Connector stress test optimization method and system based on reinforcement learning

Abstract

The invention relates to the technical field of electronic connector testing, and discloses a connector stress test optimization method and system based on reinforcement learning, and the method comprises the steps: constructing a multi-scale graph structure model of a connector; constructing a physical constraint space-time diagram neural network, and introducing physical constraint recursive cells to ensure that a prediction result accords with a material mechanics basic principle; constructing a microscopic stress field representation model by using a meta-enhanced contrast learning method, and dynamically adjusting the model precision of the local region according to the stress gradient; dynamically optimizing test parameters and test point positions by utilizing a reinforcement learning algorithm; continuously optimizing parameters of each model based on the test data optimized by the reinforcement learning algorithm to form a closed-loop optimization system; according to the invention, the test efficiency and accuracy are improved, the test cost is reduced, and more reliable data support is provided for connector design and quality control.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic connector testing, and more particularly, to a connector stress testing optimization method and system based on reinforcement learning. Background Art

[0002] Connectors are key components for connecting circuit components in electronic devices, and their reliability directly impacts the stability and lifespan of the entire electronic system. Stress testing is an important means of assessing connector reliability. Traditional connector stress testing relies primarily on finite element analysis and empirically based fixed parameter configurations. This leads to issues such as high computational resource consumption, lack of adaptability in test parameters, insufficient microstructural stress analysis, weak time-dimensional prediction capabilities, and a lack of intelligent optimization of test point placement.

[0003] As connector structures become increasingly complex and miniaturized, higher requirements are placed on the accuracy and efficiency of stress testing. Traditional testing methods cannot meet the needs of modern connector R&D and quality control, especially in the areas of accurate prediction of microscopic stress fields, dynamic optimization of test parameters, and long-term reliability assessment.

[0004] In existing technologies, connector stress testing typically uses finite element analysis to simulate the stress distribution of connectors under different operating conditions. However, this method is computationally expensive and has limited ability to characterize microstructures. Another approach involves fatigue testing using fixed parameters, but parameter configuration often relies on manual experience and lacks dynamic optimization capabilities. Although some studies have attempted to apply machine learning techniques to electronic component testing, the results remain unsatisfactory in terms of multi-scale stress analysis and optimal configuration of test resources. These methods generally suffer from the following problems: difficulty in simultaneously processing macro- and micro-scale stress analysis; lack of effective prediction of stress evolution in the time dimension; lack of adaptive optimization mechanisms for test point locations and parameter settings; and inability to form a closed-loop test-optimization system.

[0005] Therefore, a connector stress test optimization method is needed that can simultaneously solve the above technical problems, improve test efficiency and accuracy, reduce test costs, and provide more reliable data support for connector design and quality control. Summary of the Invention

[0006] The present invention provides a connector stress test optimization method and system based on reinforcement learning, which solves the technical problems of low efficiency, insufficient precision and high resource consumption of connector stress testing in related technologies.

[0007] The present invention provides a connector stress test optimization method based on reinforcement learning, comprising the following steps: Constructing a multi-scale graph structure model of the connector, the multi-scale graph structure model includes a macrograph representing the macrostructure of the connector, a micrograph representing the microstructure of the connector, and a cross-scale connection connecting the macrograph and the micrograph; Based on the multi-scale graph structure model, a physically constrained spatiotemporal graph neural network is constructed. The physically constrained spatiotemporal graph neural network includes a spatial propagation layer and a temporal evolution layer. Physically constrained recursive cells are introduced to ensure that the prediction results conform to the basic principles of material mechanics. Based on the output of the physical constraint spatiotemporal graph neural network, a meta-enhanced contrastive learning method is used to construct a microscopic stress field representation model, and the model accuracy of the local area is dynamically adjusted according to the stress gradient; Based on the prediction results of the micro-stress field representation model, the reinforcement learning algorithm is used to dynamically optimize the test parameters and test point locations; Based on the test data optimized by the reinforcement learning algorithm, the parameters of each model are continuously optimized to form a closed-loop optimization system.

[0008] In a preferred embodiment, constructing a multi-scale graph structure model of a connector includes: Get the structural data of the connector; Construct a macrograph structure, where nodes represent the key structural components of the connector and edges represent the physical connection relationships between components; Construct a microscopic graph structure, where nodes represent the microscopic components of the material and edges represent the interaction between microscopic units; Establish the connection between macro-graph structure and micro-graph structure; Initialize the node features and edge features in the graph structure.

[0009] In a preferred embodiment, the physical constraints in the step of introducing physical constraint recursive cells to ensure that the prediction results comply with the basic principles of material mechanics include: force balance constraints, material constitutive relationship constraints and boundary condition constraints.

[0010] In a preferred embodiment, the meta-enhanced contrastive learning method includes: Generate multiple views of the same microscopic region sample through data augmentation; Mapping multiple viewpoints into latent feature space; Make different views from the same sample close together in the feature space, while views from different samples move away; The micro-stress field modeling task is decomposed into multiple subtasks, and a small number of sample learning strategy is adopted for each subtask.

[0011] In a preferred embodiment, the step of dynamically adjusting the model accuracy of the local area according to the stress gradient includes: Graph coarsening is performed on areas with smaller stress gradients, and adjacent nodes are merged to reduce computational complexity. Graph refinement is performed on areas with large stress gradients to increase node density and improve model accuracy.

[0012] In a preferred embodiment, the reinforcement learning algorithm includes: Define the state space, including the connector's graph structure, stress distribution, and test configuration; Define the action space, including test parameters and test point locations; Define the reward function, designed based on information gain and resource consumption; Use deep reinforcement learning algorithms to learn optimal testing strategies.

[0013] In a preferred embodiment, the reinforcement learning algorithm is a hierarchical reinforcement learning architecture, comprising: High-level strategy is responsible for determining the overall test strategy, including prioritization of test areas and selection of test types; The low-level strategy is responsible for the fine-tuning of specific test parameters, including the exact location of test points within the selected area and the precise magnitude of the test force.

[0014] In a preferred embodiment, the step of continuously optimizing the parameters of each model based on the test data optimized by the reinforcement learning algorithm includes: Integrated sensor technology collects real-time data on connectors during stress testing; Compare the stress distribution predicted by the model with the real-time collected data and calculate the prediction error; Based on the prediction error, the model parameters are continuously updated using online learning methods; An elastic weight integration method is used to prevent overfitting and catastrophic forgetting.

[0015] In a preferred embodiment, a connector stress test optimization method based on reinforcement learning further includes an expert knowledge integration step, which includes: Initialize the policy from expert demonstration data via imitation learning; Design rule-based safety constraints to prevent the agent from exploring dangerous test parameters; Use rule-based reward shaping to guide agents to learn effective policies faster.

[0016] In a preferred embodiment, a connector stress test optimization system based on reinforcement learning is used to perform a connector stress test optimization method based on reinforcement learning, including: Multi-scale graph structure modeling module, used to build a multi-scale graph structure model of the connector; A physical constraint spatiotemporal graph neural network module predicts the stress distribution of connectors based on a multi-scale graph structure model; Meta-enhanced contrastive learning module, used to construct a microscopic stress field representation model; Reinforcement learning optimization module, used to dynamically optimize test parameters and test point locations; The real-time data feedback and model optimization module is used to continuously optimize the prediction model parameters based on real-time test data.

[0017] The beneficial effects of the present invention are: The present invention achieves a dual improvement in multi-scale modeling and computing efficiency. By constructing a multi-scale graph structure model of the connector, comprehensive modeling of the connector from macrostructure to micromaterial is achieved, overcoming the limitations of traditional methods; at the same time, the stress field prediction method based on graph neural network significantly reduces the computing resource requirements, greatly shortens the calculation time compared with the traditional finite element analysis method, and realizes near real-time stress field prediction and dynamic optimization.

[0018] The present invention improves prediction accuracy through physically constrained recursive cells and adaptive precision adjustment, ensuring that the prediction results conform to the basic principles of material mechanics and avoiding the physically unreasonable results that may be produced by purely data-driven methods. At the same time, the model accuracy of local areas is dynamically adjusted according to the stress gradient, achieving a reasonable allocation of computing resources, and optimizing the overall computing efficiency while ensuring the prediction accuracy of key areas.

[0019] The present invention optimizes the configuration of test resources and enhances the ability to analyze microscopic stress fields. It dynamically optimizes test parameters and test point locations through a reinforcement learning algorithm, maximizes the ratio of information acquisition to resource consumption, and reduces unnecessary testing. The meta-enhanced contrastive learning method enables the system to learn microscopic stress field representation from limited test data, improves the ability to predict stress distribution at the material microstructure level, and provides a basis for identifying early failure mechanisms.

[0020] The present invention realizes the fusion of time dimension prediction, knowledge accumulation and expert knowledge. The spatiotemporal graph neural network model realizes the accurate prediction of stress evolution over time, and improves the accuracy of long-term reliability assessment. Through the closed-loop optimization system, a virtuous cycle of "testing, prediction, optimization, and retesting" is realized, and the system performance continues to improve with the accumulation of test data. At the same time, it effectively integrates the knowledge of domain experts, accelerates the system learning process, and improves the reliability and security of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flow chart of a connector stress test optimization method based on reinforcement learning of the present invention; Figure 2 is a bar graph comparing stress prediction errors of different methods of the present invention; Figure 3 is a line graph showing the trend of the test efficiency of the present invention changing with connector complexity; Figure 4 It is a scatter plot comparing the micro stress prediction and actual measurement of the present invention; Figure 5 It is a radar chart for comprehensive performance evaluation of the method of the present invention. DETAILED DESCRIPTION

[0022] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0023] At least one embodiment of the present invention discloses a connector stress test optimization method based on reinforcement learning, such as Figure 1 As shown, the following steps are included: Step 1: Construct a multi-scale graph structure model of the connector. The multi-scale graph structure model includes a macrograph representing the macrostructure of the connector, a micrograph representing the microstructure of the connector, and a cross-scale connection connecting the macrograph and the micrograph. The specific steps include: Step 1.1, obtain the structural data of the connector, including information such as geometry, material properties, and boundary conditions; The data comes from Computer Aided Design (CAD) models, Computed Tomography (CT) data or 3D laser scanning data.

[0024] Step 1.2: Construct a macrograph structure, where the nodes in the graph represent the key structural components of the connector (such as contact terminals, insulators, housings, etc.), and the edges represent the physical connection relationships between these components; According to an embodiment of the present application, for physically adjacent nodes, an edge is added to the graph to connect them. The feature vector of each node contains the geometric characteristics and material properties of the component.

[0025] Step 1.3: Construct a microscopic graph structure and further refine the graph structure to the material microscale for key component areas (such as stress concentration areas and contact interfaces). The nodes of the microscopic graph structure can represent the microscopic constituent units of the material (such as grains, grain boundaries, etc.), and the edges represent the interaction relationships between these microscopic units.

[0026] In some embodiments, the microscopic graph structure can be constructed based on a crystallographic model in materials science. For example, for a metal connector, a Voronoi tessellation model can be used to generate a grain structure, where each grain is represented as a node and the grain boundaries between adjacent grains are represented as edges.

[0027] For composite connectors, a multiphase mixture model can be used, in which materials of different phases are represented as different types of nodes and phase interfaces are represented as special types of edges.

[0028] Step 1.4: Establish the connection between the macrograph structure and the micrograph structure to form a multi-scale graph model; Specifically, a macro node can connect to multiple micro nodes in the corresponding area to establish a cross-scale information transmission channel.

[0029] Step 1.5, initialize the node features and edge features in the graph structure; It should be noted that for nodes in the micrograph structure, their eigenvectors can contain microstructural properties (such as grain size, orientation, etc.) and initial stress state.

[0030] The output of this step is a multi-scale graph structure, which contains a set of nodes (including macro nodes and micro nodes), an edge set, and a node feature matrix.

[0031] This graph structure can simultaneously capture the connector's macrostructure and material microscopic characteristics, providing a data basis for subsequent stress analysis.

[0032] Step 2: Based on the multi-scale graph structure model, a physical constraint space-time graph neural network is constructed. The physical constraint space-time graph neural network includes a spatial propagation layer and a temporal evolution layer. Physical constraint recursive cells are introduced to ensure that the prediction results conform to the basic principles of material mechanics. The specific steps include: Step 2.1, build the spatiotemporal graph neural network architecture; According to an embodiment of the present application, the network represents connectors as a dynamic graph sequence, where each time step corresponds to a graph structure. The network consists of a spatial propagation layer and a temporal evolution layer.

[0033] The spatial propagation layer is responsible for transmitting node information along the graph structure within a single time step and capturing the stress transmission relationship in space.

[0034] For each node, its feature update is based on a weighted combination of its own features and the features of neighboring nodes, and then processed by a nonlinear activation function.

[0035] Step 2.2, construct physically constrained recursive cells; In order to ensure that the stress propagation prediction complies with the basic principles of material mechanics, physical constraint recursive cells are introduced, the core of which is to embed physical laws into recursive cells.

[0036] For stress propagation, the following physical constraints are mainly considered: Force balance constraint: For any node, the sum of the internal force and external force must be balanced; Material constitutive relationship constraints: stress and strain must satisfy the material constitutive equation; Boundary condition constraints: At specific boundary nodes, the stress or displacement must satisfy the preset boundary conditions.

[0037] The specific implementation of a physically constrained recurrent cell can use a Gated Recurrent Unit (GRU) or a Long Short-Term Memory (LSTM) as the basic structure, with a physical constraint layer added. Taking the GRU as an example, the physically constrained recurrent cell includes the calculation process of the update gate, reset gate, candidate hidden state, and output hidden state.

[0038] Physical constraints are implemented by adding additional constraints during the update process. For example, for a force balance constraint, a correction step can be added that adjusts the hidden state according to the gradient direction of a physical constraint loss function, where the physical constraint loss function measures the degree of physical constraint violation, and the constraint strength parameter controls the magnitude of the adjustment.

[0039] In the stress analysis of connector elastic contact terminals, physical constraint recursive cells can effectively simulate the stress propagation process in metal materials and accurately capture stress concentration areas and potential failure points.

[0040] Step 2.3, implementation of the temporal attention mechanism; To capture short-term and long-term stress evolution patterns, a temporal attention mechanism is introduced to enable the network to focus on key information in historical states.

[0041] The temporal attention mechanism generates a temporal context vector by calculating the attention weights between the current time step and the historical time steps.

[0042] The attention weight is calculated based on the attention score function, which measures the correlation between the current hidden state and the historical hidden states.

[0043] Step 2.4, end-to-end training network; The spatiotemporal graph neural network is trained using historical stress test data, with the goal of minimizing the error between the predicted stress and the actual stress while satisfying physical constraints.

[0044] The total loss function consists of data fitting loss and physical constraint loss, and the importance of the two is weighed by hyperparameters.

[0045] ; in, Represents the total loss function, which is used to measure the overall error of the model prediction; Represents the data fitting loss, which is used to measure the difference between the model prediction value and the actual measurement value; Represents the physical constraint loss, which is used to measure the degree to which the model prediction results violate the physical laws; Is a trade-off hyperparameter used to adjust the relative importance of data fitting loss and physical constraint loss. Larger values ​​make the model more focused on satisfying physical constraints, while smaller values ​​make it more focused on fitting the real data.

[0046] The output of this step is a trained spatiotemporal graph neural network model that can accurately predict the dynamic stress distribution evolution of the connector under different stress conditions, and the prediction results are consistent with the basic principles of material mechanics.

[0047] Step 3: Based on the output of the physical constraint spatiotemporal graph neural network, a meta-enhanced contrastive learning method is used to construct a microscopic stress field representation model, and the model accuracy of the local area is dynamically adjusted according to the stress gradient; The specific steps include: Step 3.1, construct a microscopic stress field representation learning framework; The framework aims to learn the relationship between microstructural features and stress distribution from limited high-resolution data.

[0048] The core idea is to use contrastive learning methods to enable the model to distinguish the differences in the performance of different microstructures in stress fields.

[0049] According to an embodiment of the present application, the contrastive learning process first generates multiple perspectives of the same microscopic region sample through data augmentation (such as transformations such as rotation and scaling), and then maps them to the latent feature space through an encoder.

[0050] The goal of contrastive learning is to make different views from the same sample close together in the feature space, while making views from different samples far apart.

[0051] The loss function for contrastive learning can also take other forms. For example, a triplet loss can be used, which makes the distance between the anchor sample and the positive sample (different perspectives of the same microstructure) smaller than the distance between the anchor sample and the negative sample (different microstructures) minus a margin value.

[0052] In some embodiments, a prototype-based contrastive learning method may also be employed. This method maintains a prototype representation for each type of microstructure and learns discriminative features by minimizing the distance between samples of the same type and the prototype while maximizing the distance between prototypes of different types.

[0053] Step 3.2, implement the meta-reinforcement learning strategy; Meta-reinforcement learning aims to enable the model to quickly adapt to new microstructures from a small number of samples. Specifically, the microstress field modeling task is decomposed into multiple subtasks, each focusing on a specific type of microstructure (such as different grain sizes, different grain boundary orientations, etc.).

[0054] According to one embodiment of the present application, for each subtask, a small number of support set samples and query set samples are used for training.

[0055] The goal of meta-learning is to find an initial model parameter that can be quickly adapted to a new task with a small number of gradient updates.

[0056] Specifically, the loss is first calculated and the parameters are updated on the support set, then the performance of the updated parameters is evaluated on the query set, and finally the initial parameters are optimized to minimize the loss on the query set.

[0057] Step 3.3, adaptive graph coarsening and refinement mechanism; In order to efficiently process multi-scale data, an adaptive graph coarsening and refinement mechanism is introduced to dynamically adjust the model accuracy of the local area according to the stress gradient: Coarsening operation: For areas with smaller stress gradients, merge adjacent nodes to reduce computational complexity.

[0058] Refinement operation: For areas with large stress gradients, increase the node density to improve model accuracy.

[0059] For the case where the stress gradient between nodes is less than a threshold, the two nodes can be merged; On the contrary, if the stress gradient in a certain area exceeds the threshold, more nodes are added in that area for refinement.

[0060] Step 3.4, integration with measured data; Combined with experimental data such as X-ray imaging or acoustic emission, the micro-stress field prediction model is calibrated.

[0061] The measured data are compared with the model predictions, and the domain adaptation loss is calculated. By minimizing this loss, the microstress field prediction model is continuously optimized.

[0062] The output of this step is a model that can accurately predict the microscopic stress field distribution of the connector. The model has the ability to learn from a small amount of data and can dynamically adjust the accuracy according to the stress gradient, providing microscopic mechanism support for refined stress testing and early failure warning.

[0063] Step 4: Based on the prediction results of the micro-stress field representation model, the reinforcement learning algorithm is used to dynamically optimize the test parameters and test point locations; The specific steps include: Step 4.1, define the reinforcement learning environment; The connector stress test process is modeled as a Markov decision process (MDP), where: The state represents the connector stress state and test configuration at the current time step, including the current graph structure, stress distribution, and test configuration.

[0064] Actions represent the adjustment of test parameters and the selection of test points, including test parameters (such as force magnitude, direction, frequency, etc.) and test point locations.

[0065] The reward is designed based on testing information gain and resource consumption, where information gain measures the amount of new information obtained by performing an action and resource consumption measures the cost of performing an action, and the two are balanced through a trade-off parameter.

[0066] The state transition is determined by the stress propagation dynamics simulated by a spatiotemporal graph neural network.

[0067] In some embodiments, the information gain function can be calculated using a Bayesian experimental design framework, which calculates the difference between the entropy of the current state and the conditional entropy after performing an action and observing the new state, intuitively measuring the amount of information obtained by performing the action.

[0068] For multi-objective test optimization scenarios, a multi-objective reinforcement learning method can be used to optimize multiple objectives (such as test accuracy, test time, resource consumption, etc.) at the same time.

[0069] In this case, Pareto frontier based methods such as multi-objective deep Q-networks or multi-objective soft actor-critic algorithms can be used.

[0070] Step 4.2, build a deep reinforcement learning algorithm; According to embodiments of the present application, a deep Q-network or a policy gradient algorithm (such as PPO or SAC) can be used to learn the optimal test policy. Taking the deep Q-network as an example, its goal is to learn the action-value function, which represents the expected long-term cumulative reward of performing an action in a given state.

[0071] The Q-network is optimized by minimizing the temporal difference error, which measures how far the current Q-value estimate is from the target Q-value. The target Q-value consists of the immediate reward plus the discounted maximum Q-value of the next state.

[0072] In practical applications, a dual-Q network can be used to improve learning stability. The network structure is as follows: Input layer: Receives state representation, including graph structural features, current stress distribution, and test configuration.

[0073] Graph convolution layer: processes graph structure data and extracts relational features between nodes.

[0074] Fully connected layer: integrates graph features and other state information.

[0075] Output layer: For discrete action space, it outputs the Q value of each action; for continuous action space, it outputs the mean and standard deviation of the action.

[0076] For example, in stress testing of automotive connectors, a reinforcement learning agent can learn to intelligently select the optimal test point locations and test parameters based on the connector's geometric characteristics and material properties.

[0077] By repeatedly trying different testing strategies, the agent is able to discover potential weaknesses that are difficult to identify with traditional methods, such as stress concentration areas during the insertion and removal process.

[0078] Step 4.3, implement the hierarchical reinforcement learning architecture; To handle the complexity of connector testing, a hierarchical reinforcement learning architecture is adopted: The high-level strategy is responsible for determining the overall testing strategy, such as prioritization of test areas, selection of test types, etc.

[0079] The low-level strategy is responsible for the fine adjustment of specific test parameters, such as the specific test point location within the selected area, the precise size of the test force, etc.

[0080] According to one embodiment of the present application, such a layered architecture can be implemented through an option framework, where high-level policies select options and low-level policies perform atomic actions within the options.

[0081] Step 4.4, integrating expert knowledge; Integrate domain expert knowledge into the reinforcement learning process to improve learning efficiency: Initialize policies from expert demonstration data via imitation learning.

[0082] Design rule-based safety constraints to prevent agents from exploring dangerous test parameters.

[0083] Use rule-based reward shaping to guide agents to learn effective policies faster.

[0084] In practical applications, expert knowledge on connector testing can be gathered from experienced test engineers, such as: High stress areas usually appear at locations where the geometry changes drastically, such as corners and around holes.

[0085] The interfaces between different materials are potential stress concentration areas and failure points.

[0086] Under cyclic loading, fatigue cracks usually propagate perpendicular to the direction of maximum principal stress.

[0087] This expert knowledge can be incorporated into the reinforcement learning process through reward function design and initial policy construction. For example, higher rewards can be given for identifying high-risk areas, and bonus rewards can be given for exploring potential weak areas identified by experts.

[0088] The output of this step is a reinforcement learning agent that can dynamically optimize the test strategy based on the connector characteristics and stress distribution, which can significantly improve test efficiency and accuracy.

[0089] Step 5: Continuously optimize the parameters of each model based on the test data optimized by the reinforcement learning algorithm to form a closed-loop optimization system; The specific steps include: Step 5.1, build a data acquisition system; Integrate X-ray imaging, acoustic emission detection or other sensor technologies to collect real-time data during the connector stress test process: X-ray imaging systems can provide high-resolution images of the connector's internal structure and microscopic cracks.

[0090] The acoustic emission detection system can capture the sound wave signals generated by tiny deformation and crack propagation inside the material.

[0091] Sensors such as strain gauges can measure local strain data at specific locations.

[0092] Step 5.2: Real-time comparison between the model and measured data; According to an embodiment of the present application, the stress distribution predicted by the model is compared with the data collected in real time, and the prediction error is calculated, which measures the difference between the model prediction value and the measured value.

[0093] Step 5.3, continuously optimize the model based on error feedback; Based on the prediction error, the model parameters are continuously updated using online learning methods; Specifically, the model parameters are updated in the opposite direction of the loss function gradient through the gradient descent method, where the learning rate controls the update step size.

[0094] To prevent overfitting and catastrophic forgetting, this application provides a flexible weight integration method. This method optimizes the current task loss while maintaining the previously learned important parameter values ​​by adding regularization terms, where parameter importance is measured by the diagonal elements of the Fisher information matrix.

[0095] In some implementations, continuous learning strategies can be employed to address time-varying data distributions (e.g., performance changes due to connector aging). For example, a sliding window approach can be used to retain only the data from the most recent N time steps for training; or task-based regularization methods, such as procedural regularization, can be used to gradually adjust model complexity over time.

[0096] Meta-learning can also be used to achieve rapid adaptation. By pre-training the meta-model on different connector test tasks, the system can quickly adapt to new connector types or test conditions, reducing the need for large amounts of real-world data.

[0097] Step 5.4, optimize the test strategy; Adjust the testing strategy of the reinforcement learning agent based on the updated model: Update the environment model, allowing the reinforcement learning agent to make decisions based on more accurate predictions.

[0098] Re-evaluate the priorities of key areas to ensure that testing resources are focused where they need the most attention.

[0099] Based on the accumulated testing experience, the reward function is dynamically adjusted to better reflect the testing objectives.

[0100] The output of this step is a closed-loop system that continuously optimizes itself, continuously improving prediction accuracy and test efficiency based on real-time test data, providing a highly intelligent solution for connector stress testing.

[0101] Application examples of this implementation: To demonstrate the practical application effect of this embodiment, an application example in stress testing of automotive electronic connectors is given below.

[0102] Application scenarios: The application scenario in this example involves stress testing a high-performance, multi-pin connector in an automotive electronic control system. This connector, with 120 pins, operates in a harsh environment characterized by high temperature, high humidity, and vibration, requiring it to withstand multiple plug-in and plug-out cycles and thermo-mechanical cycling. Traditional testing methods suffer from low test efficiency and difficulty identifying hidden fault points.

[0103] Implementation process example: Multi-scale graph structure modeling implementation: First, a multiscale graph structure was constructed based on the connector's 3D CAD model and computed tomography data. For the macrostructure, components such as the connector housing, terminal array, positioning mechanism, and seal were represented as nodes, and the physical connections between components were represented as edges. Particular attention was paid to the contact area between the terminal and the housing, which was used as a microstructural refinement area.

[0104] For example, for a 120-pin automotive connector, the macrograph contains approximately 30 nodes (representing the main structural components) and 60 edges (representing the connections between components). For the terminal contact area, a micrograph structure is constructed, with each microregion containing approximately 500 nodes, representing the material microstructure units (such as copper alloy grains and plating interfaces).

[0105] Physically constrained spatiotemporal graph neural network implementation: A graph neural network architecture consisting of five spatial propagation layers and two temporal evolution layers was constructed. The spatial propagation layers used graph convolutional networks to transfer information between nodes, while the temporal evolution layers employed physically constrained recursive cells to predict stress evolution over time.

[0106] To ensure physical constraints, physical properties such as elastic modulus, Poisson's ratio, and yield strength of the connector material are incorporated into the network, and force balance constraints are applied. For example, for copper alloy terminals, the elastic modulus is set to =110GPa, Poisson's ratio = 0.34, and the physical constraint layer ensures that the predicted stress distribution satisfies Hooke's law and the force balance equation.

[0107] Meta-enhanced contrastive learning of micro stress field implementation: Thirty sets of high-resolution X-ray tomography data were collected as a training set. Each set contained the connector's microstructure and stress distribution under different load conditions. This data was expanded to 300 sets of training samples through data augmentation, and a meta-enhanced contrastive learning model was constructed.

[0108] During the implementation process, the micro-stress field modeling task is decomposed into multiple subtasks: Terminal contact surface stress analysis, sealing interface stress analysis, plastic housing stress analysis, etc.

[0109] For each subtask, a 5-shot learning method is adopted (i.e., only 5 samples per class are used for adaptive learning).

[0110] In areas of high stress gradient (such as the root of the terminal and the corner of the housing), the node density is automatically increased, and the mesh accuracy is refined from the standard 0.5mm to 0.05mm to ensure accurate capture of stress concentration phenomena.

[0111] Reinforcement learning test strategy optimization implementation: A reinforcement learning environment is constructed with a state space dimension of 256 (including graph features and the current test state), and an action space including the test force magnitude (5 discrete levels), force direction (8 discrete directions), and test point location (15 key areas on the connector).

[0112] A dual-Q network is used to implement the reinforcement learning algorithm. The network consists of three graph convolutional layers and two fully connected layers. The reward function is designed based on a combination of information gain (70%) and resource consumption (30%).

[0113] The system integrates the expertise of five senior test engineers, including 20 test experience rules, such as "stress concentration is prone to occur in the contact area between the terminal and the shell" and "micro cracks are prone to occur at the connection between the plastic shell and the metal parts after thermal cycling". Through reward shaping, the intelligent agent is guided to explore these potential weak areas.

[0114] Real-time data feedback optimization implementation: A real-time X-ray imaging system and an 8-channel acoustic emission detection system were deployed to collect real-time data on connector microstructural changes and stress distribution during testing. Data was collected every five minutes and compared with model predictions to calculate the prediction error.

[0115] Use elastic weight integration method to continuously optimize the model and set penalty coefficients for important parameters =0.4, retaining historical training experience while adapting to new data. Use the sliding window method to retain the most recent =50 sets of test data for online learning.

[0116] Technical effect verification: Improved stress distribution prediction accuracy: Compared with the traditional finite element method, this method shows significant advantages in predicting the stress distribution of connectors.

[0117] In 30 sets of verification tests, the average stress prediction error was reduced from 15.8% of the traditional method to 6.3%. In particular, in the stress concentration area, the prediction accuracy was improved by 61.5%.

[0118] At the microstructural level, the stress concentration points at the terminal plating interface that could not be detected by traditional methods were successfully predicted, with a degree of consistency of 92% with the X-ray observation results.

[0119] Improved testing efficiency: This method shortened connector testing cycles from the traditional 14 days to 4 days, and improved test point placement efficiency by 78%. In particular, during thermal cycling testing, the test points automatically selected by the agent maximized information with the minimum number of tests (an average reduction of 65%).

[0120] For a batch of 1,000 connectors, the traditional method requires randomly sampling 50 samples for comprehensive testing, while this method only requires 18 samples to achieve the same detection reliability, significantly reducing testing costs and time.

[0121] Fault Prediction: This method can predict potential connector failure points over long-term use. In a six-month validation test, the system successfully predicted 87% of actual fault locations, a 42 percentage point improvement over traditional empirical predictions.

[0122] Especially for the early detection of tiny cracks, this method can issue an early warning when the crack length is less than 0.2mm (traditional visual inspection usually requires the crack to develop to more than 0.5mm), providing sufficient time for preventive maintenance.

[0123] like Figures 2 to 5 As shown, the stress prediction error comparison of different methods; the trend of test efficiency changing with connector complexity; the comparison between micro stress prediction and actual measurement; and the comprehensive performance evaluation.

[0124] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A connector stress test optimization method based on reinforcement learning, characterized in that: The following steps are involved: Constructing a multi-scale graph structure model of the connector, the multi-scale graph structure model includes a macrograph representing the macrostructure of the connector, a micrograph representing the microstructure of the connector, and a cross-scale connection connecting the macrograph and the micrograph; Based on the multi-scale graph structure model, a physically constrained spatiotemporal graph neural network is constructed. The physically constrained spatiotemporal graph neural network includes a spatial propagation layer and a temporal evolution layer. Physically constrained recursive cells are introduced to ensure that the prediction results conform to the basic principles of material mechanics. Based on the output of the physical constraint spatiotemporal graph neural network, a meta-enhanced contrastive learning method is used to construct a microscopic stress field representation model, and the model accuracy of the local area is dynamically adjusted according to the stress gradient; Based on the prediction results of the micro-stress field representation model, the reinforcement learning algorithm is used to dynamically optimize the test parameters and test point locations; Based on the test data optimized by the reinforcement learning algorithm, the parameters of each model are continuously optimized to form a closed-loop optimization system.

2. The connector stress test optimization method based on reinforcement learning according to claim 1, characterized in that: The multi-scale graph structure model for building a connector includes: Get the structural data of the connector; Construct a macrograph structure, where nodes represent the key structural components of the connector and edges represent the physical connection relationships between components; Construct a microscopic graph structure, where nodes represent the microscopic components of the material and edges represent the interaction between microscopic units; Establish the connection between macro-graph structure and micro-graph structure; Initialize the node features and edge features in the graph structure.

3. The connector stress test optimization method based on reinforcement learning according to claim 1, characterized in that: The physical constraints in the step of introducing physical constraint recursive cells to ensure that the prediction results comply with the basic principles of material mechanics include: force balance constraints, material constitutive relationship constraints and boundary condition constraints.

4. The connector stress test optimization method based on reinforcement learning according to claim 1, characterized in that: Meta-enhanced contrastive learning methods include: Generate multiple views of the same microscopic region sample through data augmentation; Mapping multiple viewpoints into latent feature space; Make different views from the same sample close together in the feature space, while views from different samples move away; The micro-stress field modeling task is decomposed into multiple subtasks, and a small number of sample learning strategy is adopted for each subtask.

5. The connector stress test optimization method based on reinforcement learning according to claim 1, characterized in that: The step of dynamically adjusting the model accuracy of the local area according to the stress gradient includes: Graph coarsening is performed on areas with smaller stress gradients, and adjacent nodes are merged to reduce computational complexity. Graph refinement is performed on areas with large stress gradients to increase node density and improve model accuracy.

6. The connector stress test optimization method based on reinforcement learning according to claim 1, characterized in that: Reinforcement learning algorithms include: Define the state space, including the connector's graph structure, stress distribution, and test configuration; Define the action space, including test parameters and test point locations; Define the reward function, designed based on information gain and resource consumption; Use deep reinforcement learning algorithms to learn optimal testing strategies.

7. The connector stress test optimization method based on reinforcement learning according to claim 6, characterized in that: The reinforcement learning algorithm is a hierarchical reinforcement learning architecture, including: High-level strategy is responsible for determining the overall test strategy, including prioritization of test areas and selection of test types; The low-level strategy is responsible for the fine-tuning of specific test parameters, including the exact location of test points within the selected area and the precise magnitude of the test force.

8. The connector stress test optimization method based on reinforcement learning according to claim 1, characterized in that: The step of continuously optimizing the parameters of each model based on the test data optimized by the reinforcement learning algorithm includes: Integrated sensor technology collects real-time data on connectors during stress testing; Compare the stress distribution predicted by the model with the real-time collected data and calculate the prediction error; Based on the prediction error, the model parameters are continuously updated using online learning methods; An elastic weight integration method is used to prevent overfitting and catastrophic forgetting.

9. The connector stress test optimization method based on reinforcement learning according to claim 1, characterized in that: It also includes the steps of integrating expert knowledge, which include: Initialize the policy from expert demonstration data via imitation learning; Design rule-based safety constraints to prevent the agent from exploring dangerous test parameters; Use rule-based reward shaping to guide agents to learn effective policies faster.

10. A connector stress test optimization system based on reinforcement learning, used to execute a connector stress test optimization method based on reinforcement learning according to any one of claims 1 to 9, characterized in that: include: Multi-scale graph structure modeling module, used to build a multi-scale graph structure model of the connector; A physical constraint spatiotemporal graph neural network module predicts the stress distribution of connectors based on a multi-scale graph structure model; Meta-enhanced contrastive learning module, used to construct a microscopic stress field representation model; Reinforcement learning optimization module, used to dynamically optimize test parameters and test point locations; The real-time data feedback and model optimization module is used to continuously optimize the prediction model parameters based on real-time test data.

Images