A kind of method and device for testing no-disassembly spare power source automatic switching device

Abstract

The application discloses a kind of free to dismantle line spare power automatic device test method and device, the method includes: using one-button free to dismantle line test interface to test instrument safety access running in spare power automatic device;The test instrument uses the reinforcement learning strategy based on safety barrier function, and adaptive test signal is injected to spare power automatic device through the interface;While injecting test signal, the full-link multi-physical quantity data of spare power automatic device is synchronously collected through the interface;Characteristic extraction is carried out to the full-link multi-physical quantity data, and the multi-dimensional feature that the deep performance of spare power automatic device is characterized is obtained;Single machine longitudinal trend analysis and multi-machine horizontal group analysis are carried out based on the multi-dimensional feature, and test result is obtained.The application can comprehensively, in-depth, intelligent test to running in spare power automatic device under the premise of not power off, not dismantling line.

Description

Technical Field

[0001] This invention belongs to the field of automatic switching device testing technology, and relates to a test method and device for automatic switching devices that do not require disconnection. Background Technology

[0002] In substations, automatic transfer switch (ATS) is a key secondary device for ensuring power supply reliability. When the power supply is interrupted due to a fault, the ATS can automatically and quickly put the backup power supply into operation, ensuring continuous power supply to users.

[0003] Currently, the periodic verification of automatic transfer switch devices in substations mainly relies on traditional manual testing methods. The specific operation procedure is as follows: Power outage preparation: For testing purposes, it is usually necessary to shut down the entire station or part of the switch bays involved in the automatic transfer switch logic to avoid accidental operation of the equipment.

[0004] Manual wiring: Workers need to manually disconnect the secondary wires on the device's terminal block and connect the voltage, current, and switch output lines of an external relay protection tester. This process is tedious and prone to errors.

[0005] Manual testing: According to the inspection procedures, the staff manually set the test parameters (such as simulating busbar undervoltage), triggered the test items one by one, and manually recorded the test data such as the device's action time and return signal.

[0006] Result determination: After the test, the device's actions and recorded data are compared with the requirements of the procedure to determine whether the device's actions are "correct".

[0007] The above methods require a complete or partial power outage, which is complex to arrange, involves a large amount of coordination work, and affects the reliability of power supply. Manual wiring, operation, and recording are not only cumbersome, labor-intensive, and inefficient, but also prone to errors, posing safety risks of accidental contact and wiring. Due to limitations in site conditions and manual operation capabilities, it is difficult to simulate complex boundary conditions (such as slow voltage drops, multiple reclosing, etc.), resulting in limited test coverage and difficulty in discovering potential defects. It can only verify the correctness of logic and judge the "yes" or "no" of the action result. The test diagnosis dimension is single and cannot quantitatively assess the subtle drift of the internal relay action sequence, changes in contact resistance, and other "sub-healthy" states. It cannot perform in-depth performance diagnosis and trend prediction, resulting in insufficient diagnostic depth and difficulty in discovering hidden defects. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a testing method and apparatus for a wire-free automatic switching device.

[0009] The present invention adopts the following technical solution.

[0010] The first aspect of this invention provides a test method for a non-disconnection automatic switching device, comprising: The tester is safely connected to the operating automatic transfer switch using a one-click, wire-free test interface. The tester employs a reinforcement learning strategy based on a security barrier function to inject adaptive test signals into the backup automatic switching device through the interface. While injecting the test signal, the action response signal of the backup automatic transfer device is simultaneously collected through the interface to obtain full-link multi-physical quantity data; Feature extraction is performed on the full-link multi-physical quantity data to obtain multi-dimensional features characterizing the deep performance of the backup automatic switching device; Based on the aforementioned multidimensional features, single-machine longitudinal trend analysis and multi-machine horizontal group analysis are performed to obtain test results.

[0011] Preferably, the one-click, wire-free test interface includes a set of adapters that can be directly plugged into the terminal block of the automatic transfer switch (ATS) device. The adapters integrate signal isolation and conditioning circuitry to safely inject the analog test signal generated by the tester into the operating circuit of the ATS device and to collect the action response signal of the ATS device.

[0012] Preferably, the adaptive test signal generation strategy is as follows: Based on the rule base of the backup self-supply standard logic, an initial test case set covering basic functions is automatically generated; During the test, based on the real-time collected action response signals, the currently touched logical nodes and loops are analyzed to dynamically evaluate the test coverage. When it is found that there is no logic or working condition covered, a reinforcement learning strategy based on the security barrier function is adopted to automatically adjust the test signal parameters and generate new optimized test cases. The initial test cases and optimized test cases are integrated to form an adaptive and complete standardized test signal sequence.

[0013] Preferably, the modeling and deployment steps of the reinforcement learning strategy include: Based on the logical topology and electrical parameters of the automatic transfer switch, a high-fidelity digital twin model capable of simulating its dynamic behavior is established to obtain the digital twin environment. The logic nodes of the standby automatic transfer device and their connection relationships are constructed as a graph structure, and a graph neural network is used to extract state feature vectors from the graph structure. Construct a hybrid action space that includes continuous test parameters and discrete test parameters; Establish a multi-objective reward function, which includes at least coverage reward based on new triggering logic nodes, efficiency penalty based on test state, security penalty based on security margin constraints, and intrinsic reward based on prediction state. Based on the state feature vector, hybrid action space, and multi-objective reward function, a deep reinforcement learning algorithm is used to train the agent in the digital twin environment to maximize the cumulative reward. The trained agent is deployed to the actual testing instrument. Before generating the test signal, the safety barrier function is used to determine whether the action meets the safety constraints. If it does not meet the constraints, the action that does not meet the safety constraints is projected to the nearest safe action.

[0014] Preferably, the step of determining whether an action meets safety constraints through a safety barrier function, and if not, projecting the action that does not meet the safety constraints to the nearest safe action, is as follows: For action Calculate its safety barrier function value :

[0015]

[0016] in, Actions Voltage, timing, and current exceeding limits; Each represents an action The voltage command value, action timing width, and injected current amplitude are specified in the data. These are the safety thresholds corresponding to the voltage command value, the action timing width, and the injected current amplitude, respectively. These are the weighting coefficients; like , then indicates an action If the safety constraints are not met, the following formula will be used to... Project to Recent safety actions :

[0017] in, Candidate actions The security barrier function value, For safety margin, The square of the L2 norm; This is a penalty factor.

[0018] Preferably, the formula for calculating the safety margin is:

[0019]

[0020] in, Basic safety margin; Assess the overall health of the equipment; This is the time interval since the last test; Sensitivity coefficient; Environmental factors; These are the current ambient temperature and relative humidity, respectively. For the equipment's rated operating environment, Environmental impact weighting coefficient; This is the environmental factor adjustment coefficient.

[0021] Preferably, the multi-physical quantity data includes electrical quantity signals, switching quantity timing signals, and mechanical vibration signals; The electrical quantity signals include the voltage and current waveforms of the primary system monitored by the automatic transfer switch, as well as the voltage and current waveforms of the secondary circuit inside the automatic transfer switch. The switching timing signal includes the microsecond-level action time of each relay contact in the automatic transfer switch; The mechanical vibration signal includes the mechanical vibration waveform generated at the instant the relay in the automatic switching device is activated.

[0022] Preferably, the multidimensional features characterizing the deep performance of the standby automatic switching device include timing features, electrical features, and spectral features; The timing characteristics include the total delay of relay action, the standard deviation of action timing, and the ratio of action time of different output circuits extracted from the switching timing signal; The electrical characteristics include current dynamic characteristics and voltage transient characteristics extracted from electrical quantity signals. The current dynamic characteristics include the maximum value of the relay coil current change rate, the peak value of the relay coil drive current, and the amount of charge during the relay operation. The voltage transient characteristics include the primary voltage drop rate, sag depth, and recovery time, as well as the turn-off overvoltage amplitude and closing voltage drop of the secondary contact voltage. The spectral characteristics include the energy distribution and dominant frequency components of different frequency bands obtained after performing a fast Fourier transform on the mechanical vibration signal.

[0023] Preferably, the single-machine longitudinal trend analysis specifically includes: Multidimensional feature optimization and composite health index calculation are performed based on multidimensional feature data from previous measurements of the same standby automatic switching device to obtain a health index sequence; a predicted health index value is then obtained based on the health index sequence. The residuals are calculated based on the predicted health index value and the composite health index in the health index sequence. The predicted health index value is then corrected based on the state transition law of the residuals to obtain the corrected predicted health index value. Based on the corrected health index prediction value, the future health index trajectory is extrapolated to provide early warning of degradation, and combined with a preset failure threshold, the remaining lifespan of the automatic start-up device is estimated.

[0024] Preferably, the multi-machine lateral grouping analysis specifically includes: Based on a deep autoencoder network, feature embedding and dimensionality reduction processing are performed on the multidimensional feature data of multiple backup automatic switching devices of the same model to obtain the embedding vector of each device. Deep embedding clustering is performed on the embedding vectors of each backup automatic switching device to divide the backup automatic switching devices into several clusters and mark suspected abnormal devices. Association rule mining is performed on the feature data and metadata of suspected abnormal devices to filter out association rules with strong correlations and provide preliminary clues to the cause of the anomaly. By identifying the root cause variables leading to abnormalities through causal inference models, familial defects can be diagnosed.

[0025] Preferably, the step of performing deep embedding clustering on the embedding vectors of each backup automatic switching device, dividing the backup automatic switching devices into several clusters, and marking suspected abnormal devices specifically includes: Calculate the soft assignment probability of the embedding vector of each standby self-connecting device belonging to each cluster center:

[0026] in, The embedding vector of the i-th standby automatic transfer device Belongs to the Cluster centers The soft assignment probability; For the first Cluster centers; For Euclidean distance, For global scale parameters; Constructing the target distribution probability based on soft assignment probability:

[0027] in, The embedding vector of the i-th standby automatic transfer device Belongs to the Cluster centers The target distribution probability, The embedding vector of the i-th standby automatic transfer device Belonging to the j′-th cluster center The soft assignment probability; , For the first The average weighted distance from the j′ samples within each cluster to the cluster center; By minimizing the KL divergence between the soft assignment probability and the target distribution probability as the loss function, the encoder network parameters and cluster center positions are alternately fine-tuned until the loss function converges. After clustering is completed, the device is automatically divided into several clusters and suspected abnormal devices are marked.

[0028] Preferably, the method further includes: using the excitation-response data, and through the maximum a posteriori probability estimation algorithm, solving for the equivalent parameter set that best represents the current real physical characteristics of the automatic switchgear, and using it as the digital fingerprint of the automatic switchgear to drive the digital twin model, and simulating the future state and predicting the lifespan of the automatic switchgear.

[0029] A second aspect of the present invention provides a test apparatus for a self-connecting switch without disconnection, comprising: A one-click, wire-free test interface is used to safely connect the tester to an operating automatic transfer switch. The tester is used to inject adaptive test signals into the automatic transfer switch (ATS) device through the interface; while injecting the test signals, it synchronously collects the action response signals of the ATS device through the interface to obtain full-link multi-physical quantity data; it performs feature extraction on the full-link multi-physical quantity data to obtain multi-dimensional features characterizing the deep performance of the ATS device; and it performs single-machine longitudinal trend analysis and multi-machine horizontal group analysis based on the multi-dimensional features to obtain test results.

[0030] A third aspect of the present invention provides a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to perform operations according to the instructions to execute the steps of the method.

[0031] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method.

[0032] Compared with the prior art, the beneficial effects of the present invention include at least the following: This invention employs a one-click, wire-free testing interface to safely connect the tester to an operating automatic transfer switch. This interface features isolation and achieves plug-and-play safe access through wire-free testing technology, completely eliminating the impact of testing on the user's power supply and enabling testing without power interruption, thus improving power supply reliability.

[0033] This invention overcomes the limitations of traditional fixed testing by using one-click access and adaptive test signal generation, and employing reinforcement learning to dynamically optimize test cases. It significantly reduces manual operation and decision-making time, can automatically explore the performance boundaries of equipment, discover hidden defects, improve test coverage, and enhance testing efficiency and intelligence.

[0034] The reinforcement learning strategy of this invention determines whether an action meets safety constraints through a safety barrier function. If not, it projects the action to the nearest safe action. Through nonlinear adaptive margin, environmental compensation, and penalized projection, it can achieve refined and dynamic management of safety constraints, significantly improving the safety and adaptability of reinforcement learning action generation in online testing scenarios of power secondary equipment. This ensures that the equipment is not accidentally activated or damaged during testing, while retaining the exploratory capabilities of reinforcement learning, thus enhancing the safety and practicality of online adaptive testing.

[0035] This invention utilizes multi-dimensional features to perform vertical trend analysis on a single machine and horizontal group analysis on multiple machines to obtain test results, enabling early warning of individual performance degradation and diagnosis of group-based defects. Through vertical trend analysis on a single machine, it can quantitatively predict the future degradation trend of device performance, providing maintenance personnel with sufficient early warning time and enabling a shift from reactive repair to condition-based maintenance, thus achieving predictive maintenance. Through horizontal group analysis on multiple machines, it can detect widespread batch and family-based defects early, avoiding large-scale power outages, improving the operational level of power grid equipment, and achieving systemic risk prevention and control.

[0036] This invention's single-machine longitudinal trend analysis integrates multi-dimensional feature data into a comprehensive quantitative indicator, namely the composite health index, by constructing a composite health index. Then, time series modeling is performed on this index to comprehensively and accurately depict the degradation trajectory of equipment performance, enabling performance degradation early warning and enhancing the depth and accuracy of diagnosis.

[0037] This invention employs a multi-machine lateral group analysis framework combining deep autoencoders, deep clustering, and association rule mining. This framework can automatically identify anomalous devices and trace their root causes within large-scale device groups. The deep embedding clustering process utilizes a self-optimization mechanism aided by soft assignment and target distribution. An arctangent kernel function provides a robust distance-probability mapping, while the cluster dispersion (the average weighted distance from samples within a cluster to the cluster center) target distribution guides the cluster centers to converge towards the main cluster. This synergy results in normal devices forming high-density, low-dispersion clusters, while anomalous devices, being far from all clusters, receive extremely low scores, effectively identifying suspected anomalous devices. Attached Figure Description

[0038] Figure 1 This is a flowchart of a test method for a self-starting device without disconnection according to the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0040] Embodiment 1 of this invention proposes a novel test method for automatic transfer switch (ATS) devices that do not require disconnection. The core of this method lies in constructing a closed-loop test system integrating secure access, intelligent excitation, holographic perception, and deep analysis. The overall process is as follows: First, through a dedicated plug-and-play interface, the ATS circuit is securely connected without altering the original wiring or affecting equipment operation. Then, a series of pre-designed test quantities and input signals simulating various operating conditions are injected into the system under test through this interface. Simultaneously, the system synchronously and at high speed collects the device's dynamic response data throughout the entire test process. Finally, through signal processing and artificial intelligence algorithms, features reflecting the deep performance of the device are extracted from massive amounts of data, compared and analyzed, and a comprehensive, quantitative assessment report of the device's health status is provided. Specifically, as shown... Figure 1 As shown, the method includes: Step 1: Use the one-click, wire-free test interface to safely connect the tester to the running automatic transfer switch. More preferably, this step establishes a safe and convenient testing channel, as follows: This invention presents a one-click, wire-free testing interface. The interface consists of a set of high-precision, electrically isolated dedicated adapters that can be directly plugged into the terminal block of an automatic transfer switch (ATS) device.

[0041] The adapter integrates signal isolation and conditioning circuitry, which can safely inject the analog signals generated by the tester into the operating circuit and collect the device's action response signals without damage, thus fundamentally eliminating the safety risks caused by disconnecting and reconnecting wires in traditional testing.

[0042] Step 2: The tester adopts a reinforcement learning strategy based on the security barrier function to inject adaptive test signals into the backup automatic switching device through the interface; More preferably, this step incorporates intelligent, adaptive test stimuli, specifically as follows: To comprehensively verify the performance of the automatic switching device, this invention abandons the fixed test mode and adopts an intelligent test signal generation strategy: ① Initial test case generation: The system has a built-in rule base based on the backup self-supply standard logic, which can automatically generate an initial test case set covering basic functions.

[0043] ② Dynamic coverage assessment: During the test, the system analyzes the logic nodes and loops that have been touched based on the real-time collected response data, and dynamically assesses the comprehensiveness of the test.

[0044] ③ Intelligent Test Case Optimization: When certain complex logic or boundary conditions are found to be uncovered, the system employs artificial intelligence strategies such as reinforcement learning to automatically adjust signal parameters (such as voltage drop rate, fault timing combinations, etc.) and generate new test cases. The complex logic refers to non-standard automatic transfer logic involving multiple interlocking conditions, adaptive switching of backup power supply, or coordination with other protection devices. The boundary conditions refer to the operating state of the automatic transfer device under critical operating conditions or extreme operating environments. Critical operating conditions include slow voltage drops and abnormal frequency fluctuations; extreme operating environments include power switching under heavy loads exceeding predetermined thresholds.

[0045] The reinforcement learning strategy specifically includes the following modeling steps: (1) Constructing a digital twin model: Based on the logical topology and electrical parameters of the backup automatic transfer device, a high-fidelity digital twin model that can simulate its dynamic behavior is established as a training environment for reinforcement learning agents; among them, a general model is established based on the nominal parameters of the equipment before the first test, and individual parameters are identified and the model is updated after the first test using the collected response data; (2) State space encoding: The logical nodes of the standby automatic transfer device and their connection relationships are constructed into a graph structure, and the state feature vectors are extracted from the graph structure using a graph neural network as the state input for reinforcement learning; The graph neural network adopts a message-passing neural network, which updates the embedding representation of each logical node by aggregating neighbor node information, and finally outputs a global state vector through a pooling layer. (3) Action space definition: Define the hybrid action space, including continuous test parameters and discrete test parameters; (4) Reward function design: Design a multi-objective reward function, including at least a coverage reward based on the number of new covered logical nodes, an efficiency penalty based on test efficiency, a security penalty based on security constraints, and an intrinsic reward based on curiosity. Among them, the coverage reward is calculated based on the number of newly triggered logical nodes, the length of the newly covered logical path, and the type of boundary condition touched for the first time. A counting mechanism (such as pseudo-counting) is used to encourage the exploration of rare states. Efficiency penalty: To avoid invalid repeated testing, negative rewards are given to signals that are highly similar to historical tests. The similarity is measured by the cosine distance of the state embedding. Safety penalties: If the generated test signal may exceed the device's tolerance range (e.g., excessive voltage, timing conflicts), a large negative reward is applied; safety constraints are dynamically verified by a barrier function to ensure that the device is not malfunctioning or damaged during online testing; Intrinsic Rewards: Introducing curiosity-driven intrinsic rewards encourages agents to explore states with large prediction errors, thereby discovering unknown defect patterns; The weight coefficients of each reward and punishment are automatically adjusted through multi-objective optimization.

[0046] (5) Agent training: The agent is trained in the digital twin environment using a deep reinforcement learning algorithm suitable for hybrid action spaces to maximize cumulative rewards; The deep reinforcement learning algorithm is a hybrid action space deep deterministic policy gradient algorithm, whose policy network simultaneously outputs the Gaussian distribution mean of continuous parameters and the softmax probability of discrete actions. (6) Safe online migration: The trained agent is deployed to the actual test instrument. Before generating the test signal, the safety of the action is verified by the safety barrier function, and the agent is fine-tuned online using the actual test data. The safety barrier function is constructed based on the electrical safety threshold and timing constraints of the backup automatic switching device, and is used to project the actions that violate the safety constraints into the feasible domain.

[0047] An optional implementation involves using a security barrier function to verify whether an action meets a security threshold. If a violation occurs, the action is projected to the nearest safe action. The specific projection formula is as follows:

[0048] in, For security barrier functions; These represent the voltage command value, action timing width, and injected current amplitude in the test signal, respectively. These are the safety thresholds corresponding to the voltage command value, the action timing width, and the injected current amplitude. For safety constraints, As candidate actions, For adaptive margin, if , indicating an action Violating safety constraints requires solving the projection formula to... Project the nearest safety action within the feasible domain .

[0049] The above formula integrates various heterogeneous safety constraints such as voltage, timing, and current into a single scalar by using the maximum normalized overlimit value. This avoids the difficulties of multi-constraint weighted parameter tuning; and through soft barriers and differentiable projection, the function is continuously differentiable, supports backpropagation, and can be directly embedded into the output layer of a neural network to achieve end-to-end safe action generation, rather than post-processing pruning; it employs adaptive margin. The margin can be dynamically adjusted according to the aging status of the equipment (e.g.) This approach strikes a flexible balance between safety and exploration. Furthermore, it involves only simple algebraic operations and a single low-dimensional projection, making it suitable for real-time online testing scenarios. The aforementioned barrier functions ensure that the device is not accidentally tampered with or damaged during testing, while retaining the exploratory capabilities of reinforcement learning, thus enhancing the safety and practicality of online adaptive testing.

[0050] An optional implementation scheme uses a safety barrier function to determine whether an action meets safety constraints. If it does not, the action that does not meet the safety constraints is projected to the nearest safe action, as follows: For action First, calculate the over-limit values ​​for voltage, timing, and current respectively:

[0051] in, Each represents an action The voltage command value, action timing width, and injected current amplitude are specified in the data. These are the safety thresholds corresponding to the voltage command value, the action timing width, and the injected current amplitude, respectively. Define the safety barrier function value as the Euclidean norm of the above three overlimit values:

[0052] like , then indicates an action If the safety constraints are not met, the following formula will be used to... Project to Recent safety actions :

[0053] in, As candidate actions, for The security barrier function value, For safety margin, The safety margin is the square of the L2 norm. Dynamically adjust based on equipment aging status: , As the initial margin, The aging effect coefficient. This refers to the equipment aging factor (with a value ranging from 0 to 1, obtained by fitting historical test data). In specific implementation, the initial margin ranges from 0.05 to 0.15, and the aging influence coefficient ranges from 0.5 to 2.0. These values ​​can be calibrated through accelerated aging tests on the equipment.

[0054] Considering the original function Focus only on the most severe over-limit value, when multiple constraints are close to but have not yet exceeded their respective thresholds (e.g., the voltage command value is slightly lower than the threshold). Timing width slightly greater than The injected current is slightly less than The original function value is 0, making it impossible to detect potential risks to the equipment under multi-stress coupling. The Euclidean norm of this scheme... It can comprehensively quantify the degree to which multiple dimensions simultaneously approach the safety boundary; even if no individual dimension exceeds the limit, as long as they are jointly approximated, That is, it is greater than 0, so as to give early warning and trigger safety projection, and avoid the test action from accidentally triggering equipment malfunction or damage under multiple critical conditions.

[0055] Euclidean norms satisfy the triangle inequality when both dimensions simultaneously exceed 50% ( )hour, , exceeding the limit by more than 70% in a single dimension This means that the system penalizes minor exceedances of multiple stresses more severely than severe exceedances of a single stress. This aligns with the risk aversion principle in the safety engineering of secondary power equipment—multiple deviations from normal conditions simultaneously are more likely to trigger cascading failures than a single severe deviation. This characteristic allows the tester to proactively avoid combinations of actions that simultaneously approach multiple safety boundaries when generating adaptive test signals, thereby significantly reducing the overall risk of field testing.

[0056] The original maximum function is not differentiable at the boundary, leading to gradient discontinuities during projection, which affects the convergence speed and stability of the optimization. The Euclidean norm is continuously differentiable throughout the entire space, and its gradient... It exists everywhere and is smooth, making the projection formula It can be solved efficiently using gradient descent or the Lagrange multiplier method, reducing computational latency by about 30% in online testing scenarios and meeting the requirements for millisecond-level real-time security verification.

[0057] In addition, this solution introduces aging factors. Dynamically adjust safety margin This automatically tightens the testing safety boundaries of the aging equipment. (Reduce) to avoid situations where equipment performance degradation renders the originally permissible actions under a fixed margin actually dangerous. For example, when relay coil aging leads to increased dispersion in operating timing, rise, The time series width is reduced accordingly, making it more sensitive to exceeding the time series width limit and forcing projection to safer time series parameters. This mechanism achieves safe adaptation throughout the entire lifecycle without sacrificing exploration efficiency.

[0058] An optional implementation scheme uses a safety barrier function to determine whether an action meets safety constraints. If it does not, the action that does not meet the safety constraints is projected to the nearest safe action, as follows: (1) Regarding the action Calculate the out-of-limit values ​​for voltage, timing, and current respectively:

[0059] in, Each represents an action The voltage command value, action timing width, and injected current amplitude are specified in the data. These are the corresponding security thresholds; (2) Define the safety barrier function value as the weighted Euclidean norm of the above three overlimit values:

[0060] Among them, the weighting coefficient These respectively reflect the importance of voltage, timing, and current constraints to equipment safety, and satisfy... The weighting coefficients are pre-calibrated or dynamically adjusted online based on the equipment type and operating conditions. (3) Define adaptive safety margin for:

[0061] in: For the basic safety margin, a value range of 0.07 to 0.12 is recommended. This range can be determined based on the maximum measured fluctuation range of each parameter (voltage, timing, current) relative to the safety threshold during factory testing or the initial health test. For example, if the maximum deviation of the measured voltage command for a healthy device is 5%, the maximum deviation of the timing command is 3%, and the maximum deviation of the current is 4%, then the maximum value of 5% should be taken, followed by a safety margin (e.g., multiplied by 1.5 to 2). Take a value of 0.075 to 0.10.

[0062] The comprehensive health score of the equipment for the current testing cycle is calculated, with a value range of [0,1] (1 indicates health and 0 indicates failure). The specific score can be measured by the composite health index of the previous testing cycle. The time interval since the last test (in years); The preset sensitivity coefficient ( ), controlling the rate of influence of health degradation and temporal decay on margin; Product term The confidence level used to quantify the current health status of the equipment is calculated by combining the equipment's overall health score S with a time decay factor. Multiplication, health equipment ( ≈1) and recent tests ( When the product is small, a large product indicates a high confidence level, and the safety margin should be reduced to ensure safety; aging equipment ( Small) or long-term untested ( When the product approaches 0, the confidence level is low, and the safety margin reverts to the baseline value to allow for necessary exploration. The tanh function is used to smoothly map the joint effect of health status and time decay on test risk to the [0,1] interval to achieve nonlinear adaptive adjustment of the safety margin; Environmental factors are defined as follows: ,in These represent the current ambient temperature (°C) and relative humidity (%). The rated operating environment for the equipment (typical values: 25°C, 50%). For the environmental impact weighting coefficient, a value of 0.3 to 0.7 is recommended. The value can be determined based on tests of the equipment's sensitivity to temperature and humidity. Environmental factor adjustment coefficient ( ); (4) If , then indicates an action If the safety constraints are not met, the following weighted projection formula will be used to... Project to the most recent safety action :

[0063] in, For the penalty factor ( This is used to minimize the adjustment range of actions while satisfying safety constraints; when there is no analytical solution, the Lagrange multiplier method or gradient projection iteration is used to solve it.

[0064] This solution takes into account the different sensitivities of various constraints (voltage, timing, current) to equipment safety, and introduces weighting coefficients. Different configurations can be made based on the type of equipment (e.g., electromagnetic relays are sensitive to current, while solid-state relays are sensitive to timing), making safety assessments more closely aligned with physical realities. For example, for aging relays with highly dispersed operating timing, the configuration can be increased... This causes its timing to exceed the limit. The timing width is prioritized during projection to avoid accidental changes.

[0065] And taking into account traditional margin linear adjustment (such as This approach fails to reflect the typical pattern of equipment health degradation, which is slow in the early stages and accelerates in the later stages. This solution adopts... The function takes a health score as its input. With exponential decay term The product. When the equipment is in good health ( And the test interval is relatively short. Hour, Input close to 0, margin close to When health scores decline and have not been tested for an extended period, input increases. It rapidly approaches 1, with the margin adaptively shrinking to near 0. This feature allows the tester to automatically tighten the safety boundary when the equipment is severely aged or has not been inspected for a long time, avoiding the risk of malfunction due to a fixed margin, while maintaining sufficient exploration space when not necessary.

[0066] Furthermore, high temperature and humidity accelerate insulation aging, contact oxidation, and coil parameter drift, reducing the actual safety threshold of the equipment. This solution introduces... Quantify the degree to which the environment deviates from the rated value, and use coefficients (Generally ≤0.2) Make minor adjustments to the margin. For example, when the ambient temperature rises to 40℃ (deviation from 15℃) and the humidity rises to 80% (deviation from 30%), If taken ,but , The safety margin has been increased by 6%. This allows the tester to appropriately increase its safety margin under high temperature and humidity conditions on site, avoiding excessive conservatism due to environmental factors that could affect test efficiency, while still ensuring intrinsic safety.

[0067] Traditional projection formulas only minimize the magnitude of motion adjustment ( The previous approach did not consider the inherent risks of the projected action. This solution adds [something] to the objective function. This encourages projected actions to have lower barrier function values ​​(i.e., further away from the safety boundary), thus avoiding projecting actions into high-risk areas close to the boundary. Parameters It strikes a balance between the goals of a small adjustment range and a large safety margin in the resulting action. When equipment is severely aged, the margin can be increased. This makes the projection results more conservative; when the device is in good working order, it reduces... This preserves the exploratory nature of reinforcement learning.

[0068] Therefore, this solution, through nonlinear adaptive margin, environmental compensation, and penalized projection, can achieve refined and dynamic management of safety constraints, significantly improving the safety and adaptability of reinforcement learning action generation in online testing scenarios of power secondary equipment.

[0069] ④ Form an optimized sequence: Integrate the initial test cases with the newly added optimized test cases to form a standardized test signal sequence with high adaptability and completeness.

[0070] Step 3: While injecting the test signal, the action response signal of the backup automatic transfer device is synchronously collected through the interface to obtain full-link multi-physical quantity data; More preferably, this step achieves end-to-end, multi-physical quantity data acquisition, as detailed below: Simultaneously with the stimulus injection, the system synchronously initiates high-speed data acquisition through the same interface. The action response signal is synchronously acquired to obtain end-to-end multi-physical quantity data, which includes: ① Electrical quantity signals: High-precision acquisition of transient waveforms of voltage and current.

[0071] ② Timing signal: Accurately records the action time of each relay contact, with an accuracy down to the microsecond level.

[0072] ③ Mechanical vibration signal: The weak mechanical vibration waveform generated at the moment of relay action is collected by a high-sensitivity vibration sensor integrated into the interface or externally (placed outside the automatic transfer device).

[0073] In this embodiment, the acquisition of electrical quantity signals is divided into two levels: Primary system electrical quantities: The voltage waveform of the bus voltage (such as the three-phase voltages A, B, and C) monitored by the automatic transfer switch is acquired through the voltage sensing module in the interface. The current waveforms of the incoming line and the sectionalizing switch are acquired through the current clamp or the secondary circuit of the CT inside the device to verify the accuracy of the device's perception of the grid operation status.

[0074] Internal secondary circuit electrical quantities: Through the high-bandwidth, high-precision current probe and differential voltage probe integrated in the interface, the drive current waveform of the relay coil inside the backup automatic transfer device and the voltage waveform across the contacts of the output relay are collected to evaluate the health status of the drive circuit (such as whether there is an inter-turn short circuit in the coil and whether the drive power supply is normal) and the contact performance (such as contact resistance and arcing).

[0075] Step 4: Extract features from the full-link multi-physical quantity data to obtain multi-dimensional features characterizing the deep performance of the backup automatic switching device; More preferably, this step extracts multidimensional features characterizing the deep performance of the device, specifically as follows: Faced with the massive amounts of raw data collected, various algorithms are used for in-depth analysis to extract fingerprint features that reflect the essential state of the device: ① Timing characteristics: Extract the total delay of relay action, the dispersion (standard deviation) of action timing, and the ratio of action time of different output circuits from the switch signals to quantify the logic execution speed and stability of the equipment.

[0076] ② Electrical characteristics: Extract dynamic current characteristics and transient voltage characteristics based on electrical quantity signals; The current dynamic characteristics include the maximum value of the relay coil current change rate, the peak value of the relay coil drive current, and the amount of charge during the relay operation process; the voltage transient characteristics include the primary voltage drop rate, sag depth, recovery time, and the turn-off overvoltage amplitude and closing voltage drop of the secondary contact voltage, etc., to assess the health status of the drive circuit.

[0077] ③ Spectral characteristics: Perform fast Fourier transform on the vibration signal to extract the energy distribution and dominant frequency components of different frequency bands, thereby determining whether the mechanical structure is loose or aging.

[0078] Step 5: Perform multi-dimensional fusion analysis and evaluation on the multi-dimensional features to obtain test results.

[0079] More preferably, this step performs multi-dimensional, intelligent fusion analysis and evaluation, such as comparing the extracted multi-dimensional features with a pre-established benchmark model to form a final diagnostic conclusion. Specifically, it also includes intelligent fusion analysis and evaluation in the following dimensions: 1. Single-machine longitudinal trend analysis: The characteristic data of this test is compared with the historical data of the equipment to establish time series prediction models such as autoregressive model and grey model, so as to depict the performance degradation trajectory of the equipment over time and issue performance degradation warnings accordingly.

[0080] Preferably, performance degradation is predicted based on a composite health index and a grey Markov model: Step ①: Multidimensional Feature Optimization and Composite Health Index Calculation 1) Select key indicator features that are strongly correlated with the health status of the device from the multidimensional features extracted from each test and perform standardization processing to obtain standardized feature data; 2) Calculate the covariance matrix of the standardized feature data and perform eigenvalue decomposition on the covariance matrix to obtain eigenvectors and eigenvalues. Take the eigenvector corresponding to the largest eigenvalue, and each component of the eigenvector corresponds to the loading coefficient of each feature in the first principal component. 3) Calculate the composite health index for each test based on the standardized feature data and load coefficients: Among them, the The composite health index of this test The calculation formula is:

[0081] in For the first The feature loading coefficients in the first principal component For the first Standardized values ​​of each feature and These are the projected values ​​under healthy conditions and failure thresholds, respectively; m is the number of key indicator features. Regarding the loading coefficient of the first principal component The calculation process is illustrated below: Collect multidimensional feature data from five consecutive tests (or group tests of healthy devices in the same batch) of this model of automatic switching device under factory health conditions. Each sample contains... Key features (e.g., total motion delay) Time series standard deviation Maximum rate of change of current Peak drive current Action charge Contact closing voltage drop Vibration dominant frequency High-frequency energy ratio ). These 5 samples 3D feature data constitutes a matrix .

[0082] First of all Standardization is performed by subtracting the mean from each column (each feature) and dividing by the standard deviation to obtain the standardized matrix. Then calculate. covariance matrix (8×8 matrix). For Perform eigenvalue decomposition to obtain eigenvalues. and its corresponding unit eigenvector Take the largest eigenvalue. Corresponding feature vector Then the first Load factor of each feature For example, suppose the largest eigenvalue... Corresponding feature vector middle, , , ..., =0.27, then the load factor of the first feature (total action delay) is 0.27. The loading factor of the second feature (time series standard deviation) ..., the load factor of the 8th characteristic (high-frequency energy ratio) =0.27. These load factors ( to The value of reflects the direction of contribution of each feature to the overall variance under healthy conditions, and this set of values ​​will be used in subsequent tests. to The value is used to calculate the health index.

[0083] Regarding projection extrema and Examples of methods for determining this are as follows: Take the standardized feature vectors of multiple samples (at least 5) under the above-mentioned health conditions. Calculate the projection value for each sample. A set of projection values ​​is obtained. ( (Number of healthy samples). The arithmetic mean of the projected values ​​of this group is taken as... ,Right now For the projected value under the failure threshold It can be determined using one of the following two methods: Method 1 (Based on equipment design life or industry standards): Consult the technical manual or industry maintenance procedures for this model of automatic transfer switch to determine the failure thresholds of key characteristics such as total operating delay and contact resistance (e.g., total operating delay > 55ms, closing voltage drop > 50mV). Then, collect characteristic data of the equipment under simulated aging or extreme operating conditions (or extract from historical fault records), calculate its projected value using the same method, and take the average of multiple measurements as the final value. .

[0084] Method 2 (based on statistical inference): If there is no actual failure data, the mean of the projected values ​​of healthy samples can be subtracted from three standard deviations as the statistical inference. ,Right now ,in This represents the standard deviation of the projected values ​​for healthy samples. This method assumes that the projected values ​​in a healthy state follow a normal distribution, and values ​​below three times the standard deviation of the mean are considered statistically significant anomalies / failure boundaries.

[0085] Specific numerical examples: Suppose we obtain a vector of loading coefficients for 8 features from 5 health devices (or 5 historical health tests on the same device). The projected values ​​for each healthy sample are calculated to be 2.35, 2.40, 2.38, 2.42, and 2.39, respectively. Take the standard deviation. According to the 3-times rule, If in a subsequent test, the standardized eigenvector projection value of a certain device... Then its health index This indicates that the equipment has severely degraded and is close to failure.

[0086] 2) Normalize the composite health index from each test to The interval is used to obtain the health index sequence. ; Step 2: Health Index Prediction Based on Grey Model 1) Health index series The accumulated sequence is obtained by performing an accumulation once. ; 2) Based on cumulative sequence Establish the differential equation of the grey model: in, and For development coefficient and grey effect quantity; 3) Solving for the development coefficient using the least squares method and gray action amount The predicted values ​​of the cumulative sequence are obtained, and then the predicted values ​​of the health index are obtained by subtracting from the cumulative sequence. ; Step 3: Correct the predicted health index value through residual calculation. 1) Calculate the residual series based on the composite health index and the predicted health index values. ; 2) Classified according to the size of the residuals Given a set of states, count the number of transitions between each state and calculate the state transition probability matrix. ; in From state Transition to state The state transition probability, From state Transition to state Number of times, In a state The total number of times.

[0087] 3) Regarding the future In this test, the predicted health index value was first obtained using the grey model. Then according to the first Given the state and transition probability matrix of the residual, determine the most likely residual state and take the mean of the residual for that state. As a correction factor, the final predicted health index value is obtained:

[0088] Step 4: Estimation of the probability distribution of remaining lifetime 1) Set equipment failure threshold Using the modified grey Markov model ( Extrapolate future health index trajectories; 2) Monte Carlo simulation is used to generate multiple possible degradation paths: In each simulation, starting from the current state, a future residual state is randomly selected based on the state transition probability, and random noise is added to obtain a predicted path; When the path is first lower At that time, record the remaining lifetime of the path; repeat Next, the probability distribution of remaining lifespan is obtained, and the mean, median, and 90% confidence interval are calculated.

[0089] 2. Multi-machine horizontal group analysis: Aggregate feature data from multiple devices of the same model and construct a group performance distribution map using unsupervised machine learning algorithms (such as DBSCAN clustering). By identifying anomalous clusters that deviate from the mainstream group and combining them with metadata such as the device's production batch and operating environment, it is possible to accurately locate family-related defects caused by common reasons such as design and process.

[0090] Preferably, familial defect diagnosis is performed based on deep embedding clustering and association rule mining: Step ①: Feature embedding and dimensionality reduction 1) For the same model Feature vector of Taiwan-made automatic switching device Standardization processing is required; 2) Construct a system containing an encoder and decoder A deep autoencoder network is used, and the network is trained with the goal of minimizing the reconstruction error. 3) After training is complete, retrieve the encoder output. As the embedding vector of the device ; Step 2: Deep Embedding Clustering An optional implementation involves performing deep embedding clustering on the embedding vectors of each backup self-starting device, dividing the backup self-starting devices into several clusters, and marking suspected abnormal devices. Specifically, this includes: 1) Calculate the embedding vector for each device. Belongs to the Cluster centers soft assignment probability :

[0091] in denoted as t, representing the degrees of freedom of the t-distribution. 2) Based on soft assignment probability Construct target distribution probability :

[0092] in Let be the soft assignment probability of the embedding vector of the i-th standby self-connecting device belonging to the j′-th cluster center; 3) By minimizing the soft assignment probability With target distribution probability The KL divergence is used as the loss function. The encoder network parameters and cluster center positions are fine-tuned alternately until the loss function converges.

[0093] 4) After clustering is completed, the device is automatically divided into several clusters and suspected abnormal devices are marked. In addition, the distance between the center of each cluster and the pre-stored health baseline vector can be calculated. If the distance exceeds the threshold, the entire cluster is marked as potentially systematically degraded.

[0094] Compared to traditional clustering (such as K-means), DEC can adaptively learn feature representations and discover more complex anomaly patterns.

[0095] Traditional data analysis methods (such as K-Means) are hard clustering methods, meaning a device either belongs to class A or class B. The deep embedding clustering method used in this invention introduces a soft assignment and target distribution-assisted self-optimization mechanism: soft assignment implements a t-distribution-based similarity measure, using the t-distribution (i.e., in the formula...) (Usually set to 1, which is the Cauchy distribution) to measure the embedding vector. To the cluster center The similarity, rather than the traditional Gaussian distribution or Euclidean distance, makes the characteristics of a device ( ) and all cluster centers ( Even when devices are far away (potentially due to a specific defect), a relatively ambiguous probability is still assigned, unlike a Gaussian distribution where the probability value rapidly approaches zero, leading to information loss. This allows the algorithm to better handle anomalous devices within the group, rather than arbitrarily assigning them to an inappropriate cluster, demonstrating strong robustness to outliers. Furthermore, the soft-assignment output... It is a probability value between 0 and 1, representing the probability of the first... What percentage "possibility" does the device have that it belongs to the first category? There are several clusters. The sum of the probabilities of all clusters is 1. The probabilistic attribution expression is output through soft assignment, preserving the uncertainty of classification. For example, a device might belong to cluster A with a probability of 0.6 and to cluster B with a probability of 0.4, which itself suggests that it may be in a sub-optimal or marginal state, providing richer information for subsequent analysis. Furthermore, this invention does not directly calculate the auxiliary target distribution from the data. Instead, it is determined by the current soft allocation result. Construct an idealized goal, through the By performing a sum-of-squares normalization operation, high-probability, high-confidence assignments are intentionally amplified, while low-probability, ambiguous assignments are suppressed. The training objective achieved through iterative self-optimization of the algorithm is to minimize... (Current forecast) and The KL divergence between (ideal targets) can guide feature embeddings and cluster centers to optimize in a clearer and more compact direction. This mechanism effectively widens the distance between different clusters and tightens the samples within the same cluster. Ultimately, normal devices will cluster together with a very high probability (e.g., 0.99), forming very dense clusters; while abnormal devices will be clearly separated, forming isolated small clusters or points, facilitating subsequent association rule mining. This is more intelligent and effective than parameter tuning in traditional clustering methods (such as DBSCAN). Based on this mechanism, familial defects can be accurately identified, hidden sub-health patterns can be discovered, and high-quality input can be provided for causal inference.

[0096] An optional implementation involves performing deep embedding clustering on the embedding vectors of each backup self-starting device, dividing the backup self-starting devices into several clusters, and marking suspected abnormal devices. Specifically, this includes: First, calculate the embedding vector for each standby automatic transfer device. Belongs to the Cluster centers soft assignment probability Using the arctangent kernel function:

[0097] in, For Euclidean distance, This is the global scale parameter (which can be preset to the median of the pairwise distances between all embedded vectors). The function range is Therefore, the molecule takes the value when the distance is zero. When the distance is infinite, it approaches 0, and it is monotonically decreasing and smooth. Then, based on the current soft assignment probability Calculate the scatter of each cluster. :

[0098] in, For the first The number of soft samples in each cluster (i.e., the sum of the soft assignment probabilities of all samples belonging to that cluster, used to normalize the weighted distance within the cluster); This reflects the average weighted distance from samples within a cluster to the cluster center; Next, an adaptive target distribution is constructed. :

[0099] in, The compactness penalty factor for clusters: the looser the cluster ( The larger ( The closer to 1, the higher the probability of the target; the denser the cluster ( (close to 0) When the value is close to 0, the target probability is suppressed. This design guides the optimization process to prioritize tightening loose clusters rather than strengthening already tight clusters; By minimizing the soft assignment probability With target distribution The KL divergence is the loss function:

[0100] Alternately fine-tune the encoder network parameters and cluster center positions until the loss function converges; after clustering is complete, As an anomaly score, devices with an anomaly score greater than a preset threshold (such as 0.2) are marked as suspected abnormal devices.

[0101] Compared to the standard t-distribution kernel, which decays power-law over large distances, this scheme uses an arctangent kernel that rapidly approaches zero (saturates) after a distance exceeding approximately 3σ. This results in anomalous devices far from all cluster centers receiving a nearly uniform, extremely low-probability assignment, avoiding the illusion of being forcibly assigned to a particular cluster. Furthermore, the arctangent kernel exhibits an approximately linear decreasing slope in the medium-distance region (0.5σ~2σ), making it sensitive to small distance changes and beneficial for distinguishing boundary samples. The kernel function calculation involves only one division and arctangent operation, eliminating the need for exponentiation, thus improving efficiency in embedded implementations.

[0102] Furthermore, the target distribution of the standard DEC relies solely on the squared normalization of the soft assignment probability, essentially amplifying the weights of high-confidence samples without distinguishing the degree of dispersion across different clusters. This scheme introduces cluster dispersion. Constructing factors Loose clusters ( The average distance from large samples to cluster centers is large. As the value approaches 1, the relative weight of that cluster in the target distribution increases, forcing the cluster centers to move closer to the samples to reduce the weight of the cluster. ; compact cluster ( The weights of smaller clusters are suppressed to prevent converged clusters from being over-optimized and distorted. This mechanism prioritizes solving non-converged loose clusters during the clustering process, accelerating overall convergence and reducing overfitting.

[0103] Furthermore, the arccut kernel provides a robust distance-probability mapping, and the cluster dispersion target distribution guides the cluster centers to tighten towards the main cluster. The synergy of these two factors enables normal equipment to form high-density, low-density clusters. Clusters, where abnormal devices receive extremely low latency due to their distance from all clusters. Experiments show that when the number of devices is 50-200, this scheme improves the F1 score of anomaly detection by about 20% compared with the standard DEC, and its sensitivity to the scale parameter σ is lower than that of the t distribution to the degree of freedom α, making engineering parameter tuning simpler.

[0104] 3. Digital twin parameter identification: Constructing a nonlinear state-space model that can describe the dynamic behavior of the automatic switching device is achieved through the following process: By utilizing the collected stimulus-response data and employing a maximum a posteriori probability estimation algorithm, the equivalent parameter set (such as equivalent time constant, pure delay, etc.) that best characterizes the current true physical properties of the device is obtained. These parameters can serve as the device's "digital fingerprint," driving a high-fidelity digital twin model to achieve accurate simulation and lifespan prediction of the device's future state.

[0105] The excitation-response data includes the injected test signal waveform and the synchronously acquired raw data of multiple physical quantities across the entire link, or features extracted from the raw data. However, using the raw data is preferred to improve accuracy.

[0106] In a non-restricted example, the nonlinear state-space model constructed to describe the dynamic behavior of the automatic switchgear is as follows: State equation: x(k+1)=f(x(k),u(k),θ)+w(k) Observation equation: y(k) = g(x(k), u(k), θ) + v(k) Where k represents the discrete-time sampling points; x(k) is a vector of state variables. For example, for a relay coil, x(k) can be defined as [i(k), d(k)]. T , where i(k) is the coil current and d(k) is the core displacement; u(k) is the input variable, i.e. the waveform of the injected adaptive test signal; y(k) represents the observation response, i.e., the collected multi-physical quantity data across the entire link; θ is the set of equivalent parameters to be identified, for example, θ=[R,L,k_spring] T Where R is the equivalent resistance of the coil, L is the equivalent inductance of the coil, and k_spring is the contact spring constant; f(·) and g(·) are functions based on physical laws. For example, the dynamic function of coil current is described by L·di / dt+R·i=u(t), and the state equation is obtained after discretization. w(k) and v(k) are the process noise and observation noise, respectively; Prior distribution setting: Assume that the parameters (such as R, L, k_spring, etc.) are independent and follow a normal distribution. For example, R N( m R , s R 2 ),in m R Take the factory-specified value of the health equipment. s R 2 Use empirical tolerance values.

[0107] Solution Algorithm: The algorithm combines Maximum A posteriori (MAP) estimation with Expectation-Maximization (EM) algorithm. In the current parameter estimation θ (t) The smooth distribution of the state variable x(k) is calculated using algorithms such as extended Kalman filter or particle filter. Maximize the expected joint log-posterior probability with respect to the state distribution, and update the parameter estimate θ. (t+1) .

[0108] θ obtained after iterative convergence MAP This is the digital fingerprint of the automatic switching device at the current moment, which can be used to drive the digital twin model to simulate future states and predict remaining lifespan.

[0109] Based on the above solution, the present invention achieves: Plug-and-play secure access: Uninterrupted testing is achieved through a one-click interface with isolation capabilities, which forms the basis for all subsequent functions.

[0110] Adaptive intelligent test generation: It uses reinforcement learning to dynamically optimize test cases, breaking through the limitations of traditional fixed tests and significantly improving test coverage.

[0111] Full-link multi-physical quantity feature extraction: By integrating electrical quantities, timing, and vibration signals, a multi-dimensional feature system characterizing the deep health status of equipment is constructed.

[0112] The fusion analysis of single-machine trends and group risks: By combining time series forecasting and unsupervised clustering, it has achieved a leap from early warning of individual performance degradation to diagnosis of group family defects.

[0113] Deep parameter identification based on digital twins: By constructing mathematical models and identifying their core parameters, a quantifiable digital image of the device is created, making accurate lifespan prediction possible.

[0114] Embodiment 2 of the present invention provides a test device for a non-disconnectable automatic switching device, comprising: A one-click, wire-free test interface is used to safely connect the tester to an operating automatic transfer switch. The tester is used to inject adaptive test signals into the automatic transfer switch (ATS) device through the interface; while injecting the test signals, it synchronously collects the action response signals of the ATS device through the interface to obtain full-link multi-physical quantity data; it extracts features from the full-link multi-physical quantity data to obtain multi-dimensional features characterizing the deep performance of the ATS device; it performs multi-dimensional fusion analysis and evaluation on the multi-dimensional features to obtain test results, which can be analyzed locally or uploaded to the cloud in specific applications.

[0115] Example 3: In-depth diagnostics of a single-station automatic transfer switch At a 110kV substation, maintenance personnel need to conduct a routine inspection of an RCS-965 automatic transfer switch that is in operation.

[0116] Step 1: First, the operator holds a portable tester (model: PDT-100), inserts one end of the dedicated test cable into the tester's interface, and the other end into the terminal block on the automatic projection cabinet labeled "Trip Output," "Close Output," and "Voltage Input." During insertion, the guide groove of the plug should align with the guide rail of the terminal block. Once fully inserted, the green indicator light on the tester's panel will illuminate, indicating a reliable connection and electrical isolation.

[0117] Step 2: Subsequently, the operator selects the "10kV Sectional Backup Transfer" test item on the tester's touchscreen. The tester internally stores the standard test sequence for this item: first, it outputs a three-phase 57.7V voltage to simulate normal operation; after 2 seconds, it reduces the voltage of one section of the bus to 0V while simultaneously sending a simulated switch occupancy signal. Furthermore, the tester has pre-set pass / fail criteria for this test item, including threshold ranges for key indicators such as trip output action time (e.g., 45-55 ms) and closing interval time (e.g., 100-110 ms), for automatic judgment after the test is completed. This sequence is injected into the automatic transfer switch via the connection line.

[0118] In this embodiment, complex logic refers to advanced backup functions beyond simple undervoltage backup, such as, but not limited to: Complex interlocking logic: When multiple conditions such as TV (voltage transformer) disconnection, manual tripping, protection interlocking, abnormal switch position, and system current are detected, the backup function is automatically interlocked or the backup mode is switched.

[0119] Adaptive backup mode: It can automatically select the correct backup mode (such as explicit backup or implicit backup) according to the current operation mode of the power grid (such as parallel operation or split operation).

[0120] When used in conjunction with other automatic devices such as reclosing and low-frequency load shedding, it can correctly execute complex timing and logic judgments.

[0121] Boundary conditions refer to test conditions that put the automatic switching unit (AS / RS) into a critical or extreme operating state, such as, but not limited to: Critical operating voltage: The accuracy and timing of the device's operation when the test bus voltage slowly drops to the operating threshold (e.g., 30% of the rated voltage).

[0122] Extreme fault timing: Simulates the automatic transfer behavior of backup under different action sequences of main protection, near backup, and far backup protection.

[0123] Multiple disturbances: Verify the reliability of the automatic transfer switch logic under the condition that the system has harmonics and frequency fluctuations.

[0124] High load switching: Simulates the ability to withstand large inrush currents and the stability of the system when a backup power supply is put into operation.

[0125] Additionally, optionally, while injecting the sequence, the intelligent analysis module inside the tester collects the action response signals of the automatic transfer switch (ATS) in real time. By comparing with the built-in ATS logic rule library, it automatically identifies the logic nodes (such as bus undervoltage detection, trip output, closing output, and blocking conditions) and loops currently touched by the test. The system dynamically evaluates the test coverage, for example, checking whether all possible fault types and ATS modes are covered. When it finds that boundary conditions such as adaptive switching of ATS modes and complex blocking logic are not covered, the tester automatically uses reinforcement learning strategies to adjust test parameters, such as changing the voltage drop rate and simulating fault timing combinations, to generate new optimized test cases. Subsequently, the tester integrates the initial standard test sequence with these newly generated optimized test cases to form a set of adaptive and complete test signal sequences for the device, and continues to inject them into the ATS for in-depth testing.

[0126] However, for RCS-965 devices with multiple interlocking logics or complex operating modes, simply executing the preset standard sequence may not cover all potential risks. Therefore, this tester activates an adaptive test signal generation engine while injecting the aforementioned basic sequence.

[0127] The engine first dynamically marks the triggered paths in the local logic topology diagram based on real-time collected device response data (such as the action timing of trip and close output relays and level changes of internal logic nodes). Once it finds that complex logic—for example, during simulated bus undervoltage, the device should determine whether the blocking conditions such as TV disconnection and manual trip signals have been correctly identified; or the adaptive switching rules between "explicit standby" and "implicit standby" in the backup mode—has not yet been traversed in this basic test, the engine immediately enters the boundary exploration phase.

[0128] At this point, the system invokes the built-in reinforcement learning policy network. This network, aiming to maximize test coverage, automatically decides and generates new combinations of test parameters: for example, adjusting the voltage drop rate from a sudden drop to 0V to a slow decrease at a rate of 5V / s to test the device's stability under critical operating voltage; or inserting a brief self-recovery process into the fault sequence to simulate a mixed condition of transient and permanent faults. These dynamically generated optimized test cases are injected into the loop in real time, and the device's response is observed.

[0129] Finally, the tester intelligently integrates and deduplicates the initial standard test sequence with these boundary test cases explored by the reinforcement learning strategy, forming a set of highly adaptive and complete standardized test signal sequences for the RCS-965 device, ensuring that all core logic nodes and boundary conditions are effectively reached.

[0130] In this preferred embodiment, the adaptive test signal generation employs a deep reinforcement learning framework based on digital twins and graph neural networks, and its core modeling and workflow are as follows: 1) Construct a digital twin model of the automatic switching device. First, a high-fidelity digital twin model is constructed using the factory parameters, logic diagram, historical test data, and real-time operating data of the automatic transfer switch (ATS). This model not only accurately reproduces the electrical characteristics of the device (such as voltage and current response), but also includes the topological connections between internal relay logic nodes (stored in a graph structure) and the dynamic timing behavior of each node. The digital twin model can simulate the device response under arbitrary test signal inputs in an offline environment, providing a safe and efficient training environment for reinforcement learning agents.

[0131] 2) State-space modeling: Feature extraction based on graph neural networks The logical topology of the automatic transfer switch is defined as a graph. , where nodes Representing measurable units such as logic gates, relay coils, and output contacts, etc. This indicates the signal transmission relationship between nodes. During the test, each node is assigned real-time status attributes (such as voltage value, contact continuity, whether it has been covered by the test, etc.).

[0132] Using Graph Neural Networks (GNNs) to analyze graphs The code is encoded, and neighbor node information is aggregated through a multi-layer message passing mechanism. Finally, a high-dimensional vector is output as the current test state. The embedded representation effectively captures the impact of uncovered logical paths, key nodes, and topology on testing requirements.

[0133] 3) Motion space design: Multidimensional continuous-discrete hybrid motion action The parameter combination defined as the next test signal includes: Continuous parameter: voltage sag rate ( ), duration of the fault ( ), frequency offset ( )wait; Discrete parameters: fault type (single-phase grounding, two-phase short circuit, three-phase short circuit), fault phase (A, B, C), backup mode selection (mandatory switching between visible and invisible backup), etc.

[0134] A hybrid action space model is adopted, in which continuous parameters are output through a parameterized policy network, and discrete parameters are selected through a softmax classifier.

[0135] 4) Reward function design: Multi-target coverage guidance and safety penalty award By integrating multiple objectives, the intelligent agent is guided to efficiently explore uncovered complex logic and boundary conditions: Coverage Bonus The calculation is based on the number of newly triggered logical nodes, the length of the newly covered logical path, and the type of the first boundary condition encountered. A counting mechanism (such as pseudo-counting) is used to encourage the exploration of rare states. In this embodiment, coverage rewards are... for:

[0136] in, This represents the logic node that is newly triggered by this test action (i.e., has never been covered by the test before); For the newly triggered set of logical nodes; The degree of the node; These are the weighting coefficients; For indicator functions, if The function value is 1 for the corresponding boundary condition; otherwise, it is 0.

[0137] This reward function incorporates a weighted average based on the inverse of node degree, with rare paths (low-degree nodes) triggering higher rewards; additional bonuses are awarded for first-time encounters with boundary conditions. This guides the agent to prioritize exploring less frequently accessed logic nodes, avoiding repeated testing of main paths and improving test completeness.

[0138] efficiency penalty To avoid invalid repeated testing, signals highly similar to historical tests are given a negative reward, with similarity measured by the cosine distance of the state embedding. In this embodiment, an efficiency penalty is applied. for:

[0139] in, For adaptive penalty intensity; Current test status Compared with historical test status The similarity between them; This is the bandwidth parameter.

[0140] The penalty function is based on the Gaussian similarity kernel function of state embedding. It continuously penalizes highly repetitive test states, avoiding gradient vanishing through soft penalty instead of hard truncation. It also suppresses invalid redundancy by adaptively adjusting the penalty strength.

[0141] Safety penalties If the generated test signal may exceed the device's tolerance range (e.g., excessive voltage, timing conflicts), a larger negative reward is applied; safety constraints are dynamically verified by a barrier function to ensure that the device is not malfunctioning or damaged during online testing. In this embodiment, the safety penalty... for:

[0142] in, For the security barrier function, the threshold For safety margin; This is the safety penalty coefficient.

[0143] The penalty function uses a barrier function. The degree of action exceeding the limit is quantified. Gradient information is provided through continuous punishment rather than discrete prohibition, making the safety punishment function continuously differentiable and able to be projected online to the safety domain, thus ensuring physical safety.

[0144] Intrinsic rewards This invention introduces curiosity-driven intrinsic rewards to encourage the agent to explore states with large prediction errors, thereby discovering unknown defect patterns. In this embodiment, the intrinsic reward... for:

[0145] in, To perform actions in an intelligent agent Then, the environment actually returns the true state of the next moment. The state at the next moment predicted by the dynamic environment model. It is the square of the L2 norm.

[0146] This intrinsic reward function is based on curiosity-driven prediction error rewards. The dynamic environment model predicts the next state and can automatically focus on unknown dynamics that the model has not mastered, promoting the exploration of potential defect patterns and avoiding artificially designed exploration noise.

[0147] Total Rewards The weighting coefficients can be automatically adjusted through multi-objective optimization.

[0148] 5) Reinforcement learning algorithms and training strategies An extended version of Deep Deterministic Policy Gradient (DDPG)—Hybrid Action DDPG—is employed to handle both continuous and discrete actions. Policy Network Output continuous parameter mean and discrete action logits, value network Estimate the value of actions. During training, the agent first undergoes offline training in a digital twin environment, accumulating an experience replay pool through interaction with the environment, and gradually optimizing the strategy to maximize the cumulative discount reward.

[0149] 6) Secure online adaptive migration After training convergence, the policy network is deployed to a real test instrument. In actual testing, before each test signal is generated, a safety barrier function is used to verify whether the action meets the safety threshold. If a violation occurs, the action is projected to the most recent safe action. Simultaneously, response data from the real device is collected, and online fine-tuning mechanisms (such as using small-batch experience replay) are used to update the model parameters. This enables transfer learning from the digital twin to the physical entity, allowing the agent to continuously adapt to device aging or individual differences.

[0150] 7) Adaptive generation of complete test sequences Through the reinforcement learning process described above, the system dynamically generates a series of optimized test cases and integrates them with the initial standard test cases. The final output test signal sequence has completeness (covering all preset logic nodes and boundary conditions), adaptability (adjusting for specific device states), and security (always operating within the allowed range), greatly improving test depth and efficiency.

[0151] Step 3: During the adaptive test signal sequence injection process, the high-precision synchronous acquisition module inside the tester records the multi-physical quantity data of the entire link in real time at a sampling rate of 1 MHz, specifically including: Electrical signals: Through the isolation voltage probe and high-frequency current clamp in the interface, the bus voltage and incoming current transient waveforms of the primary system, as well as the drive current waveform of the relay coil inside the automatic transfer switch and the voltage waveform at both ends of the output contact are synchronously acquired. Switching timing signals: Through optocoupler isolation input, the action time of each output relay (such as trip relay, closing relay) contact is accurately recorded with a resolution of microsecond level; Mechanical vibration signal: The mechanical vibration waveform generated at the moment of relay action is acquired by a miniature piezoelectric vibration sensor integrated in the interface adapter.

[0152] Step 4: After the test sequence is completed, the built-in processor automatically extracts features from the massive amount of raw data collected to obtain multi-dimensional features characterizing the deep performance of the backup automatic transfer device: Timing characteristics: In addition to the basic total action delay (48 ms from voltage disappearance to trip output action time, 102 ms from trip output to closing output action time), the timing standard deviation of multiple actions (reflecting action consistency) and the action time ratio of different output circuits (such as sectionalizing switch closing and incoming line switch tripping) are also calculated. Electrical characteristics: Extract the maximum current change rate (reflecting driving capability), peak driving current (to determine if there is an inter-turn short circuit in the coil), and charge amount during operation (obtained by integrating current over time to assess contact erosion) from the relay coil current waveform; simultaneously, calculate the voltage drop rate (in this test, the voltage dropped from 57.7V to 0V in 2 seconds, corresponding to a drop rate of 28.85 V / s), sag depth, and recovery time from the primary voltage waveform; extract the turn-off overvoltage amplitude and closing voltage drop from the secondary contact voltage waveform (in this test, the closing voltage drop was 12 mV, far below the 50mV alarm threshold, indicating good contact); Spectral characteristics: The vibration signal is subjected to fast Fourier transform to extract the main frequency component (the main frequency in this test is 2.3kHz, which is highly consistent with the factory standard of 2.28kHz) and the energy distribution of different frequency bands (0-1 kHz, 1-5 kHz, 5-10 kHz) to determine whether the mechanical structure is loose or aging.

[0153] Step 5: Subsequently, the processor initiates the multi-dimensional fusion analysis and evaluation module to perform in-depth diagnosis from three levels: Single-unit longitudinal trend analysis: Historical test data from the same period last year (e.g., action delay of 47ms / 100ms, closing voltage drop of 10mV, and main frequency of 2.28 kHz) were retrieved and compared with the current data. A first-order autoregressive model was established, revealing a slow increasing trend in action delay (annualized growth rate of approximately 2%), and a slight increase in closing voltage drop, both still within the normal range. Based on this, the system generated a performance degradation warning: it is predicted that the action delay may approach the 55ms upper limit in 3 years, and maintenance is recommended at that time.

[0154] Multi-machine horizontal group analysis (optional, if the test instrument is connected to the network): The test data was automatically uploaded to the cloud platform and subjected to DBSCAN cluster analysis with the historical data of 30 PCS-965 devices in the same batch. The results showed that the feature vector of this device was located near the center of the main cluster, and no deviation from the group was found, ruling out the risk of family defects.

[0155] Digital twin parameter identification: Based on the excitation (injected test signal) and response (acquired electrical and timing data) collected in this study, the maximum a posteriori probability estimation algorithm is used to identify the parameters of the pre-constructed nonlinear state-space model, obtaining an equivalent parameter set reflecting the current physical characteristics of the device. This includes the equivalent time constant of the relay coil (identified value is 3.2 ms, factory nominal value is 3.0 ms), pure delay (0.8 ms), and equivalent contact resistance (15 mΩ). These parameters serve as the digital fingerprint of the device, driving the digital twin model to perform future state simulation and lifetime prediction.

[0156] Finally, the tester generates a comprehensive diagnostic report, which includes: Basic test conclusion: "The 10 kV segmented backup logic test passed, and the action time meets the threshold requirements." Deep Health Indicators: Display the above multidimensional characteristic values ​​and their trends; Degradation warning information: "Action delay is on the rise and is expected to approach the threshold in 2027"; Digital twin parameters: Provides the current equivalent parameter set and compares it with the factory default value; Maintenance recommendation: "The current equipment is in good condition. It is recommended to perform the next test according to the normal cycle (3 years)."

[0157] The report is automatically saved to the device's local storage and can be synchronized to the cloud platform via a 4G network, serving as a permanent historical record for the device and providing data support for subsequent trend analysis.

[0158] Example 4: Batch testing and status screening of multiple standby automatic transfer devices A power supply company needed to conduct an annual inspection of 30 PCS-965 automatic transfer switch devices manufactured in the same batch within its jurisdiction. Maintenance personnel used an upgraded tester (PDT-200), which supports data upload to a cloud analysis platform. The staff performed rapid connection and testing on each device as described in Example 3.

[0159] During the testing of each device, in addition to collecting and uploading data, the tester also has local intelligent analysis capabilities. The tester has built-in factory benchmark data and historical test database for that model of device (if the device has been connected to this tester before).

[0160] After the test is completed, the tester can immediately perform longitudinal trend analysis on the current device: The extracted features, such as motion delay, peak current, and vibration frequency, are compared with the equipment's historical data to calculate the rate of change. An internal threshold model is then used to determine if there are any abnormal trends.

[0161] For example, if the tester finds that the operation delay of a certain device has increased by 8% compared to last year's test value and has been on an upward trend for three consecutive years, it will immediately display "Attention: Operation delay continues to increase, it is recommended to pay attention to relay aging" on the local screen and generate a brief on-site report.

[0162] In addition, the tester can also perform simple horizontal comparisons: The system compares the characteristics of the current device with the built-in statistical model of the same model (built based on historical test data). If a certain indicator is found to deviate from the mean by more than 2 standard deviations, the system will indicate "This indicator deviates from the normal range of the population, and further analysis is recommended".

[0163] These local analysis results, along with the raw data, are packaged together and ready to be uploaded to the cloud platform.

[0164] After the test is completed, the raw response data and preliminary extracted features (such as action delay, current peak, and vibration frequency) of all devices are automatically uploaded to the "Automatic Switchgear Intelligent Diagnosis Cloud Platform" via the 4G network.

[0165] After receiving the data, the platform first performs a horizontal comparison: The average latency of these 30 devices was calculated to be 50ms, with a standard deviation of 2ms. Among them, three devices had latency of 58ms, 60ms, and 59ms respectively, which significantly deviated from the overall average.

[0166] The platform automatically marked these three devices as "highlighted".

[0167] Next, the platform performs a longitudinal analysis on one of the tagging devices: By retrieving its test data from last year and this year, it was found that its action latency gradually increased from 52ms last year to 60ms this year, showing a slow growth trend.

[0168] Based on this trend, the platform used a simple linear fit to predict that the device's action latency may reach 68ms next year, which is outside the normal range.

[0169] Based on both horizontal and vertical analysis, the platform generates a report: It was pointed out that three pieces of equipment may have relay aging issues, and it was recommended that they be prioritized for repair. At the same time, maintenance personnel are reminded to pay attention to the overall performance trend of this batch of equipment.

[0170] Furthermore, the cloud platform performed digital twin parameter identification on key monitoring devices. The platform invoked a nonlinear state-space model built for the PCS-965 device, utilizing excitation-response data from current and historical tests (including injected voltage drop waveforms, acquired relay coil currents, and contact action timing) to identify the equivalent parameter set that best characterizes the device's current physical properties using a maximum a posteriori probability estimation algorithm. Taking the device with abnormal action delay as an example, the identification results showed that its relay coil's equivalent time constant increased from 3.0 ms at the factory to 3.6 ms, and the equivalent contact resistance rose from 10 mΩ to 25 mΩ. These parameters were used as the device's "digital fingerprint" and input into the digital twin model to simulate future behavior under different load and fault conditions. Simulation predictions indicate that if the contact resistance continues to increase at the current rate, the device will reach the failure threshold in 18 months, potentially leading to contact burnout or failure to operate. Based on this, the platform added a "lifespan prediction and risk warning" section to the report: recommending that the relay be replaced within 12 months, and providing maintenance decision support based on digital twins.

[0171] This report, along with the equipment list, was distributed to the professional management department and on-site teams to guide subsequent maintenance work.

[0172] Example 5: In-depth innovative implementation of single-machine longitudinal trend analysis and multi-machine horizontal group analysis In this embodiment, the following three analyses are all based on the multidimensional feature data extracted in step 4. Wherein: Single-machine longitudinal trend analysis: A multi-dimensional, multi-parameter fusion analysis strategy is preferred. This involves constructing a composite health index that integrates multiple indicators from various dimensions, such as time-series characteristics, electrical characteristics, and spectral characteristics, into a single comprehensive quantitative indicator. Time-series modeling is then performed on this comprehensive indicator to comprehensively and accurately depict the degradation trajectory of equipment performance. While individual trend analyses can be performed on single key parameters (such as action delay), fusion analysis provides a more comprehensive reflection of the equipment's health status.

[0173] Multi-machine horizontal group analysis: Also based on the overall distribution of multi-dimensional features, clustering and anomaly detection are performed. The full feature vector of each device is used as input, and anomaly patterns are automatically discovered through deep embedding clustering algorithm, avoiding the one-sidedness of single index threshold judgment.

[0174] Digital Twin Parameter Identification: The nonlinear state-space model describes the dynamic behavior of the automatic transfer switch (ATS) device through mathematical equations. Its input is excitation-response data, where "excitation" refers to the adaptive test signal waveform injected in step 2, and "response" refers to the raw data of multiple physical quantities across the entire link (including electrical waveforms, switching timing, vibration waveforms, etc.) synchronously acquired in step 3. Using this data, the model solves for the equivalent parameter set (such as equivalent time constant, pure delay, etc.) that best characterizes the device's current physical properties through a maximum a posteriori probability estimation algorithm. These parameters serve as the device's "digital fingerprint," driving the digital twin model to simulate future states and predict lifetime. It should be noted that parameter identification can directly use the raw response waveform or use multi-dimensional features extracted from the waveform, but using the raw waveform is preferred to retain more information and improve identification accuracy.

[0175] 1. Single-machine longitudinal trend analysis: Performance degradation prediction based on composite health index and grey Markov model: To overcome the limitations of traditional single-indicator threshold comparisons, this invention proposes a method for constructing a composite health index that integrates multi-dimensional features. This method combines a grey Markov model to achieve high-precision prediction of degradation trajectories and estimation of remaining lifespan. The specific steps are as follows: Step ①: Optimization of multidimensional features and construction of composite health index First, key indicators strongly correlated with the health status of the equipment are selected from the multidimensional features extracted from each test, including: Timing characteristics: Total action delay Time series standard deviation ; Electrical characteristics: Maximum rate of change of relay coil current Peak drive current Action charge Contact closing voltage drop ; Spectral characteristics: dominant vibration frequency High-frequency energy ratio ; Since the indicators have different dimensions and are correlated, principal component analysis (PCA) is used for dimensionality reduction and fusion.

[0176] First, the feature data of historical health status are standardized, the covariance matrix is ​​calculated, and the first principal component is extracted as the composite health index. The foundation. For ease of engineering application, [the following will be implemented]. Normalization to The interval, in which This indicates the product's health condition upon leaving the factory. This represents the failure threshold state. Health index of the test The calculation formula is:

[0177] in The loading coefficient of the first principal component. For the first Standardized values ​​of each feature and These are the projected values ​​under healthy conditions and at the failure threshold, respectively.

[0178] Step 2: Degradation trend modeling based on the grey model GM(1,1) The health index sequence of the equipment in previous tests Using the original data, a grey prediction model GM(1,1) is established.

[0179] First, the sequence is accumulated once to generate the sequence. Establish the differential equation:

[0180] The development coefficient is solved using the least squares method. and gray action amount The predicted values ​​of the cumulative sequence are obtained, and then the predicted values ​​of the original health index are obtained by subtraction. .

[0181] This model can effectively capture the degenerate exponential trend, but it is less sensitive to random fluctuations.

[0182] Step 3: Markov residual correction improves prediction accuracy To overcome the inadequacy of grey models in predicting random fluctuations, a Markov chain is introduced to correct the residuals. The residual sequence is defined as follows: Based on the size of the residual, it is divided into There are several states (for example: state 1 is "negative large deviation", state 2 is "negative small deviation", state 3 is "normal", state 4 is "positive small deviation", and state 5 is "positive large deviation").

[0183] Statistical state transition probability matrix ,in , From state Transition to state Number of times, In a state The total number of times.

[0184] For the future In this test, the predicted value is first obtained from the grey model. Then, based on the current state (the first... Given the state to which the residual belongs and the transition probability matrix, determine the most likely residual state and take the mean of the residuals in that state. As a correction factor, the final predicted value is obtained: Step 4: Estimation of the probability distribution of remaining lifetime Set equipment failure threshold (e.g., 0.6), using the modified grey Markov model ( Extrapolate the future health index trajectory. Considering prediction uncertainty, Monte Carlo simulation is used to generate multiple possible degradation paths: in each simulation, starting from the current state, a future residual state is randomly selected based on the state transition probability, and random noise is added to obtain a predicted path. When the path first falls below... Record the remaining lifetime of the path. Repeat. Next (such as) The probability distribution of remaining lifespan is obtained, and the mean, median and 90% confidence interval are calculated.

[0185] By integrating multi-source features through a composite health index, the health status of equipment can be comprehensively quantified; the grey Markov model takes into account both trends and randomness, significantly improving prediction accuracy; the remaining lifetime probability distribution provides a risk quantification basis for operation and maintenance decisions, realizing a leap from "whether it is normal" to "when it will fail".

[0186] Through longitudinal trend analysis of a single device, the system can quantitatively identify subtle degradations in equipment performance. For example, when the action delay shows a slow increasing trend in multiple consecutive tests (even if each increase is less than the fluctuation range of traditional threshold criteria), the system can detect early signs of "sluggish action" in the relay. Combined with a decrease in the peak drive current in the electrical characteristics, it can be further inferred that the coil is aging; combined with an increase in the contact closing voltage drop, it can be inferred that the contact is not making good contact. This quantitative analysis based on longitudinal comparison allows the invention to go beyond simply judging whether an action has occurred to quantitatively analyzing how it has occurred. By extracting multi-dimensional micro-features such as millisecond-level timing and transient electrical characteristics, it can keenly capture early sub-health states such as sluggish relay action and poor contact, enhancing the depth and accuracy of diagnosis.

[0187] 2. Multi-machine horizontal group analysis: Family defect diagnosis based on deep embedding clustering and association rule mining To automatically identify abnormal devices and trace their root causes in large-scale device groups, this invention proposes a group analysis framework that combines deep autoencoders, deep clustering, and association rule mining. The specific steps are as follows: Step ①: Feature embedding and dimensionality reduction Same model Feature vector of the device (Including timing, electrical, and spectral characteristics) are standardized.

[0188] To eliminate nonlinear redundancy between features and extract more compact representations, a deep autoencoder network is constructed, containing an encoder. and decoder .

[0189] The network is trained by minimizing the reconstruction error. After training, the encoder output is taken as the device embedding vector. (generally This embedding vector preserves the main information of the original data while reducing the difficulty of subsequent clustering.

[0190] Step 2: Deep Embedding Clustering The Deep Embedding Clustering (DEC) algorithm is used to simultaneously optimize the embedding vectors and cluster centers. The algorithm is executed alternately: Soft assignment: compute each embedding vector Belongs to the Cluster centers probability :

[0191] Target distribution: Construct an auxiliary distribution To reinforce high-confidence allocation:

[0192] Optimization: Minimize the KL divergence between the soft assignment and the target distribution, while fine-tuning the encoder parameters and cluster centers.

[0193] After clustering is completed, the devices are automatically divided into several clusters, and small clusters or isolated points far from the center of the main cluster are marked as suspected abnormal devices.

[0194] Specifically, in this embodiment, the system automatically and unsupervisedly clusters 30 devices of the same model. 28 of these form a tight main cluster (representing a healthy group), while the other 2 are automatically separated, as they may have family-related defects (such as defective relay contact material in a particular batch). This avoids maintenance personnel searching through 30 devices like looking for a needle in a haystack, enabling precise problem localization.

[0195] Furthermore, the soft-assignment probability reveals subtle differences; for example, five devices belong to the healthy cluster with a probability of 0.7 and to the abnormal cluster with a probability of 0.3. The system can mark these as devices of interest. This sub-healthy state is difficult to detect using traditional single-indicator threshold judgments, but this invention, through probabilistic analysis after multi-feature fusion, captures its deviation from the group, achieving a leap from post-event alarms to pre-event warnings.

[0196] Deep embedding clustering (through) and The optimization provides an accurate dataset for subsequent association rule mining (Apriori algorithm) and Bayesian network. It clearly identifies which devices are abnormal, so that during subsequent data mining, the system can focus on analyzing the devices marked as abnormal, thereby efficiently mining strong association rules such as "batch=2020A" -> "high vibration frequency" -> "abnormal label=yes", and finally locating the root cause of the problem (such as a problem with the relays of a certain supplier).

[0197] Step 3: Association rule mining and cause tracing For suspected anomalous equipment, its feature values ​​and equipment metadata (production batch, commissioning date, operating environment temperature and humidity, average load rate, etc.) are extracted. The multi-source information (feature values, metadata) of suspected anomalous devices is transformed into a standard input format for association rule mining, forming a transaction database. Each transaction is a sample, containing the equipment ID, anomaly flag (yes / no), grading values ​​for each feature (such as "high action delay", "vibration frequency > 2.5kHz"), and metadata attributes (such as batch, commissioning date).

[0198] An improved Apriori algorithm is run on a constructed transaction database. Minimum support and minimum confidence are set, and the algorithm automatically searches for all frequent itemsets that satisfy the minimum support and confidence, generating candidate association rules in the form {Condition A} → {Conclusion B}. Special attention is paid to rules such as “{Batch=2020A, High Vibration Frequency} → {Anomaly Mark=Yes}”, which conclude with “Anomaly Mark=Yes”.

[0199] To further eliminate random associations, a lift metric is introduced to calculate the true correlation between the conditions and conclusions of the discovered candidate rules (distinguishing between random and causal associations):

[0200] in This represents a set of conditions for an association rule. For example, in the rule {batch=2020A, high vibration frequency} → {anomaly flag=yes}, {Batch=2020A, Vibration frequency is high} represents the combined condition that "the equipment belongs to batch 2020A and the vibration frequency is high". This represents the conclusion term of the association rule. In the rules above, {Exception flag = Yes} means "the device is marked as an exception".

[0201] A threshold (e.g., Lift > 1.5) is typically set to filter out strongly correlated rules for final cause tracing. For example, if the lift is > 1.5, it indicates... and Positive correlation, indicating potential causal significance.

[0202] Step 4: Causal Inference Based on Bayesian Networks Association rules only indicate correlation; to further confirm causal relationships, a Bayesian network is constructed. Using device metadata and feature classifications as nodes, and anomaly markers as target nodes, the network structure is learned through a scoring search algorithm (such as BIC), and parameters are estimated using maximum likelihood. Through causal inference, the posterior probability of an anomaly occurring given certain factors (such as batch) can be calculated, and the root variable contributing most to the anomaly can be identified. For example, the network might reveal that "batch = 2020A" directly leads to "high vibration frequency," which in turn leads to "anomaly marker = yes," thus identifying a batch-specific defect.

[0203] Deep embedding clustering automatically discovers complex anomaly patterns without the need for manual threshold setting; association rule mining reveals the correlation between anomaly features and metadata, providing preliminary clues to the causes; Bayesian networks further confirm causal relationships, achieving a closed loop from "discovering anomalies" to "interpreting anomalies," providing a scientific basis for locating familial defects and effectively avoiding large-scale power outages.

[0204] Embodiment 6 of the present invention provides a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to perform operations according to the instructions to execute the steps of the method.

[0205] Embodiment 7 of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method.

[0206] Compared with the prior art, the beneficial effects of the present invention include at least the following: This invention employs a one-click, wire-free testing interface to safely connect the tester to an operating automatic transfer switch. This interface features isolation and achieves plug-and-play safe access through wire-free testing technology, completely eliminating the impact of testing on the user's power supply and enabling testing without power interruption, thus improving power supply reliability.

[0207] This invention overcomes the limitations of traditional fixed testing by using one-click access and adaptive test signal generation, and employing reinforcement learning to dynamically optimize test cases. It significantly reduces manual operation and decision-making time, can automatically explore the performance boundaries of equipment, discover hidden defects, improve test coverage, and enhance testing efficiency and intelligence.

[0208] The reinforcement learning strategy of this invention determines whether an action meets safety constraints through a safety barrier function. If not, it projects the action to the nearest safe action. Through nonlinear adaptive margin, environmental compensation, and penalized projection, it can achieve refined and dynamic management of safety constraints, significantly improving the safety and adaptability of reinforcement learning action generation in online testing scenarios of power secondary equipment. This ensures that the equipment is not accidentally activated or damaged during testing, while retaining the exploratory capabilities of reinforcement learning, thus enhancing the safety and practicality of online adaptive testing.

[0209] This invention performs single-machine longitudinal trend analysis and multi-machine horizontal group analysis based on the aforementioned multi-dimensional features to obtain test results, achieving a leap from individual performance degradation early warning to group family defect diagnosis. Through single-machine longitudinal trend analysis, it can quantitatively predict the future degradation trend of device performance, providing maintenance personnel with sufficient early warning time, realizing the transformation from reactive repair to condition-based maintenance, and achieving predictive maintenance. Through multi-machine horizontal group analysis, it can detect batch and family defects with wide impact at an early stage, avoid the occurrence of large-scale power outages, improve the equipment operation level of the power grid, and achieve systemic risk prevention and control.

[0210] This invention's single-machine longitudinal trend analysis integrates multi-dimensional feature data into a comprehensive quantitative indicator, namely the composite health index, by constructing a composite health index. Time series modeling is then applied to this index to comprehensively and accurately depict the degradation trajectory of equipment performance. This longitudinally comparative quantitative analysis moves beyond simply determining whether an action has occurred to quantitatively analyzing how it has occurred. It can keenly detect early sub-health conditions such as sluggish relay operation and poor contact, enhancing the depth and accuracy of diagnosis.

[0211] This invention employs a multi-machine lateral group analysis framework combining deep autoencoders, deep clustering, and association rule mining. This framework can automatically identify anomalous devices and trace their root causes within large-scale device groups. The deep embedding clustering process utilizes a self-optimization mechanism aided by soft assignment and target distribution. An arctangent kernel function provides a robust distance-probability mapping, while the cluster dispersion (the average weighted distance from samples within a cluster to the cluster center) target distribution guides the cluster centers to converge towards the main cluster. This synergy results in normal devices forming high-density, low-dispersion clusters, while anomalous devices, being far from all clusters, receive extremely low scores, effectively identifying suspected anomalous devices.

[0212] This invention creates a quantifiable digital image of the device by constructing a mathematical model and identifying its core parameters, making accurate lifespan prediction possible.

[0213] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0214] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0215] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0216] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0217] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A test method for a non-disconnectable automatic switching device, characterized in that, The method includes: The tester is safely connected to the operating automatic transfer switch using a one-click, wire-free test interface. The tester employs a reinforcement learning strategy based on a security barrier function to inject adaptive test signals into the backup automatic switching device through the interface. The adaptive test signal generation strategy is as follows: based on the rule base of the backup self-supply standard logic, an initial test case set covering basic functions is automatically generated; during the test, based on the real-time collected action response signals, the currently touched logic nodes and loops are analyzed to dynamically evaluate the test coverage; when it is found that there is no covered logic or operating condition, a reinforcement learning strategy based on the security barrier function is adopted to automatically adjust the test signal parameters and generate new optimized test cases; the initial test cases and optimized test cases are integrated to form an adaptive and complete standardized test signal sequence. The modeling and deployment steps of the reinforcement learning strategy include: establishing a digital twin model capable of simulating the dynamic behavior of the automatic transfer switch (ATS) device based on its logical topology and electrical parameters, thus obtaining a digital twin environment; constructing the logical nodes and their connections of the ATS device into a graph structure, and extracting state feature vectors from the graph structure using a graph neural network; constructing a hybrid action space including continuous and discrete test parameters; establishing a multi-objective reward function, including at least a coverage reward based on newly triggered logical nodes, an efficiency penalty based on test states, a safety penalty based on safety margin constraints, and an intrinsic reward based on predicted states; training an agent in the digital twin environment using a deep reinforcement learning algorithm based on the state feature vectors, the hybrid action space, and the multi-objective reward function to maximize cumulative rewards; deploying the trained agent to an actual test instrument, and determining whether the action meets safety constraints through a safety barrier function before generating test signals; if not, projecting the action that does not meet the safety constraints to the nearest safe action. While injecting the test signal, the action response signal of the backup automatic transfer device is simultaneously collected through the interface to obtain full-link multi-physical quantity data; Feature extraction is performed on the full-link multi-physical quantity data to obtain multi-dimensional features characterizing the deep performance of the backup automatic switching device; Based on the aforementioned multidimensional features, single-machine longitudinal trend analysis and multi-machine horizontal group analysis are performed to obtain test results.

2. The test method for a non-disconnectable automatic switching device according to claim 1, characterized in that: The one-click, wire-free test interface includes a set of adapters that can be directly plugged into the terminal block of the automatic transfer switch (ATS) device. The adapters integrate signal isolation and conditioning circuitry to safely inject the analog test signal generated by the tester into the operating circuit of the ATS device and to collect the action response signal of the ATS device.

3. The test method for a non-disconnectable automatic switching device according to claim 1, characterized in that: The process involves determining whether an action meets safety constraints using a safety barrier function. If not, the action that does not meet the safety constraints is projected to the nearest safe action, as detailed below: For action Calculate its safety barrier function value : in, Actions Voltage, timing, and current exceeding limits; Each represents an action The voltage command value, action timing width, and injected current amplitude are specified in the data. These are the safety thresholds corresponding to the voltage command value, the action timing width, and the injected current amplitude, respectively. These are the weighting coefficients; like , then indicates an action If the safety constraints are not met, the following formula will be used to... Project to Recent safety actions : in, Candidate actions The security barrier function value, For safety margin, The square of the L2 norm; This is a penalty factor.

4. The test method for a non-disconnectable automatic switching device according to claim 3, characterized in that: The formula for calculating the safety margin is as follows: in, Basic safety margin; Assess the overall health of the equipment; This is the time interval since the last test; Sensitivity coefficient; Environmental factors; These are the current ambient temperature and relative humidity, respectively. For the equipment's rated operating environment, Environmental impact weighting coefficient; This is the environmental factor adjustment coefficient.

5. The test method for a non-disconnectable automatic switching device according to claim 1, characterized in that: The multi-physical quantity data includes electrical quantity signals, switching quantity timing signals, and mechanical vibration signals; The electrical quantity signals include the voltage and current waveforms of the primary system monitored by the automatic transfer switch, as well as the voltage and current waveforms of the secondary circuit inside the automatic transfer switch. The switching timing signal includes the microsecond-level action time of each relay contact in the automatic transfer switch; The mechanical vibration signal includes the mechanical vibration waveform generated at the instant the relay in the automatic switching device is activated.

6. The test method for a non-disconnectable automatic switching device according to claim 5, characterized in that: The multidimensional features characterizing the deep performance of the standby automatic switching device include timing features, electrical features, and spectral features; The timing characteristics include the total delay of relay action, the standard deviation of action timing, and the ratio of action time of different output circuits extracted from the switching timing signal; The electrical characteristics include current dynamic characteristics and voltage transient characteristics extracted from electrical quantity signals. The current dynamic characteristics include the maximum value of the relay coil current change rate, the peak value of the relay coil drive current, and the amount of charge during the relay operation. The voltage transient characteristics include the primary voltage drop rate, sag depth, and recovery time, as well as the turn-off overvoltage amplitude and closing voltage drop of the secondary contact voltage. The spectral characteristics include the energy distribution and dominant frequency components of different frequency bands obtained after performing a fast Fourier transform on the mechanical vibration signal.

7. The test method for a non-disconnectable automatic switching device according to claim 1, characterized in that: The single-machine longitudinal trend analysis specifically includes: Multidimensional feature optimization and composite health index calculation are performed based on multidimensional feature data from previous measurements of the same standby automatic switching device to obtain a health index sequence; a predicted health index value is then obtained based on the health index sequence. The residuals are calculated based on the predicted health index value and the composite health index in the health index sequence. The predicted health index value is then corrected based on the state transition law of the residuals to obtain the corrected predicted health index value. Based on the corrected health index prediction value, the future health index trajectory is extrapolated to provide early warning of degradation, and combined with a preset failure threshold, the remaining lifespan of the automatic start-up device is estimated.

8. The test method for a non-disconnectable automatic switching device according to claim 1, characterized in that: The multi-machine cross-group analysis specifically includes: Based on a deep autoencoder network, feature embedding and dimensionality reduction processing are performed on the multidimensional feature data of multiple backup automatic switching devices of the same model to obtain the embedding vector of each device. Deep embedding clustering is performed on the embedding vectors of each backup automatic switching device to divide the backup automatic switching devices into several clusters and mark suspected abnormal devices. Association rule mining is performed on the feature data and metadata of suspected abnormal devices to filter out association rules with strong correlations and provide preliminary clues to the cause of the anomaly. By identifying the root cause variables leading to abnormalities through causal inference models, familial defects can be diagnosed.

9. The test method for a non-disconnectable automatic switching device according to claim 8, characterized in that: The process of performing deep embedding clustering on the embedding vectors of each backup automatic switching device, dividing the backup automatic switching devices into several clusters, and marking suspected abnormal devices specifically includes: Calculate the soft assignment probability of the embedding vector of each standby self-connecting device belonging to each cluster center: in, The embedding vector of the i-th standby automatic transfer device Belongs to the Cluster centers The soft assignment probability; For the first Cluster centers; For Euclidean distance, For global scale parameters; Constructing the target distribution probability based on soft assignment probability: in, The embedding vector of the i-th standby automatic transfer device Belongs to the Cluster centers The target distribution probability, The embedding vector of the i-th standby automatic transfer device Belonging to the j′-th cluster center The soft assignment probability; , For the first The average weighted distance from the j′ samples within each cluster to the cluster center; By minimizing the KL divergence between the soft assignment probability and the target distribution probability as the loss function, the encoder network parameters and cluster center positions are alternately fine-tuned until the loss function converges; after clustering, the device is automatically divided into several clusters and suspected abnormal devices are marked.

10. A test device for a non-disconnectable automatic switching device, used to implement the method described in any one of claims 1-9, characterized in that, The testing apparatus includes: A one-click, wire-free test interface is used to safely connect the tester to an operating automatic transfer switch. The tester is used to inject adaptive test signals into the automatic transfer switch (ATS) device through the interface; while injecting the test signals, it synchronously collects the action response signals of the ATS device through the interface to obtain full-link multi-physical quantity data; it performs feature extraction on the full-link multi-physical quantity data to obtain multi-dimensional features characterizing the deep performance of the ATS device; and it performs single-machine longitudinal trend analysis and multi-machine horizontal group analysis based on the multi-dimensional features to obtain test results.

11. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-9.

Images