An antenna assembly strain prediction method, system, device, and medium
By combining the strain prediction model of GNN and PINN with the meta-learning optimization strategy, the problem of accurate strain prediction during antenna assembly was solved, achieving efficient and accurate strain prediction and improving production yield and equipment reliability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTHWEST CHINA RES INST OF ELECTRONICS EQUIP
- Filing Date
- 2026-02-10
- Publication Date
- 2026-06-12
AI Technical Summary
During antenna assembly, the strain caused by mechanical fastening operations is difficult to predict accurately, leading to a decrease in product yield and a reduction in equipment reliability.
A strain prediction model combining GNN and PINN is adopted. Through meta-learning optimization strategy and fusion meta-learning training, accurate prediction of antenna assembly strain is achieved.
It improves the accuracy and reliability of antenna assembly strain prediction, reduces model training time and computing cost, and improves production yield and equipment stability.
Smart Images
Figure CN122196871A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of deep learning and industrial assembly technology, and in particular to an antenna assembly strain prediction method, system, device and medium. Background Technology
[0002] With the rapid development of fifth-generation mobile communication technology, related equipment continues to evolve towards high frequency, high speed, and high density integration. This trend is driving the design and manufacturing of antenna components towards greater precision, while also making the structural stability of these components more fragile.
[0003] In the actual assembly of antennas, mechanical fastening is an essential process. The forces generated during this process can easily cause strain within the printed circuit board. This strain directly impacts the product's production yield.
[0004] Meanwhile, the potential damage caused by strain will gradually become apparent during long-term use of the product, thereby reducing the operational reliability and lifespan of the equipment. Therefore, how to accurately predict the strain of the printed circuit board during antenna assembly has become a key problem that urgently needs to be solved in the industry. Summary of the Invention
[0005] To address the aforementioned issues, this invention proposes an antenna assembly strain prediction method, system, device, and medium. By combining GNN (Graph Neural Network) and PINN (Physical Information Neural Network) and integrating meta-learning training, it satisfies both data fitting requirements and adheres to mechanical equilibrium constraints, enabling accurate prediction of antenna assembly strain.
[0006] The technical solution adopted in this invention is as follows: A method for predicting strain in antenna assembly, comprising: The assembly process parameters of the antenna printed circuit board nodes are obtained and mapped to a distribution space of the same order of magnitude through standardization processing; the assembly process parameters include the three-dimensional coordinate data of the nodes, the load data of different preload conditions, and the node strain data corresponding to each condition. The standardized assembly process parameters are input into the strain prediction model, which then outputs the strain prediction results for the antenna assembly. The strain prediction model includes a coordinate feature extraction branch, a preload feature extraction branch, a feature fusion and decoding unit, and a physical constraint unit. The coordinate feature branch performs a nonlinear transformation on the three-dimensional coordinates of the nodes. The preload feature branch maps the preload vector into a high-dimensional feature embedding. The feature fusion and decoding unit concatenates and fuses the features output from the two feature extraction branches with the original standardized coordinates and decodes the prediction results. The physical constraint unit forces the prediction results to satisfy the constitutive equations in elasticity.
[0007] Furthermore, the coordinate feature extraction branch includes a feature encoding layer and a graph convolutional layer connected in sequence. The feature encoding layer uses a fully connected network to map the input standardized three-dimensional coordinates into low-dimensional feature vectors. The graph convolutional layer aggregates the features of nodes and their neighbors layer by layer based on the spatial association of the node neighbor list to extract spatial topological association features.
[0008] Furthermore, the preload feature branch maps the preload vector into a high-dimensional feature embedding, including: using a fully connected network to map the input standardized preload into a global feature vector, and then expanding the global feature vector into a feature matrix matching the number of nodes through tensor replication, so as to achieve alignment between the preload and node features.
[0009] Furthermore, the feature fusion and decoding unit concatenates and fuses the features output from the two feature extraction branches with the original standardized coordinates and decodes and outputs the prediction result, including: concatenating the spatial topological correlation features output from the coordinate feature extraction branch, the global working condition features output from the preload feature extraction branch, and the original standardized coordinates in dimensions to form a fused feature vector; and then decoding the fused feature vector through a fully connected network to output standardized prediction values of multiple strain components.
[0010] Furthermore, the physical constraint unit forces the prediction result to satisfy the constitutive equation in elasticity by: automatically differentiating and calculating the divergence residuals of the strain field in the x, y, and z directions, and converting the divergence residuals into optimizable loss terms to force the prediction result to satisfy the constitutive equation in elasticity.
[0011] Furthermore, the strain prediction model is trained using a meta-learning optimization strategy, which includes: Inner loop adaptation: For a single prestressed load case task, starting from the initial weights of the strain prediction model, the data loss and physical loss are calculated using the support set of the task, and the weights are updated through gradient descent to obtain fast weights that adapt to the current task. Outer loop generalization: The initial weights are optimized based on the adaptation effect of multiple pre-tightening working conditions. For each task, after obtaining the fast weights, the generalization loss is calculated using the query set of that task. The query set losses of multiple tasks are aggregated as the meta-loss, and the initial weights of the strain prediction model are updated through backpropagation.
[0012] Furthermore, in the meta-learning optimization strategy, each pre-tightening condition task is a meta-task, and the dataset of the meta-task includes the three-dimensional coordinates of all nodes, the adjacency relationships between nodes, and the labels of the real strain field under the corresponding condition.
[0013] An antenna assembly strain prediction system, comprising: The data acquisition and processing module is configured to acquire the assembly process parameters of the antenna printed circuit board nodes and map them to a distribution space of the same magnitude through standardization processing; the assembly process parameters include the three-dimensional coordinate data of the nodes, the load data of different preload conditions, and the node strain data corresponding to each condition. The strain prediction model is configured to output antenna assembly strain prediction results based on standardized assembly process parameters. The strain prediction model includes a coordinate feature extraction branch, a preload feature extraction branch, a feature fusion and decoding unit, and a physical constraint unit. The coordinate feature branch performs nonlinear transformation on the three-dimensional coordinates of the nodes. The preload feature branch maps the preload vector into a high-dimensional feature embedding. The feature fusion and decoding unit concatenates and fuses the features output from the two feature extraction branches with the original standardized coordinates and decodes and outputs the prediction result. The physical constraint unit forces the prediction result to satisfy the constitutive equation in elasticity.
[0014] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the antenna assembly strain prediction method.
[0015] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the antenna assembly strain prediction method.
[0016] The beneficial effects of this invention are as follows: 1. The strain prediction model constructed in this invention achieves accurate prediction of strain on antenna printed circuit boards through a multi-branch feature extraction and fusion mechanism. The coordinate feature extraction branch can effectively mine the spatial topological correlation features of nodes, while the preload feature extraction branch can capture global features under different load conditions. The splicing and fusion of these two branches with the original standardized coordinates can fully integrate multi-dimensional process parameter information and improve the model's ability to represent complex assembly scenarios.
[0017] 2. This invention introduces a physical constraint unit, which calculates the multi-directional divergence residual of the strain field and transforms it into an optimization loss term, forcing the prediction results to follow the constitutive equation of elasticity. This makes the prediction output not only fit the actual measurement data, but also conform to the physical laws, effectively avoiding the result deviation caused by simple data fitting, and improving the reliability and engineering applicability of the prediction results.
[0018] 3. This invention employs a meta-learning optimization strategy. The inner loop rapidly adapts to a single preload condition task, while the outer loop optimizes the model's initial weights, significantly enhancing the model's generalization ability under different working conditions. This strategy eliminates the need for separate model training for each working condition, greatly reducing training time and computational costs, and improving the model's engineering adaptation efficiency.
[0019] In summary, this invention enables efficient and accurate prediction of printed circuit board strain during antenna assembly, and can identify potential strain hazards caused by mechanical fastening and other operations in advance. This provides data support for optimizing assembly process parameters, thereby improving product yield and ensuring the long-term stability and reliability of equipment operation. Attached Figure Description
[0020] Figure 1 This is a schematic diagram of the antenna assembly strain prediction method in Embodiment 2 of the present invention.
[0021] Figure 2 This is an antenna structure diagram of Embodiment 3 of the present invention.
[0022] Figure 3 This is a schematic diagram of the array antenna assembly and diagonal tightening in Embodiment 3 of the present invention.
[0023] Figure 4 This is a strain simulation result diagram of Embodiment 3 of the present invention.
[0024] Figure 5 This is a flowchart of the inner loop training process in Embodiment 3 of the present invention.
[0025] Figure 6 This is the flowchart of the outer loop training in Embodiment 3 of the present invention. Detailed Implementation
[0026] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments are now described. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0027] Example 1 like Figure 1 As shown, this embodiment provides an antenna assembly strain prediction method, including: obtaining the assembly process parameters of the antenna printed circuit board nodes and mapping them to a distribution space of the same magnitude through standardization processing; inputting the standardized assembly process parameters into a strain prediction model and outputting the antenna assembly strain prediction result through the strain prediction model.
[0028] The assembly process parameters include three-dimensional coordinate data of nodes, load data under different preload conditions, and corresponding node strain data under each condition. The strain prediction model includes a coordinate feature extraction branch, a preload feature extraction branch, a feature fusion and decoding unit, and a physical constraint unit. The coordinate feature branch performs nonlinear transformation on the three-dimensional coordinates of the nodes, the preload feature branch maps the preload vector into a high-dimensional feature embedding, the feature fusion and decoding unit splices and fuses the features output by the two feature extraction branches with the original standardized coordinates and decodes and outputs the prediction result, and the physical constraint unit forces the prediction result to satisfy the constitutive equation in elasticity.
[0029] It should be noted that this method can avoid interference from process parameters of different magnitudes on model training through standardization, thereby improving the quality of model input data; the multi-branch model structure can fully extract the features of different types of process parameters, the feature fusion operation can integrate multi-dimensional information, and the physical constraint unit ensures the physical rationality of the prediction results, thus comprehensively improving the accuracy and reliability of antenna assembly strain prediction.
[0030] Preferably, the coordinate feature extraction branch includes a feature encoding layer and a graph convolutional layer connected in sequence. The feature encoding layer uses a fully connected network to map the input standardized three-dimensional coordinates into low-dimensional feature vectors. The graph convolutional layer aggregates the features of nodes and their neighbors layer by layer based on the spatial association of the node neighbor list to extract spatial topological association features.
[0031] Specifically, when constructing the coordinate feature extraction branch, a feature encoding layer is first set up. This layer adopts a fully connected network structure, inputting the standardized 3D coordinate data into the fully connected network. Through hierarchical computation of the network, the high-dimensional coordinate data is mapped into low-dimensional feature vectors. After the feature encoding layer, a graph convolutional layer is connected. A node neighbor list is pre-constructed to clarify the spatial relationship between each node. Based on this neighbor list, the graph convolutional layer performs aggregation operations on the node's own features and the features of its neighboring nodes layer by layer, thereby extracting the spatial topological relationship features between nodes.
[0032] It should be noted that the fully connected network of the feature encoding layer can achieve dimensionality reduction of coordinate data, simplifying the complexity of subsequent calculations. The graph convolutional layer, based on the feature aggregation method of the neighbor list, can effectively capture the spatial topology information of the antenna printed circuit board nodes, making the extracted features more consistent with the actual physical spatial distribution, and providing more valuable spatial feature support for subsequent strain prediction.
[0033] Preferably, the preload feature branch maps the preload vector to a high-dimensional feature embedding, including: using a fully connected network to map the input standardized preload to a global feature vector, and then expanding the global feature vector into a feature matrix matching the number of nodes through tensor replication, so as to achieve alignment between the preload and the node features.
[0034] Specifically, during the operation of the preload feature branch, the standardized preload data is first input into a pre-defined fully connected network, and a global feature vector is generated through the mapping calculation of the fully connected network. The number of nodes on the antenna printed circuit board is obtained, and a tensor copy operation is performed on the global feature vector based on this number, expanding the global feature vector into a feature matrix with dimensions matching the number of nodes, thus completing the alignment between the preload feature and the node feature.
[0035] It should be noted that the mapping process of the fully connected network can fully explore the deep features of the preload data, while the tensor copying and feature alignment operations solve the matching problem between the global features of the preload and the local features of the nodes, so that the preload features can be better integrated into the subsequent fusion calculation and improve the model's adaptability to different preload conditions.
[0036] Preferably, the feature fusion and decoding unit concatenates and fuses the features output from the two feature extraction branches with the original standardized coordinates and decodes and outputs the prediction result, including: concatenating the spatial topological correlation features output from the coordinate feature extraction branch, the global working condition features output from the preload feature extraction branch, and the original standardized coordinates in dimensions to form a fused feature vector; and then decoding the fused feature vector through a fully connected network to output standardized prediction values of multiple strain components.
[0037] Specifically, during the feature fusion and decoding unit's operation, the spatial topological correlation features output from the coordinate feature extraction branch and the global working condition features output from the preload feature extraction branch are first collected, while the original standardized three-dimensional coordinate data is retrieved simultaneously. A dimensional concatenation operation is performed on these three types of data to integrate features from different dimensions into a unified fusion feature vector. A fully connected network for decoding is constructed, and the fusion feature vector is input into this network. Through hierarchical decoding calculations within the network, standardized predicted values of strain components in multiple directions for the corresponding nodes are output.
[0038] It should be noted that the splicing and fusion of multiple features can integrate multi-dimensional information such as spatial topology, preload conditions, and original coordinates, allowing the model to comprehensively consider the impact of various factors on assembly strain. The decoding process of the fully connected network can transform the fused high-dimensional features into intuitive strain prediction results, improving the accuracy and effectiveness of the prediction results.
[0039] Preferably, the physical constraint unit forces the prediction result to satisfy the constitutive equation in elasticity, including: automatically differentiating the divergence residuals of the strain field in the x, y, and z directions, and converting the divergence residuals into optimizable loss terms to force the prediction result to satisfy the constitutive equation in elasticity.
[0040] Specifically, after the physical constraint unit starts working, it acquires the predicted strain components output by the feature fusion and decoding unit and constructs the corresponding strain field model. Using an automatic differentiation method, the divergence residuals of the strain field in the x, y, and z spatial directions are solved. The calculated divergence residuals are transformed into loss terms that can participate in model optimization. These loss terms are added to the overall loss function of the model, and during model training, backpropagation optimization is performed to force the strain prediction results to meet the relevant requirements of the elasticity constitutive equations.
[0041] It should be noted that the automatic differentiation calculation method can efficiently and accurately obtain the divergence residuals of the strain field. Transforming the residuals into loss terms and incorporating them into model optimization can effectively constrain the physical rationality of the prediction results, avoid prediction values that do not conform to the laws of elasticity, and make the strain results output by the model more in line with actual engineering application scenarios.
[0042] Preferably, the strain prediction model is trained using a meta-learning optimization strategy, which includes: Inner loop adaptation: For a single prestressed load case task, starting from the initial weights of the strain prediction model, the data loss and physical loss are calculated using the support set of the task, and the weights are updated through gradient descent to obtain fast weights that adapt to the current task. Outer loop generalization: The initial weights are optimized based on the adaptation effect of multiple pre-tightening working conditions. For each task, after obtaining the fast weights, the generalization loss is calculated using the query set of that task. The query set losses of multiple tasks are aggregated as the meta-loss, and the initial weights of the strain prediction model are updated through backpropagation.
[0043] Specifically, during the training phase of the strain prediction model, a meta-learning optimization strategy is employed. In the inner loop adaptation step, a single preload condition is selected as a training task. The model's initial weights are retrieved, and the data loss and physical loss are calculated using the support set data corresponding to that task. Based on the loss results, the model weights are updated using the gradient descent algorithm to obtain fast weights adapted to that single condition task. In the outer loop generalization step, multiple different preload condition tasks are selected. Inner loop adaptation is performed for each task to obtain corresponding fast weights. The generalization loss is then calculated using the query set data for each task. The generalization losses of all tasks are aggregated to obtain the meta-loss. Based on this meta-loss, the model's initial weights are updated using the backpropagation algorithm, completing one round of meta-learning training. This process is repeated multiple times until the model converges.
[0044] It should be noted that inner loop adaptation allows the model to quickly adapt to a single preload condition task, improving the model's prediction performance under specific conditions. Outer loop generalization, on the other hand, is based on the optimization of multiple condition tasks, improving the model's generalization ability. This enables the model to output accurate strain prediction results even when facing untrained preload conditions, shortening the model's adaptation cycle and enhancing its engineering applicability.
[0045] Preferably, in the meta-learning optimization strategy, each pre-tightening condition task is a meta-task, and the dataset of the meta-task includes the three-dimensional coordinates of all nodes, the adjacency relationships between nodes, and the labels of the real strain field under the corresponding condition.
[0046] Specifically, when constructing the training dataset for the meta-learning optimization strategy, each preload condition is set as an independent meta-task. A corresponding dataset is configured for each meta-task, which includes the three-dimensional coordinate data of all nodes on the antenna printed circuit board, as well as the adjacency relationship data between each node. It also needs to collect the actual strain field label data measured under the preload condition. The three types of data are integrated to form the complete dataset of the meta-task, which is used for the inner loop adaptation and outer loop generalization training of the model.
[0047] It should be noted that the meta-task dataset contains the coordinates and adjacency relationships of all nodes, providing complete spatial topological information to support the model. The real strain field labels provide accurate supervision information for model training, ensuring that the loss calculations for the inner and outer loops have a reliable basis, thereby improving the effectiveness of meta-learning training and allowing the model to better learn the strain characteristics under different preload conditions.
[0048] Accordingly, this embodiment also provides an antenna assembly strain prediction system, including: The data acquisition and processing module is configured to acquire the assembly process parameters of the antenna printed circuit board nodes and map them to a distribution space of the same magnitude through standardization processing; the assembly process parameters include the three-dimensional coordinate data of the nodes, the load data of different preload conditions, and the node strain data corresponding to each condition. The strain prediction model is configured to output antenna assembly strain prediction results based on standardized assembly process parameters. The strain prediction model includes a coordinate feature extraction branch, a preload feature extraction branch, a feature fusion and decoding unit, and a physical constraint unit. The coordinate feature branch performs nonlinear transformation on the three-dimensional coordinates of the nodes. The preload feature branch maps the preload vector into a high-dimensional feature embedding. The feature fusion and decoding unit concatenates and fuses the features output from the two feature extraction branches with the original standardized coordinates and decodes and outputs the prediction result. The physical constraint unit forces the prediction result to satisfy the constitutive equation in elasticity.
[0049] It should be noted that the data acquisition and processing module ensures the consistency and validity of the data input to the model, laying a solid data foundation for model prediction. The modular strain prediction model has a clear structure, with each unit having a well-defined division of labor and working collaboratively. It can efficiently complete the entire process of feature extraction, fusion, decoding, and physical constraints, achieving rapid and accurate prediction of antenna assembly strain and providing reliable system support for antenna assembly process optimization.
[0050] Example 2 This embodiment provides a method for predicting antenna assembly strain. A PGINN model combining GNN (Graph Neural Network) and PINN (Physical Information Neural Network) is designed and integrated with meta-learning training. This satisfies both data fitting requirements and adheres to mechanical equilibrium constraints, thereby enabling strain prediction of the antenna printed circuit board (PCB). Figure 1 As shown. In Figure 1 In the algorithm, input variables F1-F9 represent the preload of nine bolts, and input variables X, Y, and Z represent spatial coordinates. Physical constraint formulas describe the relationship between stress and strain in a material in three-dimensional space. The network must not only fit the data, but the output strain must also satisfy these mechanical equations. In the meta-learning structure, the data is divided into a support set and a query set, where the support set is used for the θ of a single preload condition task in the inner loop. i The training and query sets are used for θ in multiple preload conditions of the outer loop. Update, T i This represents a specific task; the inner loop loss = data-driven loss + physical constraint loss, and the outer loop loss is the average of the inner loop loss.
[0051] This embodiment focuses on the PCB strain prediction task and constructs a fully functional dataset processing module to support subsequent model training and inference. The dataset processing module first loads three types of core data: 3D coordinate data of PCB nodes (coordinates.csv), 9D load data of different preload conditions (F.csv), and corresponding nodal strain data (strain.csv) for each condition.
[0052] Preferably, in order to eliminate the influence of different physical dimensions, the dataset processing module uses standardization to normalize the coordinate, preload, and strain data respectively, mapping them to a distribution space of the same magnitude, so as to improve the convergence efficiency and prediction stability of the model.
[0053] The PGINN model in this embodiment integrates graph structure learning and physical constraint mechanisms. The model comprises two parallel feature extraction branches: a coordinate branch uses a multilayer perceptron to perform nonlinear transformations on the 3D coordinates of nodes, uncovering the implicit geometric features in the node's spatial location; and a preload branch maps the 9-dimensional preload vector into a high-dimensional feature embedding to capture load differences under different working conditions. The features output from both branches are concatenated with the original coordinates and input into the decoder, ultimately outputting a 6-dimensional strain component prediction result.
[0054] Preferably, to ensure that the prediction results conform to physical laws, the model incorporates the PINN physical constraint mechanism: by automatically differentiating and calculating the divergence residuals of the strain field in the x, y, and z directions, the prediction results are forced to satisfy the constitutive equations in elasticity, thus avoiding physical contradictions that may occur in a purely data-driven model.
[0055] Preferably, during the model training phase, the inner loop is responsible for the rapid adaptation of a single preload condition task. Starting from the initial weights of the model, it calculates the data loss and physical loss using the support set of that task, and updates the weights through gradient descent to obtain fast weights adapted to the current task. The outer loop optimizes the initial weights based on the adaptation effects of multiple tasks: for each task, after obtaining the fast weights, it calculates the generalization loss using the query set of that task, aggregates the query set losses of multiple tasks as the meta-loss, and updates the initial weights of the model through backpropagation. Finally, a fusion meta-learning strategy is used to improve the adaptation capability for few-sample tasks.
[0056] Example 3 This embodiment is based on embodiment 2: This embodiment provides a method for predicting antenna assembly strain, including finite element simulation, PGINN model design, meta-learning optimization strategy, and loss function design, as detailed below.
[0057] I. Finite Element Simulation The array antenna is composed of components made of various materials. To achieve accurate finite element simulation of the array antenna assembly process, it is necessary to accurately define the mechanical properties of its multi-material components. The key material parameters involved in this embodiment are shown in Table 1.
[0058] Table 1 - Material Properties of Array Antenna Components
[0059] A 3D model identical to the actual antenna PCB was constructed using finite element simulation software, with the structure as follows: Figure 2 As shown.
[0060] In the assembly process of array antennas, the assembly process parameters involved have a crucial impact on the mechanical performance and reliability of the entire structure. Among these, the most critical parameters include the precise control of bolt tightening torque and the optimized design of the tightening sequence. As a core parameter that directly affects the preload of the connection structure, bolt tightening torque can lead to loosening of the connection and affect performance stability if the torque is too small, while excessive torque may cause local stress concentration, resulting in substrate damage or bolt plastic deformation.
[0061] This embodiment converts the tightening torque into preload and uses its magnitude as the input variable. The tightening sequence, as another key process parameter, plays a crucial role in controlling strain distribution during assembly. The specific tightening sequence is as follows: Figure 3 As shown, preload is applied sequentially in the order of numbers 1 to 9. This aims to ensure a smooth and gradual transfer and redistribution of strain within the structure, avoiding stress concentration caused by localized overload. After each tightening operation, the strain distribution on the PCB board under the current state is calculated through mechanical simulation analysis. Figure 4 As shown in the diagram. The label LE on the left represents strain; "Average: 75%" indicates that the software smoothed the calculation results of adjacent element nodes when rendering the contour map, and 75% is a common node average threshold; the change in contour map color from blue to red represents the distribution of values from smallest to largest, with the red area being the location of the most concentrated strain. In mechanical analysis, strain is a dimensionless physical quantity representing the change per unit length.
[0062] II. PGINN Model Design The model adopts a dual-branch structure to integrate the spatial topological features and preload features of the PCB, and embeds physical constraints of equilibrium equations to achieve the dual goals of data fitting accuracy and compliance with physical laws. The specific structure is as follows.
[0063] (1) Coordinate feature branches This branch focuses on extracting the spatial topological association features of PCB nodes to capture the spatial dependencies of strain.
[0064] MLP Feature Encoding Layer: Input nodes have three-dimensional normalized coordinates. The coordinates are mapped into low-dimensional feature vectors through a fully connected network, thus achieving the initial encoding of coordinate information.
[0065] Graph convolutional layer: Based on the spatial association of the neighbor list, it aggregates the features of nodes and their neighbors layer by layer to achieve deep extraction of spatial topological associations. The formula is:
[0066] in, hv (l) For nodes v In thel Features of the layer Wself , Wneigh For the weights, N( v ) is a node v The set of neighbors, |N( v ) | represents the number of neighbors. If the number of neighbors is 0, then the feature itself is taken. If the number of neighbors is insufficient, then sampling is repeated.
[0067] (2) Preload coordinate branch This branch focuses on extracting global features of the preload condition, addressing the issue that the preload, as a global loading condition, needs to be aligned with the local features of each node, thus achieving an effective association between the preload and the node topological features. It takes a 9-dimensional normalized preload as input, maps the preload to a global feature vector through a fully connected network, and expands it into a feature matrix matching the number of nodes through tensor replication, achieving alignment between the preload and node features.
[0068] (3) Feature fusion and decoding This module is responsible for integrating multi-source features and mapping them to strain prediction results, ensuring that the fused features fully cover the influencing factors of PCB strain distribution.
[0069] Feature fusion: The topological correlation features output by the coordinate feature branch and the global working condition features output by the preload branch are concatenated with the original standardized coordinates to form a fused feature vector. Decoder: The fused features are decoded through a fully connected network, and the standardized prediction values of the six strain components are finally output.
[0070] (4) Physical constraints To avoid the problem of predictions conforming to data but violating physical laws that may occur with purely data-driven models, the model embeds constitutive equations from elasticity mechanics. This transforms physical laws into optimizable loss terms, achieving synergistic optimization of data fitting and physical compliance. Based on elasticity mechanics, strain must satisfy the constitutive equations, namely:
[0071]
[0072]
[0073]
[0074]
[0075]
[0076] in, σ For stress, ε γ is the normal strain, and γ is the shear strain.x , y , z These are the node coordinates.
[0077] III. Meta-learning Optimization Strategies Meta-learning, through a two-layer optimization mechanism of inner loop adaptation and outer loop generalization, enables the model to quickly adapt to strain prediction tasks under new preload conditions in PCB strain prediction scenarios with only a small amount of data and gradient updates. Its core is to optimize the initial parameters of the model through multi-task learning, giving the model the ability to learn quickly.
[0078] Task Definition: In the PCB strain prediction scenario, each meta-task corresponds to a strain prediction task under a specific preload combination, denoted as Ti. Each meta-task contains complete data for that preload condition: the 3D coordinates of all nodes, the adjacency relationships between nodes, and the actual strain field label for that condition. The differences between tasks are determined by the variation in the preload combination. The distribution of meta-tasks is constructed from diverse preload combinations, ensuring that the model can learn general strain prediction rules across conditions.
[0079] Inner loop: For a single task, fine-tune the initial parameters with a small amount of data to obtain fast parameters.
[0080] Outer loop: Based on the performance of fast parameters of multiple tasks on the query set, update the initial parameters to make them more adaptable to new tasks.
[0081] 1. Internal circulation The inner loop is the process by which the model is rapidly fine-tuned for a single meta-task. Its core objective is to adjust the initial parameters to quickly adapt them to the current task based on a small amount of data. This is achieved by tightly integrating graph structure characteristics with physical constraints, such as... Figure 5 As shown.
[0082] (1) Parameter initialization Extract the current set of initial parameters from the model to form the starting parameters for fine-tuning the inner loop. This process ensures that updates to the inner loop parameters do not directly interfere with the optimization basis of the outer loop's initial parameters by isolating the initial parameters from the fine-tuning parameters.
[0083] (2) Multi-step fine-tuning The inner loop adopts an iterative fine-tuning mode of subgraph sampling, loss calculation, and parameter update, making full use of the inductive learning characteristics of graph structures and the physical constraint capabilities of PINN.
[0084] Subgraph Sampling: Core nodes are randomly selected from all nodes in the current task, and a local subgraph is constructed based on the node adjacency list using a multi-layer neighbor sampling strategy. This subgraph not only contains the coordinates and strain labels of the core nodes, but also preserves the topological connections between nodes. An adaptive input structure is formed through adjacency matrix reconstruction, reducing the computational overhead of the entire graph and preserving the spatial correlation information of local strain.
[0085] Loss Calculation: The loss function is a weighted average of data loss and physical loss, achieving a synergistic constraint between data-driven and physical priors. Specifically, the data loss uses mean squared error to measure the deviation between the model's predicted strain and the actual strain; the physical loss is calculated automatically using the divergence residual of the strain field to verify whether the prediction results satisfy the elasticity equilibrium equations, ensuring that the fine-tuned parameters both fit the data and conform to physical laws.
[0086] Gradient calculation and parameter update: The parameter gradient is calculated based on the total loss, and the current parameter set is updated using gradient descent. To support backpropagation optimization of the initial parameters in the outer loop, the computation graph structure must be preserved during the gradient calculation in the inner loop to ensure the gradient path is complete and traceable.
[0087] 2. External circulation The outer loop uses the adaptation results of multiple meta-tasks as a basis, and optimizes the initial parameters by aggregating the generalization errors of multiple tasks. This enables the initial parameters to have meta-learning capabilities that allow for low-cost fine-tuning to adapt to new tasks, such as... Figure 6 As shown.
[0088] (1) Initialization The Adam optimizer is initialized for outer loop parameter updates, and its optimization target is the model's initial parameter set. Preprocessed full data is loaded, where coordinates, preload, and strain are normalized using independent normalizers to ensure that data of different dimensions and magnitudes can collaboratively participate in meta-learning, avoiding local features dominating parameter optimization.
[0089] (2) Meta-iteration Task Sampling: Randomly sample a batch of meta-tasks from the meta-task distribution to ensure that the sampled tasks cover diverse combinations of preload forces, so that the external loop optimization can take into account the common patterns of different working conditions.
[0090] Inner Loop Adaptation: An inner loop fine-tuning process is executed for each sampling task, outputting fast parameters adapted to that task. This process fully reuses the subgraph sampling, loss calculation, and parameter update logic of the inner loop, ensuring that fine-tuning of different tasks does not interfere with each other only through a parameter isolation mechanism.
[0091] Query set loss calculation: A subgraph is independently sampled for each task as the query set to simulate the model's generalization scenario on that task. The query set is then used for strain prediction using fast parameters. A bi-term loss calculation logic consistent with the inner loop is employed to obtain the query set loss, which measures the generalization performance of the fast parameters.
[0092] Meta-loss aggregation and parameter update: The meta-loss is obtained by averaging the query set losses of all tasks within the batch. This loss reflects the generalization ability of the initial parameters on new tasks after fine-tuning. After clearing the optimizer gradients, the gradient of the meta-loss is propagated to the initial parameters through backpropagation, and gradient descent is performed to update the initial parameters, making them iterate in a direction that is easier to fine-tune to the optimal solution.
[0093] IV. Loss Function Design In this embodiment, the model's loss function is a weighted sum of the data loss function and the physical loss function, ensuring both data consistency and physical compliance in the model's predictions. The formula is as follows:
[0094] In the formula, Loss DATA It is a data loss function. Loss PHY It is a loss function based on physical knowledge. λ It represents the weights of data loss and physical loss functions in the overall loss, enabling the neural network learning process to make a more accurate and reasonable trade-off under the guidance of physical knowledge.
[0095] (1) Data loss The data loss is the mean square error between the predicted strain and the actual strain, which forces the model prediction to align with the labeled data to ensure the accuracy of the basic prediction. The formula is:
[0096] In the formula, y i The neural network predicted the value. y For label values, n This represents the number of data points.
[0097] (2) Physical loss The physical loss is the mean square error of the strain divergence residuals. A mechanical equilibrium equation is introduced as prior knowledge to prevent the model from merely fitting the data and violating physical laws, thus improving the reliability of the prediction. The formula is:
[0098] In the formula, These are values calculated based on physical mechanisms.
[0099] Example 4 This embodiment is based on embodiment 1: This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the antenna assembly strain prediction method of Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form.
[0100] Specifically, when constructing this computer device, a matching memory and processor are configured, and the written computer program is stored in the memory. This computer program contains relevant instructions for implementing the antenna assembly strain prediction method of Embodiment 1. The processor calls and executes the computer program in the memory, sequentially performing steps such as acquiring and standardizing assembly process parameters, building and running each functional unit of the strain prediction model, and training the model using a meta-learning optimization strategy, ultimately realizing the entire process of the antenna assembly strain prediction method.
[0101] It should be noted that, with the hardware support of computer equipment, the relevant algorithms for antenna assembly strain prediction can be executed efficiently, improving the computational speed and processing efficiency of the prediction process. The storage function of the memory ensures the stable storage of programs and data, while the computing power of the processor ensures the smooth completion of model training and prediction tasks, providing a hardware platform for the engineering implementation of this strain prediction method.
[0102] Example 5 This embodiment is based on embodiment 1: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the antenna assembly strain prediction method of Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form. The storage medium includes any entity or device capable of carrying computer program code, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.
[0103] Specifically, a computer program containing the execution logic of the antenna assembly strain prediction method of Embodiment 1 is pre-stored in the computer-readable storage medium. After the storage medium establishes a data connection with the processor, the processor reads and executes the computer program in the storage medium, and performs operations such as acquiring and standardizing assembly process parameters, constructing and calculating the strain prediction model, and optimizing the model according to the program instructions, thereby realizing the corresponding antenna assembly strain prediction method.
[0104] It should be noted that computer-readable storage media enables stable storage and convenient transmission of the prediction method program, facilitating its deployment and execution on different computer devices. The portability of the storage medium allows the antenna assembly strain prediction method to be flexibly applied to various engineering scenarios, lowering the application threshold and improving the efficiency of technology dissemination.
[0105] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed in this embodiment and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be modified within the scope of the concept described in this embodiment through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.
[0106] It should be noted that, for the sake of simplicity, the foregoing method embodiments are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.
Claims
1. A method for predicting strain in antenna assembly, characterized in that, include: The assembly process parameters of the antenna printed circuit board nodes are obtained and mapped to a distribution space of the same order of magnitude through standardization. The assembly process parameters include three-dimensional coordinate data of nodes, load data for different preload conditions, and corresponding node strain data for each condition. The standardized assembly process parameters are input into the strain prediction model, which then outputs the strain prediction results for the antenna assembly. The strain prediction model includes a coordinate feature extraction branch, a preload feature extraction branch, a feature fusion and decoding unit, and a physical constraint unit. The coordinate feature branch performs a nonlinear transformation on the three-dimensional coordinates of the nodes. The preload feature branch maps the preload vector into a high-dimensional feature embedding. The feature fusion and decoding unit concatenates and fuses the features output from the two feature extraction branches with the original standardized coordinates and decodes the prediction results. The physical constraint unit forces the prediction results to satisfy the constitutive equations in elasticity.
2. The antenna assembly strain prediction method according to claim 1, characterized in that, The coordinate feature extraction branch includes a feature encoding layer and a graph convolutional layer connected in sequence. The feature encoding layer uses a fully connected network to map the input standardized three-dimensional coordinates into low-dimensional feature vectors. The graph convolutional layer aggregates the features of nodes and their neighbors layer by layer based on the spatial association of the node neighbor list to extract spatial topological association features.
3. The antenna assembly strain prediction method according to claim 1, characterized in that, The preload feature branch maps the preload vector into a high-dimensional feature embedding, including: using a fully connected network to map the input standardized preload into a global feature vector, and then expanding the global feature vector into a feature matrix matching the number of nodes through tensor copying, so as to achieve alignment between the preload and node features.
4. The antenna assembly strain prediction method according to claim 1, characterized in that, The feature fusion and decoding unit concatenates and fuses the features output from the two feature extraction branches with the original standardized coordinates and decodes and outputs the prediction result. This includes: concatenating the spatial topological correlation features output from the coordinate feature extraction branch, the global working condition features output from the preload feature extraction branch, and the original standardized coordinates to form a fused feature vector; and then decoding the fused feature vector through a fully connected network to output standardized prediction values for multiple strain components.
5. The antenna assembly strain prediction method according to claim 1, characterized in that, The physical constraint unit forces the prediction result to satisfy the constitutive equation in elasticity by: automatically differentiating and calculating the divergence residuals of the strain field in the x, y, and z directions, and converting the divergence residuals into optimizable loss terms to force the prediction result to satisfy the constitutive equation in elasticity.
6. The antenna assembly strain prediction method according to claim 1, characterized in that, The strain prediction model is trained using a meta-learning optimization strategy, which includes: Inner loop adaptation: For a single prestressed load case task, starting from the initial weights of the strain prediction model, the data loss and physical loss are calculated using the support set of the task, and the weights are updated through gradient descent to obtain fast weights that adapt to the current task. Outer loop generalization: The initial weights are optimized based on the adaptation effect of multiple pre-tightening working conditions. For each task, after obtaining the fast weights, the generalization loss is calculated using the query set of that task. The query set losses of multiple tasks are aggregated as the meta-loss, and the initial weights of the strain prediction model are updated through backpropagation.
7. The antenna assembly strain prediction method according to claim 6, characterized in that, In the meta-learning optimization strategy, each pre-tightening condition task is a meta-task. The dataset of the meta-task includes the three-dimensional coordinates of all nodes, the adjacency relationships between nodes, and the labels of the actual strain field under the corresponding condition.
8. An antenna assembly strain prediction system, characterized in that, include: The data acquisition and processing module is configured to acquire the assembly process parameters of the antenna printed circuit board nodes and map them to a distribution space of the same magnitude through standardization processing. The assembly process parameters include three-dimensional coordinate data of nodes, load data for different preload conditions, and corresponding node strain data for each condition. The strain prediction model is configured to output antenna assembly strain prediction results based on standardized assembly process parameters. The strain prediction model includes a coordinate feature extraction branch, a preload feature extraction branch, a feature fusion and decoding unit, and a physical constraint unit. The coordinate feature branch performs nonlinear transformation on the three-dimensional coordinates of the nodes. The preload feature branch maps the preload vector into a high-dimensional feature embedding. The feature fusion and decoding unit concatenates and fuses the features output from the two feature extraction branches with the original standardized coordinates and decodes and outputs the prediction result. The physical constraint unit forces the prediction result to satisfy the constitutive equation in elasticity.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the antenna assembly strain prediction method according to any one of claims 1-7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the antenna assembly strain prediction method according to any one of claims 1-7.