-specific prediction models output from steps 5.2 to 5.4 are used as base learners. For a given connecting plate design scheme, the four base learners are input respectively to obtain the performance prediction results for the four conditions. The multi-condition performance is integrated through a weighted fusion strategy. Specifically, the weight coefficients are determined according to the importance of different conditions in the vehicle evaluation system. The static load condition is set to a weight of 0.3 as a basic constraint, the frontal collision condition is set to a weight of 0.35 as a safety regulation requirement, the side collision condition is set to a weight of 0.25, and the fatigue vibration condition is set to a weight of 0.1. The normalized performance scores of each condition are weighted and summed to obtain a comprehensive performance score. The comprehensive score can comprehensively reflect the overall performance of the connecting plate under multiple conditions, avoid the performance degradation of other conditions caused by single-condition optimization, and obtain robust multi-condition performance evaluation results.

[0061] Step 5.6, Quantitative analysis of coupling relationships between operating conditions; Based on the integrated prediction framework constructed in step 5.5, the coupling relationship between different working conditions is analyzed. Specifically, for the 15 design schemes in the Pareto front solution set output in step 3.7, their performance indicators under four working conditions are calculated, and a working condition performance correlation matrix is ​​plotted. The linear correlation between the performance of any two working conditions is quantified by the Pearson correlation coefficient. It is found that the maximum stress of the static bearing working condition is negatively correlated with the energy absorption of the frontal collision working condition, with a correlation coefficient of -0.62. This indicates that increasing the static stiffness often leads to a decrease in the collision energy absorption capacity, which reflects the contradiction between rigidity and ductility. At the same time, it is found that the side impact performance is positively correlated with the fatigue performance, with a correlation coefficient of 0.54. This indicates that the structural form that resists side impact is usually also beneficial to fatigue life. Through correlation analysis, quantitative analysis results of the coupling relationship between working conditions are obtained, providing engineers with decision-making basis when making multi-objective trade-offs. When the performance of a certain working condition is a key constraint, the changing trend of the performance of other working conditions can be predicted to avoid neglecting one aspect for another.

[0062] In some embodiments, since different operating conditions have different sensitivities to design parameters, a hierarchical ensemble strategy can be adopted to improve prediction accuracy based on the ensemble prediction in step 5.5. The aim is to fully utilize the advantages of each base learner. Specifically, the first layer base learner is trained, which includes the aforementioned four models: random forest, gradient boosting tree, support vector regression, and deep neural network. Each model is trained for all operating conditions, resulting in four sets of multi-operating condition prediction results. The prediction results of the first layer are used as new features and input into the second layer meta-learner along with the original design parameters. The meta-learner adopts a ridge regression model and outputs the final performance prediction by learning the trade-off relationship between the prediction results of each base learner. The regularization parameter of the ridge regression is optimized through cross-validation. Through hierarchical ensemble, the meta-learner can automatically identify the advantages and disadvantages of each base learner under different operating conditions and dynamically allocate weights. Compared with simple weighted averaging, the prediction error of hierarchical ensemble is reduced by 18%, resulting in a more accurate multi-operating condition performance evaluation. At the same time, the weight coefficients of the meta-learner are interpretable. Engineers can understand the reliability of each base learner under different operating conditions by analyzing the weights, providing guidance for model improvement.

[0063] Step 6: Based on the unified representation vector of multi-source heterogeneous data, candidate schemes of initial geometric configuration, Pareto front solution set, surrogate model of adaptive optimization and quantitative analysis results of coupling relationship between working conditions, integrate development and visualization implementation to obtain intelligent design scheme of vehicle connection plate. Step 6.1, Modular intelligent design system architecture development; Based on the unified representation vector of multi-source heterogeneous data output in step 1, the candidate schemes of initial geometric configuration output in step 2, the Pareto front solution set output in step 3, the adaptive surrogate model output in step 4, and the quantitative analysis results of the coupling relationship between working conditions output in step 5, an integrated intelligent design system for automotive connection plates is developed. This system adopts a modular architecture, comprising five core modules: data management, feature engineering, model training, optimization solution, and result visualization. Each module interacts with data through standardized interfaces, supporting flexible expansion and independent upgrades. The system backend is developed using Python, employing the PyTorch deep learning framework to implement neural network models and the Scikit-learn library to implement traditional machine learning algorithms. The system frontend is developed using a Web technology stack, using the React framework to build the user interface and communicating with the backend via a RESTful API. The system is deployed on a high-performance computing server within the enterprise, configured with an 8-core CPU, 64GB of memory, and an NVIDIA RTX 3090 graphics card to meet the computational requirements for model training and inference.

[0064] Step 6.2, Data management module implements storage and retrieval; The data management module is responsible for storing, indexing, and retrieving the historical design data collected in step 1.1. It uses a relational database, MySQL, to store structured data, including design parameter tables, performance index tables, material property tables, and operating condition tables. These tables are linked via foreign keys, supporting complex queries. An unstructured data file system is used to store unstructured data, including 3D geometric model files, finite element mesh files, and simulation result files. File paths are recorded in the database for associated management. The data management module provides a data import interface, supporting batch import of historical data from the product data management system. It also provides a data cleaning interface, calling the aforementioned data quality assessment and intelligent repair algorithms. Finally, it provides a data retrieval interface, supporting multi-dimensional sample filtering based on vehicle model, material, performance range, and other criteria to obtain an efficient design knowledge base. Currently, the knowledge base stores over 2,000 connection plate design samples, covering the company's product development history over the past eight years.

[0065] Step 6.3, implementation of multimodal extraction in the feature engineering module; The feature engineering module is responsible for extracting and constructing feature vectors required for machine learning from raw data. This module integrates the feature extraction algorithms from steps 1.3 to 1.5 and includes a geometric feature extraction submodule, a mechanical feature extraction submodule, and a structured feature construction submodule. The geometric feature extraction submodule calls the aforementioned graph neural network to convert the 3D model into graph data and extract topological features. The mechanical feature extraction submodule calls the aforementioned convolutional neural network to extract mechanical response features from the stress-strain cloud map. The structured feature construction submodule encodes and normalizes material parameters, working conditions, and boundary constraints. The outputs of the three submodules are integrated into a unified high-dimensional representation vector through the feature fusion submodule. The feature engineering module provides a parameter configuration interface, allowing users to select which feature submodules to use, adjust feature dimensions and fusion strategies to meet the needs of different tasks, and obtain a flexible feature extraction pipeline.

[0066] Step 6.4: Integrate the model training module into the management platform; The model training module integrates the initial geometric configuration candidate schemes output in step 2, the Pareto front solution set output in step 3, the adaptive optimization surrogate model output in step 4, and the quantitative analysis results of the coupling relationship between operating conditions output in step 5. It provides model training, validation, and deployment functions. Users can select the training task type through the interface, including initial configuration generation, parameter optimization, performance prediction, etc., upload the training dataset, and set the model hyperparameters, including network structure, learning rate, batch size, number of training epochs, etc. After clicking "Start Training," the module calls the backend training script to train the model in parallel on the GPU. During the training process, the loss curve and performance index curve are displayed in real time, and users can monitor the training progress and model convergence. After training is completed, the module automatically evaluates the model performance on the validation set and generates a performance report, including indicators such as prediction error, correlation coefficient, and confusion matrix. After the user confirms that the model performance meets the requirements, the model can be deployed to the production environment for the optimization and solution module to call, resulting in an efficient model management platform that supports the coexistence and switching of multiple model versions.

[0067] Step 6.5: Optimize the human-computer interaction interface of the solver module; The optimization solution module provides an interactive design optimization interface, integrating the optimization algorithms from steps 2 to 3. Users input design requirement parameters, including target mass range, strength and stiffness requirements, geometric constraints, and material selection. They then select an optimization algorithm, including deep reinforcement learning, genetic algorithms, and particle swarm optimization, and set optimization parameters, including iteration steps, population size, and convergence threshold. After clicking "Start Optimization," the module calls the initial geometric configuration candidate schemes output in step 2.8 and the deep Q-network output in step 3.5 or the surrogate model output in step 4.4 for iterative optimization. The optimization process is executed asynchronously in the background, while the front-end interface updates the optimization progress and intermediate results in real time, including the parameter values, performance indicators, and 3D model preview of the current optimal solution. Users can pause optimization at any time to view the current scheme or adjust optimization parameters and continue. After optimization is complete, the module displays the Pareto front solution set. Users can interactively browse various schemes in the multi-objective trade-off graph, select the scheme of interest to view detailed information, and obtain a human-machine collaborative optimization platform that fully leverages engineers' domain knowledge and machine learning computational capabilities.

[0068] Step 6.6: The results visualization module implements multi-dimensional display; The results visualization module presents design schemes and performance analysis results in multiple formats. For the geometric model, WebGL technology is used to render the 3D model in the browser. Users can rotate, zoom, and pan the viewpoint with the mouse to observe various details of the connection plate. It supports displaying boundary conditions and load application locations to help users understand the simulation settings. For stress and strain results, a contour plot is used for coloring, with colors ranging from blue to red to indicate values ​​from low to high. The contour plot range and color mapping can be adjusted, and stress concentration areas are highlighted. For changes in performance indicators with optimization iterations, a line graph is used, with the horizontal axis representing the number of iterations and the vertical axis representing the indicator value. Multiple indicators can be overlaid on the same graph to facilitate observation of the trade-offs between objectives. For the Pareto front solution set output in step 3.7, a scatter plot is used to display it in a 2D or 3D target space. Users can click on a scatter point to view detailed information about the corresponding scheme. It supports exporting the 3D model file and performance report document of the scheme to obtain the visualization interface.

[0069] Through multi-dimensional visualization, the aforementioned functional modules are organically integrated to obtain an intelligent design solution for automotive connection boards. This solution realizes intelligent design capabilities throughout the entire process, from historical data learning, configuration generation, parameter optimization to performance prediction.

[0070] A machine learning-based intelligent design system for automotive connection boards, such as Figure 3 As shown, a machine learning-based intelligent design method for automotive connection boards, as described above, includes: The multi-source heterogeneous data fusion module is used to collect historical design data of automotive connection boards, perform data preprocessing and multi-source heterogeneous feature fusion, and obtain a unified representation vector of multi-source heterogeneous data. The intelligent configuration generation module is used to obtain the user's design requirement condition vector, combine it with the unified representation vector of multi-source heterogeneous data, perform initial configuration generation and topology optimization coupling, and obtain initial geometric configuration candidate schemes. The multi-objective collaborative optimization module extracts design parameters based on the initial geometric configuration candidate schemes, constructs an adaptive multi-objective reward function to optimize the design parameters, and obtains the Pareto front solution set; The adaptive agent modeling module constructs an agent model based on a unified representation vector of multi-source heterogeneous data, performs agent model training and online updates, and obtains an adaptively optimized agent model. The multi-condition performance prediction module, based on an adaptive optimization surrogate model, performs multi-condition performance prediction and evaluation on the service process of automotive connecting plates, and obtains quantitative analysis results of the coupling relationship between conditions. An integrated visualization implementation module is developed and visualized based on a unified representation vector of multi-source heterogeneous data, candidate schemes for initial geometric configurations, Pareto front solution set, surrogate model for adaptive optimization, and quantitative analysis results of coupling relationships between working conditions, resulting in an intelligent design scheme for automotive connection plates.

[0071] In one embodiment of the present invention, a specific example is provided: Taking a mid-to-large-sized pure electric SUV as an example, this paper demonstrates the practical application process and effects of intelligent design methods.

[0072] The vehicle's battery pack has a capacity of 100 kWh and is located under the chassis. The total weight of the battery pack is 600 kg. The subframe connecting plate needs to transfer the weight of the battery pack to the longitudinal beams of the vehicle body while ensuring collision safety performance. The design requirements are: the connecting plate weight does not exceed 3.5 kg; the maximum stress under the combined conditions of battery pack gravity load and road excitation does not exceed 80% of the material's yield strength (i.e., a safety factor of not less than 1.25); the first-order modal frequency is not lower than 80 Hz to avoid resonance; the intrusion of the passenger compartment under frontal collision conditions does not exceed 50 mm; and the manufacturing cost does not exceed the budget.

[0073] Using the intelligent design system, the boundary condition data of the vehicle model is imported, including the battery pack installation interface location, suspension connection point coordinates, and body longitudinal beam interface dimensions. The system automatically retrieves design cases of similar models from the historical design database as references to obtain the initial dataset. The feature engineering module is started to extract multi-source heterogeneous features, and the generative adversarial network is started to generate 5 initial configuration candidate schemes. These schemes present different forms in terms of stiffener layout and hole distribution. The two schemes with the highest topological rationality scores are selected to enter the parameter optimization stage.

[0074] Deep reinforcement learning was employed for multi-objective parameter optimization, setting four optimization objectives: lightweighting, strength, stiffness, and cost. The optimization process iterated for 200 steps, with each step calling a neural network surrogate model for performance prediction. Finite element simulation verification was triggered at 10 key points with high uncertainty, ultimately yielding 12 Pareto front solutions. Engineers viewed the 3D models and performance indicators of each solution in a visualization interface and selected solution A, which had the best overall performance, as the final design.

[0075] The main parameters of Scheme A are: main plate thickness 2.8 mm, stiffener height 15 mm, stiffener thickness 2.5 mm, material is high-strength steel with a yield strength of 500 MPa, and total mass of connecting plates 3.35 kg. Through high-precision finite element simulation verification, Scheme A has a maximum stress of 385 MPa under static load conditions, a safety factor of 1.30, a first-order modal frequency of 85 Hz, and a passenger compartment intrusion of 42 mm under frontal collision conditions. All performance indicators meet the design requirements.

[0076] To illustrate the obtained data, the key parameters and performance indicators of five design schemes are shown in Table 1: Table 1: Key parameters and performance indicators of the five design schemes; Table 2 shows the performance of five design schemes under different operating conditions: Table 2: Performance of the 5 design schemes under different working conditions; As can be seen from the application examples, the intelligent design method can achieve lightweight design of the connecting plate while meeting the performance constraints of multiple operating conditions. Compared with the traditional method, it shortens the design cycle, improves the design quality, and provides strong technical support for the development of new energy vehicle products.

[0077] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for intelligent design of a vehicle connection plate based on machine learning, characterized in that, Includes the following steps: Historical design data of automotive connection plates are collected, and data preprocessing and multi-source heterogeneous feature fusion are performed to obtain a unified representation vector of multi-source heterogeneous data. Obtain the user's design requirement condition vector, combine it with the unified representation vector of multi-source heterogeneous data, perform initial configuration generation and topology optimization coupling, and obtain initial geometric configuration candidate schemes; Design parameters are extracted based on the initial geometric configuration candidate schemes, and an adaptive multi-objective reward function is constructed to optimize the design parameters, thereby obtaining the Pareto front solution set; A proxy model is constructed based on a unified representation vector of multi-source heterogeneous data. The proxy model is trained and updated online to obtain an adaptively optimized proxy model. Based on an adaptive optimization surrogate model, the performance of automotive connecting plates under multiple operating conditions is predicted and evaluated, and quantitative analysis results of the coupling relationship between operating conditions are obtained. Based on the unified representation vector of multi-source heterogeneous data, candidate schemes of initial geometric configuration, Pareto front solution set, surrogate model of adaptive optimization, and quantitative analysis results of coupling relationship between working conditions, an integrated development and visualization implementation is carried out to obtain the intelligent design scheme of automotive connection plate. 2.The method of claim 1, wherein, The data preprocessing includes: The original dataset is constructed by collecting historical design data of automotive connection plates. A data integrity detection algorithm is used to scan each data field of each sample in the original dataset one by one, and the proportion of missing values, data type consistency and value range rationality are statistically analyzed to obtain a data quality assessment report. For missing fatigue test data, a collaborative filtering method based on similar samples is used for intelligent filling. By calculating the cosine similarity between the sample to be filled and other complete samples in terms of geometric features and material properties, the neighboring sample with the highest similarity is selected, and the weighted average of its fatigue test data is taken as the filling value. The isolated forest algorithm is used for outlier detection. The performance index of each sample is used to form a feature vector. Multiple isolated trees are constructed by randomly dividing the feature space. The average path length required for each sample to be isolated is calculated, and samples that are incorrectly identified are removed.

3. The intelligent design method for automotive connecting plates based on machine learning according to claim 1, characterized in that, The multi-source heterogeneous feature fusion includes: A mesh simplification algorithm is used to convert the 3D geometric model into a graph structure representation. Feature points of the geometric model are extracted as graph nodes. Edges are established between connected feature points according to geometric topology. Each node is assigned spatial coordinates, local curvature and the number of adjacent edges. Graph convolutional neural networks are used for node feature learning. Multiple graph convolutional layers are constructed so that the final embedding of each node contains topological features within a multi-order neighborhood range. Global average pooling is used to average the embeddings of all nodes to obtain a global topological feature vector. A multimodal feature fusion network is used for high-dimensional unified representation learning. The topological feature vector, mechanical response feature vector and structured feature vector are nonlinearly transformed by independent fully connected layers. The correlation weights between modes are calculated by a self-attention layer. The weighted modal features are then concatenated to obtain a unified representation vector of multi-source heterogeneous data.

4. The intelligent design method for automotive connecting plates based on machine learning according to claim 1, characterized in that, The initial configuration generation and topology optimization coupling process includes: The user's design requirement condition vector is obtained, and a generative adversarial network is constructed to generate the initial configuration. It includes two sub-networks: a generator and a discriminator. The input of the generator is a concatenation of a unified representation vector of multi-source heterogeneous data, the user's input design requirement condition vector, and a random noise vector. A multi-layer deconvolutional neural network is used to map low-dimensional vectors to high-dimensional topological density fields. The output value at each position is mapped to a preset range to represent the material density through multiple deconvolutional blocks and an activation function. Based on the thickness range of the connecting plate specified in the user's design requirements, the density map is stretched along the vertical direction to form a three-dimensional voxel model. The isosurface of the density threshold is extracted using the moving cube algorithm to generate a preliminary three-dimensional mesh model.

5. The intelligent design method for automotive connecting plates based on machine learning according to claim 4, characterized in that, The generative adversarial network also includes: A discriminator network is constructed to evaluate the authenticity and physical rationality of the initial configuration. The discriminator adopts a dual-branch structure, with the first branch being the authenticity discrimination branch and the second branch being the physical constraint discrimination branch. The two branches share the first few layers of convolutional feature extraction layers. In the physical constraint discrimination branch, the compliance minimization criterion of topology optimization is introduced as a physical constraint to perform fast finite element analysis on the input three-dimensional mesh model. The displacement field is solved under given load and boundary conditions, the total compliance value of the structure is calculated, and the stiffness score is obtained after normalizing the compliance value. An alternating training strategy is adopted to optimize the generator and discriminator. In each training iteration, the generator parameters are fixed and the discriminator is trained for several steps, while the discriminator parameters are fixed and the generator is trained, so that the generator and discriminator reach Nash equilibrium.

6. The intelligent design method for automotive connecting plates based on machine learning according to claim 1, characterized in that, The optimization of design parameters by constructing an adaptive multi-objective reward function includes: The design parameter optimization problem of the connecting plate is modeled as a Markov decision process, and the state space is defined to include two parts: the parameter state and the performance state of the current design. A dual-deep Q-network is constructed as a value function approximator, consisting of two neural networks with identical structures: a current network and a target network. The current network is used to select actions, and the target network is used to calculate the target Q-value. An adaptive multi-objective reward function is designed, which comprehensively considers the lightweight objective, strength constraints, stiffness constraints, cost constraints, and manufacturing process constraints, and dynamically adjusts the weights of each objective based on the position of the current solution in the objective space.

7. The intelligent design method for automotive connecting plates based on machine learning according to claim 6, characterized in that, The adaptive multi-objective reward function includes: Define the normalized reward components for each single objective: the lightweight reward is defined as the reduction ratio of the current mass to the initial mass, the strength reward is defined as a normalized function of the safety factor, the stiffness reward is defined as a normalized function of the maximum displacement, and the cost reward is defined as the percentage of manufacturing cost savings relative to the budget. Increase the reward weight of manufacturing process constraints, conduct manufacturing feasibility assessment of the current design scheme, stamping constraints are achieved by detecting the draft angle of the connecting plate surface, and welding assembly constraints are achieved by detecting the interface matching degree between the connecting plate and adjacent parts; Design adaptive weighting coefficients to dynamically adjust the weight of each objective based on the current solution's position in the objective space.

8. The intelligent design method for automotive connecting plates based on machine learning according to claim 1, characterized in that, The construction of the proxy model, the training of the proxy model, and the online updating of the proxy model include: The Latin hypercube sampling method was used to select representative sample points. For each sample point, a three-dimensional model of the connecting plate was constructed based on the parameter values ​​of each sample point. The model was then imported into finite element analysis software for static structural analysis and modal analysis. A deep feedforward neural network is constructed as a surrogate model, with design parameters as input and performance index vector as output; An uncertainty estimation mechanism is introduced, and the standard deviation of the prediction results is calculated as an uncertainty index. When the uncertainty exceeds a preset threshold, finite element simulation is triggered for evaluation.

9. The intelligent design method for automotive connecting plates based on machine learning according to claim 1, characterized in that, The process of training and updating the proxy model online also includes: When a new sample is obtained by triggering a finite element simulation, the new sample is added to the training set of the surrogate model, and the surrogate model is retrained periodically. The update process adopts an incremental learning strategy, which fine-tunes the original model parameters using new samples. The learning rate is set to a preset proportion of the initial learning rate, and the number of training cycles is set to a preset value. Through incremental learning, the surrogate model absorbs information from new samples and gradually expands the range of the parameter space.

10. A machine learning-based intelligent design system for automotive connection plates, characterized in that, A machine learning-based intelligent design method for automotive connection plates, as described in any one of claims 1-9, includes: The multi-source heterogeneous data fusion module is used to collect historical design data of automotive connection boards, perform data preprocessing and multi-source heterogeneous feature fusion, and obtain a unified representation vector of multi-source heterogeneous data. The intelligent configuration generation module is used to obtain the user's design requirement condition vector, combine it with the unified representation vector of multi-source heterogeneous data, perform initial configuration generation and topology optimization coupling, and obtain initial geometric configuration candidate schemes. The multi-objective collaborative optimization module extracts design parameters based on the initial geometric configuration candidate schemes, constructs an adaptive multi-objective reward function to optimize the design parameters, and obtains the Pareto front solution set; The adaptive agent modeling module constructs an agent model based on a unified representation vector of multi-source heterogeneous data, performs agent model training and online updates, and obtains an adaptively optimized agent model. The multi-condition performance prediction module, based on an adaptive optimization surrogate model, performs multi-condition performance prediction and evaluation on the service process of automotive connecting plates, and obtains quantitative analysis results of the coupling relationship between conditions. An integrated visualization implementation module is developed and visualized based on a unified representation vector of multi-source heterogeneous data, candidate schemes for initial geometric configurations, Pareto front solution set, surrogate model for adaptive optimization, and quantitative analysis results of coupling relationships between working conditions, resulting in an intelligent design scheme for automotive connection plates.

Images