Silicone button conductive property reinforcement learning test system and method

Abstract

The present application relates to the technical field of silicone key conductivity test, and discloses a silicone key conductivity reinforcement learning test system and method, wherein the silicone key conductivity reinforcement learning test method comprises the following steps: constructing a silicone key conductivity test environment; establishing a test knowledge base, extracting general feature representation of key conductivity performance through feature mapping and domain adaptation algorithm; constructing a test parameter optimization model, modeling the test process as a Markov decision process, and learning the optimal test strategy based on a reward function; adjusting the test parameters in real time during the test process, performing reliability analysis after obtaining the test results, and feeding back the new test experience to the knowledge base to realize the accumulation and reuse of test knowledge; the present application realizes the self-adaptive optimization of test parameters and the cross-key knowledge transfer by fusing reinforcement learning, transfer learning and meta-learning technology.

Description

Technical Field

[0001] This invention relates to the field of silicone button conductivity testing technology, and more specifically, to a silicone button conductivity reinforcement learning testing system and method. Background Technology

[0002] Silicone buttons are widely used human-computer interaction components in modern electronic devices, and the stability and reliability of their conductivity directly affect the user experience and quality of the devices. As electronic products develop towards miniaturization and high integration, the requirements for the accuracy and efficiency of conductivity testing for silicone buttons are constantly increasing.

[0003] Currently, the conductivity testing of silicone buttons mainly employs static testing methods based on preset parameters. These methods typically set fixed parameters such as test current, measurement time, and sampling frequency according to the button's rated parameters and empirical values, and then execute the test according to a predetermined procedure.

[0004] The main problems with existing technologies include: First, the issue of standardized test parameters. Traditional testing methods use uniform test parameter configurations, which cannot be customized for silicone buttons with different materials, structures, and working environments. This "one-size-fits-all" testing approach is prone to insufficient or excessive testing when dealing with new silicone materials or buttons with special structures, affecting the accuracy and efficiency of test results. Second, low knowledge utilization efficiency. Existing testing methods lack effective experience accumulation and knowledge reuse mechanisms. Each time a new type of button is encountered, test benchmarks need to be re-established and parameters adjusted, failing to fully utilize past testing experience and resulting in repetitive work and wasted resources. Third, poor environmental adaptability. In actual production... In the environment, factors such as temperature, humidity, and electromagnetic interference can affect the stability of test results. Existing technologies lack the ability to perceive and compensate for environmental changes in real time, resulting in poor repeatability and consistency of test results. The testing process lacks intelligence. Traditional testing methods execute according to preset procedures and cannot judge and adjust subsequent testing strategies based on intermediate test results. This testing mode, which lacks adaptability, is prone to testing blind spots or inefficiency in complex testing scenarios. It is also highly dependent on data. For testing new silicone buttons, existing methods require a large amount of labeled data to establish test models and calibration parameters. The data acquisition process is time-consuming and costly, which seriously restricts the development cycle and testing efficiency of new products.

[0005] Therefore, there is a need for an intelligent silicone button conductivity testing method that can adaptively adjust test parameters, make full use of historical test experience, and has environmental adaptability, in order to solve the above-mentioned problems in the existing technology. Summary of the Invention

[0006] This invention provides a silicone button conductivity reinforcement learning testing system and method, which solves the technical problems in related technologies such as fixed test parameters, low knowledge utilization efficiency, poor environmental adaptability, lack of intelligence in the testing process, and strong data dependence.

[0007] This invention provides a method for testing the conductivity reinforcement learning of silicone buttons, including:

[0008] A conductivity testing environment for silicone buttons was constructed, which included a resistance measurement unit, a test current control unit, an environmental factor monitoring unit, and an actuation force control unit.

[0009] A test knowledge base is established based on button test data collected in the test environment. General feature representations of button conductivity are extracted through feature mapping and domain adaptation algorithms.

[0010] Based on the feature representation output by the test knowledge base, a test parameter optimization model is constructed using a reinforcement learning algorithm. The test process is modeled as a Markov decision process, and the optimal test strategy is learned based on the reward function.

[0011] The output of the test parameter optimization model is used to adjust the test parameters in real time during the test, including the test current, measurement time and sampling frequency. After obtaining the test results, reliability analysis is performed, and new test experience is fed back to the knowledge base to realize the accumulation and reuse of test knowledge.

[0012] Furthermore, the steps for establishing the test knowledge base include:

[0013] Feature extraction is performed on the collected button test data to form a feature vector containing button physical characteristics, test conditions, and test results;

[0014] Build a basic neural network model to predict test results based on key features and test conditions;

[0015] Implement a feature mapping network to map the features of different button types to a unified feature space;

[0016] By applying domain adversarial training, the feature representation can reduce the distribution differences between different key types while preserving the ability to predict key performance.

[0017] Furthermore, the establishment of the test knowledge base also includes meta-learning optimization, specifically including:

[0018] Treat each button type as an independent task to build a multi-task learning environment;

[0019] The model-independent meta-learning algorithm is used to optimize the initial parameters of the model, enabling the model to quickly adapt to new button types under conditions of a small number of samples;

[0020] For the new button model, with optimized initial parameters, only a small number of test samples are needed to complete model adaptation and achieve high-precision prediction.

[0021] Furthermore, in the step of constructing a test parameter optimization model using a reinforcement learning algorithm:

[0022] The state space includes current test parameters, measured conductivity indicators, environmental factors, and key characteristics;

[0023] The action space is defined as the adjustment method of the test parameters, including test current adjustment, measurement time adjustment, and sampling frequency adjustment;

[0024] The reward function is defined as a weighted combination of test accuracy and time efficiency, used to balance test quality and efficiency.

[0025] Furthermore, the reinforcement learning algorithm includes the following steps:

[0026] Construct a main network and a target network, with the same structure but different parameters;

[0027] A priority-based empirical replay mechanism is adopted to assign a priority to each sample based on the temporal difference error.

[0028] By using an ε-greedy strategy to select actions, a balance is achieved between exploration and exploitation;

[0029] The main network parameters are updated using a dual-delay Q-learning method to reduce the problem of overestimation of Q-values.

[0030] Furthermore, the step of adjusting test parameters in real time during the test includes:

[0031] Deploy the trained reinforcement learning model to the test environment;

[0032] Real-time acquisition of current test status information, including current test parameters, measurement results, and environmental conditions;

[0033] Based on the action decisions output by the reinforcement learning model, adjust the test current, measurement time, and sampling frequency.

[0034] By combining confidence interval estimation, an adaptive balance between exploration and utilization is achieved during the adjustment of test parameters.

[0035] Furthermore, the steps for performing reliability analysis include:

[0036] Calculate the uncertainty of the measurement results, including Type A uncertainty and Type B uncertainty;

[0037] Statistical methods are used to identify abnormal measurement data, including box plots, Z-scores, and density-based spatial clustering.

[0038] The reliability score of the test results is determined by taking into account both uncertainty and the proportion of outliers.

[0039] Furthermore, the step of feeding new testing experience back to the knowledge base includes:

[0040] Based on the reliability analysis results, select high-quality test data;

[0041] Update the association mapping of test results for button type test parameters, including creating mapping records for new types of buttons;

[0042] Update the key feature representation and feature transformation network parameters using the new data;

[0043] Incremental learning is performed on the prediction model based on new data, and the parameters of the domain adaptation algorithm are adjusted.

[0044] Furthermore, the feeding of new testing experience back to the knowledge base also includes knowledge distillation, specifically including:

[0045] Use a well-trained reinforcement learning model as a teacher model;

[0046] Design a student model with a simpler structure and lower computational cost;

[0047] Through hard-objective learning, soft-objective learning, and feature distillation, knowledge from the teacher model is transferred to the student model;

[0048] The performance of the student model after knowledge distillation was evaluated to ensure that the core decision-making capabilities of the teacher model were preserved.

[0049] This invention provides a silicone button conductivity reinforcement learning testing system for performing the aforementioned silicone button conductivity reinforcement learning testing method, comprising:

[0050] The test environment module includes a resistance measurement unit, a test current control unit, an environmental factor monitoring unit, and a pressure control unit, which are used to construct a controllable button conductivity test environment;

[0051] The knowledge base module is used to store button test data and establish a general feature representation library of button conductivity performance through feature mapping and domain adaptation algorithms.

[0052] The reinforcement learning optimization module constructs a test parameter optimization model based on the feature representation of the knowledge base, and learns the optimal test strategy by using Markov decision process modeling and reward function mechanism.

[0053] The adaptive control and feedback module adjusts test parameters in real time based on the output of the optimization model, performs reliability analysis on the test results, and updates the knowledge base with test experience feedback.

[0054] The beneficial effects of this invention are: it can dynamically adjust test parameters based on real-time feedback during the test process, avoiding redundant test steps and effectively improving test efficiency;

[0055] By using transfer learning and meta-learning techniques, this invention can transfer existing button testing experience to new buttons, reducing the sample data required for testing new buttons and lowering data collection costs.

[0056] It can learn the common characteristics between different button types, quickly adapt to various new silicone buttons, and realize the reuse and transfer of test knowledge;

[0057] Through the adaptive strategy of reinforcement learning and the continuous updating of the knowledge base, good test consistency was maintained under different environmental conditions, and the fluctuation of test results was reduced.

[0058] Without human intervention, this invention can automatically adjust the optimal testing strategy according to the testing process, reducing the impact of human factors and improving the level of testing automation.

[0059] By using a structured knowledge base and knowledge distillation technology, this invention enables the systematic accumulation and efficient reuse of testing experience, allowing testing knowledge to be inherited and shared across different testing tasks. Attached Figure Description

[0060] Figure 1 This is a flowchart of the silicone button conductivity reinforcement learning test method in this invention;

[0061] Figure 2 This is a bar chart comparing the test times of traditional testing methods and the method of this application on different button types;

[0062] Figure 3 This is a radar chart comparing the traditional method with the method of this application on five key indicators;

[0063] Figure 4 This is a scatter plot showing the relationship between the number of samples and the test accuracy in a novel button test.

[0064] Figure 5 It is a line graph showing the changing trends of average reward value, test accuracy, and test time as the number of training steps increases during reinforcement learning training.

[0065] Figure 6 This is an area plot comparing the cumulative test throughput of the traditional method and the method of this application within 8 hours. Detailed Implementation

[0066] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0067] At least one embodiment of the present invention discloses a method for reinforcing learning the conductivity of silicone buttons, such as... Figure 1 As shown, it includes:

[0068] Step 1: Construct a conductivity test environment for silicone buttons. The test environment includes a resistance measurement unit, a test current control unit, an environmental factor monitoring unit, and a pressure control unit.

[0069] The purpose of this step is to build a controllable and stable testing environment, providing the hardware and software foundation for subsequent conductivity testing of silicone buttons.

[0070] Step 1.1, build the test hardware system;

[0071] Build a test hardware system that includes the following components:

[0072] High-precision resistance measurement unit: Used for accurate measurement of the resistance of silicone buttons under different conditions, with a resolution better than 0.01Ω and a measurement range of 0.1Ω to 10MΩ. This unit uses a four-wire measurement method to eliminate the influence of wire resistance and improve measurement accuracy.

[0073] Variable test current control unit: capable of generating an adjustable and stable test current ranging from 0.1mA to 100mA with an accuracy better than ±0.5%. This unit adopts a constant current source design to ensure that the test current remains stable when the resistance of the button under test changes.

[0074] Ambient temperature monitoring unit: Real-time monitoring of temperature changes in the test environment, with an accuracy better than ±0.5℃ and a monitoring range of -10℃ to 80℃. This unit senses temperature through a thermistor or thermocouple and converts the analog signal into a digital signal through an analog-to-digital converter.

[0075] The pressure control unit is capable of applying adjustable pressure from 0.1N to 10N to silicone buttons, with an accuracy better than ±1% and repeatability better than ±0.5%. This unit uses a precision pressure application mechanism driven by a stepper motor, which can accurately control the pressing depth and speed.

[0076] Step 1.2: Deploy the data acquisition system;

[0077] Deploy a high-speed, multi-channel data acquisition system with the following functions:

[0078] Real-time acquisition of resistance measurement data, with a sampling rate of no less than 1000Hz and a resolution of no less than 16 bits;

[0079] Simultaneously record parameters such as test current, ambient temperature, and pressure.

[0080] Achieve time synchronization of data from multiple sensors with an error of no more than 1ms;

[0081] It has anti-interference processing capabilities and a signal-to-noise ratio of no less than 60dB.

[0082] The data acquisition system adopts a modular design, including signal conditioning circuits, analog-to-digital converters, and data buffer units, and can be expanded to more than 16 parallel measurement channels.

[0083] Step 1.3: Establish a data processing system;

[0084] Construct a data processing system to preprocess the collected raw data, including:

[0085] Filtering: A low-pass filter algorithm is used to remove high-frequency noise, and the cutoff frequency can be automatically adjusted according to test requirements;

[0086] Anomaly detection: Identifying outlier data points through statistical methods to improve data quality;

[0087] Feature extraction: Extracting key features from raw measurement data, such as average resistance, rate of change, and stability;

[0088] Data standardization: Standardize parameters of different dimensions to facilitate subsequent processing by machine learning algorithms.

[0089] The data processing system adopts a pipeline architecture, supports real-time data processing, and has a latency of no more than 10ms.

[0090] Step 1.4, integrate the test environment;

[0091] The above hardware and software systems are integrated into a complete testing environment to achieve the following functions:

[0092] The work of each unit is coordinated through a unified control interface;

[0093] Establish a database of test parameters and test results to support historical data query and analysis;

[0094] It provides a visual interface to display the test process and results in real time;

[0095] It supports both manual setting and automatic adjustment modes for test parameters.

[0096] This integrated testing environment adopts a modular and scalable design, facilitating subsequent functional upgrades and maintenance. Once built, the testing environment outputs a fully functional and stable silicone button conductivity testing platform, providing fundamental support for subsequent reinforcement learning testing.

[0097] Step 2: Establish a test knowledge base based on the button test data collected in the test environment, and extract general feature representations of button conductivity performance through feature mapping and domain adaptation algorithms;

[0098] The purpose of this step is to build a knowledge base that can store and utilize experience in testing the conductivity of silicone buttons, providing a knowledge foundation for subsequent optimization of test parameters.

[0099] Step 2.1, Build the basic model;

[0100] The basic model is built using existing button test data, as follows:

[0101] Data collection and processing: Collect conductivity data of different types of silicone buttons under various test conditions, including button type information (material, structure, size, etc.), test parameters (current, pressure, etc.) and test results (resistance value, stability, etc.).

[0102] Feature engineering: Extracting effective features from raw data, including:

[0103] Button physical characteristics: dimensions, material ratio, structural type, etc.;

[0104] Test conditions include: test current, ambient temperature, and pressure applied.

[0105] Test result characteristics: resistance value, resistance change rate, response time, etc.

[0106] Establishing a basic model: A deep neural network is used to construct a prediction model for the conductivity of buttons. The network structure includes:

[0107] Input layer: Receives key press features and test condition features;

[0108] Hidden layers: Multiple fully connected layers, using ReLU activation function;

[0109] Output layer: Predicts test result features.

[0110] This basic model can predict possible test results by inputting key features and test conditions, providing a reference for subsequent tests.

[0111] Step 2.2: Implement the feature mapping and domain adaptation algorithm;

[0112] To address the differences between different button types, feature mapping and domain adaptation algorithms are implemented:

[0113] Feature mapping: A feature transformation network maps the features of different buttons to the same feature space, reducing inter-domain differences. The feature transformation network employs a multilayer perceptron structure, and its implementation includes:

[0114] Input layer: Receives the original features of the key presses, with dimension d. in ;

[0115] Hidden layer 1: Contains 128 neurons, with ReLU activation function;

[0116] Hidden layer 2: Contains 64 neurons, with ReLU activation function;

[0117] Output layer: Outputs the transformed feature representation, with dimension d. out .

[0118] The feature transformation network is trained by minimizing the difference between the feature distributions of the source and target domains. In the conductivity testing scenario of silicone buttons, the feature mapping network can map the features of ordinary silicone buttons to a feature space similar to that of highly elastic silicone buttons, realizing knowledge transfer between buttons made of different materials.

[0119] Domain Adaptation: A domain adversarial training method is employed to reduce the distribution differences between different key types while preserving the key performance prediction capability of the feature representation. The specific implementation of domain adversarial training includes:

[0120] Feature extractor: transforms input features into a neighborhood-invariant feature representation.

[0121] Task Classifier: Predicting Key Conductivity Based on Feature Representation

[0122] Domain classifier: Attempts to distinguish which key type (domain) a feature comes from. During training, the feature extractor must minimize both the task classification loss and the domain classification loss, achieving adversarial training through a gradient inversion layer. The domain adaptation algorithm is achieved through the following objective function:

[0123]

[0124] in Let represent the total training loss function, and represent the overall objective of adversarial training; This represents the performance prediction loss, which is the loss function of the task classifier. It is calculated using mean squared error and measures the accuracy of the button conductivity prediction. λ represents the domain classification loss, which is the loss function of the domain classifier. It is calculated using cross-entropy and measures the performance of the features generated by the feature extractor on the domain classification task. λ represents the tradeoff coefficient, which is used to control the relative importance between the performance prediction loss and the domain classification loss. It is initially set to 0.1 and dynamically adjusted during training.

[0125] In practical applications, domain adaptation algorithms can handle key data tested under different pressure ranges. For example, by aligning key features tested under low pressure conditions (0.1N to 1N) with key features tested under high pressure conditions (1N to 10N), the model can make accurate predictions across pressure conditions.

[0126] Through feature mapping and domain adaptation algorithms, the base model can transfer knowledge learned from one type of button to another, improving its adaptability to new types of buttons.

[0127] Step 2.3, implement meta-learning optimization;

[0128] According to embodiments of this application, a meta-learning method is employed to further improve the model's ability to quickly adapt to novel buttons:

[0129] Task Construction: Treating each button type as an independent task, a multi-task learning environment is constructed. Specific implementation includes:

[0130] Key task classification: Keys are classified into different task categories based on their material, structure, and conductivity.

[0131] Task sampling strategy: A stratified sampling method is adopted to ensure a balanced distribution of various key press tasks during training;

[0132] Support set and query set partitioning: The data within each task is divided into a support set (for rapid adaptation) and a query set (for evaluating the adaptation effect).

[0133] Meta-learning algorithm implementation: This application adopts the Model-Agnostic Meta-Learning (MAML) algorithm, which optimizes the initial parameters of the model through the following steps:

[0134] Initialize the model parameters; the model structure is a four-layer fully connected neural network.

[0135] Input layer: Receives key press features and test condition features, with dimensions consistent with the feature vector;

[0136] Hidden layer 1: 256 neurons, using the ReLU activation function;

[0137] Hidden layer 2: 128 neurons, using the ReLU activation function;

[0138] Output layer: Predicts conductivity performance indicators, with the same number of dimensions as the performance indicators.

[0139] For each task (key type):

[0140] First, starting with the initial parameters, gradient updates are performed using the support set data and the SGD optimizer, with the learning rate set to 0.01.

[0141] Then, the loss is calculated on the query set for the task, using mean squared error as the loss function.

[0142] Update the initial parameters to minimize the average loss across all tasks. This step uses the Adam optimizer with an initial learning rate of 0.001.

[0143] It should be understood that this process is essentially a "learning how to learn" approach, with the goal of finding a set of initial parameters that are highly adaptable to new tasks, enabling the model to quickly adapt to new button types with a small number of samples.

[0144] Meta-learning optimization process:

[0145] Training phase: In each meta-training iteration, 5 to 10 tasks with different button types are randomly sampled for updates;

[0146] Validation phase: Test the model's adaptability on button types that were not involved in training;

[0147] Early stopping strategy: Monitor the performance on the validation set and stop training when there is no improvement after 5 consecutive iterations;

[0148] Hyperparameter selection: Determine the optimal values ​​of the internal learning rate and the meta-learning rate through grid search.

[0149] Through meta-learning optimization, the model can learn a set of initial parameters that allow for rapid adaptation to new tasks, enabling the performance prediction model to be adjusted with only a small number of novel button samples. Furthermore, this capability is particularly important in scenarios involving rapid iteration of button products and small-batch customized production, improving the flexibility and adaptability of the testing system.

[0150] Step 2.4: Establish a structured knowledge base;

[0151] Build an association mapping database to form a structured knowledge base:

[0152] Data structure design: Design a database structure that includes the following entities and relationships:

[0153] Key type entity: Stores the physical characteristics and classification information of the keys;

[0154] Test parameter entity: Stores test conditions and parameter configurations;

[0155] Test Result Entity: Stores the various metrics obtained from the test;

[0156] Association Relationships: Establish multi-dimensional associations between button type test parameters and test results.

[0157] Knowledge organization: Organize test knowledge according to the principles of domain ontology to form a conceptual hierarchy and relational network.

[0158] Query Interface Implementation: Provides a flexible query interface, supporting functions such as similarity-based key search and condition-based parameter recommendation.

[0159] Incremental update mechanism: Design an incremental update mechanism for the knowledge base to support the dynamic addition of new test data and the automatic adjustment of the knowledge structure.

[0160] By establishing a structured knowledge base, the system can efficiently store, retrieve, and utilize past testing experience, providing knowledge support for subsequent testing parameter optimization.

[0161] Once the test knowledge base is established, a comprehensive knowledge system is output, which includes the basic model, feature mapping and domain adaptation algorithms, meta-learning optimization methods, and a structured database, providing a knowledge foundation for the reinforcement learning optimization testing process.

[0162] Step 3: Based on the feature representation output by the test knowledge base, a reinforcement learning algorithm is used to construct a test parameter optimization model, modeling the test process as a Markov decision process, and learning the optimal test strategy based on the reward function.

[0163] The purpose of this step is to build a test parameter optimization system based on reinforcement learning, which learns the optimal test strategy to achieve a balance between test efficiency and accuracy.

[0164] Step 3.1, Modeling the Test Process;

[0165] According to one embodiment of this application, the conductivity testing process of silicone buttons is modeled as a Markov Decision Process (MDP), specifically implemented as follows:

[0166] State space definition: The test state includes the following components:

[0167] Current test parameters: test current, measurement time, sampling frequency, etc.

[0168] Measured conductivity indicators: the sequence of resistance values ​​acquired up to the current moment and their statistical characteristics;

[0169] Environmental factors: temperature, humidity, etc.;

[0170] Key features: The current key feature vector obtained from the test knowledge base.

[0171] It should be noted that the state can be represented as a combination vector of these parameters, containing all the information needed for decision-making.

[0172] In some implementations, the state space may be further extended to include the following elements:

[0173] Historical test trajectory: Record the state-action pair sequence of the past N1 time steps to form a trajectory vector;

[0174] Uncertainty estimation: The assessment of the uncertainty of the current measurement result can be based on the variance of multiple consecutive measurements;

[0175] Button operation history: Includes information such as the number of times the button was used and the cumulative working time.

[0176] Alternatively, for scenarios with limited computing resources, state space compression methods can be used, employing principal component analysis (PCA) or an autoencoder to map high-dimensional states to low-dimensional representations, thereby reducing computational complexity.

[0177] Action space definition: Actions are defined as the ways to adjust test parameters, including:

[0178] Test current adjustment: increase, decrease, or remain unchanged;

[0179] Measurement time adjustment: extend, shorten, or remain unchanged;

[0180] Sampling frequency adjustment: increase, decrease, or remain unchanged;

[0181] Other parameter adjustments: Defined according to specific test requirements.

[0182] These adjustments can be expressed as changes in the corresponding parameters, such as increasing the test current by 0.1mA or shortening the measurement time by 10ms.

[0183] For example, the test current adjustment has three options: {-0.1mA, 0, +0.1mA}, the measurement time adjustment has three options: {-10ms, 0, +10ms}, and the sampling frequency adjustment has three options: {-100Hz, 0, +100Hz}.

[0184] In some implementations, the action space can be represented continuously rather than discretely, and optimized using Deep Deterministic Policy Gradient (DDPG) or Soft Actor-Critic (SAC) algorithms to achieve finer parameter tuning.

[0185] Reward Function Definition: In this application, the reward function is defined as a weighted combination of test accuracy and time efficiency. Specifically, after each test step, the system calculates an immediate reward value based on the accuracy of the current measurement and the time consumed. A higher reward value indicates a better testing strategy.

[0186] Optionally, the reward function can be extended to a multi-objective form, simultaneously considering factors such as test accuracy, time efficiency, energy consumption, and measurement stability. Furthermore, the weights of each factor can be adjusted for different scenarios; for example, increasing the weight of energy consumption in portable testing equipment, while increasing the weight of stability in high-precision testing scenarios.

[0187] State transition process: The system transitions to the next state based on the current state and the selected action, following the inherent laws of the physical system. State transitions have a certain degree of randomness, reflecting the uncertainties in actual testing.

[0188] In some implementations, an explicit state transition model can be learned simultaneously, using a neural network to predict the distribution of the next state. This model-based reinforcement learning approach can support internal planning and simulation, reduce the number of interactions with the real environment, and accelerate the learning process.

[0189] Through the above modeling, this application transforms the complex problem of test parameter optimization into a sequential decision-making problem within a reinforcement learning framework, laying the foundation for subsequent learning of the optimal test strategy. Furthermore, this modeling method has good versatility, applicable not only to the conductivity testing of silicone buttons but also extending to other types of electronic component testing scenarios.

[0190] Step 3.2, construct the reinforcement learning algorithm;

[0191] The optimal testing strategy is learned by employing a DoubleDQN algorithm combined with a priority-based experience replay mechanism.

[0192] Network architecture: Construct two neural networks with the same structure but different parameters:

[0193] The main network Q(s, a; θ1) represents the action selection, with parameter θ1.

[0194] The target network Q(s, a; θ2) represents the network used to calculate the target Q value, with parameter θ2. The specific implementation of the network structure includes:

[0195] Input layer: Receives the state vector The dimension is the size of the state space (usually 20 to 30 dimensions, including test parameters, environmental factors, etc.);

[0196] Shared feature extraction layer: Two fully connected layers, containing 128 and 64 neurons respectively, both using the ReLU activation function;

[0197] Dominance flow: A fully connected layer that outputs the dominance value for each action; the number of neurons is equal to the size of the action space.

[0198] Value stream: A single fully connected layer that outputs state value;

[0199] Output layer: Combining advantage stream and value stream, using the DuelingDQN architecture, the Q-value of each action is calculated.

[0200]

[0201] Where Q(s, a) represents the Q value (action value function) of taking action a in state s; V(s) represents the state value function of state s, which represents the expected reward in state s; A(s, a) represents the advantage function of action a in state s, which represents the advantage of this action relative to the average action; |A| represents the size of the action space, i.e. the total number of actions to choose from; ∑ represents the summation symbol; a′ represents any action in the action space; and A(s, a′) represents the advantage function value of action a′ in state s.

[0202] This network architecture can separate state value assessment and action selection in silicone button testing, improving learning efficiency and stability.

[0203] Priority Experience Replay: Construct a priority replay buffer to store transfer samples. And assign sample priorities based on TD error:

[0204]

[0205] in This represents the priority value of the i1th sample; Represents the absolute value of the TD error of the i1th sample, |·| represents the absolute value operator; ∈1 is a small positive constant, set to 0.01, used to ensure that all samples have a chance to be sampled; α1 is the priority index parameter, set to 0.6, used to control the degree of influence of priority, and the value range is usually from 0 to 1.

[0206] The TD (Time Difference) error of the i1th sample is calculated using the following formula:

[0207]

[0208] in γ1 represents the immediate reward obtained in the i1th sample; γ1 represents the discount factor used to balance the importance of current rewards and future rewards. Indicates the target network's next state And the Q-value estimation of the optimal action, argmax a′ Indicates to make The action a′ that achieves the maximum value; θ1 represents the parameters of the main network; θ2 represents the parameters of the target network; This indicates the main network's view of the current state. and actions Q-value estimation.

[0209] The sampling probability is directly proportional to the priority:

[0210]

[0211] Where P(i1) represents the probability that the i1th sample is sampled; ∑ represents the priority value of the i1th sample, k1 represents the index of the sample in the buffer, and ∑ represents the summation symbol.

[0212] Use importance sampling weights to compensate for non-uniform sampling bias:

[0213]

[0214] in The importance sampling weight of the i1th sample is represented by N2; the total number of samples in the buffer is N2; and the sampling probability of the i1th sample is P(i1). β1 represents the sampling probability of each sample under uniform sampling; β1 represents the importance sampling index parameter, which is used to control the degree of compensation for bias, and increases linearly from 0.4 to 1.

[0215] In silicone button testing applications, the priority experience playback buffer size is set to 10000 to ensure sufficient historical experience diversity, while prioritizing the learning of samples with large prediction errors to accelerate the learning process.

[0216] Learning algorithm implementation: Initialize the main network parameters θ1 and the target network parameters θ2 (initial θ2 = θ1);

[0217] For each training iteration: from the current state Start by selecting an action based on the ε greedy strategy. ε decreases linearly from 1.0 to 0.1;

[0218] Execute action Observation Rewards and the next state

[0219] Transfer sample Store in the priority replay buffer;

[0220] Sample a batch of transfer samples of size 64 from the buffer according to priority;

[0221] Update the main network parameter θ1 using double-delay Q-learning:

[0222] Calculate the target Q value:

[0223]

[0224] in This represents the target Q value of the i2th sample; γ represents the immediate reward obtained by the i2th sample; γ2 represents the discount factor, used to balance the importance of current rewards and future rewards, and is set to 0.99; Indicates the target network's next state Q-value estimation; Indicates the next state of the i2th sample; argmax a′ Indicates the choice to make The action a′ to obtain the maximum value; a′ represents the action that can be selected for the next state; θ1 represents the parameters of the main network; θ2 represents the parameters of the target network;

[0225] Calculate the loss function:

[0226]

[0227] Where L(θ1) represents the loss function with the main network parameter θ1 as the variable; This indicates taking the average value; N3 represents the total number of samples in the sampling batch; ∑ represents the summation symbol. The importance sampling weight of the i2th sample is represented by the weight. This represents the target Q value of the i2th sample; This represents the state of the main network for the i2th sample. and actions Q-value estimation; This represents the squared error between the target Q value and the predicted Q value;

[0228] Update θ1 using the Adam optimizer with a learning rate of 0.0001.

[0229] Update the target network parameters every 100 steps:

[0230] θ2=τ1θ1+(1-τ1)θ2;

[0231] Where θ2 represents the parameters of the target network; τ1 represents the soft update coefficient, which is used to control the speed of updating the target network parameters and is set to 0.01; θ1 represents the parameters of the main network; (1-τ1) represents the proportion of the original target network parameters retained;

[0232] In practical silicone button testing applications, this reinforcement learning algorithm exhibits different policy adaptation capabilities for buttons made of different materials. For example:

[0233] For highly elastic silicone buttons, the algorithm learns to first use a low test current (0.5mA) to quickly detect the resistance range, and then selects an appropriate test current and sampling frequency based on the preliminary results, thus avoiding the overtesting problem in traditional methods.

[0234] For buttons with unevenly distributed conductive carbon particles, the algorithm automatically learns to add a multi-point measurement strategy, measuring and analyzing at different positions on the button to improve test accuracy.

[0235] In environments with fluctuating temperatures, the algorithm can dynamically adjust measurement parameters and compensation coefficients based on feedback from temperature sensors to maintain consistency in measurement results.

[0236] Through the algorithm described above, the system can gradually find the optimal test parameter configuration strategy under different test conditions through repeated trials and learning.

[0237] Step 3.3, Strategy Evaluation and Optimization;

[0238] The learned testing strategies are evaluated and optimized to ensure their effectiveness and robustness:

[0239] Evaluation Indicator Design: Define the following evaluation indicators:

[0240] Cumulative Rewards: The total reward value earned during the entire test;

[0241] Test accuracy: How close the measurement result is to the true value;

[0242] Test time: The total time required to complete the test;

[0243] Parameter adjustment frequency: The ratio of the number of parameter adjustments to the total number of steps during the test.

[0244] Policy evaluation: Policy performance is evaluated in the following ways:

[0245] Execute the learned policy on the test set and calculate the evaluation metrics;

[0246] Compare with benchmark strategies (such as fixed parameter strategies or human experience strategies);

[0247] Analyze the performance differences of different types of buttons and evaluate their generalization ability.

[0248] Strategy optimization: Based on the evaluation results, the strategy is further optimized.

[0249] Adjust the weighting coefficients in the reward function to balance accuracy and efficiency;

[0250] Optimize network structure and hyperparameters to improve learning efficiency and policy quality;

[0251] Increase the randomness of environmental variables to improve the robustness of the strategy;

[0252] By evaluating and optimizing strategies, we continuously improve the performance of reinforcement learning algorithms to ensure that they can run effectively in real-world testing environments.

[0253] After the reinforcement learning optimization test process is completed, an intelligent system that can adaptively adjust test parameters according to the test environment and button characteristics is output, providing core algorithm support for achieving efficient and accurate conductivity testing of silicone buttons.

[0254] Step 4: Based on the output of the test parameter optimization model, the test parameters, including test current, measurement time and sampling frequency, are adjusted in real time during the test. After obtaining the test results, a reliability analysis is performed, and the new test experience is fed back to the knowledge base to realize the accumulation and reuse of test knowledge.

[0255] The purpose of this step is to achieve real-time adaptive adjustment of test parameters based on the aforementioned reinforcement learning model, thereby ensuring the efficiency and accuracy of the testing process.

[0256] Step 4.1, Agent Deployment;

[0257] Deploy the trained reinforcement learning agent into a real-world testing environment:

[0258] Model conversion: Converts neural network models from the training phase into a format suitable for running on test hardware, supporting real-time inference.

[0259] Interface Implementation: Develop and test the communication interface between the system and the agent, including:

[0260] Status information acquisition interface: Collects current status information from the test system;

[0261] Action execution interface: Converts the actions selected by the agent into test parameter adjustment instructions;

[0262] Reward calculation interface: Calculates instant rewards based on test results;

[0263] Parallel processing: Implements a parallel processing mechanism for status acquisition, action decision-making, and parameter adjustment to ensure the system's real-time responsiveness.

[0264] By deploying intelligent agents, the decision-making capabilities of reinforcement learning are introduced into the actual testing system, providing an intelligent core for adaptive parameter adjustment.

[0265] Step 4.2, Real-time parameter adjustment;

[0266] Real-time adjustment of test parameters is achieved based on the decision-making of the reinforcement learning agent:

[0267] Parameter adjustment process:

[0268] The system collects the current test status s t ;

[0269] The agent, based on state s t Select the optimal action a t ;

[0270] System executes action a t Adjust the test parameters;

[0271] The system enters a new state. t+1 Continue testing.

[0272] The adjustment strategy is implemented as follows:

[0273] Gradual adjustment: The magnitude of parameter changes gradually decreases as the testing progresses, with initial exploration followed by minor fine-tuning in the later stages;

[0274] Condition-triggered adjustment: Parameter adjustments are only made when the measurement stability or accuracy reaches a threshold;

[0275] Multi-parameter coordinated adjustment: Considering the mutual influence between parameters, multi-parameter coordinated optimization is achieved;

[0276] By adjusting parameters in real time, the system can dynamically optimize test parameters based on the actual situation during the test, thereby improving test efficiency and accuracy.

[0277] Step 4.3: Predict the optimal parameter combination for the next step;

[0278] Using a reinforcement learning model and a test knowledge base, predict the optimal combination of test parameters for the next step:

[0279] State prediction: based on the current state s t and possible actions a t Predict the next possible state s t+1 ;

[0280] Q-value calculation: For each possible state-action pair (s) t+1 a t+1 Calculate the Q-value to assess its long-term value;

[0281] Parameter combination generation: Selecting the optimal action based on the Q value Generate the corresponding test parameter combinations;

[0282] Uncertainty estimation: Calculate the uncertainty of the forecast to provide a reference for parameter selection.

[0283] By predicting the optimal combination of parameters for the next step, the system can plan the test path in advance, reduce invalid exploration, and accelerate the testing process.

[0284] Step 4.4, Balancing exploration and utilization;

[0285] In the process of adjusting test parameters, this application needs to balance the relationship between exploration and exploitation to ensure the effectiveness and robustness of the test strategy. According to an embodiment of this application, this balance is achieved in the following way:

[0286] Confidence interval estimation: A confidence interval is calculated for the value estimate of each action, taking into account both the average of the predictions and its uncertainty. Specifically, the system calculates a value interval for each possible action, with the interval width proportional to the uncertainty of that action. In this way, the system understands both which actions are likely to bring high value and which actions' estimates are not yet sufficiently certain.

[0287] Alternatively, the uncertainty estimate can be obtained in one of the following ways:

[0288] Monte Carlo sampling: Random noise is added to the parameters of the main network to generate multiple network variants, and the standard deviation of their outputs is calculated;

[0289] Bayesian neural networks: use variational inference methods to learn the posterior distribution of network weights and directly estimate the uncertainty of predictions;

[0290] Ensemble approach: Train multiple Q-networks and use their predicted mean and variance as estimates of value and uncertainty, respectively;

[0291] Upper Confidence Boundary Strategy: This application employs an upper confidence boundary strategy for action selection. It should be understood that this strategy considers not only the expected value of an action but also its uncertainty, tending to select actions that are "likely to be good," even if their average expected value is not the highest. This strategy effectively balances the utilization of known high-value actions with the exploration of uncertain actions.

[0292] In some implementations, different exploration strategies can be used instead of the upper confidence bound method:

[0293] Thompson sampling: Sampling from the posterior distribution of values ​​to select the action with the largest sample value;

[0294] Noise injection method: Add derelated noise during the action selection process, such as parameter space noise or the NoisyNet method;

[0295] Count-based exploration rewards: Provide additional rewards for rare state-action pairs based on the frequency of their access.

[0296] For example, noise injection can be used for extensive exploration in the early stages of testing, and then the upper confidence bound strategy can be switched to for more targeted exploration as learning progresses.

[0297] Adaptive exploration strategy: The exploration level in this application will be dynamically adjusted according to the testing progress and model uncertainty.

[0298] In the initial testing phase: the system adopts a high degree of exploration, encouraging experimentation with different test parameters;

[0299] Later in the testing phase: The system gradually reduces the level of exploration and makes more use of the effective testing strategies that have been discovered;

[0300] High uncertainty area: When the system encounters unfamiliar buttons or test conditions, it will automatically increase the exploration level.

[0301] Optionally, the exploration level can be adjusted using adaptive rules, combining time decay and uncertainty perception. This way, the system gradually reduces exploration as the number of tests increases, but appropriately increases exploration when encountering high uncertainty situations.

[0302] It should be noted that in some implementations, the exploration strategy can be combined with environmental familiarity. For example, a lower level of exploration can be used for button types for which systems have accumulated rich experience, such as highly elastic silicone buttons; while a higher level of exploration can be used for buttons made of novel conductive materials.

[0303] By balancing exploration and utilization, the system can fully explore the parameter space while efficiently utilizing the learned knowledge, ensuring the stability and convergence of the testing process.

[0304] After the test parameters are adaptively adjusted, an adaptive system that can respond to changes in test status in real time and intelligently adjust test parameters is output, providing efficient and accurate parameter control capabilities for the conductivity testing of silicone buttons.

[0305] Step 4.5: Reliability analysis of test results and knowledge feedback;

[0306] This step involves performing a reliability analysis on the obtained test results and feeding high-quality testing experience back into the test knowledge base, thereby accumulating and reusing test knowledge.

[0307] First, a reliability analysis of the test results is conducted to ensure the quality and credibility of the test data. Type A uncertainty of the measurement results is calculated by dividing the standard deviation of multiple consecutive measurements by the square root of the number of measurements. Type B uncertainty is calculated by combining the root mean square of factors such as instrument accuracy, ambient temperature coefficient, and contact resistance. Outliers exceeding 1.5 times the interquartile range are identified using box plots, data points exceeding 3 times the standard deviation are removed using the Z-score method, and local anomaly patterns are detected using a density-based spatial clustering algorithm. A reliability score is calculated by combining the percentage of uncertainty to the measured values ​​and the proportion of outliers to the total data, using a weighted average method. The score ranges from 0 to 100, with data scoring above 80 points marked as high-quality data.

[0308] Next, high-quality test data, filtered through reliability analysis, is fed back into the test knowledge base. A reliability threshold of 80 points is set, and the filtered data undergoes an integrity check to ensure complete records containing button features, test parameters, and results. A hash mapping structure is used to update the associated database of button type test parameters and test results. New buttons automatically create mapping records, while existing mappings retain the latest 1000 records using a sliding window approach. Gradient descent is used to update the feature transformation network parameters, with a learning rate of 0.001 and a batch size of 32. The prediction model is incrementally updated using online learning, with model parameter adjustments performed every 100 new samples collected, while simultaneously updating the discriminator weights of the domain adaptation network.

[0309] Finally, knowledge distillation optimization is implemented to improve model deployment efficiency. The trained four-layer fully connected reinforcement learning network is used as the teacher model, and a student model with two hidden layers (64 and 32 neurons) is designed. Knowledge transfer is achieved by minimizing the mean squared error between the student model output and the teacher model output. The loss function combines hard-label loss (weight 0.3) and soft-label loss (weight 0.7), with a temperature coefficient set to 4.

[0310] Through the four steps described above, this invention achieves automation, intelligence, and efficiency in testing the conductivity of silicone buttons. This method organically combines test environment construction, knowledge base establishment, reinforcement learning optimization, and adaptive adjustment to form a complete testing solution.

[0311] A silicone button conductivity reinforcement learning testing system, used to perform the aforementioned silicone button conductivity reinforcement learning testing method, including:

[0312] The test environment module includes a resistance measurement unit, a test current control unit, an environmental factor monitoring unit, and a pressure control unit, which are used to construct a controllable button conductivity test environment;

[0313] The knowledge base module is used to store button test data and establish a general feature representation library of button conductivity performance through feature mapping and domain adaptation algorithms.

[0314] The reinforcement learning optimization module constructs a test parameter optimization model based on the feature representation of the knowledge base, and learns the optimal test strategy by using Markov decision process modeling and reward function mechanism.

[0315] The adaptive control and feedback module adjusts test parameters in real time based on the output of the optimization model, performs reliability analysis on the test results, and updates the knowledge base with test experience feedback.

[0316] Here, the present invention provides an implementation example:

[0317] The reinforcement learning testing method for the conductivity of silicone buttons proposed in this application has been practically applied on the production line of a smart home device manufacturer. This manufacturer produces various models of smart home control panels, each containing silicone buttons of different materials and structures. Specific application scenarios are as follows:

[0318] Product type: Smart home control panel, including function buttons for temperature adjustment, lighting control, scene switching, etc.;

[0319] Button type:

[0320] Standard silicone button: made of conventional conductive silicone material, circular structure;

[0321] High-elasticity silicone buttons: Special elastic material is added to provide a better tactile feel; rectangular structure.

[0322] Waterproof silicone buttons: featuring a special sealing structure for waterproofing, with an oval shape;

[0323] Miniature silicone buttons: smaller in size, used in compact control panels, with a diameter of less than 5mm.

[0324] Testing requirements:

[0325] Each batch of products must undergo full testing to ensure that the conductivity of the buttons meets the standards.

[0326] New button models require rapid development of testing methods to reduce time to market.

[0327] The testing process needs to be efficient and accurate to reduce production line downtime.

[0328] Test results must be traceable and support the quality management system.

[0329] The method implemented in this application on the smart home control panel production line is as follows:

[0330] First, a conductivity testing environment for silicone buttons was built on the production line:

[0331] Hardware configuration:

[0332] Resistance measurement unit: adopts a four-wire measurement method, with a resolution of 0.005Ω and a measurement range of 0.1Ω to 1MΩ;

[0333] Test current control unit: adjustable range 0.1mA to 50mA, accuracy ±0.3%;

[0334] Temperature monitoring unit: operating range 15℃ to 35℃, accuracy ±0.3℃;

[0335] The pressure control unit has a pressure range of 0.5N to 5N and an accuracy of ±0.8%, and is driven by a servo motor.

[0336] Test workstation: Quad-core 3.2GHz processor, 16GB RAM, 256GB SSD storage.

[0337] Software configuration:

[0338] Operating system: Linux;

[0339] Deep learning framework: TensorFlow 2.5;

[0340] Database: MongoDB is used to store test data and models;

[0341] Communication interface: The Modbus protocol is used to connect the various measurement units;

[0342] This testing environment supports simultaneous testing of 16 buttons, with a sampling rate of 2000Hz and a data preprocessing latency of less than 5ms, meeting the high-efficiency testing requirements of the production line.

[0343] This smart home device manufacturer has accumulated test data on 32 models of silicone buttons across 4 types, and has built a test knowledge base based on this data.

[0344] Basic model construction:

[0345] Collect 50,000 sets of test data for buttons of various models under different conditions;

[0346] Twenty-eight features were extracted from each button, including physical properties, material properties, and historical test performance.

[0347] A four-layer neural network model was constructed, with 28 neurons in the input layer, 256 and 128 neurons in the hidden layers, and 8 test result indicators in the output layer.

[0348] Feature mapping and domain adaptation:

[0349] Implement a feature transformation network to map the features of different button types to a unified feature space;

[0350] Applying domain adversarial training allows buttons with similar characteristics to cluster in the feature space, reducing differences between models.

[0351] The verification results show that the cross-model prediction accuracy has improved from 68% to 89%.

[0352] Meta-learning optimization:

[0353] The 32 models were divided into 24 training tasks and 8 testing tasks;

[0354] The initial parameters were optimized using the MAML algorithm, and converged after 10 iterations.

[0355] For the new model of buttons, only 8 to 12 samples are needed to achieve a prediction accuracy of over 85%.

[0356] During production testing, reinforcement learning is applied to optimize test parameters:

[0357] Test process modeling:

[0358] State space: Contains 22 dimensions, covering current test parameters, environmental conditions, and key characteristics;

[0359] Action space: Defines 9 adjustment combinations for test current, measurement time, and sampling frequency;

[0360] Reward function: combines test accuracy (weight 0.7) and time efficiency (weight 0.3).

[0361] Implementation of reinforcement learning algorithms:

[0362] The architecture adopted is DoubleDQN, and both networks are Layer 3 fully connected networks;

[0363] The experience replay buffer size is 20,000, and the batch size is 64.

[0364] After training in a production environment for 500 hours, the model converged to a stable performance.

[0365] Test Case: Demonstration of the testing process for highly elastic buttons;

[0366] Initial test parameters: test current 1mA, measurement time 100ms, sampling frequency 500Hz;

[0367] Step 1: The system detects that the initial resistance of the button is too high, and increases the test current to 2mA;

[0368] Step 2: The measurement results fluctuated greatly, so the system increased the sampling frequency to 1000Hz;

[0369] Step 3: Based on the resistance stability, the system shortens the measurement time to 50ms;

[0370] Steps 4 to 6: The system fine-tunes each parameter based on real-time measurement results to achieve the optimal combination.

[0371] Final test parameters: test current 2.3mA, measurement time 45ms, sampling frequency 1200Hz.

[0372] The testing process was optimized from the original fixed parameter method (taking 22 seconds) to the adaptive parameter method (taking 8.5 seconds), while the measurement accuracy was improved by 12%.

[0373] In production applications, the specific manifestations of adaptive adjustment of test parameters are as follows:

[0374] Agent deployment:

[0375] Convert the trained DoubleDQN model to TensorFlow Lite format;

[0376] A parallel processing architecture is implemented, with an average latency of only 7ms for state acquisition, decision-making, and execution.

[0377] Real-time parameter adjustment:

[0378] Waterproof button test example: The system automatically identifies the influence of the waterproof membrane and adopts a "low current-long time" test strategy;

[0379] Miniature button test example: The system recognizes the small contact area and adopts a "high-frequency sampling-multi-point averaging" strategy;

[0380] New model button test example: The system first uses exploration mode to quickly find the optimal parameter range, and then makes fine adjustments.

[0381] Balance between exploration and utilization:

[0382] In the initial stage of production, a high exploration rate (0.8) was adopted for the new model buttons;

[0383] As the number of tests increased, the exploration rate gradually decreased to 0.1;

[0384] When the system detects an ambient temperature change exceeding 5°C, it automatically increases the exploration rate to 0.4 and searches for the optimal parameters again.

[0385] Through the above adaptive adjustments, the system can optimize test parameters in real time according to different button characteristics and environmental conditions, thereby improving test efficiency and accuracy.

[0386] After implementing the method of this application on a smart home control panel production line for 12 months, its technical effectiveness was verified through production data. This is mainly reflected in the following two aspects:

[0387] The method described in this application has improved testing efficiency:

[0388] Comparison of test times for different button types:

[0389] Standard button: Reduced from 18 seconds using the traditional method to 7.2 seconds, a reduction of 60%;

[0390] Highly responsive buttons: reducing the time from 22 seconds using the traditional method to 8.5 seconds, a reduction of 61.4%;

[0391] Waterproof button: Reduced from 25 seconds using the traditional method to 9.8 seconds, a reduction of 60.8%;

[0392] Miniature button: Reduces the time from 15 seconds using the traditional method to 5.4 seconds, a reduction of 64%.

[0393] Batch test throughput:

[0394] Traditional method: Test 180 buttons per hour;

[0395] This application method tests 450 keys per hour, resulting in a 150% improvement.

[0396] Impact on production line efficiency:

[0397] Bottleneck time in the testing phase was reduced by 63%;

[0398] The overall production cycle was shortened by 22%;

[0399] Production downtime related to button testing was reduced by 78%.

[0400] Test cost analysis:

[0401] Labor costs: Reduces tester operation time by approximately 65%;

[0402] Equipment utilization rate increased from 42% to 78%;

[0403] The average cost of testing each button has been reduced by 56%.

[0404] The above data shows that the method described in this application improves testing efficiency and reduces testing costs through adaptive optimization of test parameters. In particular, it increases production line throughput by reducing testing time.

[0405] The verification results of the method in this application in terms of testing accuracy and adaptability to new types of buttons are as follows:

[0406] Comparison of test accuracy:

[0407] Accuracy of traditional methods: 89.5% (compared to offline precision measurement equipment);

[0408] The accuracy of the method in this application is 97.3% (compared with offline precision measurement equipment);

[0409] Accuracy improved by 7.8 percentage points.

[0410] Stability analysis of the test:

[0411] The standard deviation of the measurement results decreased from ±8.2% using the traditional method to ±3.6%.

[0412] The difference in test results between different operators decreased from ±6.5% to ±1.2%.

[0413] The deviation of test results under different temperature conditions (15℃ to 35℃) decreased from ±7.8% to ±2.4%.

[0414] New button adaptability:

[0415] The traditional method for establishing a test baseline for a new model takes an average of 5 days.

[0416] The method applied to the new model takes an average of 4 hours, an acceleration of approximately 30 times.

[0417] After adapting to the new model, the accuracy rate reached 94% on the first day and over 97% on the third day.

[0418] Fault detection capability:

[0419] Detection rate of conductivity anomalies: increased from 78% using traditional methods to 96%;

[0420] False alarm rate: reduced from 5.3% with traditional methods to 1.7%;

[0421] The detection rate of minute defects (resistance fluctuation <5%) has increased from 42% to 87%.

[0422] These results validate that the proposed method, through reinforcement learning and knowledge transfer techniques, not only improves testing accuracy and stability but also significantly enhances the system's adaptability to new types of buttons. In particular, the improved testing stability under different environmental conditions provides a more reliable guarantee for product quality control.

[0423] Based on the above test efficiency and accuracy verification results, the silicone button conductivity reinforcement learning test method proposed in this application has achieved results in the actual production environment and meets the expected technical effect in the technical solution.

[0424] like Figures 2 to 6The figures show: a bar chart comparing the test time of the traditional testing method and the method of this application on different button types; a radar chart comparing the traditional method and the method of this application on five key indicators; a scatter plot showing the relationship between the number of samples and the test accuracy in the novel button test; a line graph showing the changing trends of average reward value, test accuracy, and test time as the number of training steps increases during reinforcement learning training; and an area plot comparing the cumulative test throughput of the traditional method and the method of this application over 8 hours.

[0425] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for testing the conductive property of a silicone rubber key, characterized in that, The application relates to a key conductive performance prediction method and device. A silica gel key conductive performance test environment is constructed, and the test environment comprises a resistance measurement unit, a test current control unit, an environmental factor monitoring unit and a pressing force control unit; Based on the key test data collected by the test environment, a test knowledge base is established, the collected key test data is subjected to feature extraction, a feature vector containing key physical characteristics, test conditions and test results is formed, a neural network basic model is constructed, the test result is predicted according to the key features and the test conditions, a feature mapping network is realized, the features of different key types are mapped to a unified feature space, domain adversarial training is applied, the feature representation is made to retain the key performance prediction ability while reducing the distribution difference between different key types, the general feature representation of the key conductive performance is extracted through the feature mapping and the field adaptation algorithm; each key type is regarded as an independent task, a multi-task learning environment is constructed, a model-independent meta-learning algorithm is used to optimize the initial parameters of the model, the model can quickly adapt to new key types under the condition of a small amount of samples, for new key models, only a small amount of test samples are needed to complete the model adaptation, and high-precision prediction is realized; Based on the feature representation output by the test knowledge base, a test parameter optimization model is constructed by using a reinforcement learning algorithm, the test process is modeled as a Markov decision process, and the optimal test strategy is learned based on a reward function; Based on the output result of the test parameter optimization model, the test parameters, including the test current, the measurement time and the sampling frequency, are adjusted in real time in the test process, reliability analysis is carried out after the test result is obtained, and new test experience is fed back to the knowledge base, so that the test knowledge is accumulated and reused; in the step of constructing the test parameter optimization model by using the reinforcement learning algorithm: the state space comprises current test parameters, measured conductive performance indexes, environmental factors and key features. The action space is defined as the adjustment mode of the test parameters, including test current adjustment, measurement time adjustment and sampling frequency adjustment; the reward function is defined as a weighted combination of test accuracy and time efficiency, for balancing test quality and efficiency; the reinforcement learning algorithm comprises the following steps: constructing a main network and a target network, the two network structures are the same but the parameters are different, the network output layer combines advantage flow and value flow, a DuelingDQN architecture is adopted, and the Q value of each action is calculated: wherein represents the Q value of performing action a in state s, i.e. the action value function, represents the state value function of state s, and represents the expected return in state s, represents the advantage function of action a in state s, and represents the advantage of the action relative to the average action; represents the size of the action space, i.e. the total number of selectable actions; represents any one action in the action space, represents the advantage function value of action in state s. The priority experience replay mechanism is adopted to assign a priority to each sample according to a time difference error; an action is selected through an epsilon-greedy strategy to balance exploration and utilization; a double-delayed Q-learning method is used to update main network parameters to reduce overestimation of Q value estimation; the step of adjusting test parameters in real time during testing includes: deploying the trained reinforcement learning model to a test environment; collecting current test state information in real time, including current test parameters, measurement results and environmental conditions; adjusting test current, measurement time and sampling frequency according to the action decision output by the reinforcement learning model, the parameter adjustment adopts a gradual adjustment strategy, the parameter change amplitude gradually decreases with the test process, a conditional trigger adjustment mechanism is used, and parameter adjustment is performed only when the measurement stability or accuracy reaches a preset threshold, realizing multi-parameter collaborative adjustment and considering the mutual influence between test current, measurement time and sampling frequency; in combination with confidence interval estimation, adaptive balance between exploration and utilization in the test parameter adjustment process is realized; the step of performing reliability analysis includes: calculating the uncertainty of the measurement results, including type A uncertainty and type B uncertainty; identifying abnormal measurement data by using statistical methods, including box plot method, Zscore method and density-based spatial clustering; considering the uncertainty and the proportion of outliers, a reliability score of the test result is obtained; the step of feeding the new test experience back to the knowledge base includes: filtering high-quality test data according to the reliability analysis result; updating the association mapping of the test result of the key type test parameter, including creating a mapping record of a new key, updating the association database by using a hash mapping structure, and retaining the latest 1000 records by using a sliding window method; updating the key feature representation and feature transformation network parameters by using new data; based on the new data, incremental learning is performed on the prediction model, the domain adaptation algorithm parameters are adjusted, and the prediction model is learned online, and the model parameter adjustment is performed once every 100 new samples.

2. The method of claim 1, wherein the method is used to test the conductive property of a silicone rubber key. The step of feeding the new test experience back to the knowledge base also includes knowledge distillation, specifically including: using a well-trained reinforcement learning model as a teacher model; designing a student model with simpler structure and smaller calculation amount; migrating the knowledge of the teacher model to the student model through hard target learning, soft target learning and feature distillation; evaluating the performance of the student model after knowledge distillation to ensure that the core decision-making ability of the teacher model is retained.

3. A system for testing the conductive strength of silicone rubber keys, characterized by The method for performing the silicone button key conductivity reinforcement learning test method of claim 1, comprising: a test environment module, including a resistance measurement unit, a test current control unit, an environmental factor monitoring unit and a pressing force control unit, for building a controllable button key conductivity test environment; a knowledge base module for storing button test data, establishing a general feature representation library of button key conductivity performance through feature mapping and domain adaptation algorithm; a reinforcement learning optimization module, based on the feature representation of the knowledge base, builds a test parameter optimization model, and learns the optimal test strategy by using Markov decision process modeling and reward function mechanism; an adaptive control and feedback module, which adjusts the test parameters in real time according to the output of the optimization model, analyzes the reliability of the test results, and updates the test experience to the knowledge base.

Images