Mechanical part optimization design method and system based on artificial intelligence

Abstract

The invention discloses a mechanical part optimization design method and system based on artificial intelligence, and belongs to the technical field of mechanical design, and the method comprises the steps: building a parameterized design space model according to collected multi-dimensional historical data of parts, extracting key design variables through feature engineering, and calculating the performance residual error of each variable and an optimal target, forming a design scheme-performance residual distribution map; in the design scheme-performance residual distribution map, identifying and rejecting invalid design schemes of which the contribution degree tends to be attenuated based on the sensitivity gradient direction of the key design variables, and screening out an effective design variable set; inputting the effective design variable set into a pre-trained optimization design network, and outputting a globally optimal design parameter set; according to the optimal design parameter set, parameterized three-dimensional model reconstruction and simulation verification are carried out, and a final optimized part three-dimensional model is generated; according to the method, the design efficiency and the performance reliability are improved.

Description

Technical Field

[0001] This invention belongs to the field of mechanical design technology, specifically an artificial intelligence-based method and system for optimizing the design of mechanical parts. Background Technology

[0002] As high-end equipment manufacturing develops towards higher precision, higher reliability, and lighter weight, the design requirements for mechanical components are becoming increasingly complex. Traditional design methods rely on empirical formulas and iterative finite element simulations, which suffer from problems such as long design cycles, difficulty in balancing conflicts in multi-objective optimization, and insufficient accuracy in matching key parameters. For example, in the design of core components such as aero-engine blades, automotive transmission gears, and precision mechanical bearings, even small deviations in structural parameters can lead to significant performance degradation, while overly conservative designs can result in material waste and increased energy consumption. Therefore, there is an urgent need for a design method that integrates 3D geometric feature extraction, AI intelligent decision-making, and multi-objective optimization to improve the efficiency and accuracy of mechanical component design. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes an artificial intelligence-based optimization design method and system for mechanical components. Based on collected multi-dimensional historical data of the components, a parametric design space model is constructed. Key design variables are extracted through feature engineering, and the performance residual between each variable and the optimal target is calculated, forming a design scheme-performance residual distribution map. In the design scheme-performance residual distribution map, invalid design schemes with diminishing contributions are identified and eliminated based on the sensitivity gradient direction of the key design variables, thus selecting a set of effective design variables. This set of effective design variables is input into a pre-trained optimization design network, which outputs a globally optimal set of design parameters. Based on the optimal set of design parameters, a parametric 3D model is reconstructed and simulated to generate the final optimized 3D model of the component. This method improves design efficiency and performance reliability.

[0004] To achieve the above objectives, the present invention provides the following technical solution: Artificial intelligence-based optimization design methods for mechanical parts include: Based on the collected multi-dimensional data of the components, a design space model is constructed; Based on the performance optimization objective of the design space model, key design variables are extracted through feature engineering, and the performance residual between each key design variable and the optimal performance objective is calculated to obtain the design scheme-performance residual distribution map. Based on the sensitivity trends of key design variables, an invalid design scheme elimination strategy is implemented in the design scheme-performance residual distribution map to obtain a set of valid design variables; The effective set of design variables is input into a preset optimization design network for processing, and the optimal set of design parameters is output. The optimal set of design parameters is then used to reconstruct a three-dimensional model. The optimization design network is trained using historical design data and performance verification data of the components.

[0005] Specifically, the process of constructing the design space model includes: The collected multi-dimensional data of the components are projected into a multi-objective optimization space with core performance indicators as reference axes. The multi-dimensional data includes three-dimensional point cloud geometric data, mechanical property test data, material property data, and manufacturing process constraint parameters. The mechanical property test data includes at least stress, strain, and fatigue life. The material property data includes at least density and elastic modulus. The manufacturing process constraint parameters include at least machining accuracy and forming limit. Based on the projected multi-dimensional data, multiple equidistant sub-regions are divided in the design variable dimension of the multi-objective optimization space, and the set of parameter points that form local optimum density in the performance response dimension in each sub-region is extracted. Using the optimal peak performance of the parameter point set as the reference constraint boundary, a fitting model of performance consistency constraint is constructed based on the rate of change of performance sensitivity in the neighborhood of the parameter point set, so as to generate the design space model.

[0006] Specifically, based on the performance optimization objective of the design space model, key design variables are extracted through feature engineering, including: Based on the performance optimization objectives and the distribution of parameters in the design space model, the sensitivity projection direction of the design variables is determined, and the parameter-performance response curve is calculated based on the sensitivity projection direction and the parameter space density; the performance optimization objectives include lightweight, high strength, and long lifespan. Based on the shape of the design space model and the density variation of the parameters in the sensitivity projection direction, a nonlinear screening operator matching the parameter coupling characteristics is generated. Using the nonlinear screening operator as a sliding window, redundant parameter removal and key parameter enhancement operations are performed on the parameter-performance response curve. The redundant parameter removal operation is processed based on the minimum local performance contribution within the sliding window; the key parameter enhancement operation is processed based on the maximum local performance contribution within the sliding window. The results of redundant parameter elimination and key parameter enhancement operations are merged into a smooth response curve, which serves as the benchmark performance curve for the parameter-performance response curve. The parameter-performance response curve and the baseline performance curve are compared by residual analysis. Design variables with performance residual amplitude greater than a preset optimization threshold are selected to generate key design variables.

[0007] Specifically, the calculation of the performance residual between each key design variable and the optimal performance target includes: Obtain the optimal performance target parameters for each key design variable; Calculate the geometric parameter residuals, mechanical property residuals, and process adaptation residuals between the key design variables and the optimal performance target parameters.

[0008] Specifically, based on the sensitivity trends of key design variables in the design scheme-performance residual distribution map, an ineffective design scheme elimination strategy is implemented to obtain a set of effective design variables, including: Calculate the spatial gradient of the performance sensitivity of each key design variable in the performance residual distribution map of the design scheme, and construct the sensitivity gradient field of the design variables; Based on the gradient direction of the performance sensitivity spatial gradient, the design schemes corresponding to the key design variables are divided into sensitivity convergence schemes and sensitivity divergence schemes. The sensitivity convergence scheme refers to a scheme in which the contribution of design variable adjustment to performance improvement gradually decreases and tends to converge. The sensitivity divergence scheme refers to a scheme in which the contribution of design variable adjustment to performance improvement continuously increases and tends to diffuse. An invalid design scheme elimination strategy is implemented, that is, the sensitivity convergence scheme is eliminated, and a set of valid design variables is generated based on the key design variables corresponding to the remaining sensitivity divergence schemes.

[0009] Specifically, the optimization design network includes a design variable encoding layer and a performance sensitivity extraction layer. The performance sensitivity extraction layer is equipped with a graph convolutional network, which includes an adjacency graph construction module, a graph convolution calculation module, and a differentiation module. The design variable encoding layer receives a set of valid design variables and constructs a design feature vector based on the geometric parameter value, performance response value, and local sensitivity coefficient of each design variable. The adjacency graph construction module uses the generated design feature vectors as nodes, calculates edge weights based on the parameter coupling degree between design variables, and constructs a parameter adjacency graph. The graph convolution calculation module performs information propagation and aggregation on the nodes in the parameter adjacency graph, and extracts the performance sensitivity change rate and parameter gradient of each design variable in its neighborhood. The differentiation module determines whether the rate of change of performance sensitivity of each design variable is greater than the weighted standard deviation of the mean rate of change of sensitivity in its neighborhood based on the rate of change of performance sensitivity and parameter gradient in its neighborhood, so as to identify the abrupt change in performance. Under the condition that the parameter gradient is less than the preset constraint threshold and the performance value is the optimal value in its neighborhood, the module identifies the local optimal performance region.

[0010] Specifically, the optimization design network further includes an optimal solution determination layer, used to output an optimal design parameter set based on the performance mutation relationship and the local performance optimal region, specifically including: Perform parameter connectivity analysis on the aforementioned performance mutation relationship and local performance optimal region to generate a set of candidate optimal parameter points; The candidate optimal parameter point set is clustered and grouped according to a multi-objective clustering algorithm to obtain the clustering results; Based on the spatial distribution density, performance gradient variation range, and process adaptability consistency of the candidate optimal parameters in the clustering results, the optimization confidence of the parameter region corresponding to the clustering results is calculated. The parameter region with an optimization confidence level greater than a preset optimization threshold is output as the optimal design parameter set.

[0011] Specifically, the process of reconstructing a three-dimensional model from the optimal set of design parameters includes: Extract the three-dimensional geometric coordinates and constraint parameters corresponding to the optimal design parameter set, and construct a parameterized three-dimensional model framework based on the three-dimensional geometric coordinates and constraint parameters; Finite element simulation verification and interpolation optimization are performed on the parametric 3D model framework to generate the final optimized 3D model of the parts.

[0012] The AI-based mechanical parts optimization design system includes: a spatial model construction module, a performance residual analysis module, a scheme elimination module, and a network processing module. The spatial model construction module is used to construct a design spatial model based on the collected multi-dimensional data of the components. The performance residual analysis module is used to extract key design variables, calculate performance residuals, and generate a design scheme-performance residual distribution map based on the performance optimization target of the design space model. The scheme elimination module is used to execute an invalid design scheme elimination strategy based on the sensitivity trend of key design variables in the design scheme-performance residual distribution map, so as to obtain a set of valid design variables. The network processing module is used to input the set of effective design variables into a preset optimization design network for processing, output the optimal set of design parameters, and reconstruct the three-dimensional model.

[0013] Compared with the prior art, the beneficial effects of the present invention are: This invention proposes an AI-based optimization design method for mechanical components. By constructing a parameter-response curve primarily based on performance sensitivity and introducing feature engineering to extract key design variables, it achieves efficient screening of high-value design parameters. Furthermore, by constructing a design scheme-performance residual distribution map through performance optimization target mapping and combining it with sensitivity gradient field directionality analysis, it accurately eliminates invalid design schemes. Based on this, it combines parameter adjacency graphs and graph-structured convolutional units to perform multi-dimensional performance correlation analysis on candidate parameters, ultimately achieving precise discovery of globally optimal design parameters. This method deeply integrates AI technology with mechanical design, effectively shortening the design iteration cycle and improving the coordination between component performance and manufacturing feasibility under multiple constraints. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the principle of the artificial intelligence-based mechanical parts optimization design method of the present invention. Figure 2 This is a system architecture diagram of the artificial intelligence-based mechanical component optimization design system of the present invention. Detailed Implementation

[0015] Example 1: Please see Figure 1 The present invention provides an embodiment of an artificial intelligence-based mechanical component optimization design method, comprising the following steps: Based on the collected multi-dimensional data of the components, a design space model is constructed; Based on the performance optimization objective of the design space model, key design variables are extracted through feature engineering, and the performance residual between each key design variable and the optimal performance objective is calculated to obtain the design scheme-performance residual distribution map. Based on the sensitivity trends of key design variables, an invalid design scheme elimination strategy is implemented in the design scheme-performance residual distribution map to obtain a set of valid design variables; The effective set of design variables is input into a preset optimization design network for processing, and the optimal set of design parameters is output. The optimal set of design parameters is then used to reconstruct a three-dimensional model. The optimization design network is trained using historical design data and performance verification data of the components.

[0016] Furthermore, the method of this application is applicable to the design of complex mechanical components requiring multi-objective optimization, including but not limited to: structural components in the aerospace field, such as engine blades and fuselage frames, which need to balance strength, lightweight, and fatigue resistance; transmission system components in the automotive industry, such as gears and drive shafts, which need to balance transmission efficiency, wear resistance, and manufacturing cost; and core components in the precision machinery field, such as bearings and lead screws, which need to meet the requirements of high precision, low friction, and long life. The application scenario selection must possess at least one of the following characteristics: numerous design variables with strong coupling relationships; conflicting performance indicators, such as strength improvement versus weight reduction; complex manufacturing process constraints, requiring a balance between design feasibility and processing cost; and performance response sensitive to parameter changes, requiring high-precision parameter matching.

[0017] For example, the process involves acquiring 3D point cloud data of the gear, test data on transmission efficiency and fatigue life under different modules / number of teeth / pressure angles, material property data, and machining process constraints. This data is then projected into a multi-objective optimization space encompassing module-number of teeth-pressure angle-transmission efficiency-fatigue life. The parameter point set of the optimal performance region is extracted, and a design space model including strength and efficiency constraints is constructed. Key parameters significantly affecting transmission efficiency and fatigue life, such as module, tooth width, and fillet radius, are selected through feature engineering. The residuals between each key parameter and the optimal performance target are calculated, generating a design scheme-performance residual distribution map. Design schemes with diminishing performance contributions, such as those with excessively large modules leading to weight overruns, are eliminated based on a sensitivity gradient field. Effective design variables are input into the optimization design network, outputting an optimal parameter set that balances transmission efficiency, fatigue life, and manufacturing cost. A 3D gear model is constructed based on the optimal parameters, and strength and transmission performance are verified through finite element simulation, ultimately generating an optimized design scheme.

[0018] The process of constructing the design space model includes: The collected multi-dimensional data of the components are projected into a multi-objective optimization space with core performance indicators as reference axes. The multi-dimensional data includes three-dimensional point cloud geometric data, mechanical property test data, material property data, and manufacturing process constraint parameters. The mechanical property test data includes at least stress, strain, and fatigue life. The material property data includes at least density and elastic modulus. The manufacturing process constraint parameters include at least machining accuracy and forming limit. Furthermore, the collected multi-dimensional data of components are projected onto a multi-objective optimization space with core performance indicators as reference axes, including: (1) Based on the core performance optimization goals of the components, such as lightweight, high strength and long life, each goal is established as an independent coordinate axis in a multi-dimensional space. Thus, these goals together span a mathematical multi-objective optimization space. The key is that each coordinate axis must be given a clear physical meaning, dimension and optimization direction. For example, the optimization direction of the mass axis is defined as the value decreases, and the optimization direction of the strength axis is defined as the value increases. Finally, a fully defined empty multi-objective optimization space framework with data to be filled is output. (2) Standardize all the collected geometric, mechanical, material and process data. Specifically, for each type of data, calculate the mean and standard deviation of its historical dataset. Then, subtract the mean from each original data value and divide by the standard deviation. This process transforms all data into dimensionless pure values ​​that conform to the standard normal distribution, i.e., standardized design parameters. The calculation of the mean and standard deviation are conventional methods that can be understood and implemented by those skilled in the art. This application is not limited to specific partitioning methods. (3) The Gaussian process regression algorithm is used as the performance mapping model. The historical dataset is used to train the performance mapping model. The input of the performance mapping model is the standardized design parameters, and the output is the core performance index value corresponding to the parameters, which has also been standardized. The trained performance mapping model is essentially a function that can predict its comprehensive performance based on any set of input parameters. The Gaussian process regression algorithm is a conventional method that can be understood and implemented by those skilled in the art. This application is not limited to a specific partitioning method. (4) The standardized design parameters corresponding to each component design scheme to be analyzed are passed as input to the trained performance mapping model. After the performance mapping model performs calculations, it outputs a multi-dimensional vector. Each component value of the vector represents the specific coordinates of the design scheme on the corresponding coordinate axis in the multi-objective optimization space, which can also be called a coordinate vector. The coordinate vector output by the performance mapping model uniquely determines the location of the design scheme in the space. The set of all design scheme locations completes the distribution in the multi-objective optimization space, and the projected multi-dimensional data is obtained.

[0019] Furthermore, the distribution of all design schemes in the multi-objective optimization space after projection can be analyzed. By observing the aggregation of point sets, the shape of the leading edge, and other characteristics, a preliminary insight into the characteristics of the overall design space can be formed, such as identifying the areas where high-performance schemes are clustered.

[0020] Based on the projected multi-dimensional data, multiple equidistant sub-regions are divided in the design variable dimension of the multi-objective optimization space, and the set of parameter points that form local optimum density in the performance response dimension in each sub-region is extracted. Furthermore, the process of extracting the parameter point set includes: (1) Obtain all data points projected into the multi-objective optimization space, select a specific design variable dimension based on the data points, and determine its numerical range; from the minimum value to the maximum value, divide the numerical range into multiple consecutive intervals with equal widths. These intervals are equidistant sub-regions. Each equidistant sub-region represents a specific value range of the design variable. The grid division of the selected design variable dimension is completed, and a sequence of equidistant sub-regions is obtained. (2) For the divided equidistant sub-regions, analyze all data points falling within this sub-region, but the focus shifts from the design variable values ​​to their performance response values. That is, analyze the distribution of all data points on the performance dimension through the density estimation algorithm, identify the location where the data points are most concentrated, i.e. the performance density peak. The performance density peak indicates the most frequently achieved or approached high performance level under the local design constraints. The density estimation algorithm is a conventional means that can be understood and implemented by those skilled in the art. This application is not limited to specific partitioning methods. (3) Take the peak performance density as the center, set a neighborhood range, and extract all the data points in this neighborhood as a whole. This extracted data set is the parameter point set. The points in the parameter point set have dual characteristics: their design variable values ​​belong to the same equidistant sub-region, and their performance values ​​are highly concentrated near the peak performance density. (4) For each equidistant sub-region under the current design variable dimension, sequentially perform performance distribution analysis and parameter point set extraction. After completing the scan of one dimension, switch to the next design variable dimension. (5) After all the equidistant sub-regions of all design variable dimensions have been processed, all parameter point sets extracted from all parts of the world will be summarized to form a complete parameter point set library. This parameter point set library systematically records the information of all local high-performance regions in the entire multi-objective optimization space.

[0021] Using the optimal peak performance of the parameter point set as the reference constraint boundary, a fitting model of performance consistency constraint is constructed based on the rate of change of performance sensitivity in the neighborhood of the parameter point set, so as to generate the design space model.

[0022] Furthermore, using the optimal peak performance of the parameter point set as the reference constraint boundary, a fitting model for performance consistency constraints is constructed based on the rate of change of performance sensitivity within the neighborhood of the parameter point set to generate a design space model, including: (1) For all data points in each parameter point set, calculate the maximum value of its performance response value. This maximum value is the optimal peak value of the performance of the point set. The optimal peak value marks the upper limit of the performance of the local area. (2) Define a neighborhood range with the parameter point set as the center, and analyze the rate of change of the performance response value caused by the small perturbation of the design variable within this neighborhood range. This rate of change is the performance sensitivity change rate. The calculated performance sensitivity change rate quantifies the stability and trend of performance near the optimal peak. (3) A fitting model with performance consistency constraints is constructed using the Gaussian process regression algorithm, including: using the parameter point set and the data in its neighborhood as training samples, and using the determined optimal performance peak as the constraint boundary condition of the fitting model in the training process. At the same time, the obtained performance sensitivity change rate information is used to guide the selection and configuration of the kernel function of the fitting model, so that the fitting model can accurately reflect the characteristics of the high-performance region. After training, a fitting model with performance consistency constraints is obtained. (4) Validate the fitting model with performance consistency constraints, including: using the reserved validation dataset to evaluate its prediction accuracy, especially to evaluate its prediction consistency near the peak boundary of optimal performance; fine-tuning and optimizing the fitting model parameters based on the validation results to ensure that the fitting model can accurately characterize the boundary and internal variation law of the high-performance region; and finally, defining the validated and optimized fitting model as the final design space model.

[0023] Based on the performance optimization objective of the design space model, key design variables are extracted through feature engineering, including: Based on the performance optimization objectives and the distribution of parameters in the design space model, the sensitivity projection direction of the design variables is determined, and the parameter-performance response curve is calculated based on the sensitivity projection direction and the parameter space density; the performance optimization objectives include lightweight, high strength, and long lifespan. Furthermore, the calculation process for the parameter-performance response curve includes: (1) Obtain the constructed design space model and clear performance optimization goals. Quantify the contribution of each design variable to the performance optimization goals through global sensitivity analysis. Select variables with a contribution higher than the preset threshold and mark them as key design variables. Then, for each key design variable, calculate the partial derivative of performance with respect to the variable. Based on the performance optimization goals, such as increasing or decreasing the performance index, determine the optimization direction of its numerical change. This direction is the sensitivity projection direction. Global sensitivity analysis is a conventional method that can be understood and implemented by those skilled in the art. This application is not limited to specific partitioning methods. (2) Scan along the sensitivity projection direction from the lower limit to the upper limit of the design variable value with a fixed step size. At each sampling point, calculate the concentration of historical data points in its neighboring area, i.e. parameter space density. Record a series of sampling points along the projection direction and their corresponding parameter space density to complete the density exploration of the variable change path. (3) Use the design space model to predict the performance response value at each sampling point. When predicting, the parameter space density is introduced as a weight to perform a weighted average of the performance prediction values ​​of the neighboring areas of the sampling point, so that the prediction of the data-rich area is more weighted. Finally, the parameter values ​​of the sampling points are associated with their corresponding weighted average performance values ​​to generate a smooth curve that reflects the impact of parameter changes on performance, namely the parameter-performance response curve. (4) For each identified key design variable, iteratively repeat the process until the corresponding parameter-performance response curve is generated for all key design variables.

[0024] Based on the shape of the design space model and the density variation of the parameters in the sensitivity projection direction, a nonlinear screening operator matching the parameter coupling characteristics is generated. Furthermore, based on the shape of the design space model and the density variation of the parameters in the sensitivity projection direction, a nonlinear screening operator matching the parameter coupling characteristics is generated, including: (1) Analyze the topological structure of the design space model in the parameter space and identify its overall morphological characteristics, such as the smoothness of the model surface, whether there are multiple peak regions or continuous slope-like structures. These overall morphological characteristics determine the basic smoothing or edge preservation capabilities that the screening operator needs to have. (2) Analyze the spatial density distribution of parameters along the sensitivity projection direction and quantify its variation characteristics, including the width and height of the density peak, the depth and breadth of the density valley, and the severity of density changes. These quantification results together define the parameter density variation characteristics and reflect the local non-uniformity of the data distribution. (3) Based on the obtained overall morphological features and parameter density change characteristics, the core behavior rules of the nonlinear screening operator are defined, and the basic mathematical form of the nonlinear screening operator is determined to match the overall morphology. Crucially, the core parameters of the nonlinear screening operator, such as the width of its effective range window, are defined as a function of the parameter density change characteristics. The specific rules are as follows: in the region where the parameter density is high and the change is gradual, the operator is given a larger window width to promote smoothness; in the region where the parameter density changes rapidly or is extremely low, the operator is given a smaller window width to preserve details, thus ensuring the operator's adaptive ability to parameter coupling features. (4) Implement the defined parameter rules in detail, including: writing program code to construct a function that can dynamically query the local parameter density change characteristics based on the input parameter point position, and calculate the window width and other parameters to be used by the operator in real time according to the established rules. This successfully constructed function, whose behavior is dynamically driven by data features, is the final nonlinear filtering operator.

[0025] Using the nonlinear screening operator as a sliding window, redundant parameter removal and key parameter enhancement operations are performed on the parameter-performance response curve. The redundant parameter removal operation is processed based on the minimum local performance contribution within the sliding window; the key parameter enhancement operation is processed based on the maximum local performance contribution within the sliding window. The results of redundant parameter elimination and key parameter enhancement operations are merged into a smooth response curve, which serves as the benchmark performance curve for the parameter-performance response curve. Furthermore, the results of redundant parameter removal and key parameter enhancement operations are fused into a smooth response curve, which serves as the benchmark performance curve for the parameter-performance response curve, including: (1) Normalize the data sequences after the redundant parameter removal operation and the key parameter enhancement operation are processed separately to make their numerical range consistent. At the same time, ensure that the sampling point positions of the two sets of data on the parameter axis are completely corresponding to achieve data alignment and prepare for fusion operation. (2) For each aligned parameter point, calculate the enhancement magnitude of its value in the key parameter enhancement operation result relative to the original value. Based on this enhancement magnitude, calculate a dynamic fusion weight for the point. The larger the enhancement magnitude, the more likely the dynamic fusion weight will be given to the enhancement operation result. The smaller the enhancement magnitude, the more likely the weight will be given to the elimination operation result. (3) For each parameter point, the normalized and aligned elimination operation result and the enhancement operation result are multiplied by the corresponding dynamic fusion weight, and then weighted and summed. This calculation is applied to all parameter points to obtain a series of initial fusion data points after weighted fusion. (4) The obtained discrete initial fusion data points are used as input, and a smooth interpolation algorithm is used to generate a continuous and smooth curve. This curve is the required benchmark performance curve. The smooth interpolation algorithm is a conventional method that can be understood and implemented by those skilled in the art. This application is not limited to a specific partitioning method.

[0026] The parameter-performance response curve and the baseline performance curve are compared by residual analysis. Design variables with performance residual amplitude greater than a preset optimization threshold are selected to generate key design variables.

[0027] The calculation of the performance residual between each key design variable and the optimal performance target includes: Obtain the optimal performance target parameters for each key design variable; Calculate the geometric parameter residuals, mechanical property residuals, and process adaptation residuals between the key design variables and the optimal performance target parameters.

[0028] Further, the geometric parameter residuals, mechanical property residuals, and process adaptation residuals between the key design variables and the optimal performance target parameters are calculated, including: (1) Clarify the specific quantitative values ​​of the optimal performance target parameters in the three dimensions of geometry, mechanics, and process. These values ​​together constitute a multi-dimensional performance target benchmark, which serves as a unified reference system for residual calculation. (2) Align and compare the current geometric model defined by the key design variables with the ideal geometric model in the multidimensional performance target benchmark, calculate the Euclidean distance of the corresponding points of the key contour and the absolute difference of the key dimensions, and aggregate these local geometric deviations into a comprehensive geometric parameter residual; Furthermore, through calculations based on the physical functional principles of components or high-fidelity simulations, such as Computational Fluid Dynamics (CFD) and Finite Element Analysis (FEA), a theoretically optimal geometric model is generated under ideal conditions. The key geometric parameters of this model, such as contour coordinates, curvature distribution, and thickness, are the theoretically optimal target values ​​for the geometric dimension. The mechanical performance target is related to the functional realization and reliability of the components, and its quantification is based on safety criteria. The process dimension target is related to the manufacturability, assemblability, and economy of the design, and its quantification is based on the limits of manufacturing capabilities and cost-benefit analysis.

[0029] (3) Based on the current geometric model, mechanical simulation is performed to obtain its mechanical response value. The simulation results are compared with the ideal mechanical response value in the multidimensional performance target benchmark. The numerical differences of key indicators, such as maximum stress and fatigue life, are calculated and aggregated into a comprehensive mechanical performance residual. The mechanical simulation process is a conventional means that can be understood and implemented by those skilled in the art. This application is not limited to a specific partitioning method. (4) Compare the various process parameters of the current design, such as minimum fillet radius and dimensional tolerance, with the process constraints defined in the multidimensional performance target benchmark, calculate the degree of deviation of each parameter from the constraint boundary, and aggregate them into a comprehensive process adaptation residual; Furthermore, process constraints refer to a series of process boundaries and rules that must be followed during the design phase to ensure that parts can be manufactured efficiently, economically, and reliably. Essentially, they are a mapping of manufacturing capabilities to the design end, transforming abstract physical principles and specific production limitations into explicit mathematical limits that design parameters must meet. These conditions typically stem from the inherent physical limits and cost considerations of specific manufacturing processes, such as casting, forging, machining, and additive manufacturing. For example, for casting, constraints include minimum allowable wall thickness, maximum allowable wall thickness, minimum draft angle, and minimum fillet radius; for machining, constraints are reflected in economically achievable dimensional tolerance grades and achievable surface roughness values; for injection molding, there are also constraints such as acceptable weld line locations. Therefore, process constraints are mathematically defined as a set of thresholds or intervals. For example, it is clearly stipulated that the minimum fillet radius is greater than or equal to 1.0 mm, the dimensional tolerance T must be between IT8 and IT10, and the draft angle is greater than or equal to 1.5°. If the design parameter values ​​fall within this threshold or range, they are considered to have good manufacturability; if they deviate from this range, they are considered to have process risks or lead to increased costs. The process of calculating the process adaptation residual is the process of quantifying the gap between the current design parameter values ​​and these hard constraint boundaries.

[0030] (5) The calculated geometric parameter residuals, mechanical property residuals and process adaptation residuals are used together as outputs to form a complete quantitative description of the gap between the key design variables and the optimal performance target parameters.

[0031] Based on the sensitivity trends of key design variables, an ineffective design scheme elimination strategy is implemented in the design scheme-performance residual distribution map to obtain a set of effective design variables, including: Calculate the spatial gradient of the performance sensitivity of each key design variable in the performance residual distribution map of the design scheme, and construct the sensitivity gradient field of the design variables; Further, the performance sensitivity spatial gradient of each key design variable in the design scheme-performance residual distribution map is calculated, and the design variable sensitivity gradient field is constructed, including: (1) Extract the data points consisting of each key design variable and its corresponding performance residual value from the design scheme-performance residual distribution map to form a discrete parameter-residual point set associated with each variable; (2) For each discrete parameter-residual point set obtained in the previous step, a continuous local parameter-residual fitting surface that reflects the mapping relationship between the key design variable and the performance residual is generated by the surface fitting algorithm. The surface fitting algorithm is a conventional means that can be understood and implemented by those skilled in the art. This application is not limited to a specific partitioning method. (3) Based on the local parameter-residual fitting surface constructed in the second step, calculate the gradient of each point on the surface. The gradient vector is the performance sensitivity spatial gradient of the corresponding point. Its direction indicates the direction of the fastest increase of the residual, and the magnitude represents the rate of change. (4) The calculated performance sensitivity spatial gradient vectors corresponding to all key design variables and all data points are set and labeled back into the original design scheme-performance residual distribution map. This map, which integrates all gradient vectors, constitutes the required design variable sensitivity gradient field. (5) The completed design variable sensitivity gradient field is used as the final output.

[0032] Based on the gradient direction of the performance sensitivity spatial gradient, the design schemes corresponding to the key design variables are divided into sensitivity convergence schemes and sensitivity divergence schemes. The sensitivity convergence scheme refers to a scheme in which the contribution of design variable adjustment to performance improvement gradually decreases and tends to converge. The sensitivity divergence scheme refers to a scheme in which the contribution of design variable adjustment to performance improvement continuously increases and tends to diffuse. An invalid design scheme elimination strategy is implemented, that is, the sensitivity convergence scheme is eliminated, and a set of valid design variables is generated based on the key design variables corresponding to the remaining sensitivity divergence schemes.

[0033] The optimized design network includes a design variable encoding layer and a performance sensitivity extraction layer. The performance sensitivity extraction layer is equipped with a graph convolutional network, which includes an adjacency graph construction module, a graph convolution calculation module, and a discrimination module. Specifically: The design variable encoding layer receives a set of valid design variables and constructs a design feature vector based on the geometric parameter value, performance response value, and local sensitivity coefficient of each design variable. Furthermore, a design feature vector is constructed based on the geometric parameter values, performance response values, and local sensitivity coefficients of each design variable, including: (1) The design variable coding layer receives the set of valid design variables from the upstream input. First, it verifies the set of valid design variables to ensure that it is not empty and that each design variable contains three complete attributes: geometric parameter value, performance response value and local sensitivity coefficient. (2) Standardize the three types of attribute data of each design variable in the valid design variable set after verification. That is, calculate the mean and standard deviation of each type of data in all schemes separately, and then subtract the mean from each original value and divide by the standard deviation to obtain the corresponding standardized geometric parameter value, standardized performance response value and standardized local sensitivity coefficient. (3) For each design scheme in the effective design variable set, according to the fixed arrangement order of its key design variables, the standardized geometric parameter values, standardized performance response values ​​and standardized local sensitivity coefficients of all variables under the scheme are spliced ​​in sequence to form an initial design feature vector representing all information of the scheme; (4) The principal component analysis method is used to reduce the dimensionality of the set of initial design feature vectors of all schemes, extract the main feature components, and project each initial design feature vector onto these principal components to generate a refined design feature vector with lower dimensionality and more concentrated information. The principal component analysis method is a conventional means that can be understood and implemented by those skilled in the art, and this application is not limited to a specific partitioning method. (5) The refined design feature vectors corresponding to all design schemes in the effective design variable set are taken as the final output and passed to the next layer of the network for processing.

[0034] The adjacency graph construction module uses the generated design feature vectors as nodes, calculates edge weights based on the parameter coupling degree between design variables, and constructs a parameter adjacency graph. Furthermore, the process of constructing the parametric adjacency graph includes: (1) The adjacency graph construction module receives the set of refined design feature vectors generated by the upstream coding layer. The adjacency graph construction module defines each refined design feature vector as a graph node. The set of all these graph nodes constitutes the node basis of the graph structure and forms a node set. (2) For any two different refined design feature vectors in the node set, extract the part representing the geometric parameter value, calculate the cosine similarity between the two parameter sub-vectors, and the cosine similarity value is defined as the parameter coupling degree between the two nodes. Perform this calculation on all node pairs to complete the full connection similarity evaluation. The cosine similarity calculation is a conventional method that can be understood and implemented by those skilled in the art. This application is not limited to a specific partitioning method. (3) Set a parameter coupling degree threshold, traverse all node pairs, and if the parameter coupling degree between any pair of nodes is greater than or equal to this parameter coupling degree threshold, then establish an edge connection between the two nodes; if it is less than the parameter coupling degree threshold, then do not establish a connection. This step is based on the similarity strength to filter out significant connection relationships. (4) Assign a weight value to each established edge connection, called edge weight. The value of the edge weight is directly taken from the parameter coupling degree between the two nodes connected by the edge. Thus, the edge weight directly quantifies the similarity of parameter patterns between nodes. (5) Integrate the defined set of nodes, the established set of edge connections, and the set of assigned edge weights to form a weighted undirected graph, which is the final output parameter adjacency graph.

[0035] The graph convolution calculation module performs information propagation and aggregation on the nodes in the parameter adjacency graph, and extracts the performance sensitivity change rate and parameter gradient of each design variable in its neighborhood. Furthermore, the graph convolution calculation module performs information propagation and aggregation on the nodes in the parameter adjacency graph, extracting the performance sensitivity change rate and parameter gradient of each design variable in its neighborhood, including: (1) The graph convolution calculation module receives the input parameter adjacency graph and the initial feature vector corresponding to each node in the graph. The graph convolution calculation module reads the topology of the graph and determines the set of direct neighbors of each node and the weight of the connecting edges. (2) For each node in the parameter adjacency graph, the graph convolution calculation module collects information from all its direct neighbor nodes. Specifically, it multiplies the initial feature vector of each neighbor node by the weight of the edge connecting it to the current node and performs weighted processing. Then, it sums all the weighted neighbor feature vectors and performs a linear combination with the feature vector of the current node itself to generate the first-order aggregated feature representation of the node. (3) The obtained first-order aggregated feature representation is used as the new node feature. The information aggregation operation is repeated. Each time a convolution layer is added, the information that the node can receive expands the range of the first-order neighbors. By stacking multiple layers of graph convolution, each node finally obtains a high-order feature representation, which aggregates the information of the nodes in its multi-hop neighborhood. (4) A well-trained fully connected network is used to take the high-order feature representation of each node as input and map it to two specific physical quantities: one is the rate of change of performance sensitivity of the design represented by the node in its graph neighborhood, and the other is the parameter gradient of the node. The fully connected network is a conventional means that can be understood and implemented by those skilled in the art, and this application is not limited to a specific partitioning method. (5) The performance sensitivity change rate and parameter gradient obtained after decoding all nodes in the parameter adjacency graph are used as the final output.

[0036] The differentiation module determines whether the rate of change of performance sensitivity of each design variable is greater than the weighted standard deviation of the mean rate of change of sensitivity in its neighborhood based on the rate of change of performance sensitivity and parameter gradient in its neighborhood, so as to identify the abrupt change in performance. Under the condition that the parameter gradient is less than the preset constraint threshold and the performance value is the optimal value in its neighborhood, the module identifies the local optimal performance region.

[0037] Furthermore, the process of identifying the region with the best local performance includes: (1) The differentiation module receives the performance sensitivity change rate and parameter gradient of each design variable node. For each node, it calculates the average and standard deviation of the performance sensitivity change rate of the node and all its first-order neighbor nodes, and multiplies the standard deviation by a preset weighting coefficient to obtain the weighted standard deviation threshold. (2) Compare the performance sensitivity change rate of each node with the average value of the sensitivity change rate of its neighbors plus the weighted standard deviation threshold. If the change rate of the node itself is greater than the value, it is determined that the node has a sudden change in performance and is recorded. (3) Evaluate whether the parameter gradient of the node is less than the preset constraint threshold, and select a subset of nodes whose parameter gradients satisfy the stability condition; (4) For a node whose parameter gradient satisfies the condition, compare its performance value with the performance values ​​of all its first-order neighbor nodes. If the condition is met, then the node is determined to have the best performance in its neighborhood. (5) For nodes that simultaneously satisfy the parameter gradient condition and the optimal performance condition, their region is determined to be a local optimal performance region. Finally, the differentiation module outputs a list of all nodes that identify performance mutation relationships and a list of all nodes that identify local optimal performance regions.

[0038] The optimization design network further includes an optimal solution determination layer, which outputs an optimal design parameter set based on the performance mutation relationship and the local performance optimal region, specifically including: Perform parameter connectivity analysis on the aforementioned performance mutation relationship and local performance optimal region to generate a set of candidate optimal parameter points; Furthermore, parameter connectivity analysis is performed on the aforementioned performance mutation relationship and the local performance optimal region to generate a candidate optimal parameter point set, including: (1) Taking the set of nodes in the local performance optimal region and the set of nodes with performance mutation relationship output by the differentiation module as the analysis object, extract the subgraph containing the above two types of nodes and the connecting edges between them in the original parameter adjacency graph, and construct the parameter connectivity analysis network for analysis. (2) In the constructed parameter connectivity analysis network, a density-based clustering algorithm is used to divide the nodes into several interconnected node sets according to the proximity and connection relationship of the nodes in the parameter space. Each set is called a parameter connectivity region. The density-based clustering algorithm is a conventional method that can be understood and implemented by those skilled in the art. This application is not limited to a specific partitioning method. (3) Evaluate and screen all the discovered parameter connected regions, set screening criteria, such as requiring that the region must contain at least one local performance optimal region node and that the internal connection of the region has sufficient strength, retain the high-quality regions that meet the criteria, and form a set of high-quality parameter connected regions. (4) Merge all nodes in the set of high-quality parameter connected regions and remove duplicate nodes to form the final candidate optimal parameter point set, which contains high-performance and key feature nodes confirmed by connectivity analysis. (5) Output the generated set of candidate optimal parameter points as the final result.

[0039] The candidate optimal parameter point set is clustered and grouped according to a multi-objective clustering algorithm to obtain the clustering result. The multi-objective clustering algorithm is a conventional method that can be understood and implemented by those skilled in the art, and this application is not limited to a specific partitioning method. Based on the spatial distribution density, performance gradient variation range, and process adaptability consistency of the candidate optimal parameters in the clustering results, the optimization confidence of the parameter region corresponding to the clustering results is calculated. Further, based on the spatial distribution density, performance gradient variation range, and process adaptability consistency of the candidate optimal parameters in the clustering results, the optimization confidence of the parameter region corresponding to the clustering results is calculated, including: (1) For a given clustering result, i.e. a candidate parameter region, calculate the volume occupied by the region in the parameter space, count the number of candidate parameter points contained in the region, and divide the number of points by the region volume to obtain the spatial distribution density of the region. (2) Extract the performance gradient value of each point in the candidate parameter region, calculate the difference between the maximum and minimum values ​​of these performance gradient values ​​to obtain the performance gradient variation range, and at the same time, calculate the standard deviation of these performance gradient values. (3) Check whether each point in the candidate parameter area meets all the preset process constraints, count the number of points that meet all constraints, and calculate the proportion of the total number of points in the area to obtain the process adaptation consistency score. (4) The obtained spatial distribution density, performance gradient variation range and process adaptation consistency score are standardized to eliminate the influence of dimensions. Weights are assigned to the three standardized indicators and weighted summation is performed to calculate the optimization confidence score of the candidate parameter region. (5) Output the calculated optimization confidence as the comprehensive evaluation index of the candidate parameter region.

[0040] The parameter region with an optimization confidence level greater than a preset optimization threshold is output as the optimal design parameter set.

[0041] The process of reconstructing a three-dimensional model from the optimal set of design parameters includes: Extract the three-dimensional geometric coordinates and constraint parameters corresponding to the optimal design parameter set, and construct a parameterized three-dimensional model framework based on the three-dimensional geometric coordinates and constraint parameters; Further, the three-dimensional geometric coordinates and constraint parameters corresponding to the optimal design parameter set are extracted, and a parameterized three-dimensional model framework is constructed based on the three-dimensional geometric coordinates and constraint parameters, including: (1) Parse and extract the three-dimensional geometric coordinate data required to define the geometric shape and the constraint parameters required to define the engineering relationship from the input optimal design parameter set; (2) In the parametric modeling environment, the extracted three-dimensional geometric coordinate data is assigned to the basic primitives of the sketch. At the same time, the extracted constraint parameters are assigned to the corresponding geometric constraints and dimensional constraints, thereby establishing the driving relationship between the parameters and the geometric elements. (3) Based on the established driving relationship of the sketch and constraints, parameterized three-dimensional features are generated through feature operations such as stretching, rotating and sweeping. These three-dimensional features are combined according to the design logic to construct a complete parameterized three-dimensional model feature structure. (4) Verify the feature structure of the constructed parametric 3D model: First, verify its geometric integrity to ensure that it is a closed entity; second, verify the integrity of the constraints to ensure that the model is fully defined; finally, verify the validity of the parameters to ensure that the parameter-driven model update is correct. (5) Define the parameterized 3D model feature structure that has passed the verification as the final parameterized 3D model framework.

[0042] Finite element simulation verification and interpolation optimization are performed on the parametric 3D model framework to generate the final optimized 3D model of the parts. Finite element simulation verification and interpolation optimization are conventional methods that can be understood and implemented by those skilled in the art, and this application is not limited to specific partitioning methods.

[0043] Example 2: Please see Figure 2 Another embodiment of the present invention provides: an artificial intelligence-based mechanical parts optimization design system, comprising: Spatial model construction module, performance residual analysis module, scheme elimination module, network processing module; The spatial model building module is used to build a design spatial model based on the collected multi-dimensional data of the components; The performance residual analysis module is used to extract key design variables, calculate performance residuals, and generate a design scheme-performance residual distribution map based on the performance optimization target of the design space model, providing a basis for screening effective design schemes. The scheme elimination module is used to execute an invalid design scheme elimination strategy based on the sensitivity trend of key design variables in the design scheme-performance residual distribution map, so as to obtain a set of effective design variables and improve the efficiency of optimization design. The network processing module is used to input the effective set of design variables into the preset optimization design network for processing, output the optimal set of design parameters, and reconstruct the three-dimensional model to realize the optimized design of parts.

[0044] The spatial model building module includes: a data acquisition unit and a model building unit; The data acquisition unit is used to collect multi-dimensional data of parts, including dimensions, material properties, manufacturing process parameters, and environmental conditions, to provide data support for building a design space model; The model building unit is used to construct a design space model using the collected multi-dimensional data. The design space model can describe the relationship between component design variables and performance.

[0045] The performance residual analysis module includes: a variable extraction unit, a residual calculation unit, and a graph generation unit; The variable extraction unit is used to filter out key design variables that have a significant impact on the performance of components from the design space model using feature engineering methods, thereby reducing the complexity of subsequent analysis. The residual calculation unit is used to calculate the performance residual between each key design variable and the optimal performance target, and to quantify the gap between each design scheme and the optimal performance. The graph generation unit is used to generate a design scheme-performance residual distribution graph based on the calculated performance residuals, which intuitively shows the relationship between different design schemes and optimal performance.

[0046] The scheme elimination module includes: a sensitivity analysis unit and a strategy execution unit; The sensitivity analysis unit is used to analyze the sensitivity of key design variables to performance residuals, and to determine which design variables have a greater impact on performance changes and which have a smaller impact. The strategy execution unit is used to execute an invalid design scheme elimination strategy based on the sensitivity analysis results, removing those design schemes that contribute little to performance improvement or may cause performance degradation, and obtaining a set of effective design variables.

[0047] The network processing module includes: an optimized network design unit and a 3D model reconstruction unit; The optimized design network unit is trained using historical design data and performance verification data of components, and can quickly and accurately output the optimal set of design parameters based on the input set of valid design variables. The 3D model reconstruction unit is used to convert the optimal design parameters output by the optimization design network into a 3D model using 3D modeling software or algorithms, so as to intuitively display the optimized component structure.

[0048] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

Claims

1. A mechanical parts optimization design method based on artificial intelligence, characterized in that, include: Based on the collected multi-dimensional data of the components, a design space model is constructed; Based on the performance optimization objective of the design space model, key design variables are extracted through feature engineering, and the performance residual between each key design variable and the optimal performance objective is calculated to obtain the design scheme-performance residual distribution map. Based on the sensitivity trends of key design variables, an invalid design scheme elimination strategy is implemented in the design scheme-performance residual distribution map to obtain a set of valid design variables; The effective set of design variables is input into a preset optimization design network for processing, and the optimal set of design parameters is output. The optimal set of design parameters is then used to reconstruct a three-dimensional model. The optimization design network is trained using historical design data and performance verification data of the components.

2. The artificial intelligence-based mechanical component optimization design method as described in claim 1, characterized in that, The process of constructing the design space model includes: The collected multi-dimensional data of the components are projected into a multi-objective optimization space with core performance indicators as reference axes. The multi-dimensional data includes three-dimensional point cloud geometric data, mechanical property test data, material property data, and manufacturing process constraint parameters. The mechanical property test data includes at least stress, strain, and fatigue life. The material property data includes at least density and elastic modulus. The manufacturing process constraint parameters include at least machining accuracy and forming limit. Based on the projected multi-dimensional data, multiple equidistant sub-regions are divided in the design variable dimension of the multi-objective optimization space, and the set of parameter points that form local optimum density in the performance response dimension in each sub-region is extracted. Using the optimal peak performance of the parameter point set as the reference constraint boundary, a fitting model of performance consistency constraint is constructed based on the rate of change of performance sensitivity in the neighborhood of the parameter point set, so as to generate the design space model.

3. The artificial intelligence-based mechanical component optimization design method as described in claim 2, characterized in that, Based on the performance optimization objective of the design space model, key design variables are extracted through feature engineering, including: Based on the performance optimization objectives and the distribution of parameters in the design space model, the sensitivity projection direction of the design variables is determined, and the parameter-performance response curve is calculated based on the sensitivity projection direction and the parameter space density; the performance optimization objectives include lightweight, high strength, and long lifespan. Based on the shape of the design space model and the density variation of the parameters in the sensitivity projection direction, a nonlinear screening operator matching the parameter coupling characteristics is generated. Using the nonlinear screening operator as a sliding window, redundant parameter removal and key parameter enhancement operations are performed on the parameter-performance response curve. The redundant parameter removal operation is processed based on the minimum local performance contribution within the sliding window; the key parameter enhancement operation is processed based on the maximum local performance contribution within the sliding window. The results of redundant parameter elimination and key parameter enhancement operations are merged into a smooth response curve, which serves as the benchmark performance curve for the parameter-performance response curve. The parameter-performance response curve and the baseline performance curve are compared by residual analysis. Design variables with performance residual amplitude greater than a preset optimization threshold are selected to generate key design variables.

4. The artificial intelligence-based mechanical component optimization design method as described in claim 3, characterized in that, The calculation of the performance residual between each key design variable and the optimal performance target includes: Obtain the optimal performance target parameters for each key design variable; Calculate the geometric parameter residuals, mechanical property residuals, and process adaptation residuals between the key design variables and the optimal performance target parameters.

5. The artificial intelligence-based mechanical component optimization design method as described in claim 4, characterized in that, Based on the sensitivity trends of key design variables, an ineffective design scheme elimination strategy is implemented in the design scheme-performance residual distribution map to obtain a set of effective design variables, including: Calculate the spatial gradient of the performance sensitivity of each key design variable in the performance residual distribution map of the design scheme, and construct the sensitivity gradient field of the design variables; Based on the gradient direction of the performance sensitivity spatial gradient, the design schemes corresponding to the key design variables are divided into sensitivity convergence schemes and sensitivity divergence schemes. The sensitivity convergence scheme refers to a scheme in which the contribution of design variable adjustment to performance improvement gradually decreases and tends to converge. The sensitivity divergence scheme refers to a scheme in which the contribution of design variable adjustment to performance improvement continuously increases and tends to diffuse. An invalid design scheme elimination strategy is implemented, that is, the sensitivity convergence scheme is eliminated, and a set of valid design variables is generated based on the key design variables corresponding to the remaining sensitivity divergence schemes.

6. The artificial intelligence-based mechanical component optimization design method as described in claim 5, characterized in that, The optimized design network includes a design variable encoding layer and a performance sensitivity extraction layer. The performance sensitivity extraction layer is equipped with a graph convolutional network, which includes an adjacency graph construction module, a graph convolution calculation module, and a discrimination module. Specifically: The design variable encoding layer receives a set of valid design variables and constructs a design feature vector based on the geometric parameter value, performance response value, and local sensitivity coefficient of each design variable. The adjacency graph construction module uses the generated design feature vectors as nodes, calculates edge weights based on the parameter coupling degree between design variables, and constructs a parameter adjacency graph. The graph convolution calculation module performs information propagation and aggregation on the nodes in the parameter adjacency graph, and extracts the performance sensitivity change rate and parameter gradient of each design variable in its neighborhood. The differentiation module determines whether the rate of change of performance sensitivity of each design variable is greater than the weighted standard deviation of the mean rate of change of sensitivity in its neighborhood based on the rate of change of performance sensitivity and parameter gradient in its neighborhood, so as to identify the abrupt change in performance. Under the condition that the parameter gradient is less than the preset constraint threshold and the performance value is the optimal value in its neighborhood, the module identifies the local optimal performance region.

7. The artificial intelligence-based mechanical component optimization design method as described in claim 6, characterized in that, The optimization design network further includes an optimal solution determination layer, which outputs an optimal design parameter set based on the performance mutation relationship and the local performance optimal region, specifically including: Perform parameter connectivity analysis on the aforementioned performance mutation relationship and local performance optimal region to generate a set of candidate optimal parameter points; The candidate optimal parameter point set is clustered and grouped according to a multi-objective clustering algorithm to obtain the clustering results; Based on the spatial distribution density, performance gradient variation range, and process adaptability consistency of the candidate optimal parameters in the clustering results, the optimization confidence of the parameter region corresponding to the clustering results is calculated. The parameter region with an optimization confidence level greater than a preset optimization threshold is output as the optimal design parameter set.

8. The artificial intelligence-based mechanical component optimization design method as described in claim 7, characterized in that, The process of reconstructing a three-dimensional model from the optimal set of design parameters includes: Extract the three-dimensional geometric coordinates and constraint parameters corresponding to the optimal design parameter set, and construct a parameterized three-dimensional model framework based on the three-dimensional geometric coordinates and constraint parameters; Finite element simulation verification and interpolation optimization are performed on the parametric 3D model framework to generate the final optimized 3D model of the parts.

9. An artificial intelligence-based mechanical component optimization design system, used to implement the artificial intelligence-based mechanical component optimization design method according to any one of claims 1-8, characterized in that, Includes: spatial model building module, performance residual analysis module, scheme elimination module, and network processing module; The spatial model construction module is used to construct a design spatial model based on the collected multi-dimensional data of the components. The performance residual analysis module is used to extract key design variables, calculate performance residuals, and generate a design scheme-performance residual distribution map based on the performance optimization target of the design space model. The scheme elimination module is used to execute an invalid design scheme elimination strategy based on the sensitivity trend of key design variables in the design scheme-performance residual distribution map, so as to obtain a set of valid design variables. The network processing module is used to input the set of effective design variables into a preset optimization design network for processing, output the optimal set of design parameters, and reconstruct the three-dimensional model.

Images