Deep learning contact stress analysis method and system for fast iteration

Abstract

The invention provides a deep learning contact stress analysis method and system oriented to rapid iteration. The method comprises a data input and simulation result generation step, a data processing step, a data set construction step, a prediction model construction step, a prediction model training verification step and an analysis calculation step. According to the method, grid division and iterative solution steps in traditional FEA are replaced, the analysis and calculation time is remarkably shortened, calculation resource consumption is reduced, and therefore the iteration efficiency is improved, and the requirement for rapid feedback is met; a visual optimization basis for further optimization design is provided, and the production efficiency and the product quality are improved; the calculation complexity and the resource demand are obviously reduced, and the method is suitable for various product development scenes related to human body contact; according to the method, the stress distribution evaluation result is rapidly generated, the calculation efficiency is improved, the design process is closely combined with personalized requirements, and powerful support is provided for large-scale production design of specific products.

Description

Technical Field

[0001] The present invention belongs to the technical field of stress analysis, and specifically, relates to a deep learning contact stress analysis method and system for rapid iteration. Background Art

[0002] In the initial stage of product design, especially when it comes to human-product contact, such as wearable devices, car seats, medical devices, etc., accurate contact stress analysis is crucial. Accurate prediction of stress distribution can not only help designers evaluate the reliability of products under different usage conditions, but also effectively avoid local overload or uncomfortable designs, thereby improving product quality and comfort. To achieve this goal, rapid stress analysis feedback is crucial for accelerating iteration and optimizing efficiency, especially during the process of multiple modifications and rapid adjustment of design schemes.

[0003] Finite element analysis (FEA) is a numerical method widely used in the analysis of physical phenomena such as structural stress, heat conduction, and fluid dynamics. FEA obtains overall stress, strain, displacement, etc. results by decomposing an object into small elements and solving the physical behavior of each element. The specific process includes steps such as mesh generation, material property assignment, boundary condition and load application, system solution, and post-processing analysis. This method can provide high-precision stress distribution prediction for product design, helping designers judge the performance of products under different working conditions.

[0004] In the prior art, the traditional finite element analysis (FEA) method has advantages in terms of accuracy, but due to its large amount of calculation, long iterative calculation time, and high demand for computing resources, it severely restricts the rapid iteration and optimization in the design stage. Especially when it is necessary to frequently modify the geometric shape or material properties of the contact area, the traditional FEA method cannot provide timely feedback, and it may take hours or even days to obtain new stress distribution results, greatly affecting the design efficiency.

[0005] Taking ANSYS as an example, the traditional process includes importing the geometric model, mesh generation, setting material properties, applying boundary conditions and loads, etc. This process is both cumbersome and time-consuming, especially when dealing with complex contact problems, the feedback cycle is long, and it cannot meet the requirements of real-time design optimization.

[0006] Some scholars have proposed a solution that combines deep learning and finite element analysis (FEA) to improve computational efficiency, using convolutional neural networks (CNNs), long short-term memory networks (LSTMs), etc. to accelerate the solution process. However, these methods mostly rely on specific deep learning models or surrogate models. Although they can improve computational efficiency to a certain extent, there are still limitations in dealing with the problem of human-product contact. There are obvious limitations in the stress distribution analysis guided by user-related features. The current deep learning models and finite element analysis methods lack the ability to effectively process personalized user data. Therefore, the existing methods cannot accurately provide large-scale production design optimization for specific products according to specific features such as the user's body shape and usage habits.

[0007] The patent document "A Fast Stress Analysis Method for Three-Dimensional Manufacturing Technology" (CN103970952A) discloses a method for quickly solving stress by dimensionality reduction, but still requires steps such as mesh generation, and the process is cumbersome.

[0008] The patent document "Mechanical Element Analysis Device Based on Deep Learning Neural Network" (KR1020230029132A) discloses learning an autoencoder-type neural network using a learning dataset on stress distribution, and can quickly and accurately interpret the effects of mechanical components using an interpretation model based on a deep learning neural network, improving computational efficiency, but still requires the cumbersome steps of FEA.

[0009] The patent document "A Method and System for Analyzing Abnormal Forces on an Automobile Structure Based on Deep Learning" (CN118313070A) discloses collecting data to construct a three-dimensional model for force analysis and distribution simulation. Although the computational efficiency is improved, additional modeling steps are required and a certain amount of computational accuracy will be lost.

[0010] Therefore, for fast iteration, especially in the application scenario of human-product contact stress analysis, a deep learning model based on a multi-layer perceptron (MLP) is proposed, which can not only quickly generate simulation results, but also significantly reduce the computational complexity and resource requirements of traditional finite element analysis (FEA) through intuitive visualization, providing efficient and flexible design optimization support. Summary of the Invention

[0011] Aiming at the defects in the prior art, the purpose of the present invention is to provide a deep learning contact stress analysis method and system for fast iteration.

[0012] According to the deep learning contact stress analysis method for fast iteration provided by the present invention, it includes:

[0013] Data input and simulation result generation step: Input analysis data and calculate the stress distribution result of the contact area.

[0014] Data processing steps: perform standardized preprocessing on the analysis data;

[0015] Dataset construction steps: Based on the input data combination, generate a stress matrix through finite element simulation, construct a dataset and divide it;

[0016] Prediction model construction steps: Construct a multi-layer perceptron model as the stress distribution prediction model;

[0017] Prediction model training and validation steps: Use the Adam optimizer and backpropagation algorithm to update the network weights, train and validate;

[0018] Analysis and calculation steps: Upload point cloud data, input relevant human body parameters and generate a pressure matrix, and adjust the parameters.

[0019] Preferably, the analysis data includes a pressure matrix, point cloud data, and material properties. Through standard finite element analysis methods, combined with mechanical boundary conditions and constraints, the stress matrix S of the contact area is calculated, S = [σ ij m×n .

[0020] The pressure matrix P is the discrete pressure distribution in the contact area between the human body and the product, P = [P ij m×n .

[0021] The point cloud data C is the three-dimensional geometric shape of the product in the contact area, obtained by converting the three-dimensional digital model in the modeling software, C = {(x i , y i , z i )}.

[0022] The material property T defines the mechanical properties of the product material, T = [t1, t2,..., t n .

[0023] Among them, both m and n represent the grid size of the contact area;

[0024] x i , y i , z i respectively represent the coordinate positions of the i-th point in the point cloud on the x, y, and z axes in three-dimensional space;

[0025] t n represents the n-th specific mechanical property;

[0026] σ ij is the stress value at the contact point at the (i, j) position.

[0027] The standardized preprocessing includes: ​​

[0028] Using a standardized method, convert the pressure matrix P into a form with zero mean and unit variance, denoted as P norm ,

[0029] Convert the point cloud data C and material properties T to a fixed scale and perform mean normalization;

[0030] where, μ P represents the mean of the pressure matrix P, and σ P represents the standard deviation of the pressure matrix P.

[0031] Preferably, the dataset construction steps include:

[0032] Step S3.1: Randomize the input parameters, randomly generate different pressure matrices P and point cloud data C, and combine them with the material properties T to form an input parameter space;

[0033] The types and quantities of the material properties are determined according to the actual requirements of the product, and the quantity and type of the material properties of each material are kept consistent.

[0034] Step S3.2: Import the input pressure matrix P, point cloud data C, and material properties T into a finite element analysis tool for simulation to solve the stress matrix S in the contact area.

[0035] Step S3.3: Divide the generated dataset into a training set, a validation set, and a test set, with proportions of 70%, 15%, and 15% respectively.

[0036] The sample organization form of the dataset is

[0037] where, D represents the dataset;

[0038] N represents the number of data samples;

[0039] P k , C k , T k respectively represent the k-th group of inputs of P, C, and T;

[0040] S k represents the stress distribution result corresponding to the k-th group of inputs.

[0041] Preferably, the network structure of the multi-layer perceptron model includes:

[0042] Input layer: Expand the pressure matrix P, point cloud data C, and material properties T into vector forms and splice them, X = [flatten(P), flatten(C), T].

[0043] Middle layer: Composed of multiple fully connected networks, the output of each layer is processed by the ReLU activation function, h (l) = ReLU(W (l) h (l-1) + b (l) ).

[0044] Output layer: Outputs the predicted stress distribution result matrix

[0045] Loss function: The mean squared error is used as the loss function

[0046] Among them, represents the model input;

[0047] d represents the feature dimension;

[0048] represents the set of real numbers;

[0049] h represents the output of the l-th layer;

[0050] l represents the number of layers;

[0051] W represents the weight;

[0052] L represents the last layer of the network;

[0053] m and n respectively represent the number of rows and columns of the stress matrix;

[0054] σ ij , respectively represent the true and predicted stresses at the (i, j) position.

[0055] Preferably, in the training and validation steps of the prediction model, a dataset generated by finite element simulation is combined to train and validate the multi-layer perceptron model;

[0056] The mean squared error, coefficient of determination and accuracy are used for evaluation, and the prediction accuracy is comprehensively judged;

[0057] Among them, represents the mean of the stress distribution;

[0058] σ ij , respectively represent the true and predicted stresses at the (i, j) position;

[0059] The human body parameters include height, weight, and posture;

[0060] In the analysis and calculation steps, the analysis results are generated, including the pressure matrix and the visualization mapping result, and the parameters are repeatedly adjusted.

[0061] A deep learning contact stress analysis system for rapid iteration provided by the present invention includes:

[0062] Data input and simulation result generation module: Input analysis data and calculate the stress distribution result of the contact area;

[0063] Data processing module: Perform standardized preprocessing on the analysis data;

[0064] Dataset construction module: Generate a stress matrix through finite element simulation based on the input data combination, construct a dataset and divide it;

[0065] Prediction model construction module: Construct a multi-layer perceptron model as the stress distribution prediction model;

[0066] Prediction model training and validation module: Use the Adam optimizer and backpropagation algorithm to update the network weights, train and validate;

[0067] Analysis and calculation module: Upload point cloud data, input relevant human body parameters and generate a pressure matrix, and adjust the parameters.

[0068] Preferably, the analysis data includes a pressure matrix, point cloud data, and material properties. Through a standard finite element analysis method, combined with mechanical boundary conditions and constraints, the stress matrix S of the contact area is calculated, S = [σ ij m×n .

[0069] The pressure matrix P is the discrete pressure distribution of the contact area between the human body and the product, P = [P ij m×n .

[0070] The point cloud data C is the three-dimensional geometric shape of the product in the contact area, obtained by converting the three-dimensional digital model in the modeling software, C = {(x i , y i , z i )}.

[0071] The material property T defines the mechanical properties of the product material, T = [t1, t2,..., t n .

[0072] Among them, both m and n represent the grid size of the contact area;

[0073] x i , y i , z i respectively represent the coordinate positions of the i-th point in the point cloud on the x, y, and z axes in three-dimensional space;

[0074] t n represents the n-th specific mechanical property;​​

[0075] σ ij is the stress value at the contact point in the (i, j) position.

[0076] The standardized preprocessing includes:

[0077] Using a standardization method, convert the pressure matrix P into a form with zero mean and unit variance P norm ,

[0078] Convert the point cloud data C and the material property T into a fixed scale and perform mean normalization.

[0079] where, μ P represents the mean of the pressure matrix P, and σ P represents the standard deviation of the pressure matrix P.

[0080] Preferably, the dataset construction module includes:

[0081] Module M3.1: Input parameter randomization, randomly generate different pressure matrices P and point cloud data C, and combine with the material property T to form an input parameter space.

[0082] The types and quantities of the material properties are determined according to the actual needs of the product, and the quantity and type of the material properties of each material are kept consistent.

[0083] Module M3.2: Import the input pressure matrix P, point cloud data C and material property T into a finite element analysis tool for simulation, and solve the stress matrix S in the contact area.

[0084] Module M3.3: Divide the generated dataset into a training set, a validation set, and a test set, with proportions of 70%, 15%, and 15% respectively;

[0085] The sample organization form of the dataset is

[0086] where, D represents the dataset;

[0087] N represents the number of data samples;

[0088] P k , C k , T k respectively represent the k-th group of inputs of P, C, and T;

[0089] S k represents the stress distribution result corresponding to the k-th group of inputs.

[0090] Preferably, the network structure of the multi-layer perceptron model includes:

[0091] Input layer: Expand the pressure matrix P, point cloud data C, and material property T into vector form and concatenate them, X = [flatten(P), flatten(C), T].

[0092] Middle layer: Composed of multiple fully connected networks, the output of each layer is processed by the ReLU activation function, h (l) = ReLU(W (l) h (l-1) + b (l) ).

[0093] Output layer: Output the predicted stress distribution result matrix

[0094] Loss function: Use the mean squared error as the loss function

[0095] Among them, represents the model input;

[0096] d represents the feature dimension;

[0097] represents the set of real numbers;

[0098] h represents the output of the l-th layer;

[0099] l represents the number of layers;

[0100] W represents the weight;

[0101] L represents the last layer of the network;

[0102] m and n respectively represent the number of rows and columns of the stress matrix;

[0103] σ ij , respectively represent the true and predicted stresses at the (i, j) position.

[0104] Preferably, in the prediction model training and verification module, a dataset generated by finite element simulation is combined to train and verify the multi-layer perceptron model;

[0105] Use the mean squared error, coefficient of determination and accuracy rate for evaluation, and comprehensively judge the prediction accuracy;

[0106] Among them, represents the mean of the stress distribution;

[0107] σ ij , respectively represent the true and predicted stresses at the (i, j) position;

[0108] The human body parameters include height, weight, and posture;

[0109] In the analysis and calculation steps, the generated analysis results include a stress matrix and a visualization mapping result, and the adjustment parameters are repeated.

[0110] Compared with the prior art, the present invention has the following beneficial effects:

[0111] 1. By introducing a multi-layer perceptron (MLP) deep learning model, the present invention replaces the mesh generation and iterative solution steps in traditional FEA, significantly shortening the analysis and calculation time and reducing the consumption of computing resources, thereby improving the iteration efficiency and meeting the requirements of rapid feedback.

[0112] 2. On the basis of ensuring high-precision stress prediction, the present invention adjusts parameters according to the simulation results, and provides an intuitive optimization basis for further optimizing the design based on the rapid generation of the simulation matrix predicted by the model and the visualization stress distribution result, thereby realizing the improvement of product performance and user experience, and improving production efficiency and product quality.

[0113] 3. The deep learning model of the present invention simplifies the cumbersome steps in the traditional FEA method, eliminates complex mesh generation and material property assignment, and can accurately analyze the contact stress between the human body and the product, significantly reducing the computational complexity and resource requirements, and is applicable to various product development scenarios involving human contact, thereby efficiently optimizing rapid iteration.

[0114] 4. The present invention can quickly generate stress distribution evaluation results according to user-related data, such as body contact characteristics, postures, material requirements, etc., not only improving the calculation efficiency, but also closely combining the design process with personalized needs, providing strong support for the large-scale production design of specific products. BRIEF DESCRIPTION OF THE DRAWINGS

[0115] By reading the following detailed description of the non-limiting embodiments with reference to the accompanying drawings, other features, objects, and advantages of the present invention will become more apparent:

[0116] Figure 1 It is a schematic diagram of the deep learning contact stress analysis process for rapid iteration;

[0117] Figure 2 It is a schematic diagram of the system architecture of the deep learning contact stress analysis for rapid iteration;

[0118] Figure 3 It is a schematic diagram of the operation of the deep learning contact stress analysis system for rapid iteration;

[0119] Figure 4 It is a schematic diagram of the system framework of the contact stress analysis platform optimized based on the multi-layer perceptron;

[0120] Figure 5 Schematic diagram of the home page interface of the contact stress analysis platform optimized based on the multi-layer perceptron;

[0121] Figure 6 Schematic diagram of the analysis interface of the contact stress analysis platform optimized based on the multi-layer perceptron;

[0122] Figure 7 Schematic diagram of the communication interface of the contact stress analysis platform optimized based on the multi-layer perceptron;

[0123] Figure 8 Schematic diagram of the main analysis interface of the contact stress analysis platform optimized based on the multi-layer perceptron. Specific implementation manners

[0124] The present invention will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that those of ordinary skill in the art can make several changes and improvements without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0125] According to a deep learning contact stress analysis method for rapid iteration provided by the present invention, by constructing a multi-layer perceptron (MLP) prediction model, combining human characteristic parameters (such as height, weight, posture), material properties (such as elastic modulus, Poisson's ratio) and three-dimensional point cloud data of the contact area, the stress distribution result can be quickly generated. By optimizing the input parameter structure, introducing the multi-layer perceptron (MLP) model and the node-based interaction process, the contact stress distribution can be quickly predicted and visualized, providing real-time feedback and optimization basis.

[0126] Taking Figure 1 as an example, it specifically includes:

[0127] Steps of data input and simulation result generation: The input data includes three main parts, the pressure matrix, the point cloud data and the material properties.

[0128] Pressure matrix P: Describing the discrete pressure distribution of the contact area between the human body and the product, defined as P = [P ij m×n , where m and n respectively represent the grid size of the contact area.

[0129] Point cloud data C: Representing the three-dimensional geometric shape of the product in the contact area, the point cloud C = {(x i , y i , z i )} is obtained by converting the three-dimensional digital model in the modeling software. Where x i , y i , z i ​They respectively represent the coordinate positions of the \(i\)-th point in the point cloud on the \(x\), \(y\), and \(z\) axes in three-dimensional space.

[0130] Material property \(T\): Defines the mechanical properties of the product material, including Young's modulus \(E\), Poisson's ratio \(\nu\), etc., expressed as \(T = [t_1, t_2,\cdots, t_n]\). n . Where \(t_n\) n represents a specific mechanical property, such as Young's modulus, Poisson's ratio, etc., and \(n = 1, 2, 3,\cdots\).

[0131] These input data are calculated through the standard finite element analysis (FEA) method, combined with mechanical boundary conditions and constraints, to obtain the stress distribution results in the contact area, that is, the simulation results, expressed as the stress matrix \(S\):

[0132] \(S=[\sigma_{ij}]\) ij m×n

[0133] where \(\sigma_{ij}\) ij is the stress value at the contact point at the \((i, j)\) position.

[0134] By constructing a pressure matrix and a material parameter input mechanism suitable for the human contact scenario, combined with a deep learning model for simulation calculation, the contact stress distribution results are quickly generated.

[0135] Data processing steps: To ensure the stability and prediction accuracy of model training, the input data needs to be preprocessed by standardization.

[0136] Specifically, the preprocessing and standardization steps of the input data include:

[0137] Using the standardization method, convert the pressure matrix \(P\) into a form with zero mean and unit variance \(P'\): norm :

[0138]

[0139] where \(\mu\) P represents the mean of the pressure matrix \(P\), and \(\sigma\) P represents the standard deviation of the pressure matrix \(P\).

[0140] Convert the three-dimensional point cloud data \(C\) to a fixed scale and perform mean normalization processing to ensure that all input data is within the same numerical range. Standardize the material property vector \(T\) using the same method.

[0141] Through the standardization processing of the input pressure matrix, point cloud data, and material property vector, combined with finite element simulation, a large-scale dataset is generated.

[0142] Steps for constructing the dataset: To enable model training, based on combinations of multiple sets of different input data (pressure matrix, point cloud data, and material properties), the corresponding stress distribution matrix is generated through finite element simulation, and a complete training dataset is constructed and divided. Specifically, it includes:

[0143] Step S3.1: Randomization of input parameters.

[0144] Randomly generate different pressure matrices P and point cloud data C;

[0145] Combined with multiple material properties T, an input parameter space is formed. The quantity and type of material properties of each selected material are the same. For example, if two material properties, elastic modulus and Poisson's ratio, are selected, then all materials should choose these two properties. The type and quantity of selected material properties are determined according to the actual requirements of the product.

[0146] Step S3.2: Finite element simulation.

[0147] Import the input parameters into the finite element analysis tool to solve the stress distribution matrix S of the contact area.

[0148] Step S3.3: Dataset division.

[0149] Divide the generated dataset into a training set, a validation set, and a test set, with proportions of 70%, 15%, and 15% respectively.

[0150] The organization form of data samples is as follows:

[0151]

[0152] Where D is the dataset, N is the number of data samples, P k , C k , T k are the k-th group inputs of P, C, and T respectively, and S k is the corresponding stress distribution result.

[0153] Steps for constructing the prediction model: Use a multi-layer perceptron (MLP) as the stress distribution prediction model. The specific network structure includes:

[0154] Input layer: Expand the pressure matrix P, point cloud data C, and material properties T into vector form and concatenate them:

[0155] X = [flatten(P), flatten(C), T] where represents the model input, and d is the feature dimension. represents the set of real numbers.

[0156] Intermediate layer: Consists of multiple fully connected networks, and the output of each layer is processed by the ReLU activation function.

[0157] h (l) = ReLU(W (l) h (l-1) + b (l) )

[0158] Among them, h represents the output of the l-th layer, l represents the number of layers, and W represents the weight.

[0159] Output layer: Output the matrix of the predicted stress distribution result

[0160]

[0161] L is the last layer of the network, and m and n are the number of rows and columns of the stress matrix.

[0162] Loss function: The mean squared error (MSE) is used as the loss function.

[0163]

[0164] σ ij , are the true and predicted stresses at the (i, j) position respectively.

[0165] Steps for training and validating the prediction model: Use the Adam optimizer and the backpropagation algorithm (BP) to update the network weights of the MLP model, and train and validate the MLP model, thereby significantly improving the efficiency and accuracy of stress prediction. During the training and validation process, the following metrics are used to evaluate the model performance:

[0166] Mean squared error (MSE), coefficient of determination (R 2 ) and accuracy (Accuracy). Based on these metrics, the training effect is comprehensively judged, and the prediction accuracy is judged. It is expected that the MSE is small, and R 2 and accuracy are high.

[0167] Coefficient of determination

[0168] is the mean of the stress distribution.

[0169] Analysis and calculation steps: Upload the 3D point cloud file, input relevant human parameters in the MLP model, that is, human characteristic parameters (such as height, weight, posture), and generate the pressure matrix, and adjust the simulation matrix parameters.

[0170] After starting the analysis, the results of the system analysis will be generated on the right side of the interface, including the simulation matrix results and the visualization mapping results. If you are not satisfied with the generated results, re-design and iterate, so repeating all the above processes can assist in achieving rapid design iteration and re-adjustment.

[0171] The result is output in an intuitive visual form, which not only simplifies the cumbersome steps in the traditional simulation process, but also ensures the accuracy and reliability of the prediction results. It is applicable to application scenarios involving human contact pressure analysis such as wearable devices, seat design, and medical devices. Through the present invention, the stress distribution in the product contact area can be efficiently analyzed and optimized, overcoming the limitations of the prior art in terms of analysis efficiency and resource requirements.

[0172] The present invention also provides a deep learning contact stress analysis system for rapid iteration. The deep learning contact stress analysis system for rapid iteration can be implemented by executing the process steps of the deep learning contact stress analysis method for rapid iteration. That is, those skilled in the art can understand the deep learning contact stress analysis method for rapid iteration as a preferred embodiment of the deep learning contact stress analysis system for rapid iteration.

[0173] According to a deep learning contact stress analysis method and system for rapid iteration provided by the present invention, Figure 3 for example, schematic diagrams of model input and output are provided. The working system architecture is taken as Figure 2 an example. The system includes:

[0174] Data input and simulation result generation module: Input analysis data and calculate the stress distribution result in the contact area;

[0175] Data processing module: Perform standardized preprocessing on the analysis data;

[0176] Dataset construction module: Based on the input data combination, generate a stress matrix through finite element simulation, construct a dataset and divide it;

[0177] Prediction model construction module: Construct a multi-layer perceptron model as the stress distribution prediction model;

[0178] Prediction model training and verification module: Use the Adam optimizer and backpropagation algorithm to update the network weights, train and verify;

[0179] Analysis and calculation module: Upload point cloud data, input relevant human parameters and generate a pressure matrix, and adjust the parameters.

[0180] In more preferred examples, the analysis data includes a pressure matrix, point cloud data, and material properties. Through a standard finite element analysis method, combined with mechanical boundary conditions and constraints, the stress matrix S in the contact area is calculated, S = [σ ij m×n .

[0181] The pressure matrix P is the discrete pressure distribution in the human-product contact area, P = [P​ij m×n 。

[0182] The point cloud data C is the three-dimensional geometric shape of the product in the contact area, which is obtained by converting the three-dimensional digital model in the modeling software. C = {(x i , y i , z i )}.

[0183] The material property T defines the mechanical properties of the product material. T = [t1, t2,..., t n .

[0184] Among them, both m and n represent the mesh size of the contact area;

[0185] x i , y i , z i respectively represent the coordinate positions of the i-th point in the point cloud on the x, y, and z axes in three-dimensional space;

[0186] t n represents the n-th specific mechanical property;

[0187] σ ij is the stress value at the contact point in the (i, j) position.

[0188] The standardized preprocessing includes:

[0189] Using the standardization method, convert the pressure matrix P into a form with zero mean and unit variance P norm ,

[0190] Convert the point cloud data C and the material property T into a fixed scale and perform mean normalization processing.

[0191] Among them, μ P represents the mean of the pressure matrix P, and σ P represents the standard deviation of the pressure matrix P.

[0192] In more preferred examples, the dataset construction module includes:

[0193] Module M3.1: Randomize the input parameters, randomly generate different pressure matrices P and point cloud data C, and combine with the material property T to form the input parameter space.

[0194] The types and quantities of the material properties are determined according to the actual needs of the product, and the number and types of the material properties of each material are kept consistent.

[0195] ​Module M3.2: Import the input pressure matrix P, point cloud data C, and material property T into the finite element analysis tool for simulation to solve the stress matrix S in the contact area.

[0196] Module M3.3: Divide the generated dataset into a training set, a validation set, and a test set, with proportions of 70%, 15%, and 15% respectively;

[0197] The sample organization form of the dataset is

[0198] where D represents the dataset;

[0199] N represents the number of data samples;

[0200] P k , C k , T k respectively represent the k-th group of inputs of P, C, and T;

[0201] S k represents the stress distribution result corresponding to the k-th group of inputs.

[0202] In more preferred examples, the network structure of the multi-layer perceptron model includes:

[0203] Input layer: Expand the pressure matrix P, point cloud data C, and material property T into vector form and splice them, X = [flatten(P), flatten(C), T].

[0204] Middle layer: Consists of multiple fully connected networks, and the output of each layer is processed by the ReLU activation function, h (l) = ReLU(W (l) h (l-1) + b (l) ).

[0205] Output layer: Output the predicted stress distribution result matrix

[0206] Loss function: Use the mean squared error as the loss function

[0207] where, represents the model input;

[0208] d represents the feature dimension;

[0209] represents the set of real numbers;

[0210] h represents the output of the l-th layer;

[0211] l represents the number of layers;

[0212] W represents the weight;

[0213] L represents the last layer of the network;

[0214] m and n respectively represent the number of rows and columns of the stress matrix;

[0215] σ ij 、 respectively represent the true and predicted stresses at the (i, j) position.

[0216] In more preferred examples, in the prediction model training and verification module, the multi-layer perceptron model is trained and verified by combining the data set generated by finite element simulation.

[0217] The mean square error, coefficient of determination and accuracy are used for evaluation, and the prediction accuracy is judged based on a threshold;

[0218] wherein, represents the mean value of the stress distribution;

[0219] σ ij 、 respectively represent the true and predicted stresses at the (i, j) position.

[0220] The analysis calculation module generates analysis results, including the pressure matrix and the visualization mapping result, and the repetition trigger module adjusts the parameters.

[0221] The contact stress analysis platform optimized based on the multi-layer perceptron (MLP) technology mainly includes functions such as intelligent analysis, case query, and community communication. The framework is as Figure 4 shown, focusing on the analysis scenario of the contact stress between the human body and the product. By introducing the multi-layer perceptron deep learning model, the calculation process of the traditional finite element analysis (FEA) is improved, and a fast and efficient analysis tool is provided.

[0222] The home page of the platform is as Figure 5 shown, mainly displaying the user's analysis records and the analysis case library for the user to view. Clicking the "New" button can immediately create a new analysis. The "My Analysis" interface is as Figure 6 described, used to view and manage all historical analyses. The "Communication Community" interface is as Figure 7 shown, used to view and share various finite element problems and cases.

[0223] The main interface for the user to perform finite element analysis is as Figure 8 shown. The left side of the interface is the input area, and the right side is the result output, which can export multiple results for comparison.

[0224] By uploading point cloud files, inputting human body parameters, and generating a pressure matrix, the simulation and prediction of contact stress can be quickly completed, which is used to solve the problems of complex calculation, long feedback cycle, and complex operation in the prior art, and improve the design efficiency and user experience.

[0225] Those skilled in the art know that in addition to implementing the system and its various devices, modules, and units provided by the present invention in the form of pure computer-readable program code, the method steps can be logically programmed to enable the system and its various devices, modules, and units provided by the present invention to be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers to achieve the same functions. Therefore, the system and its various devices, modules, and units provided by the present invention can be regarded as a kind of hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as the structure within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as software modules for implementing the method and the structure within the hardware component.

[0226] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. Without conflict, the embodiments of the present application and the features in the embodiments can be combined arbitrarily.

Claims

1. A deep learning contact stress analysis method for rapid iteration, characterized in that Including: Data input and simulation result generation step: Input analysis data and calculate the stress distribution result of the contact area; Data processing step: Perform standardized preprocessing on the analysis data; Dataset construction step: Based on the input data combination, generate a stress matrix through finite element simulation, construct a dataset and divide it; Prediction model construction step: Construct a multi-layer perceptron model as the stress distribution prediction model; Prediction model training and validation step: Use the Adam optimizer and backpropagation algorithm to update the network weights, train and validate; Analysis and calculation step: Upload point cloud data, input relevant human parameters and generate a pressure matrix, and adjust the parameters.

2. The method for analyzing contact stress in deep learning for rapid iteration according to claim 1, wherein The analysis data includes a pressure matrix, point cloud data, and material properties. By using standard finite element analysis methods and combining mechanical boundary conditions and constraints, the stress matrix S of the contact area is calculated, where S = [σ ij m×n ;​ The pressure matrix P is the discrete pressure distribution in the area where the human body contacts the product, P = [P ij m×n ;​ The point cloud data C is the three-dimensional geometric shape of the product in the contact area, which is obtained by converting the three-dimensional digital model in the modeling software, and C = {(x i , y i , z i )}; The material property T defines the mechanical properties of the product material, T = [t1, t2,..., t n ; Wherein, both m and n represent the grid size of the contact area; x i 、y i 、z i respectively represent the coordinate positions of the i-th point in the point cloud on the x, y, and z axes in three-dimensional space; t n represents the nth specific mechanical property; σ ij is the stress value at the contact point in the (i, j) position; The said standardized preprocessing includes: Using a standardization method, the pressure matrix P is transformed into the form P with zero mean and unit variance norm , The point cloud data C and the material property T are transformed into a fixed scale and processed by mean normalization; Among them, μ P represents the mean of the pressure matrix P, and σ P represents the standard deviation of the pressure matrix P.

3. The deep learning contact stress analysis method for rapid iteration according to claim 1, characterized in that The said dataset construction step includes: Step S3.1: Randomize the input parameters, randomly generate different pressure matrices P and point cloud data C, and combine with the material property T to form an input parameter space; The type and quantity of the said material properties are determined according to the actual requirements of the product, and the quantity and type of the material properties of each material are kept consistent; Step S3.2: Import the input pressure matrix P, point cloud data C and material property T into the finite element analysis tool for simulation, and solve the stress matrix S of the contact area; Step S3.3: Divide the generated dataset into a training set, a validation set and a test set, with the proportions being 70%, 15% and 15% respectively; The sample organization form of the said data set is Wherein, D represents the dataset; N represents the number of data samples; P k ,C k ,T k respectively represent the k-th group inputs of P, C, and T; S k Indicates the stress distribution result corresponding to the k-th group of inputs.

4. The deep learning contact stress analysis method for rapid iteration according to claim 1, characterized in that The network structure of the said multi-layer perceptron model includes: Input layer: Expand the pressure matrix P, point cloud data C and material property T into vector forms and splice them, X = [flatten(P), flatten(C), T]; Middle layer: Composed of multiple fully-connected networks, the output of each layer is processed by the ReLU activation function, h (l) = ReLU(W (l) h (l-1) + b (l) ; Output layer: Output the result matrix of the predicted stress distribution Loss function: Mean squared error is used as the loss function Among them, represents the model input; d represents the feature dimension; denotes the set of real numbers; h represents the output of the l-th layer; l represents the number of layers; W represents the weight; L represents the last layer of the network; m and n respectively represent the number of rows and columns of the stress matrix; σ ij 、 respectively represent the true and predicted stresses at the (i, j) position.

5. The method for analyzing contact stress in deep learning for rapid iteration according to claim 1, characterized in that In the said prediction model training and validation step, combine the dataset generated by finite element simulation to train and validate the multi-layer perceptron model; Using mean square error, coefficient of determination and accuracy to evaluate, and comprehensively judge the prediction accuracy; Among them, represents the mean value of the stress distribution; σ ij and represent the true and predicted stresses at the (i, j) position, respectively. The said human parameters include height, weight and posture; In the said analysis and calculation step, generate the analysis results, including the pressure matrix and the visualization mapping result, and repeatedly adjust the parameters.

6. A deep learning contact stress analysis system for rapid iteration, characterized in that, Including: Data input and simulation result generation module: Input analysis data and calculate the stress distribution result of the contact area; Data processing module: Perform standardized preprocessing on the analysis data; Dataset construction module: Based on the input data combination, generate a stress matrix through finite element simulation, construct a dataset and divide it; Prediction model construction module: Construct a multi-layer perceptron model as the stress distribution prediction model; Prediction model training and validation module: Use the Adam optimizer and backpropagation algorithm to update the network weights, train and validate; Analysis and calculation module: Upload point cloud data, input relevant human parameters and generate a pressure matrix, and adjust the parameters.

7. The deep learning contact stress analysis system for rapid iteration according to claim 6, characterized in that The analysis data includes a pressure matrix, point cloud data, and material properties. By using a standard finite element analysis method and combining mechanical boundary conditions and constraints, the stress matrix S of the contact area is calculated, where S = [σ ij m×n ;​ The pressure matrix P is the discrete pressure distribution in the area where the human body contacts the product, and P = [P ij m×n ;​ The point cloud data C is the three-dimensional geometric shape of the product in the contact area, which is obtained by converting the three-dimensional digital model in the modeling software, C = {(x i , y i , z i )}; The material property T defines the mechanical properties of the product material, T = [t1, t2,..., t n ; Wherein, both m and n represent the grid size of the contact area; x i 、y i 、z i respectively represent the coordinate positions of the i-th point in the point cloud on the x, y, and z axes in three-dimensional space; t n represents the nth specific mechanical property; σ ij is the stress value at the contact point in the (i, j) position; The said standardized preprocessing includes: Using a standardized method, the pressure matrix P is transformed into a form with zero mean and unit variance, P norm , The point cloud data C and the material property T are transformed into a fixed scale and processed by mean normalization; Among them, μ P represents the mean of the pressure matrix P, and σ P represents the standard deviation of the pressure matrix P.

8. The deep learning contact stress analysis system for rapid iteration according to claim 6, characterized in that The said dataset construction module includes: Module M3.1: Randomize the input parameters, randomly generate different pressure matrices P and point cloud data C, and combine with the material property T to form an input parameter space; The types and quantities of the material properties are determined according to the actual requirements of the product, and the quantity and type of the material properties of each material are kept consistent; Module M3.2: Import the input pressure matrix P, point cloud data C, and material properties T into a finite element analysis tool for simulation to solve the stress matrix S of the contact area; Module M3.3: Divide the generated data set into a training set, a validation set, and a test set, with proportions of 70%, 15%, and 15% respectively; The sample organization form of the said data set is Among them, D represents the data set; N represents the number of data samples; P k ,C k ,T k respectively represent the k-th group of inputs of P, C, and T; S k Represents the stress distribution result corresponding to the k-th group of inputs.

9. The deep learning contact stress analysis system for rapid iteration according to claim 6, wherein The network structure of the multi-layer perceptron model includes: Input layer: Expand the pressure matrix P, point cloud data C, and material properties T into vector forms and splice them, X = [flatten(P), flatten(C), T]; Intermediate layer: Composed of multiple fully connected networks, the output of each layer is processed by the ReLU activation function, h (l) = ReLU(W (l) h (l-1) + b (l) ); Output layer: Output the result matrix of the predicted stress distribution Loss function: Mean squared error is used as the loss function Among them, represents the model input; d represents the feature dimension; denote the set of real numbers; h represents the output of the l-th layer; l represents the number of layers; W represents the weight; L represents the last layer of the network; m and n respectively represent the number of rows and columns of the stress matrix; σ ij and respectively represent the true and predicted stresses at the (i, j) position.

10. The deep learning contact stress analysis system for rapid iteration according to claim 6, characterized in that, In the prediction model training and validation module, the multi-layer perceptron model is trained and validated by combining the data set generated by the finite element simulation; Using mean square error, coefficient of determination and accuracy for evaluation, comprehensively judging the prediction accuracy; Among them, represents the mean value of the stress distribution; σ ij and represent the true and predicted stresses at the (i, j) position, respectively; The human body parameters include height, weight, and posture; In the analysis and calculation steps, the analysis results are generated, including the pressure matrix and the visualization mapping result, and the parameters are repeatedly adjusted.

Images