A Multidisciplinary Optimization Method and System for Vehicle Collision and Lightweighting Based on Agent Model
By employing a surrogate model-based multidisciplinary optimization method for vehicle collision and lightweighting, the problems of high computational cost and difficulty in multidisciplinary collaboration in vehicle body structure optimization are solved, achieving a balance between efficient lightweighting and safety performance, and providing scientific design guidance.
Patent Information
- Application Number
- CN202611602816.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-16
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies face challenges in optimizing vehicle body structures due to high computational costs, difficulties in multidisciplinary collaboration, and limited design space, making it difficult to achieve an effective balance between lightweighting and safety performance.
A multidisciplinary optimization method for vehicle collision and lightweighting based on a surrogate model is adopted. By training the surrogate model, a multidisciplinary optimization framework is constructed, design variables are unified and cross-disciplinary collaborative evaluation is carried out, and the optimization problem is solved by combining a multi-objective evolutionary algorithm, which supports flexible trade-offs between lightweighting and safety performance.
It significantly reduces computing costs, improves optimization efficiency, enables multidisciplinary collaborative optimization, supports a flexible trade-off between lightweight design and security performance, provides scientific design guidance, avoids design blind spots, and shortens the development cycle.
Smart Images

Figure CN123116872A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of automotive body structure optimization and vehicle computer-aided engineering simulation technology, and in particular relates to a multidisciplinary optimization method and system for vehicle body collision and lightweighting based on a proxy model. Background Technology
[0002] With the rapid development of the automotive industry, the passive safety performance and lightweighting level of the vehicle body structure have become important indicators for measuring the overall performance of the vehicle. In the early structural forward design of passenger car body-in-white, it is usually necessary to conduct simulation analysis of safety conditions such as frontal rigid barrier collision (FRB) and moving barrier collision (MDB) to evaluate the crashworthiness and occupant protection capabilities of the vehicle body, while also taking into account the lightweighting goal of reducing the weight of the vehicle body.
[0003] Existing vehicle body safety and lightweighting optimization typically employs a technical approach of "Design of Experiments (DOE) - Finite Element Simulation - Manual Post-processing - Single-condition Local Optimization". Specifically, engineers first set samples for geometric position, plate thickness, and material combination, and then run whole-vehicle finite element simulations of FRB frontal rigid wall collision and MDB moving barrier collision, respectively. The results, such as intrusion amount, velocity, acceleration, and mass, are then exported as tagged files for analysis.
[0004] However, the aforementioned existing technologies have the following limitations in practical applications: First, the computational cost is high. A single FRB or MDB full-vehicle finite element explicit dynamics simulation typically requires several hours or even tens of hours of computation time, and simulation verification must be performed again after each design change, resulting in a lengthy development cycle and making it difficult to meet the needs of rapid vehicle model iteration.
[0005] Secondly, multidisciplinary collaboration is difficult. When considering multiple disciplines' objectives such as FRB, MDB, and lightweighting, traditional methods require iterative iterations and trade-offs across multiple operating conditions, resulting in extremely low optimization efficiency and making it difficult to achieve an effective balance between lightweighting and safety performance.
[0006] Finally, the design space exploration is limited. Due to the cost of simulation calculations, engineers can usually only perform local searches on a small number of design variables, making it difficult to find the optimal solution in the global design space. This can easily lead to getting stuck in local optima and prevent the full exploitation of the vehicle body structure's performance potential.
[0007] Therefore, how to provide a vehicle body structure optimization method and system that can significantly reduce computational costs, achieve efficient collaborative optimization across multiple operating conditions and disciplines, and quickly find the optimal solution within the global design space is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0008] In view of this, the present invention aims to propose a multidisciplinary optimization method and system for vehicle body collision and lightweighting based on a proxy model, in order to solve the technical problems existing in the prior art, such as high simulation calculation cost, difficulty in multi-condition and multi-disciplinary collaboration, and limited design space exploration that is prone to getting trapped in local optima.
[0009] To achieve the above objectives, the technical solution of the present invention is implemented as follows: This invention provides a multidisciplinary optimization method for vehicle body collision and lightweighting based on a surrogate model, characterized by the following steps: Step S1, Data preparation and preprocessing: Read the design input feature files and label files for frontal rigid wall collision and moving barrier collision respectively, match samples according to vehicle type, filter valid label columns and remove abnormal samples, and divide the training set, validation set and test set. Step S2, Proxy Model Training and Loading: Train or load proxy models for frontal rigid wall collision and moving barrier collision, verify the input and output dimensions of the model, and configure the inverse normalization method. Step S3, Multidisciplinary Optimization Framework Construction and Design Variable Analysis: Scan the input header, identify continuous variables, thickness column and material column, combine thickness variable and material grade variable to form discrete combination variable, remove duplicates of the same variable to form the union of cross-disciplinary design variables, and reconstruct the exclusive input vector according to the original header of each discipline; Step S4, interdisciplinary collaborative evaluation: The same candidate design is fed into the surrogate models of frontal rigid wall collision and moving barrier collision respectively, the predicted response is restored to physical dimensions, the response direction is unified, and the global quality is formed by the maximum value of the predicted quality of the two disciplines. Step S5, Multi-objective optimization solution: Establish objective function and inequality constraints, including fixed-value performance constraints and cross-disciplinary quality consistency constraints, and use multi-objective evolutionary algorithm to obtain candidate solution set; Step S6, Result Screening and Sensitivity Analysis: Recommended schemes are screened from the last generation population according to the priority of feasibility, constraint violation, and objective function, and sensitivity analysis is performed on the recommended schemes.
[0010] Furthermore, in step S1, the input vector for frontal rigid wall collision is 79-dimensional, and the input vector for moving barrier collision is 83-dimensional; a fixed random seed is used to divide the training set, validation set, and test set. In step S2, a multilayer perceptron, a one-dimensional convolutional neural network, a bidirectional long short-term memory network, and a self-attention model are selected for parallel training, and a surrogate model is selected based on the performance on the validation set.
[0011] Furthermore, in step S3, 53 actual design variables for frontal rigid wall collision and 57 actual design variables for moving barrier collision are obtained through analysis. In the discrete combination variables, each thickness-material legal pair is treated as an indivisible discrete unit, and the thickness value is mapped to the material grade through a combination index. For geometrically or positionally continuous variables, normalization is performed using the center and half-range to preserve the input scale, in order to handle extreme values in the data distribution and improve the accuracy of physical quantity recovery.
[0012] Furthermore, in step S4, the frontal rigid wall collision and moving barrier collision responses are unified into an optimization direction where smaller values are better. global quality ,in and The quality of the predictions for the two disciplines are respectively; In interdisciplinary collaborative assessments, each discipline uses its own independent center and half-range to recover to the physical domain, and the cross-use of scale parameters is prohibited.
[0013] Furthermore, in step S5, the interdisciplinary quality consistency constraint limits the absolute value of the difference between the predicted quality of the two disciplines to no more than a preset tolerance. Specifically, it is set as a two-sided inequality constraint; The multi-objective evolutionary algorithm selects the NSGA-II algorithm when the number of objectives is less than or equal to 5, and selects the NSGA-III algorithm when the number of objectives is greater than 5. The method supports a minimum mass mode and a target tracking mode. In the target tracking mode, a dimensionless deviation target is calculated and upper and lower boundary constraints are generated.
[0014] Furthermore, in step S6, when selecting recommended solutions, a three-level sorting rule is adopted: the first level prioritizes feasible solutions, the second level prioritizes solutions with lower constraint violation, and the third level prioritizes solutions with lower objective function sum. The sensitivity analysis employs the Morris method, which perturbs the continuous variables and discrete combination indices, calculates the absolute mean and standard deviation of the basic effects, and outputs a variable-response ranking table.
[0015] Furthermore, in the interdisciplinary collaborative evaluation in step S4 and the multi-objective optimization solution in step S5, batch evaluation technology is adopted to package multiple candidate designs in the population and input them into the surrogate model in parallel, and to accelerate parallel computing using a graphics processor.
[0016] Based on the same concept, the present invention also provides a multidisciplinary optimization system for vehicle collision and lightweighting based on a surrogate model, including: a data preparation module, a surrogate model training and loading module, a design variable analysis and coupling module, a cross-disciplinary collaborative evaluation module, a multi-objective optimization module, and a result screening and sensitivity analysis module. The data preparation module is used to read the design inputs and labels of frontal rigid wall collision and moving barrier collision respectively, align them by samples and divide the dataset; The proxy model training and loading module is used to train or load proxy models for two disciplines and verify the dimensions. The design variable analysis and coupling module is used to identify continuous variables and thickness-material discrete combination variables, form a cross-disciplinary variable union and reconstruct the input vector; The interdisciplinary collaborative evaluation module is used to feed candidate designs into two surrogate models respectively, unify the response direction, and form the global quality by the maximum value of the predicted quality of the two disciplines. The multi-objective optimization module is used to establish the objective function and the inequality constraints that include cross-disciplinary quality consistency constraints, and to call the multi-objective evolutionary algorithm to solve it; The result screening and sensitivity analysis module is used to screen recommended schemes according to priority and perform sensitivity analysis.
[0017] Based on the same concept, the present invention also provides an electronic device, including a memory and a processor, characterized in that the memory stores a computer program, and the processor executes the computer program to implement the aforementioned surrogate model-based multidisciplinary optimization method for vehicle collision and lightweighting.
[0018] Based on the same concept, the present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the aforementioned multidisciplinary optimization method for vehicle collision and lightweighting based on a proxy model.
[0019] Compared with existing technologies, the multidisciplinary optimization method and system for vehicle collision and lightweighting based on surrogate models described in this invention has the following beneficial effects: 1. Significantly reduces computational costs and improves optimization efficiency. This invention replaces time-consuming finite element simulation with a deep learning proxy model, achieving millisecond-level performance prediction. This allows multidisciplinary optimization to be completed within an acceptable timeframe, improving optimization efficiency by several orders of magnitude and effectively shortening the development cycle of vehicle body structures.
[0020] 2. Achieving true multidisciplinary collaborative optimization. This invention establishes an integrated multidisciplinary optimization architecture, unifies the three types of design variables—geometry, plate thickness, and materials—and adds cross-disciplinary quality consistency constraints. It unifies the modeling of two safety disciplines—frontal rigid wall collision and moving barrier collision—with the lightweight objective, overcoming the shortcomings of traditional methods in collaborative iteration.
[0021] 3. Supports flexible trade-offs between lightweight design and safety performance. This invention combines multi-objective optimization algorithms to quickly solve for Pareto optimal solutions and provides Pareto solution sets. This allows engineers to flexibly balance vehicle lightweighting and safety performance according to actual engineering needs, ensuring efficient lightweight design of the vehicle body while ensuring that the overall vehicle crash safety performance meets standards.
[0022] 4. Provides scientific design guidance to avoid design blind spots. This invention supports fixed-value target constraints to meet specific performance requirements, and combined with parameter sensitivity analysis, it can help engineers accurately identify key design variables and conflicting parameters, preventing them from getting trapped in local optima, and providing scientific guidance for subsequent structural design adjustments and improvements. Attached Figure Description
[0023] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the data preparation and label selection process in the multidisciplinary optimization method for vehicle collision and lightweighting based on the proxy model provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the multidisciplinary optimization overall architecture provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the definition of a multi-objective optimization problem provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the sensitivity analysis process provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the Pareto solution output process provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the overall architecture of the FRB-MDB multidisciplinary optimization platform provided in this embodiment of the invention; Figure 7 This is a schematic diagram of the FRB-MDB data preparation module provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the FRB-MDB proxy model training and loading module provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the FRB-MDB design variable parsing and coupling module provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of the FRB-MDB interdisciplinary collaborative assessment module provided in an embodiment of the present invention; Figure 11 This is a schematic diagram of the FRB-MDB multi-objective optimization and constraint module provided in an embodiment of the present invention; Figure 12 This is a schematic diagram of the FRB-MDB result filtering and output module provided in an embodiment of the present invention; Figure 13 This is a schematic diagram of the FRB-MDB Morris sensitivity module provided in an embodiment of the present invention. Detailed Implementation
[0024] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.
[0025] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.
[0026] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0027] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0028] like Figures 1 to 13 As shown, a multidisciplinary optimization method and system for vehicle collision and lightweighting based on a surrogate model is presented. The multidisciplinary optimization method for vehicle collision and lightweighting based on a surrogate model includes the following steps: Step S1, Data preparation and preprocessing: Read the design input feature files and label files for frontal rigid wall collision and moving barrier collision respectively, match samples according to vehicle type, filter valid label columns and remove abnormal samples, and divide the training set, validation set and test set. Step S2, Proxy Model Training and Loading: Train or load proxy models for frontal rigid wall collision and moving barrier collision, verify the input and output dimensions of the model, and configure the inverse normalization method. Step S3, Multidisciplinary Optimization Framework Construction and Design Variable Analysis: Scan the input header, identify continuous variables, thickness column and material column, combine thickness variable and material grade variable to form discrete combination variable, remove duplicates of the same variable to form the union of cross-disciplinary design variables, and reconstruct the exclusive input vector according to the original header of each discipline; Step S4, interdisciplinary collaborative evaluation: The same candidate design is fed into the surrogate models of frontal rigid wall collision and moving barrier collision respectively, the predicted response is restored to physical dimensions, the response direction is unified, and the global quality is formed by the maximum value of the predicted quality of the two disciplines. Step S5, Multi-objective optimization solution: Establish objective function and inequality constraints, including fixed-value performance constraints and cross-disciplinary quality consistency constraints, and use multi-objective evolutionary algorithm to obtain candidate solution set; Step S6, Result Screening and Sensitivity Analysis: Recommended schemes are screened from the last generation population according to the priority of feasibility, constraint violation, and objective function, and sensitivity analysis is performed on the recommended schemes.
[0029] The surrogate model-based multidisciplinary optimization system for vehicle collision and lightweighting includes: a data preparation module, a surrogate model training and loading module, a design variable analysis and coupling module, a cross-disciplinary collaborative evaluation module, a multi-objective optimization module, and a result screening and sensitivity analysis module. The data preparation module is used to read the design inputs and labels of frontal rigid wall collision and moving barrier collision respectively, align them by samples and divide the dataset; The proxy model training and loading module is used to train or load proxy models for two disciplines and verify the dimensions. The design variable analysis and coupling module is used to identify continuous variables and thickness-material discrete combination variables, form a cross-disciplinary variable union and reconstruct the input vector; The interdisciplinary collaborative evaluation module is used to feed candidate designs into two surrogate models respectively, unify the response direction, and form the global quality by the maximum value of the predicted quality of the two disciplines. The multi-objective optimization module is used to establish the objective function and the inequality constraints that include cross-disciplinary quality consistency constraints, and to call the multi-objective evolutionary algorithm to solve it; The result screening and sensitivity analysis module is used to screen recommended schemes according to priority and perform sensitivity analysis.
[0030] The specific implementation method is as follows: For ease of description, in this specification, frontal rigid wall collision is abbreviated as FRB, and moving barrier collision is abbreviated as MDB.
[0031] Step S1, Data Preparation and Preprocessing like Figure 1 and Figure 7 As shown, this step aims to provide a high-quality dataset for subsequent model training. This step specifically includes the following sub-steps: First, the comma-separated input feature file is read. This input feature file contains vehicle body geometric parameters, plate thickness parameters, and material property parameters. The vehicle body geometric parameters include cross-sectional dimensions, structural locations, etc., corresponding to the geometric or position columns in the data table. Geometric columns typically represent geometric shape parameters, such as the cross-sectional height, width, and stiffener locations of beams. Position columns typically represent structural position parameters, such as the longitudinal coordinates of crossbeams and the bending angle of longitudinal beams. The plate thickness parameters correspond to the thickness column, representing the thickness of various body panels and structural components. The material property parameters correspond to the material column, representing the material's yield strength, tensile strength, elastic modulus, and other mechanical properties, usually indicated by material grade or code.
[0032] Secondly, the spreadsheet tag file is read. This tag file contains key performance indicators such as cross-sectional force, displacement, velocity, and energy. These indicators are the results of finite element simulation calculations, reflecting the dynamic response of the vehicle body during the collision. For elongation and displacement outputs, only columns ending with the character "final" are retained as valid tags to characterize the final state at the moment the collision ends. This is because displacement and elongation change over time during the collision, while optimization design typically focuses on the residual deformation after the collision, i.e., the final state, which directly determines the integrity of the occupant survival space.
[0033] Subsequently, samples are matched by vehicle type. The system matches samples based on vehicle type (sedan or SUV) and sample numbering rules to ensure a one-to-one correspondence between input features and output labels. For the FRB and MDB datasets, the input file line number is calculated using the vehicle number and sequence number in the sample folder, ensuring that each line of design variables aligns with the performance response of the corresponding operating condition. Since different vehicle types have significantly different body structures, processing by vehicle type can improve the prediction accuracy of the surrogate model.
[0034] Next, the valid label columns are filtered. For elongation and displacement outputs, only columns ending with the character "final" are retained as valid labels. For example, the final value column for each spring elongation. In crash safety assessments, the final intrusion amount determines the ease of rescue; therefore, retaining these key columns is crucial for accurately assessing safety.
[0035] Finally, null samples and outlier handling are performed. The system removes samples containing null or missing values, performing outlier detection and handling to ensure the quality of the training data. Outliers may be caused by simulation non-convergence or mesh distortion; removing these samples prevents the surrogate model from learning incorrect physical laws.
[0036] like Figure 7 As shown, upon system startup, the system checks the project directory, input files, label files, weight files, and normalization files item by item. After reading the FRB and MDB design input files, it retains the original table headers, column names, and fixed column order. It then checks if the FRB input has 79 columns and the MDB input has 83 columns; if not, it outputs the actual number of columns and stops. If they are, it reads the label tables for both subjects and aligns the input and output according to the sample numbers. Missing, non-numerical labels, and identified anomalous samples are then deleted, forming the FRB and MDB datasets. Finally, a fixed random seed of 2025 is used, and the training, validation, and test sets are divided into 70%, 20%, and 10% groups, respectively. The datasets, table header order, and independent normalization parameters are then output.
[0037] Step S2, Proxy Model Training and Loading like Figure 8 As shown, this step uses multiple neural network models for parallel training, selects the best surrogate model based on the performance on the validation set, and configures an inverse normalization method to map the prediction results back to the physical quantity domain.
[0038] In this embodiment, four types of models—multilayer perceptron, one-dimensional convolutional neural network, bidirectional long short-term memory network, and self-attention model—are selected for parallel training to cover different types of nonlinear relationships in the input features. Complex coupling effects exist among the vehicle body structural parameters, making it difficult for a single model to capture all features.
[0039] The multilayer perceptron (MLP) model is suitable for establishing a fundamental nonlinear mapping from high-dimensional parameters to performance responses. This model consists of an input layer, two hidden layers with 512 and 256 neurons respectively, employing a modified linear unit activation function, including a dropout regularization layer to prevent overfitting, and an output layer. Through multiple layers of nonlinear transformations, the MLP can fit arbitrarily complex functional relationships.
[0040] One-dimensional convolutional neural network (CNN) models are suitable for capturing the combination patterns between locally adjacent features in an input column. This model includes an up-dimensional linear layer, two one-dimensional convolutional layers with max-pooling, a flattening layer, and a fully connected output layer, making it suitable for extracting local correlations of input features. In vehicle body design, adjacent components often share loads; CNNs, through sliding windows of the convolutional kernels, can effectively extract these local topological features.
[0041] The bidirectional Long Short-Term Memory (LSTM) network model employs a bidirectional structure to extract temporal dependencies from input features. It is suitable for parameter inputs with sequential characteristics and is beneficial for expressing feature dependencies arranged along part or structural paths. LTM networks control information flow through gating mechanisms, effectively addressing the gradient vanishing problem in long sequence data.
[0042] The self-attention model is based on a transformer encoder structure and includes a multi-head self-attention mechanism, residual connections, layer normalization, and feedforward networks. It can capture global dependencies between input features and capture global correlations across components and variable types through multi-head self-attention.
[0043] like Figure 8 As shown in the flowchart, first select either "Retrain" or "Reload". If retraining is selected, a candidate deep network is constructed, and the loss function, optimizer, and training batch are set. The training process uses an adaptive moment estimation optimizer with an initial learning rate of 0.0001 and a weight decay coefficient of 0.0001. The batch size is 16 for the FRB case and 8 for the MDB case. The training epochs are 100, gradient clipping is used, and the maximum norm is set to 1.0 to prevent gradient explosion. Backpropagation is then performed to update the parameters, the optimal weights are saved using the validation loss, and the error is output on the test set. After training, the prefix of the parallel computing module name is cleaned and the weights are loaded. If direct loading is selected, the model is instantiated and the weight file is read.
[0044] Next, dimension verification is performed. The FRB model is checked to see if it satisfies 79 inputs and 15 outputs; otherwise, a weight mismatch is reported. If it does, the MDB model is checked to see if it satisfies 83 inputs and 15 outputs; otherwise, a weight mismatch is reported. After successful verification, both models are switched to evaluation mode and gradients are disabled. The model, input header, output name, direction, and denormalization parameters are then registered.
[0045] The inverse normalization method reads the normalization upper and lower bounds table and performs a linear mapping from the negative 1 to 1 interval to the physical quantity domain by matching column names. The inverse normalization formula is as follows: (1); in, Represents physical values. Indicates the lower limit. Indicates the predicted value. Indicates the upper limit.
[0046] Step S3: Construction of Multidisciplinary Optimization Framework and Analysis of Design Variables Based on the high-quality dataset partitioned in step S1 and the high-precision surrogate model trained in step S2, this step first defines the design variables, then inputs them into the FRB subject component and the MDB subject component respectively, and finally converges them into the unified output of the quality model.
[0047] like Figure 9 As shown in the flowchart, the FRB 79 list header and MDB 83 list header are scanned first. Continuous columns, thickness columns, and material columns are identified. Actual variables are mapped using a mapping dictionary. This process yields 53 actual variables for the FRB and 57 actual variables for the MDB. Subsequently, duplicate variables with the same name are removed, and a cross-disciplinary variable union is established. The registration type, upper and lower bounds, combination sets, and placement rules are then defined.
[0048] Determine if the current variable is a discrete combination. If yes, decode both thickness and material using the combined index. If not, write it to a continuous column. For uncovered input columns, retain the baseline values for each discipline. Finally, construct the input vector in a fixed order and output the joint design variable definition and the dual-discipline input tensor.
[0049] The design variable analysis is based on the actual input headers of two disciplines. The first category consists of geometric or positional continuous variables, which are formed into real-valued optimization variables based on configured physical upper and lower bounds. The second category consists of thickness-material combination variables, where each combination is represented by an integer index for a valid material-thickness pair. The third category consists of material input columns, which share the same combination index with the corresponding thickness column.
[0050] Furthermore, each thickness-material valid pair is treated as an indivisible discrete unit, and the composite index is mapped to engineering parameters according to the following formula: (2); in, Indicates the first Thickness - Material legal pairing, This indicates the corresponding thickness value. Indicates the corresponding material grade. For composite indexes, This represents the total number of valid pairings.
[0051] During input vector reconstruction, for geometric or positional columns, normalization is performed to the interval between -1 and 1 based on the upper and lower limits of the geometric boundaries. The normalization formula is as follows: (3); in, This represents the normalized value. Represents the original value. Indicates the lower limit. This represents the upper limit. Normalization to the interval between -1 and 1, where the gradient is symmetrically distributed, helps the gradient propagation of the neural network stably, prevents gradient vanishing or exploding, and thus accelerates model convergence.
[0052] Furthermore, the platform preserves the input scale using the center and half-range, and performs normalization and physical quantity recovery according to the formula: (4); in, Indicates the first The center value of the column, Indicates the upper limit value. Indicates the lower limit value. Indicates half-range. This represents the normalized value. This represents the original value.
[0053] Step S4, Interdisciplinary Collaborative Assessment Using the bidisciplinary input vector reconstructed in step S3, this step is used to feed the same candidate design into the surrogate models of the two disciplines respectively, restore the predicted response to physical dimensions, unify the response direction, and form the global quality with the maximum value of the predicted quality of the two disciplines.
[0054] like Figure 10 As shown in the flowchart, the system receives joint candidate designs and reconstructs the bidisciplinary input. This is done through two channels: FRB and MDB. A surrogate model is invoked without gradients to obtain standardized responses. Subsequently, the physical responses are reconstructed using their respective independent parameters and prefixed with a discipline name. The system checks for duplicate response names; if duplicates are found, the process is blocked and a configuration error is reported. Finally, the qualities of both disciplines are obtained, and the global quality is calculated. The bidisciplinary objective, constraint values, and global quality are returned.
[0055] The formula for target direction conversion is as follows: (5); in, To unify the target value into the minimized direction, The original network output is the normalized domain. This is the sign direction coefficient. For safety science, The value is +1.
[0056] Furthermore, all participating responses are unified to the direction of minimization through directional coefficients, as shown in the formula: (6); in, Indicate subject The The response value after unifying the direction of the items. For direction coefficients, These are the standardized predictions from the original network output.
[0057] In interdisciplinary collaborative assessment, disciplines The The standardized predictions are recovered to the physical domain using discipline-independent centers and half-ranges, as shown in the formula: (7); in, Indicate subject The The term is restored to the predicted value in the physical quantity domain. This represents the center value of the response. This represents the half-range of the response. Represents standardized predicted values. Represents the FRB or MDB discipline.
[0058] Step S5, Multi-objective optimization solution like Figure 3 and Figure 11 As shown, this step sets optimization objectives including non-quality and safety response objectives and global quality objectives, sets constraints including fixed-value performance constraints and quality consistency constraints, and uses a multi-objective optimization algorithm to solve for Pareto candidate solutions.
[0059] like Figure 11 As shown in the flowchart, the response roles are first read. All participating responses are standardized to "smaller is better" based on their direction coefficients. It is then determined whether the minimum mass mode is in effect. If so, the global mass is set as the primary objective, and response boundaries are configured. If not, the dimensionless deviation of the target is calculated, generating lower and upper boundary constraints for the target tracking response.
[0060] Then, a bilateral constraint for quality consistency is added. It is determined whether the number of targets is greater than 5. If so, the NSGA-III algorithm with reference direction is used; otherwise (i.e., the number of targets is less than or equal to 5), the mixed-variable NSGA-II algorithm is used. After configuration and startup, sampling, joint evaluation, feasibility selection, crossover, and mutation are performed cyclically.
[0061] The formula for normalizing quality objectives is as follows: (8); in, Indicates the normalized quality. Indicates the original mass. Indicates the lower limit of quality. Indicates the upper limit of quality.
[0062] The formula for the constant value performance constraint is as follows: (9); and (10); in, This represents the actual predicted value of any performance index. This indicates the target value of the indicator. Indicates tolerance.
[0063] For the target tracking response, the constraint formula is as follows: (11); in, Indicates the first The dimensionless bias objective of the response. Indicates the predicted value. Indicates the target value. This represents the tolerance of the physical quantity domain. To prevent the minimum value where the denominator is zero, and These are the lower boundary and upper boundary constraint functions, respectively.
[0064] The formula for quality consistency constraint is as follows: (12); and (13); in, , The predicted quality values for the two disciplines are respectively. Tolerance for quality consistency.
[0065] The formula is as follows when calculating global quality and constraining assembly: (14); in, Indicates global conservative quality. and Let be the quality consistency constraint function. For quality consistency tolerance, where and The quality of the predictions for the two disciplines are respectively.
[0066] Step S6, Result Screening and Sensitivity Analysis like Figure 4 , Figure 5 , Figure 12 and Figure 13 As shown, this step is used to select recommended schemes from the last generation population according to priority and perform sensitivity analysis on the recommended schemes.
[0067] like Figure 12 As shown, read the last generation population data. Calculate the infeasibility flag. Construct the sorting key. Select the first candidate in lexicographical ascending order. Determine if the candidate is feasible; if not, retain the scheme with the minimum constraint violation and explicitly mark it as infeasible; if so, continue decoding the combined index into actual thickness and material, run the joint model again for consistency verification, save the data, and output the recommended scheme.
[0068] like Figure 5As shown, the selection rules employ a three-level sorting. The first level prioritizes feasible solutions. The second level prioritizes solutions with lower constraint violation rates. The third level prioritizes solutions with lower objective function values.
[0069] The recommended selection formula is as follows: (15); in, Represent design variables Total constraint violations For the first A constraint function, For comprehensive sorting key, For indicator functions, For the first The objective function value, This is the final recommended solution. This lexicographical sorting mechanism ensures that in engineering applications, satisfying constraints (feasibility) is the primary prerequisite, followed by performance, which aligns with the logic of actual engineering design.
[0070] like Figure 4 and Figure 13 As shown, the sensitivity analysis employed the Morris method. Figure 13 As shown in the flowchart, the process reads data such as the recommended scheme. A random trajectory is generated centered on the recommended scheme. The variable type is determined: continuous variables are subjected to tiered perturbations, while discrete variables are ablated between indices. Only one variable is changed at a time. The joint evaluation is re-executed. The basic effects and statistics are calculated.
[0071] The formula for calculating sensitivity is as follows: (16); in, Indicates the first The basic effects of each variable Output the response function for the model. For the current design variable vector, For the first The perturbation step size of each variable, For the first A unit vector of 1 dimension For the average absolute basic effect, The standard deviation is denoted as .
[0072] Example 1 In a preferred embodiment of the present invention, a sample of a sports utility vehicle model will be used as an example for specific description.
[0073] In this embodiment, the FRB uses a 79-column input file, and the MDB uses an 83-column input file. The system parses the two headers separately, removes duplicate actual design variables with the same name to form a joint variable space; during each evaluation, the inputs are reconstructed according to their original column order, thus ensuring that the design is coupled but the model inputs are not mixed.
[0074] Next, proxy models are trained separately, and the four network architectures are compared and the best is selected. During the optimization phase, the system reads the upper and lower limits of geometric variables and the list of material thickness combinations, and merges the variable dictionaries into a unique global design space.
[0075] When the optimization algorithm generates a candidate design, it first writes the design into a global variable. The two discipline components then call their respective reconstruction methods to reconstruct the input vectors according to the original header order. Finally, they call their pre-trained models to obtain response predictions. The optimization problem iteratively solves the non-quality and safety responses of the two operating conditions, the global quality objective, and the constraint violation degree. The algorithm then outputs the top N Pareto candidate solutions in the order of feasible solution priority, lower constraint violation degree, and lower objective sum.
[0076] For the candidate solutions output, the system further performs sensitivity analysis to obtain the top few geometric or thickness combinations that have the greatest impact on collision and mass objectives, which engineers can refer to when confirming the final solution and adjusting the structural design.
[0077] Example 2 In another preferred embodiment of the present invention, taking a car model as an example, the number of optimization targets is set to 6 (including FRB intrusion amount, MDB acceleration, global mass, etc.).
[0078] At this point, in step S5, the system automatically determines that the number of targets is greater than 5 and switches to the NSGA-III algorithm with reference directions. The system generates reference direction vectors that match the 6 targets, guiding the population to converge in multiple directions of the Pareto front, effectively solving the problem of sparse population distribution in high-dimensional target space.
[0079] Furthermore, this embodiment adopts a "target tracking mode" and sets the target value of the global mass to 1200 kg with a tolerance of 20 kg. The system calculates the dimensionless deviation using formula (11), so that the optimization result not only pursues the minimum mass, but also gets as close as possible to the target value set by the engineering, avoiding excessive cost or manufacturing difficulties caused by excessive weight reduction.
[0080] In this embodiment, the population evaluation during the surrogate model training and multi-objective optimization process is deployed on computing nodes equipped with high-performance graphics processing units (GPUs). The interdisciplinary collaborative evaluation module utilizes batch evaluation technology to package multiple candidate designs from the population and input them into the surrogate model in parallel. By leveraging the parallel computing capabilities of the GPU, the evaluation time for a single iteration is compressed from hours to milliseconds, significantly improving industrial applicability.
[0081] In another preferred embodiment of the present invention, an electronic device is provided, including a memory and a processor; the memory stores platform programs and configurations, and the processor executes the program to construct and call the modules of the above-mentioned multidisciplinary optimization system and execute the steps of the above-mentioned optimization method.
[0082] In another preferred embodiment of the present invention, a computer-readable storage medium is provided, wherein a program stored thereon, when executed by a processor, constructs modules for data preparation, model loading, variable coupling, collaborative evaluation, multi-objective optimization, and result output, and performs the steps of the above-described optimization method.
[0083] Beneficial effects Through the above embodiments, the present invention has the following beneficial effects: 1. Significantly reduces computational costs and improves optimization efficiency. This invention replaces time-consuming finite element simulation with a deep learning proxy model, achieving millisecond-level performance prediction. This allows multidisciplinary optimization to be completed within an acceptable timeframe, improving optimization efficiency by several orders of magnitude and effectively shortening the development cycle of vehicle body structures.
[0084] 2. Achieving true multidisciplinary collaborative optimization. This invention establishes an integrated multidisciplinary optimization architecture, unifies the three types of design variables—geometry, plate thickness, and materials—and adds cross-disciplinary quality consistency constraints. It unifies the modeling of the two safety disciplines, FRB and MDB, with the lightweight objective, overcoming the shortcomings of traditional methods in collaborative iteration.
[0085] 3. Supports flexible trade-offs between lightweighting and safety. This invention combines multi-objective optimization algorithms to quickly solve for Pareto optimal solutions and provides Pareto solution sets. This allows engineers to flexibly balance vehicle lightweighting and safety performance according to actual engineering needs, ensuring efficient lightweighting design while ensuring the overall vehicle crash safety performance meets standards.
[0086] 4. Provides scientific design guidance. This invention supports setpoint constraints to meet specific performance requirements, and combined with parameter sensitivity analysis, it can help engineers accurately identify key design variables and conflicting parameters, avoiding the blind exploration of the design space, preventing getting trapped in local optima, and providing scientific guidance for subsequent structural design adjustments and improvements.
[0087] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A multidisciplinary optimization method for vehicle body collision and lightweighting based on a surrogate model, characterized in that, Includes the following steps: Step S1, Data preparation and preprocessing: Read the design input feature files and label files for frontal rigid wall collision and moving barrier collision respectively, match samples according to vehicle type, filter valid label columns and remove abnormal samples, and divide the training set, validation set and test set. Step S2, Proxy Model Training and Loading: Train or load proxy models for frontal rigid wall collision and moving barrier collision, verify the input and output dimensions of the model, and configure the inverse normalization method. Step S3, Multidisciplinary Optimization Framework Construction and Design Variable Analysis: Scan the input header, identify continuous variables, thickness column and material column, combine thickness variable and material grade variable to form discrete combination variable, remove duplicates of the same variable to form the union of cross-disciplinary design variables, and reconstruct the exclusive input vector according to the original header of each discipline; Step S4, interdisciplinary collaborative evaluation: The same candidate design is fed into the surrogate models of frontal rigid wall collision and moving barrier collision respectively, the predicted response is restored to physical dimensions, the response direction is unified, and the global quality is formed by the maximum value of the predicted quality of the two disciplines. Step S5, Multi-objective optimization solution: Establish objective function and inequality constraints, including fixed-value performance constraints and cross-disciplinary quality consistency constraints, and use multi-objective evolutionary algorithm to obtain candidate solution set; Step S6, Result Screening and Sensitivity Analysis: Recommended schemes are screened from the last generation population according to the priority of feasibility, constraint violation, and objective function, and sensitivity analysis is performed on the recommended schemes.
2. The multidisciplinary optimization method for vehicle body collision and lightweighting based on a surrogate model as described in claim 1, characterized in that, In step S1, the input vector for frontal rigid wall collision is 79-dimensional, and the input vector for moving barrier collision is 83-dimensional; a fixed random seed is used to divide the training set, validation set, and test set. In step S2, a multilayer perceptron, a one-dimensional convolutional neural network, a bidirectional long short-term memory network, and a self-attention model are selected for parallel training, and a surrogate model is selected based on the performance on the validation set.
3. The multidisciplinary optimization method for vehicle body collision and lightweighting based on a surrogate model according to claim 1, characterized in that, In step S3, 53 actual design variables for frontal rigid wall collision and 57 actual design variables for moving barrier collision are obtained through analysis. In the discrete combination variables, each thickness-material legal pair is treated as an indivisible discrete unit, and the thickness value is mapped to the material grade through a combination index. For geometrically or positionally continuous variables, normalization is performed using the center and half-range to preserve the input scale, in order to handle extreme values in the data distribution and improve the accuracy of physical quantity recovery.
4. The multidisciplinary optimization method for vehicle body collision and lightweighting based on a surrogate model according to claim 1, characterized in that, In step S4, the frontal rigid wall collision and moving barrier collision responses are unified into an optimization direction where smaller values are better. global quality ,in and The quality of the predictions for the two disciplines are respectively; In interdisciplinary collaborative assessments, each discipline uses its own independent center and half-range to recover to the physical domain, and the cross-use of scale parameters is prohibited.
5. The multidisciplinary optimization method for vehicle body collision and lightweighting based on a surrogate model according to claim 1, characterized in that, In step S5, the interdisciplinary quality consistency constraint limits the absolute value of the difference between the predicted quality of the two disciplines to no more than a preset tolerance. Specifically, it is set as a two-sided inequality constraint; The multi-objective evolutionary algorithm selects the NSGA-II algorithm when the number of objectives is less than or equal to 5, and selects the NSGA-III algorithm when the number of objectives is greater than 5. The method supports a minimum mass mode and a target tracking mode. In the target tracking mode, a dimensionless deviation target is calculated and upper and lower boundary constraints are generated.
6. The multidisciplinary optimization method for vehicle body collision and lightweighting based on a surrogate model according to claim 1, characterized in that, In step S6, when selecting recommended solutions, a three-level sorting rule is adopted: the first level prioritizes feasible solutions, the second level prioritizes solutions with smaller constraint violations, and the third level prioritizes solutions with smaller objective function sums. The sensitivity analysis employs the Morris method, which perturbs the continuous variables and discrete combination indices, calculates the absolute mean and standard deviation of the basic effects, and outputs a variable-response ranking table.
7. The multidisciplinary optimization method for vehicle collision and lightweighting based on a surrogate model according to claim 1, characterized in that, In the interdisciplinary collaborative evaluation in step S4 and the multi-objective optimization solution in step S5, batch evaluation technology is adopted to package multiple candidate designs in the population and input them into the surrogate model in parallel, and use a graphics processor for parallel computing acceleration.
8. A multidisciplinary optimization system for vehicle collision and lightweighting based on a surrogate model, characterized in that, include: The module includes data preparation, surrogate model training and loading, design variable analysis and coupling, interdisciplinary collaborative evaluation, multi-objective optimization, and result screening and sensitivity analysis. The data preparation module is used to read the design inputs and labels of frontal rigid wall collision and moving barrier collision respectively, align them by samples and divide the dataset; The proxy model training and loading module is used to train or load proxy models for two disciplines and verify the dimensions. The design variable analysis and coupling module is used to identify continuous variables and thickness-material discrete combination variables, form a cross-disciplinary variable union and reconstruct the input vector; The interdisciplinary collaborative evaluation module is used to feed candidate designs into two surrogate models respectively, unify the response direction, and form the global quality by the maximum value of the predicted quality of the two disciplines. The multi-objective optimization module is used to establish the objective function and the inequality constraints that include cross-disciplinary quality consistency constraints, and to call the multi-objective evolutionary algorithm to solve it; The result screening and sensitivity analysis module is used to screen recommended schemes according to priority and perform sensitivity analysis.
9. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and when the processor executes the computer program, it implements the multidisciplinary optimization method for vehicle collision and lightweighting based on the surrogate model as described in any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multidisciplinary optimization method for vehicle collision and lightweighting based on the surrogate model as described in any one of claims 1 to 7.
