Part manufacturing-oriented lattice structure mechanical property prediction method

Abstract

The invention discloses a part manufacturing-oriented lattice structure mechanical property prediction method. Aiming at the bottlenecks of the traditional method in precision, efficiency and speed, the method is optimized through four modules: a structure generation module generates VTK format data; the feature extraction module analyzes five types of features including point coordinate statistics, unit topology and the like; the deep crossover network prediction module outputs porosity and compression modulus through double-path processing of a crossover network and a deep network; and the optimization module optimizes the printing parameters according to the prediction result. The method can accurately capture the interaction effect of the microstructure and the technological parameters, accelerates the whole process of'design-prediction-optimization ', and is suitable for performance prediction of core components of high-end equipment.

Description

Technical Field

[0001] The present invention relates to the fields of additive manufacturing, deep learning, and computational mechanics, and in particular to a method for predicting the mechanical properties of lattice structures for parts manufacturing. Background Art

[0002] In the field of robotics and intelligent manufacturing, the lattice structure is a three-dimensional lightweight structure composed of periodically arranged micro units, which forms a spatial network through rods and nodes. As a key technology for achieving a balance between lightweight and high performance, it is widely used in core components such as collaborative robot end effectors, precision guide rails and sliders, and automation equipment frames. This type of application scenario has strict requirements for performance prediction. For example, the end effector of the robotic arm must meet the comprehensive indicators of lightweight, high stiffness and high-precision positioning; the high-speed guide rail slider must have fast dynamic response capabilities and maintain structural stability under long-term cyclic loads. Traditional performance prediction methods are at a bottleneck in the three dimensions of accuracy, efficiency, and speed, and it is difficult to meet the technical requirements of intelligent manufacturing for flexibility and rapid iteration.

[0003] Insufficient prediction accuracy directly affects the performance of high-end equipment. When dealing with complex lattice topologies, finite element analysis suffers from significant deviations in stress distribution predictions due to the conflict between meshing accuracy and computational cost, especially in the unit connection area. The empirical formula oversimplifies the structure-performance relationship and uses a single density parameter for linear fitting, which cannot reflect the differences in mechanical behavior of different topological configurations. Ordinary machine learning models are limited by feature dimensions and only use a small number of geometric parameters. They cannot capture the microscopic configuration characteristics of the lattice structure, such as unit arrangement patterns and bounding box ratios, resulting in poor stability in performance prediction results.

[0004] Inefficient prediction will extend the product development cycle. Traditional parameter optimization uses orthogonal experimental design, which can only adjust a small number of process parameters at a time and cannot reveal the coupling effects between multiple parameters. When multiple process parameters change in coordination, nonlinear performance responses may occur, and a simple control variable method cannot capture this complex relationship. Multi-objective optimization requires the construction of separate prediction models for different performance indicators, making it impossible to coordinate the trade-off between lightweight and dynamic performance, causing the design process to fall into a cycle of trial and error. This inefficient optimization process leads to lengthy process debugging cycles for core robot components, making it difficult to adapt to customized production needs.

[0005] Slow prediction speeds hinder the real-time response of intelligent manufacturing. Traditional numerical simulation methods consume significant computing resources for single performance evaluations of complex lattice models and cannot support rapid comparisons of multiple solutions. Physical experimental methods also face lengthy process cycles, requiring multiple steps from structural preparation to mechanical testing, and are unable to process multiple parameter solutions in parallel. This speed makes real-time feedback development difficult, limiting the flexible production capabilities of intelligent manufacturing.

[0006] Targeting the performance prediction needs of robotics and intelligent manufacturing, this invention addresses the core issues of traditional methods, including insufficient accuracy, low efficiency, and slow speed, through multi-dimensional feature extraction, deep cross-network modeling, and ultra-fast closed-loop optimization. Its technical advantages lie in achieving precise characterization of microstructures, capturing the interactive effects of process parameters, and ultimately accelerating the entire "design-prediction-optimization" process. This innovative solution is particularly suitable for the rapid design of high-end equipment such as collaborative robots and precision automation equipment, providing key technical support for flexible production in intelligent manufacturing. Summary of the Invention

[0007] This paper proposes a method for predicting the mechanical properties of lattice structures based on a deep cross-network. The method targets truss-type lattice structures and aims to achieve high-precision prediction of mechanical properties in the fields of robotics and intelligent manufacturing. First, the proposed method for predicting the mechanical properties of lattice structures for parts manufacturing integrates four modules: a structure generation module, a feature extraction module, a deep cross-network prediction module, and an optimization module. This design upgrades traditional experience-driven lattice structure design to a data-driven, precision development model. Secondly, key process parameters are input based on the operating characteristics of the robot's core components to generate lattice structure data in VTK format. Geometric and topological information is parsed from the VTK data, and five key features are extracted to comprehensively quantify the physical properties of the lattice structure, providing a data foundation for part quality control. Next, the cross-network explicitly models the nonlinear interactions of process parameters, generating high-order cross terms while retaining interpretability. The deep network then learns local features such as unit arrangement and connectivity through fully connected layers, achieving dimensionality reduction and noise filtering. Finally, the dual-path outputs are concatenated and fed into a linear layer, which simultaneously outputs porosity and compression modulus, directly linking structural performance to manufacturing parameters. This design combines feature engineering and deep learning to provide a high-precision, high-efficiency, innovative, and easy-to-implement lattice structure design solution for the robotics and intelligent manufacturing fields. It promotes the upgrade of truss parts from traditional empirical design to data-driven precision manufacturing, meeting the core needs of the field for lightweight, high strength, and personalization. To achieve the above goals, the following steps are implemented:

[0008] Step 1: The structure generation module generates an Octet truss lattice suitable for STL format based on the line spacing, layer height, printing speed, and printing temperature requirements of the robot's core components, and then converts the STL mesh into a string in VTK format.

[0009] Step 2: The feature extraction module first performs point coordinate analysis and unit topology analysis on the robot part VTK format string converted by the structure generation module, and then obtains point coordinate statistical features, unit topology features, bounding box geometric features, structure density features, and complexity features;

[0010] Step 2.1: Perform point coordinate analysis and cell topology analysis using VTK data, and then obtain point coordinate statistical features, bounding box geometric features, cell topology features, structural density features, and complexity features. The expressions involved are as follows:

[0011] ,

[0012] in is the statistical feature of point coordinates, yes The mean value of the coordinates, yes The mean value of the coordinates, yes The mean value of the coordinates, yes The standard deviation of the coordinates, yes The standard deviation of the coordinates, yes The standard deviation of the coordinates, All points The minimum value of the coordinates, All points The minimum value of the coordinates, All points The minimum value of the coordinates, All points The maximum value of the coordinates, All points The maximum value of the coordinates, All points The maximum value of the coordinates;

[0013] ,

[0014] in is the bounding box geometry, is the bounding box again The length in the axial direction, is the bounding box again The length in the axial direction, is the bounding box again The length in the axial direction, is the volume of the bounding box;

[0015] ,

[0016] in is the cell topological characteristic, is the average of all side lengths of all elements, is the standard deviation of all edge lengths of all cells, is the average of all face areas of all cells, is the sum of all face areas of all cells;

[0017] ,

[0018] in is the structural density characteristic, is the total number of points in the lattice structure, that is, the number of all points in the VTK data, is the sum of all surface areas of all units in the lattice structure, is the minimum axis-aligned bounding box volume enclosing the lattice structure;

[0019] ,

[0020] in is the complexity feature, is the total number of points in the lattice structure, is the total number of faces in the lattice structure;

[0021] ,

[0022] in is the initial input feature vector of the cross network.

[0023] Step 3: The deep cross network prediction module first receives the 3D structural data of the robot part and extracts the feature vector from it. The feature vector is then input into the cross network and the deep network. The outputs of the cross network and the deep network are then concatenated into a combined feature vector. Finally, the combined feature vector is input into the linear output layer to obtain the two mechanical properties of porosity and compression modulus that match the manufacturing requirements of the robot part.

[0024] Step 3.1: Input the feature vector into the cross network. The weights and biases of the cross layer are initialized using a truncated normal distribution. The cross network explicitly captures the interaction effects of line spacing, layer height, speed, and temperature. At the same time, the three-layer cross network can generate fourth-order feature cross terms, focusing on processing global features of robot parts such as density and bounding box. The relevant expressions involved are as follows:

[0025] ,

[0026] in It is a cross network The output feature vector of the layer, is the initial input feature vector of the cross network, represents multiplication, It is a cross network the transpose of the layer weight matrix, It is a cross network The input feature vector of the layer, It is a cross network The bias vector of the layer;

[0027] ,

[0028] ,

[0029] in It is a cross network The weight matrix of the layer, It is a cross network The bias vector of the layer, is a truncated normal distribution, is the mean of the truncated normal distribution, is the standard deviation of the truncated normal distribution;

[0030] Step 3.2: Input the feature vector into the deep network. The deep network consists of three fully connected layers, each followed by a ReLU activation function, and the dimensions between layers decrease. It focuses on learning local features of robot parts such as unit topology. The relevant expressions involved are as follows:

[0031] ,

[0032] ,

[0033] ,

[0034] ,

[0035] in is the hidden layer output of the first layer of the deep network, , is the hidden layer output of the second layer of the deep network, , is the hidden layer output of the third layer of the deep network, , is the weight matrix of the first layer of the deep network, , is the weight matrix of the second layer of the deep network, , is the weight matrix of the third layer of the deep network, , is the bias vector of the first layer of the deep network, , is the bias vector of the second layer of the deep network, , is the bias vector of the third layer of the deep network, , is the input feature vector of the deep network, is the input feature dimension;

[0036] Step 3.3: Concatenate the outputs of the cross network and the deep network into a combined feature vector:

[0037] ,

[0038] in is the final output of the cross network, is the final output of the deep network, is the fused feature vector;

[0039] Step 3.4: Input the combined feature vector into the linear output layer to obtain the predicted values ​​of the two mechanical properties of porosity and compression modulus:

[0040] ,

[0041] ,

[0042] ,

[0043] in is the original prediction of porosity, is the raw prediction of the compressibility modulus, is the weight matrix of the output layer, is the fused feature vector, is the bias vector of the output layer, is the final predicted value of porosity, is the final predicted value of the compression modulus, is the Sigmoid activation function, is the Softplus activation function.

[0044] Step 4: An optimization module is configured to optimize printing parameters based on the predicted performance and the manufacturing constraints of the robotic part. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is an overall flow chart of an embodiment of the present invention; DETAILED DESCRIPTION

[0046] In order to more clearly illustrate the purpose, technical solutions and advantages of the present invention, the present invention will be described in detail below with the help of the accompanying drawings and through specific embodiments.

[0047] Figure 1 This is a flow chart of an embodiment. This embodiment provides a method for predicting the mechanical properties of lattice structures for parts manufacturing. The specific process includes: generating truss lattice structure data in VTK format through a structure generation module, extracting five types of key features from VTK data through a feature extraction module to generate a multi-dimensional feature vector, performing dual-path processing through a deep cross network prediction module, explicitly modeling the interaction effect of process parameters through the cross network path, and finally realizing the prediction of mechanical properties porosity and compression modulus through feature fusion output.

[0048] A method for predicting mechanical properties of lattice structures for parts manufacturing includes the following steps:

[0049] Step 1: The structure generation module generates an Octet truss lattice suitable for STL format based on the line spacing, layer height, printing speed, and printing temperature requirements of the robot's core components, and then converts the STL mesh into a string in VTK format.

[0050] Step 2: The feature extraction module first performs point coordinate analysis and unit topology analysis on the robot part VTK format string converted by the structure generation module, and then obtains point coordinate statistical features, unit topology features, bounding box geometric features, structure density features, and complexity features;

[0051] Step 2.1: Perform point coordinate analysis and cell topology analysis using VTK data, and then obtain point coordinate statistical features, bounding box geometric features, cell topology features, structural density features, and complexity features. The expressions involved are as follows:

[0052] ,

[0053] in is the statistical feature of point coordinates, yes The mean value of the coordinates, yes The mean value of the coordinates, yes The mean value of the coordinates, yes The standard deviation of the coordinates, yes The standard deviation of the coordinates, yes The standard deviation of the coordinates, All points The minimum value of the coordinates, All points The minimum value of the coordinates, All points The minimum value of the coordinates, All points The maximum value of the coordinates, All points The maximum value of the coordinates, All points The maximum value of the coordinates;

[0054] ,

[0055] in is the bounding box geometry, is the bounding box again The length in the axial direction, is the bounding box again The length in the axial direction, is the bounding box again The length in the axial direction, is the volume of the bounding box;

[0056] ,

[0057] in is the cell topological characteristic, is the average of all side lengths of all elements, is the standard deviation of all edge lengths of all cells, is the average of all face areas of all cells, is the sum of all face areas of all cells;

[0058] ,

[0059] in is the structural density characteristic, is the total number of points in the lattice structure, that is, the number of all points in the VTK data, is the sum of all surface areas of all units in the lattice structure, is the minimum axis-aligned bounding box volume enclosing the lattice structure;

[0060] ,

[0061] in is the complexity feature, is the total number of points in the lattice structure, is the total number of faces in the lattice structure;

[0062] ,

[0063] in is the initial input feature vector of the cross network.

[0064] Step 3: The deep cross network prediction module first receives the 3D structural data of the robot part and extracts the feature vector from it. The feature vector is then input into the cross network and the deep network. The outputs of the cross network and the deep network are then concatenated into a combined feature vector. Finally, the combined feature vector is input into the linear output layer to obtain the two mechanical properties of porosity and compression modulus that match the manufacturing requirements of the robot part.

[0065] Step 3.1: Input the feature vector into the cross network. The weights and biases of the cross layer are initialized using a truncated normal distribution. The cross network explicitly captures the interaction effects of line spacing, layer height, speed, and temperature. At the same time, the three-layer cross network can generate fourth-order feature cross terms, focusing on processing global features of robot parts such as density and bounding box. The relevant expressions involved are as follows:

[0066] ,

[0067] in It is a cross network The output feature vector of the layer, is the initial input feature vector of the cross network, represents multiplication, It is a cross network the transpose of the layer weight matrix, It is a cross network The input feature vector of the layer, It is a cross network The bias vector of the layer;

[0068] ,

[0069] ,

[0070] in It is a cross network The weight matrix of the layer, It is a cross network The bias vector of the layer, is a truncated normal distribution, is the mean of the truncated normal distribution, is the standard deviation of the truncated normal distribution;

[0071] Step 3.2: Input the feature vector into the deep network. The deep network consists of three fully connected layers, each followed by a ReLU activation function, and the dimensions between layers decrease. It focuses on learning local features of robot parts such as unit topology. The relevant expressions involved are as follows:

[0072] ,

[0073] ,

[0074] ,

[0075] ,

[0076] in is the hidden layer output of the first layer of the deep network, , is the hidden layer output of the second layer of the deep network, , is the hidden layer output of the third layer of the deep network, , is the weight matrix of the first layer of the deep network, , is the weight matrix of the second layer of the deep network, , is the weight matrix of the third layer of the deep network, , is the bias vector of the first layer of the deep network, , is the bias vector of the second layer of the deep network, , is the bias vector of the third layer of the deep network, , is the input feature vector of the deep network, is the input feature dimension;

[0077] Step 3.3: Concatenate the outputs of the cross network and the deep network into a combined feature vector:

[0078] ,

[0079] in is the final output of the cross network, is the final output of the deep network, is the fused feature vector;

[0080] Step 3.4: Input the combined feature vector into the linear output layer to obtain the predicted values ​​of the two mechanical properties of porosity and compression modulus:

[0081] ,

[0082] ,

[0083] ,

[0084] in is the original prediction of porosity, is the raw prediction of the compressibility modulus, is the weight matrix of the output layer, is the fused feature vector, is the bias vector of the output layer, is the final predicted value of porosity, is the final predicted value of the compression modulus, is the Sigmoid activation function, is the Softplus activation function.

[0085] Step 4: An optimization module is configured to optimize printing parameters based on the predicted performance and the manufacturing constraints of the robotic part.

Claims

1. A method for predicting mechanical properties of lattice structures for parts manufacturing, characterized in that: The steps include: Step 1: The structure generation module generates an Octet truss lattice suitable for STL format according to the line spacing, layer height, printing speed, and printing temperature requirements of the robot's core components, and then converts the STL mesh into a string in VTK format; Step 2: The feature extraction module first performs point coordinate analysis and unit topology analysis on the robot part VTK format string converted by the structure generation module, and then obtains point coordinate statistical features, unit topology features, bounding box geometric features, structure density features, and complexity features; Step 3: The deep cross network prediction module first receives the 3D structural data of the robot part and extracts the feature vector from it. The feature vector is then input into the cross network and the deep network. The outputs of the cross network and the deep network are then concatenated into a combined feature vector. Finally, the combined feature vector is input into the linear output layer to obtain the two mechanical properties of porosity and compression modulus that match the manufacturing requirements of the robot part. Step 4: An optimization module is configured to optimize printing parameters based on the predicted performance and the manufacturing constraints of the robotic part.

2. A method for predicting mechanical properties of lattice structures for parts manufacturing according to claim 1, characterized in that: The feature extraction module described in step 2 first performs point coordinate analysis and unit topology analysis on the robot part VTK format string converted by the structure generation module, and then obtains point coordinate statistical features, unit topology features, bounding box geometric features, structure density features and complexity features. The specific implementation is as follows: Step 2.1: Perform point coordinate analysis and cell topology analysis using VTK data, and then obtain point coordinate statistical features, bounding box geometric features, cell topology features, structural density features, and complexity features. The expressions involved are as follows: ， in is the statistical feature of point coordinates, yes The mean value of the coordinates, yes The mean value of the coordinates, yes The mean value of the coordinates, yes The standard deviation of the coordinates, yes The standard deviation of the coordinates, yes The standard deviation of the coordinates, All points The minimum value of the coordinates, All points The minimum value of the coordinates, All points The minimum value of the coordinates, All points The maximum value of the coordinates, All points The maximum value of the coordinates, All points The maximum value of the coordinates; ， in is the bounding box geometry, is the bounding box again The length in the axial direction, is the bounding box again The length in the axial direction, is the bounding box again The length in the axial direction, is the volume of the bounding box; ， in is the cell topological characteristic, is the average of all side lengths of all elements, is the standard deviation of all edge lengths of all cells, is the average of all face areas of all cells, is the sum of all face areas of all cells; ， in is the structural density characteristic, is the total number of points in the lattice structure, that is, the number of all points in the VTK data, is the sum of all surface areas of all units in the lattice structure, is the minimum axis-aligned bounding box volume enclosing the lattice structure; ， in is the complexity feature, is the total number of points in the lattice structure, is the total number of faces in the lattice structure; ， in is the initial input feature vector of the cross network.

3. The method for predicting mechanical properties of lattice structures for parts manufacturing according to claim 1, characterized in that: The deep cross network prediction module described in step 3 first receives the three-dimensional structure data of the robot part and extracts the feature vector from it. Then, the feature vector is input into the cross network and the deep network. The outputs of the cross network and the deep network are then spliced ​​into a combined feature vector. Finally, the combined feature vector is input into the linear output layer to obtain the two mechanical properties of porosity and compression modulus that match the manufacturing requirements of the robot part. The specific implementation is as follows: Step 3.1: Input the feature vector into the cross network. The weights and biases of the cross layer are initialized using a truncated normal distribution. The cross network explicitly captures the interaction effects of line spacing, layer height, speed, and temperature. At the same time, the three-layer cross network can generate fourth-order feature cross terms, focusing on processing global features of robot parts such as density and bounding box. The relevant expressions involved are as follows: ， in It is a cross network The output feature vector of the layer, is the initial input feature vector of the cross network, represents multiplication, It is a cross network the transpose of the layer weight matrix, It is a cross network The input feature vector of the layer, It is a cross network The bias vector of the layer; ， ， in It is a cross network The weight matrix of the layer, It is a cross network The bias vector of the layer, is a truncated normal distribution, is the mean of the truncated normal distribution, is the standard deviation of the truncated normal distribution; Step 3.2: Input the feature vector into the deep network. The deep network consists of three fully connected layers, each followed by a ReLU activation function, and the dimensions between layers decrease. It focuses on learning local features of robot parts such as unit topology. The relevant expressions involved are as follows: ， ， ， ， in is the hidden layer output of the first layer of the deep network, , is the hidden layer output of the second layer of the deep network, , is the hidden layer output of the third layer of the deep network, , is the weight matrix of the first layer of the deep network, , is the weight matrix of the second layer of the deep network, , is the weight matrix of the third layer of the deep network, , is the bias vector of the first layer of the deep network, , is the bias vector of the second layer of the deep network, , is the bias vector of the third layer of the deep network, , is the input feature vector of the deep network, is the input feature dimension; Step 3.3: Concatenate the outputs of the cross network and the deep network into a combined feature vector: ， in is the final output of the cross network, is the final output of the deep network, is the fused feature vector; Step 3.4: Input the combined feature vector into the linear output layer to obtain the predicted values ​​of the two mechanical properties of porosity and compression modulus: ， ， ， in is the original prediction of porosity, is the raw prediction of the compressibility modulus, is the weight matrix of the output layer, is the fused feature vector, is the bias vector of the output layer, is the final predicted value of porosity, is the final predicted value of the compression modulus, is the Sigmoid activation function, is the Softplus activation function.

Images