Relay protection device closed loop logic verification system based on data direct acquisition

Abstract

The invention discloses a relay protection device closed-loop logic verification system based on data direct collection, and particularly relates to the technical field of relay protection device verification. Action logic messages and multi-mode physical quantity data of a relay protection device are synchronously collected through hardware synchronization pulses; carrying out feature extraction on the multi-modal physical quantity data by adopting a transfer learning strategy, matching the multi-modal physical quantity data with a hidden fault feature library, and identifying an abnormal feature vector; performing space-time alignment on the abnormal feature vector and the action logic message, quantitatively evaluating a risk entropy value and judging a risk level; dynamically generating a test sequence priority based on a relay protection device function unit association topological graph, and optimizing test excitation parameters in real time through a digital twinborn model; fusing a physical failure mechanism and a data driving trend to predict the remaining service life of the key element; early warning and predictive intelligent maintenance of the hidden fault of the relay protection device are realized, and the problems that the internal state of the relay protection device cannot be deeply perceived and the self-adaptive optimization capability is lacked in the existing verification technology are solved.

Description

Technical Field

[0001] This invention relates to the field of relay protection device verification technology, and more specifically, to a closed-loop logic verification system for relay protection devices based on direct data acquisition. Background Technology

[0002] Currently, the verification of relay protection devices mainly focuses on whether their final action is correct, such as whether the relay protection device issues a trip command within the specified time limit after a fault current is applied. This verification method is like judging a person's health status solely by whether they have reached their destination, ignoring any signs of sub-health such as unsteady gait or slow reaction during their journey. Relay protection devices are composed of numerous electronic components, such as sampling chips, logic processing units, and power modules. Their performance degradation over long-term operation is a gradual process. This degradation often manifests in early stages as some latent fault symptoms, such as harmonic spectrum distortion in analog sampling channels, abnormally high DSP operation cycle utilization, and increased power bus ripple coefficient. These deep-seated multimodal physical quantity changes are difficult to effectively capture and analyze in traditional verification methods, allowing relay protection devices to operate in a sub-healthy state. Ultimately, when the power grid encounters disturbances, these latent faults are activated, leading to maloperation or failure to operate, triggering a cascading failure.

[0003] These potential, latent fault symptoms cannot be detected through traditional operational logic correctness checks. Most existing testing systems lack the capacity to collect and analyze this deep-seated, multi-dimensional internal operational data of relay protection devices, making it impossible to provide early warnings of device degradation trends during periodic inspections. As a result, relay protection devices may operate in the power grid for extended periods with undetected defects, eventually failing at a critical moment due to component performance deteriorating to a critical point, leading to significant power outages and economic losses. Existing post-incident assessment mechanisms cannot achieve predictive health diagnosis and early warning. Summary of the Invention

[0004] To overcome the aforementioned deficiencies in the prior art, this invention provides a closed-loop logic verification system for relay protection devices based on direct data acquisition. This system directly acquires multimodal physical quantity data through hardware synchronous pulses, combines prior knowledge of the relay protection device topology with transfer learning to identify latent faults, quantifies risk entropy values ​​for dynamic evaluation, and optimizes and predicts maintenance based on digital twin-driven testing strategies to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a closed-loop logic verification system for relay protection devices based on direct data acquisition, comprising: Synchronous acquisition module: During the same period of executing the test sequence, the high-speed data acquisition card of the hardware synchronous pulse trigger tester works in conjunction with the internal waveform recording function of the relay protection device to synchronously acquire the action logic message generated by the relay protection device, as well as multi-modal physical quantity data including the digital signal processor operation cycle occupancy rate and the harmonic spectrum components of the analog sampling channel. Feature recognition module: It integrates physical laws and data-driven feature extraction of multimodal physical quantity data to form real-time feature vectors. By combining the transfer learning strategy of the topological prior knowledge of the relay protection device with the latent fault feature library, it identifies abnormal feature vectors associated with the degradation mode. Risk assessment module: It uses a unified timestamp to align the abnormal feature vector with the action logic message in time and space, and obtains the risk entropy value based on the quantification of risk intensity, trend instability and correlation coupling; it adjusts the dynamic threshold in combination with the operation scenario, and determines the risk level by comparing the risk entropy value with the adjusted dynamic threshold; it performs failure propagation analysis based on the relay protection device topology, and predicts the risk development trajectory by combining the time gradient of the risk entropy value, and generates a comprehensive assessment report including the current risk level and development expectations. Strategy optimization module: Analyzes the risk propagation path in the comprehensive assessment report, constructs a topology diagram of the functional units of the relay protection device, dynamically generates test sequence priorities based on critical path analysis, optimizes test excitation parameters in real time through a digital twin model, uses composite test waveforms to simultaneously excite the response characteristics of multiple related nodes, and dynamically adjusts the test strategy according to the difference between the actual response and the expectation. The explanation is that the edges of the functional unit association topology diagram of the relay protection device have directional attributes, and the direction represents the dominant direction of fault, risk or signal transmission. Predictive Maintenance Module: Based on real-time test data and historical operation records, it calculates the remaining service life of key components by integrating physical failure mechanism analysis and data-driven trend prediction; it generates maintenance recommendations using a multi-objective decision-making method that comprehensively considers reliability, maintenance costs, and power outage losses; and it generates predictive diagnostic reports that include health status, maintenance plan comparisons, and decision-making basis through structured templates. Among them, key components refer to specific physical parts located within one or more relay protection device functional units, whose failure will directly lead to the loss of the core function of the unit, and are obtained based on reliability engineering analysis or historical fault data.

[0006] Preferably, the synchronous acquisition includes: generating a hardware synchronization pulse at the start of the test excitation signal output by the tester; this pulse simultaneously triggers the high-speed data acquisition card of the tester and the internal waveform recording function of the relay protection device; the occupancy rate of the acquired digital signal processor is obtained by reading the internal status register of the relay protection device; the harmonic spectrum components of the analog sampling channel are obtained by accessing the fault waveform recording file of the relay protection device; the ripple coefficient of the input and output circuit power bus is obtained by measuring the high impedance probe of the tester; the real-time junction temperature data is acquired by the built-in temperature sensor of the relay protection device; and all acquired data are embedded with a unified timestamp to form a spatiotemporal correlated dataset.

[0007] Preferably, a dual-path fusion network is used for feature extraction; the first path is an interpretable path based on physical laws, which uses a short-time Fourier transform layer to extract the energy distribution of the recorded data in a predefined key frequency band, and integrates a statistical variance calculation layer to obtain the temporal stability features of the action time series; the second path is a data-driven path based on a multi-scale one-dimensional convolutional neural network, which automatically learns deep nonlinear features with multiple convolutional kernels of different durations; the gated attention fusion module dynamically weights and fuses the output features of the two paths according to the signal-to-noise ratio and stationarity of the input data to form a robust real-time feature vector.

[0008] Preferably, the construction and matching process of the latent fault feature library includes: A graph-structured prior knowledge base is constructed based on historical failure cases and domain expert experience. The prior knowledge base defines the association between degradation modes and topological relationship nodes of functional units of relay protection devices. In the process of cross-domain feature distribution alignment, the prior knowledge of the relay protection device topology is transformed into structured constraint terms that characterize the physical connection logic, and a topology constraint loss function is constructed. The topology constraint loss function is used to quantify the degree of deviation between the feature matching result and the physical topology logic of the relay protection device. An alternating optimization strategy is adopted. First, the parameters of the matching network are fixed, and the topological embedding representation of the feature space is optimized by minimizing the deviation. Then, the topological embedding representation is fixed, and the main loss function of the matching network is optimized. The topology constraint loss function and the main loss function of the deep matching network jointly participate in adversarial training so that the matching results output by the matching network conform to the physical connection and signal flow logic inside the relay protection device.

[0009] Preferably, the construction of the graph structure prior knowledge base includes: First, perform triple extraction of unstructured data: retrieve historical operation and maintenance and fault reports in electronic document format, parse the text using the integrated converter bidirectional encoding representation, bidirectional long short-term memory network and conditional random field named entity recognition model, extract component, phenomenon and fault mode entities and generate standardized entity relationship triples. Secondly, the implicit expert experience is weighted and graphed: the empirical rules of the domain experts are transformed into a set of weighted directed graph edges, where the explicit physical signal flow and connection logic are defined as irreversible unidirectional hard constraint edges with a weight of 1.0, and the fuzzy causal relationships described in the expert experience (such as "may trigger") are quantified into conditional probability weights with values ​​between 0 and 1, and the corresponding connection edges are assigned to form a probability-dependent network. Finally, the vector space embedding step of the graph structure is performed: a fault mechanism topology graph of the relay protection device is constructed in the graph database based on entity relations and weighted directed graph edge sets, and a graph convolutional network is used to perform representation learning on the topology graph. By minimizing the structural loss function, the discrete nodes and edges in the graph are mapped into continuous low-dimensional dense embedding vectors. The embedding vectors retain the structural information and probabilistic dependencies in the original topology and serve as the mathematical benchmark for subsequent feature similarity calculation.

[0010] Preferably, the risk level classification adopts a dynamic clustering method based on Mahalanobis distance, constructs a feature space based on risk intensity, trend instability and correlation coupling, and determines the boundary conditions of the risk level through iterative optimization; wherein the number of clusters is adaptively determined according to the distribution characteristics of historical data, the cluster centers are dynamically updated through the mean shift algorithm, and the level boundary is described by a confidence ellipse.

[0011] Preferably, the failure propagation analysis adopts a probabilistic prediction method based on a random walk model. In the constructed functional unit association topology diagram of the relay protection device, the node corresponding to the abnormal feature vector is taken as the starting point, and the risk propagation process is simulated by a random walk algorithm with restart. The walk transfer probability is dynamically adjusted according to the edge set weight and node state. The hit probability of the walk path is used as a probability estimate of the propagation risk. Combined with the time gradient of the risk entropy value, the expected time window of each propagation path is predicted. Wherein, the edge set weight Defined from functional units to The ease of conduction is calculated according to the formula. ,in Select the electrical connection impedance between the two units (including the DC resistance value of the PCB traces or the standardized communication impedance of the bus). Select the logic transmission delay of the signal between units. and To pre-determine the normalization coefficients, a basic topology is constructed in which the lower the electrical impedance or the shorter the logic delay, the greater the weight. The node state is defined as the normalized risk entropy value calculated by the risk assessment module based on risk intensity, trend instability, and correlation coupling degree. During the dynamic adjustment of the walk transition probability, a transition probability matrix is ​​constructed based on the risk tendency principle, and then expressed using the formula... Calculate from node Transfer to node transition probability , represents the probability that, during a random walk, the risk will transfer from the current source node i to a specific adjacent target node j. The larger this value, the easier it is for the fault to propagate to node j; i is the source node index, representing the relay protection device functional unit where the fault risk is located at the current moment (i.e., the starting point of the walk), and j is the target node index, representing a specific downstream functional unit that has a direct physical or logical connection with the source node i (i.e., the potential endpoint of the walk); 1 is a baseline bias constant, used to prevent the risk from spreading when the target node is in a completely healthy state (i.e., ...). When (=0), the transition probability becomes 0, ensuring that even if the target node is not at risk, as long as a physical connection exists ( >0), the fault still has the potential to propagate naturally through physical paths; among which The real-time normalized risk entropy value of the target node is calculated by the following calculation mechanism: when the risk entropy value of a specific downstream functional unit is detected to increase due to an anomaly, the gain term in the formula automatically increases the transfer probability from the adjacent node to the high-risk node. This simulates the physical characteristics of the fault tending to spread to the vulnerable area at the mathematical model level, and realizes the dynamic reconstruction of the probability matrix according to the real-time status of the relay protection device. This indicates that the summation operation is performed on all the directly adjacent nodes of node i, and k represents the neighbor node iteration index, which represents any directly adjacent node of node i (k traverses all the neighbors of node i, including node j). Represents the weight of the neighbor edge set, the weight of the edge set from the source node i to any neighboring node k; Represents the state of neighboring nodes, specifically the normalized risk entropy value of any adjacent node k.

[0012] When the random walk algorithm with restart is specifically executed, a restart probability is set. (The preferred value is 0.15) is used as the core constraint parameter, forcing the wandering particles to move in a certain direction during the iteration process. The probability forces a jump back to the seed node corresponding to the initial abnormal feature vector, and the iteration process follows the formula. Update, among which Let be the probability distribution vector at the current time. This is the transpose of the aforementioned transition probability matrix after dynamic adjustment based on node states and edge set weights. The initial state vector is set to 1 for the seed node corresponding to the abnormal feature vector, and 0 for the rest of the nodes. This calculation process introduces... The project constrains the risk propagation simulation within the topological local neighborhood centered on the seed node, and continues to iterate until the Euclidean distance between probability vectors at adjacent time points is less than a preset convergence threshold. At this point, the output steady-state probability vector serves as the hit probability estimate of each associated functional unit affected by the specific fault source.

[0013] To explain, in a preferred embodiment, the probability prediction method based on the random walk model is implemented through the following steps: A topology graph of relay protection device functional units is constructed based on the physical connection and logical dependency relationship of the relay protection device functional units. The node weights are dynamically adjusted according to the criticality of the corresponding relay protection device functional units, and the edge set weights are determined based on the electrical connection impedance and the logical delay of signal transmission. Using the relay protection device functional unit where the abnormal feature vector is located as the propagation source node, the propagation path is calculated by combining the graph traversal algorithm with the edge set weight, and the propagation conditional probability on each path is statistically analyzed based on historical failure data. A time prediction model is established by using regression algorithms. This model takes the real-time monitored risk entropy gradient, the weight of the current propagation path, and the conditional probability as inputs, and integrates the time series characteristics of historical failure cases to deduce the expected time window for the risk to develop into a failure along each key propagation path.

[0014] The topology graph is a directed graph, where nodes represent functional units of the relay protection device, edges represent electrical connections or signal flow relationships between functional units, and the weight of the edges is determined based on electrical impedance, signal transmission delay, or historical fault correlation strength; the elements in the transition probability matrix represent the probability of transitioning from node i to the adjacent node j. Secondly, a random walk algorithm with restart is executed, using the functional unit node corresponding to the identified abnormal feature vector as the starting node (i.e., the seed node) of the walk; the restart probability γ is set, with a value ranging from 0.1 to 0.3, and the preferred value being 0.1; this restart probability means that at each step of the walk, the algorithm has a probability of γ to directly jump back to the seed node and restart the walk, thereby ensuring that the risk propagation simulation always revolves around the abnormal source point; in the initial probability vector of the walk, the probability of the seed node is set to 1, and the probability of the other nodes is 0; Next, iterative calculations are performed until convergence. The walk process is iteratively updated using the formula: next probability vector = (1-γ) transpose of the transition probability matrix, current probability vector + γ initial probability vector. After each iteration, the Euclidean distance between the current probability vector and the previous probability vector is calculated. When this distance is less than a preset convergence threshold (e.g., on the order of 10 to the power of -10), the walk is considered to have reached a stable state, and the probability distribution at this time is the steady-state probability of risk propagation. Finally, the walk results are analyzed for prediction; the probability value corresponding to each node in the probability vector after reaching a steady state is used as an estimate of the probability of risk propagation to that node. Simultaneously, combining the time gradient of the risk entropy value calculated in previous steps (i.e., the rate of risk change) and the sum of the weights on each potential propagation path (representing the resistance of the propagation path), a linear regression model is used to extrapolate the expected time window for risk development to reach failure. The coefficients of this regression model are determined by fitting historical failure case data.

[0015] Preferably, the dynamic threshold adjustment adopts a multi-factor coupled regression model, and establishes an adaptive adjustment mechanism by comprehensively considering ambient temperature, power grid load rate and cumulative operating time of the relay protection device; wherein the ambient temperature is collected by the built-in temperature sensor of the relay protection device, the power grid load rate is obtained through the monitoring system interface, and the cumulative operating time of the relay protection device is extracted from the equipment commissioning record.

[0016] In one possible embodiment, the risk level is achieved through the following steps: Normalized values ​​for three feature parameters—risk intensity, trend instability, and cross-dimensional correlation coupling—were calculated separately. Risk intensity normalization employed a hyperbolic tangent function to handle the relative deviation between the feature values ​​and a dynamic threshold. Trend instability normalization was achieved by calculating the approximate entropy of the time series, which consisted of continuous sampling points of a preset length. The approximate entropy calculation used a tolerance threshold related to the statistical characteristics of the series. Cross-dimensional correlation coupling normalization was evaluated using the contribution rate of the first principal component in principal component analysis. A confidence-based variable weight allocation mechanism is adopted, which multiplies the initial weight of each feature parameter by its confidence coefficient, and then normalizes it using the softmax function so that the sum of the three weight coefficients is one. The risk entropy value is a continuous value between 0 and 1, which is used to quantify the overall risk level of the relay protection device. The confidence level is evaluated by the coefficient of variation within a sliding window of a preset size. The risk level is determined by comparing and mapping the calculated risk entropy value with a preset threshold range.

[0017] Preferably, the test sequence priority generation includes: constructing a topology diagram of the relay protection device functional units based on the risk propagation path, calculating the test priority weight of each path through critical path analysis, the test priority weight being determined by the ratio of the number of high-risk units in the path to the path length, and generating a test sequence order by arranging the test priority weights in descending order. The test optimization includes: predicting the expected response waveform using a digital twin model, calculating the root mean square error between the measured response and the predicted response, iteratively updating the response using a gradient descent algorithm based on real-time difference feedback, and introducing a momentum term to accelerate convergence and avoid local optima.

[0018] Preferably, a closed-loop feedback link is established between the dynamic threshold used for risk level determination and the digital twin model used for test stimulus parameter optimization. Specifically, the parameter adjustment trajectory generated by the digital twin model based on the difference between the measured response and the expected response is fed back to the regression model in real time as an additional feature input reflecting the sensitivity of the relay protection device to test stimuli under the current specific operating conditions. The regression model integrates the additional features on the basis of the original ambient temperature, grid load rate and cumulative operating time parameters, and performs dynamic secondary calibration of the threshold through an online learning mechanism, thereby improving the identification accuracy of latent fault risk patterns induced by specific test conditions and having operating condition correlation in subsequent risk assessments.

[0019] Preferably, when calculating the remaining service life of key components, the data-driven trend prediction model on which it relies has input features that, in addition to historical performance parameter decay trajectories, also deeply integrate prediction results from risk development trajectories. Specifically, the set of affected relay protection device functional units, the trigger probability of each propagation path, and the expected time window information contained in the risk development trajectory are encoded into a weighted time decay graph network. This graph network is used as a priori structural constraint and trained together with the time series data of historical performance parameters to form a graph convolutional recurrent neural network. This network uses topological structure (referring to the fixed electrical connections and signal interface relationships between various physical hardware components inside the relay protection device, such as analog input modules, digital signal processors, input / output boards, power supply modules, etc.) constraints to guide feature propagation, thereby simultaneously modeling the cascading performance decay effect caused by the risk propagation of adjacent units when predicting component life, generating a remaining service life distribution that is more consistent with the internal physical failure chain law of the relay protection device. In the functional unit association topology graph of the relay protection device, two functional units directly connected by an edge are called adjacent units.

[0020] In a preferred embodiment, the data-driven trend prediction model employs a long short-term memory network model to predict the remaining service life of key components, and its specific implementation process is as follows: The selection of the Long Short-Term Memory Network model is based on the fact that its gating mechanism (including input gate, forget gate, and output gate) can effectively capture long-term dependencies in time series, and is particularly suitable for predicting the trend of slow decay of the performance parameters of relay protection devices. The model's structure is specifically configured. The length of the input sequence (time step) is set according to the data acquisition cycle and performance degradation characteristics of the relay protection device. For example, a historical performance parameter sequence of thirty consecutive time units (such as days or weeks) is selected as the input. Each long short-term memory network layer contains several hidden units (such as fifty), which can be connected to a fully connected layer. Finally, a scalar is output, representing the predicted performance parameter value for the next time unit. The model's input features are time series of key performance parameters extracted from historical operating data, such as the daily average of digital signal processor cycle utilization and the weekly maximum of power bus ripple coefficient. These data need to be normalized before being input into the model. The training process of the model is as follows: historical full life cycle data is used as the training set, mean squared error is used as the loss function, and an adaptive moment estimation optimizer is used to optimize the parameters; an early stopping method is introduced during the training process to prevent overfitting, that is, training is stopped when the loss on the validation set no longer decreases in several consecutive training cycles. The remaining useful life prediction method is as follows: input the latest performance parameter sequence of key components into the trained long short-term memory network model and perform multi-step forward iterative prediction; when the performance parameter value predicted by the model exceeds the preset failure threshold for the first time (for example, the ripple coefficient is greater than 5%), the difference between the corresponding prediction time point and the current time is the remaining useful life prediction value calculated by the data-driven trend prediction model.

[0021] Preferably, the calculation of the remaining service life of key components includes: calculating the theoretical life loss through the electrothermal stress accelerated aging equation, and simultaneously performing data-driven prediction by analyzing the degradation trajectory of historical performance parameters, taking the smaller of the two values ​​as the final remaining service life; the generation of maintenance recommendations includes: establishing a weighted scoring system for reliability improvement effect, maintenance cost, and power outage loss, and ranking different maintenance schemes based on their comprehensive benefits; the generation of predictive diagnostic reports includes: automatically generating a structured report containing health status, maintenance scheme comparison, and decision-making basis through predefined templates, and providing auxiliary decision support by associating with similar historical cases.

[0022] Preferably, the remaining service life of the key components is calculated by integrating physical models and data-driven methods: theoretical service life loss estimates are obtained based on the electrothermal stress accelerated aging equation, while data-driven predicted service life is obtained by analyzing the degradation trajectory of historical performance parameters. Finally, the smaller of the two values ​​is taken as the result of determining the remaining service life. The maintenance recommendations are generated using a multi-objective decision-making mechanism, constructing a comprehensive evaluation system covering reliability improvement, maintenance costs, and power outage losses. The benefits of different maintenance schemes are quantitatively ranked through weighted scoring. The predictive diagnostic report is generated based on a structured template, automatically outputting content including equipment health status assessment, comparative analysis of multiple maintenance schemes, and relevant decision-making basis, and providing auxiliary support for maintenance decisions by associating with similar historical cases.

[0023] The technical effects and advantages of this invention are as follows: (1) The present invention adopts a hardware synchronous pulse triggering mechanism to synchronously collect the action logic message of the relay protection device and directly read multi-modal physical quantity data such as the operation cycle occupancy rate of the digital signal processor and the harmonic spectrum component of the analog sampling channel; by introducing the topological prior knowledge of the relay protection device as a constraint transfer learning strategy, it matches and identifies with the latent fault feature library, thereby realizing in-depth mining and early accurate identification of latent fault symptoms inside the relay protection device, overcoming the limitation of the existing verification method in its insufficient ability to capture the initial symptoms of latent faults.

[0024] (2) This invention constructs a test logic relationship network that integrates transient characteristics and transforms it into an event-driven state machine. During test execution, it synchronously calculates the logic path consistency index and the transient feature fidelity index, and integrates them with the abnormal features matched by the latent fault feature library. The results are then input into a test risk quantification model based on support vector machine regression to dynamically generate a degradation risk coefficient. Based on the propagation probability in the topology diagram associated with the functional unit of the relay protection device, the invention makes intelligent decisions to generate optimized test instructions, switch backup branches, or terminate warnings. This enables the test process to transition from fixed sequence execution to adaptive dynamic adjustment based on real-time risk perception. This effectively solves the technical problems of traditional verification methods, which are unable to quantify and evaluate logical and transient composite anomalies and lack association with latent fault knowledge, resulting in insufficient test depth and difficulty in discovering the progressive degradation risk inside the relay protection device. Attached Figure Description

[0025] Figure 1 This is a block diagram of the closed-loop logic verification system of the relay protection device of the present invention. Detailed Implementation

[0026] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0027] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0028] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0029] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0030] Example 1, see Figure 1 The present invention provides a block diagram of a closed-loop logic verification system for a relay protection device. Figure 1 The relay protection device closed-loop logic verification system shown includes: Synchronous acquisition module: When executing the test sequence, it acquires the action logic messages generated by the relay protection device, as well as multi-modal physical quantity data including the digital signal processor's operation cycle occupancy rate and the harmonic spectrum components of the analog sampling channel; The explanation is as follows: Action logic messages refer to the set of status information generated by the protection device during the testing process, reflecting the results of its logical judgment. Specifically, they include the start signal of the protection element, action output events, setting area switching records, soft pressure plate activation / deactivation status, self-diagnostic alarms of the protection device, and fault report summaries. They are obtained in real time from the logic processing unit of the protection device through the protection device communication module using standardized communication protocols (such as IEC61850 MMS service).

[0031] Feature recognition module: Extracts features from multimodal physical quantity data, and identifies abnormal feature vectors associated with degradation modes by matching them with a latent fault feature library through a transfer learning strategy that combines the topological prior knowledge of the relay protection device. The explanation is that the degradation mode refers to the performance degradation of specific components or functional units inside the relay protection device due to aging, wear, or inherent defects. This state has not yet triggered a logic error in the protection function, but has already shown observable changes in physical characteristics.

[0032] Risk assessment module: It performs spatiotemporal alignment of abnormal feature vectors and action logic messages, and obtains risk entropy value based on risk intensity, trend instability and correlation coupling quantification assessment; it compares the risk entropy value with the adjusted dynamic threshold to determine the risk level; it performs failure propagation analysis based on the relay protection device topology, and predicts the risk development trajectory by combining the time gradient of the risk entropy value, and generates a comprehensive assessment report including the current risk level and development expectations. The explanation is as follows: Risk intensity is obtained by dividing the absolute difference between the current value of the feature vector and the dynamic steady-state operating threshold by the threshold range width; Trend instability is quantified by time series curvature analysis, calculated by the mean of the second derivatives of the feature vector's change trajectory over the most recent ten sampling periods; In this embodiment of the invention, if the mean exceeds 0.1, the trend is determined to be unstable; Correlation coupling is obtained by calculating the Pearson correlation coefficient of different feature vectors in the same time period; Specifically, the calculation process for quantifying trend instability includes the following steps: Step 1: Construct the feature vector time series matrix; map the multimodal physical quantity data of the relay protection device at continuous sampling time t into m-dimensional feature vectors. , where m represents the feature dimension (such as DSP utilization, harmonic amplitude, etc.). The normalized component values ​​are used; data from the ten most recent sampling periods, including the current time t and past times, are selected to form a time series set. ; The second step is to calculate the second-order rate of change vector of the eigenvectors based on the discrete finite difference method; given the discrete characteristics of the digital sampling system, the backward difference formula is used to approximate the mathematical second derivative; for each time k in the time series (where... ), calculate its second-order difference vector The calculation formula is defined as follows: The second-order difference vector In a physical sense, it represents the acceleration or curvature of the trajectory of the feature vector in the multidimensional feature space, that is, the degree of drastic change in state. The third step is to calculate the scalar magnitude of the second-order difference vector. To transform a multidimensional vector into a comparable scalar index, the Euclidean norm (L2 norm) is used to calculate the second-order difference vector. Length of the module The calculation formula is: ,in For vectors The component in the i-th dimension; the magnitude It can comprehensively reflect the overall second-order change intensity of all feature dimensions at time k; Step 4: Calculate the trend instability index; calculate the modulus set obtained from the most recent ten sampling periods. Calculate the arithmetic mean, which is the mean of the second derivative of the trajectory of the feature vector change.

[0033] Strategy optimization module: Constructs a topology diagram of the functional units of the relay protection device, dynamically generates test sequence priorities based on critical path analysis, optimizes test excitation parameters in real time through a digital twin model, uses composite test waveforms to simultaneously excite the response characteristics of multiple associated nodes, and dynamically adjusts the test strategy according to the difference between the actual response and the expected response. In this invention, test excitation parameters refer to the set of adjustable input variables generated by the digital twin model in a simulation environment to drive the relay protection device to produce a specific response. Test excitation parameters include basic electrical properties such as amplitude, phase, and frequency of voltage and current signals applied to the relay protection device, injection ratio and phase of specific harmonic components, duration and switching sequence of each step in the test sequence, and fault initial angle and transition resistance value for simulating specific fault types (such as short circuit and open circuit). Test excitation parameters may cover composite test waveform characteristics simulating complex operating scenarios, such as the time constant of the decaying DC component and the voltage dip depth at the oscillation center. The process of the digital twin model optimizing the above test excitation parameters in real time is a closed-loop feedback control process. It uses the current parameters of the model to predict the expected response waveform of the relay protection device under a given excitation. It compares the actual response of the relay protection device collected in the actual test with the predicted response and calculates the difference (for example, using the root mean square error as a quantitative indicator). If the difference exceeds a preset threshold, it triggers an optimization algorithm (such as gradient descent). By calculating the gradient (partial derivative) of the difference with respect to each excitation parameter, it iteratively adjusts the parameter values ​​in the opposite direction of the gradient with the goal of reducing the difference.

[0034] Predictive maintenance module: Based on real-time test data and historical operation records, it calculates the remaining service life of key components by integrating physical failure mechanism analysis and data-driven trend prediction, and generates predictive diagnostic reports.

[0035] Furthermore, the specific implementation process corresponding to the synchronous acquisition module includes: The synchronous acquisition module is completed collaboratively by the relay protection device, the tester, the protection relay protection device communication module, and the tester control module. The tester, as the execution unit, receives the test sequence instructions from the tester control module, generates and outputs current and voltage excitation signals that conform to the preset timing and amplitude characteristics to the relay protection device. While outputting the excitation signal, the tester monitors the quality of the output waveform in real time through its built-in high-speed data acquisition card, including waveform distortion rate, amplitude accuracy, and phase accuracy. The explanation explains that the tester is responsible for generating and outputting current and voltage excitation signals that conform to preset timing and amplitude characteristics to the relay protection device. Simultaneously, it monitors the quality of the output waveform (including waveform distortion rate, amplitude accuracy, and phase accuracy) in real time through its built-in high-speed data acquisition card. The tester's control module, as the local control core, is responsible for parsing the test sequence, generating high-precision hardware synchronization pulses, and coordinating the timing of data acquisition. The relay protection device's communication module is dedicated to data interaction with the relay protection device's internal systems, directly reading the device's internal status data and fault waveform files through a standard protocol interface. The automation control module, acting as the upper-level scheduling center, coordinates the testing process, dynamically generates test sequence priorities based on risk assessment results, and interacts with the digital twin model to optimize the testing strategy.

[0036] In practice, while the driving tester applies current and voltage excitation signals that vary according to a preset sequence to the protection relay device, a synchronous data acquisition task is initiated. This task is completed by two logically independent but coordinated functional units. The protection relay device communication module interacts with the protection relay device's logic processing unit using a standardized communication protocol, detecting and recording in real time the information constituting the action logic message, such as the generated action events, setting groups, and soft pressure plate activation / deactivation status. To achieve synchronous acquisition and a precise time reference, the tester control module, at the start of outputting each set of excitation signals, will... A high-precision hardware synchronization pulse is generated. This pulse simultaneously triggers the high-speed data acquisition card of the tester and the internal waveform recording function of the protection relay device. The hardware synchronization pulse is output to the time synchronization or trigger interface of the protection relay device through the dedicated trigger signal line on the back panel of the tester. Alternatively, when hard-wiring is not available, a network time synchronization method based on the IEEE 1588 Precision Clock Protocol (PTP) is used to synchronize the clocks of the tester and the protection relay device to the microsecond level. The precise issuance of trigger events is achieved in the form of software commands, ensuring that the data streams obtained from the two independent sources have a unified time tag.

[0037] The collected multimodal physical quantity data specifically includes, but is not limited to: 1) The occupancy rate of the digital signal processor's operation cycle, which reflects the core load status of the calculation, is obtained through the internal status register of the direct-read relay protection device; 2) Reveal the instantaneous voltage and current waveforms and their harmonic spectrum components that demonstrate the linearity and stability of the analog sampling channel. This data is derived from the fault recording files of the protection relay device or a dedicated diagnostic data interface. 3) The bus ripple coefficient and noise amplitude of the input and output circuit power supply, which characterize the filtering performance of the power module, need to be measured at the designated test points of the relay protection device using the high impedance measurement probe of the tester. 4) Real-time junction temperature data indicating the operating environment and load of key integrated circuits, which is obtained through the temperature sensor built into the relay protection device or the external red leaf temperature measurement unit.

[0038] Furthermore, the specific implementation process corresponding to the feature recognition module includes: Based on the latent fault feature library, specifically, by collecting and analyzing historical fault case reports from field operations, accelerated aging test data from laboratory simulations, and the experience and knowledge of domain experts, reproducible degradation patterns with clear physical meaning are digitally modeled. Each identified degradation pattern, such as the deterioration of the differential nonlinearity of an analog-to-digital converter at a specific input level, the transmission delay drift of an optocoupler in a high-temperature environment, or the increase in the equivalent series resistance of an electrolytic capacitor under high ripple current, is deconstructed and mapped into one or more sets of quantifiable mathematical feature vectors. These mathematical feature vectors constitute the standard templates in the feature library, which not only include the normal range boundaries of feature values ​​but also record the typical trajectory of feature changes as the degree of degradation intensifies.

[0039] In one possible embodiment, after obtaining timestamped multimodal physical quantity data from the synchronous acquisition module, a multi-dimensional feature extraction process is initiated; differentiated signal processing and feature engineering strategies are adopted for different types of data. For continuous waveform data, the energy distribution within a specific frequency band is extracted by short-time Fourier transform or wavelet packet analysis. The short-time Fourier transform uses a Hanning window function to balance frequency resolution and spectral leakage, and the window length is adaptively selected according to an integer multiple of the power frequency period. The total harmonic distortion rate is calculated to evaluate waveform purity, and the phase jitter at the zero-crossing point of the signal is analyzed. For discrete state sequence data, such as action time records, calculate their statistical variance to measure the timing stability of logic execution. For slowly changing physical quantities such as chip temperature, focus on the temperature rise slope and steady-state value during the test cycle. All extracted real-time features are assembled into a comprehensive feature vector; a multi-level matching strategy is adopted to compare the comprehensive feature vector with templates in the latent fault feature library; fast screening is performed based on Euclidean distance or cosine similarity to lock in several candidate degradation modes; dynamic time warping algorithm or discriminant analysis based on Mahalanobis distance is introduced to calculate the degree of fit between real-time features and each candidate template, and output a list of potential degradation modes sorted by confidence.

[0040] In one possible implementation, the feature extraction process employs a parallel dual-path architecture: The first path is an interpretable path based on physical laws. This path is implemented through a differentiable signal processing layer integrated into the forward propagation computation graph. Specifically, it includes a differentiable short-time Fourier transform layer to extract energy distribution features of predefined key frequency bands, and a differentiable statistical variance calculation layer to quantify temporal stability from action time series. The second path is a data-driven path based on a multi-scale one-dimensional convolutional neural network. This path deploys multiple parallel convolutional branches with different temporal receptive fields to adaptively capture deep nonlinear features from local transients to global trends. The output features of the two paths are integrated through a gated attention fusion module. This module dynamically calculates the weights of each path's contribution based on the real-time signal-to-noise ratio and stationarity metric of the input data, achieving adaptive weighted fusion and ultimately outputting a more robust real-time feature vector.

[0041] Furthermore, the specific implementation process of the risk assessment module includes: Step 301: Risk assessment aligns the abnormal feature vectors and action logic messages spatiotemporally based on a unified timestamp to establish a time-series correlated dataset. The aligned data is then processed using a multidimensional risk assessment matrix, which comprehensively quantifies risk intensity, trend instability, and correlation coupling. Risk intensity is obtained by calculating the relative deviation between the current value of the feature vector and the dynamic operating threshold. Trend instability is quantified using nonlinear dynamics analysis to assess the complexity of the time series. Correlation coupling is analyzed using multivariate statistical methods to examine the intrinsic correlation characteristics between different feature vectors. The explanation is as follows: In specific implementation, the risk intensity is calculated by dividing the absolute difference between the current value of the feature vector and the dynamic steady-state operating threshold by the threshold range width; trend instability is quantified through time series curvature analysis (this indicator reflects the severity of fluctuations in the acceleration of state change, and its magnitude directly characterizes the swiftness and instability of the relay protection device's state change), and the mean of the second derivative of the feature vector change trajectory within a preset number of sampling periods is calculated for evaluation; the correlation coupling degree is calculated based on the Pearson correlation coefficient of different feature vectors in the same time period; Step 302: The dynamic threshold adjustment adopts a multi-factor coupled regression method, and establishes an adaptive adjustment model by comprehensively considering ambient temperature, power grid load rate and cumulative operating time of relay protection devices; among them, the temperature coefficient is calculated by the Arrhenius accelerated model, the load coefficient is calculated by the inverse power law model, and the aging coefficient is obtained by fitting the Weibull distribution. Step 303: Risk Entropy Calculation and Classification: The risk entropy value calculation adopts a confidence-based variable weight allocation mechanism, which integrates the evaluation results of three dimensions: risk intensity, trend instability, and correlation coupling. The initial weight of each dimension is determined according to its contribution to historical failures, and the confidence coefficient is evaluated through data stability within a sliding window. The weight allocation process uses normalization to ensure the coordination of the weights of each dimension, and finally generates a continuous risk entropy value in the range of 0 to 1. In the preferred implementation scheme, the risk level classification adopts a dynamic clustering method based on Mahalanobis distance. A feature space is constructed based on risk intensity, trend instability, and correlation coupling degree. The boundary conditions of the risk level are determined through iterative optimization. The number of clusters is adaptively determined according to the distribution characteristics of historical data. The cluster centers are dynamically updated through the mean shift algorithm. The level boundary is described by a confidence ellipse to ensure that the risk level classification reflects both the essential structure of the data and has statistical significance. For ease of understanding, the following example illustrates the mapping relationship between risk entropy value and risk level: When the risk entropy value is less than the first threshold T1, it is determined to be a low-risk level; when the risk entropy value is greater than or equal to the first threshold T1 and less than the second threshold T2, it is determined to be a medium-risk level; when the risk entropy value is greater than or equal to the second threshold T2, it is determined to be a high-risk level; where thresholds T1 and T2 are preset critical values, and T2 is greater than T1; for example, a risk entropy value below 0.3 is low risk, 0.3 to 0.7 is medium risk, and above 0.7 is high risk; the risk level, as a key judgment result, is directly written into the comprehensive assessment report to trigger different levels of early warning and guide the formulation of subsequent deep diagnostic test sequence generation and maintenance strategies; In one possible embodiment, the risk entropy value is calculated as follows: Risk Entropy Value = Normalized Risk Intensity × 0.5 + Normalized Trend Instability × 0.3 + Normalized Correlation Coupling Degree × 0.2; where the trend instability coefficient is the value after maximum-minimum normalization, and the correlation coupling degree is the proportion of strongly correlated feature vectors; where all three parameters are dimensionless values, the normalized risk intensity refers to the value obtained by mapping the absolute value of the relative deviation between the current value of the feature vector and the dynamic threshold using the Sigmoid function; the normalized trend instability refers to the value after maximum-minimum normalization of the second-order difference magnitude mean (i.e., discrete curvature) of the feature vector change trajectory, which is positively correlated with the trajectory curvature, the larger the curvature, the more drastic the state change, and the higher the instability value; the normalized correlation coupling degree refers to the ratio of the number of strongly correlated feature vectors with a correlation coefficient exceeding a preset threshold (e.g., 0.8) to the total number of features in the same period.

[0042] Step 304 Failure Propagation Analysis and Trajectory Prediction: A directed topology graph is constructed based on the physical connection and logical dependency of the functional units of the relay protection device. The node weights in the graph are dynamically adjusted according to the criticality of the functional units, and the edge set weights are determined based on electrical characteristics and signal transmission characteristics. The functional units corresponding to the abnormal feature vectors are taken as the propagation source nodes, and potential propagation paths and impact ranges are identified through graph theory analysis.

[0043] In practice, the failure propagation analysis employs a probabilistic prediction method based on a random walk model. Starting from the node corresponding to the abnormal feature vector in the constructed directed topology graph, a random walk algorithm with restart is used to simulate the risk propagation process. The walk transition probability is dynamically adjusted based on the edge set weights and node states. The hit probability of the walk path serves as a probability estimate of the propagation risk, and the expected time window for each propagation path is predicted by combining the time gradient of the risk entropy value. Furthermore, a directed topology graph is constructed based on the physical connections and logical dependencies of the functional units of the relay protection device. Node weights are dynamically adjusted according to the criticality of the corresponding functional unit, and edge set weights are determined based on electrical connection impedance and signal transmission logical delay. A graph traversal algorithm combined with edge set weights is used to calculate propagation paths, and the propagation conditional probabilities on each path are statistically analyzed based on historical failure data to determine the potential impact range and path trigger probability. A time prediction model is established using a regression algorithm, taking the real-time monitored risk entropy gradient, the weights and conditional probabilities of the current propagation path as inputs, and incorporating the time series characteristics of historical failure cases to extrapolate the expected time window for risk to develop into a fault along each critical propagation path. The final output risk development trajectory includes the set of affected functional units, the trigger probability of each propagation path, and the corresponding time prediction results, providing accurate input for the subsequent in-depth diagnostic testing of the strategy optimization module.

[0044] Furthermore, the specific implementation process corresponding to the strategy optimization module includes: Step 401: Risk Path Analysis and Test Sequence Generation The risk propagation path in the comprehensive assessment report is analyzed, and a directed topology graph with functional units as nodes and electrical connections as edges is constructed. Based on the critical path analysis algorithm in graph theory, the test priority weight of each path is calculated. The weight value is determined by the ratio of the number of high-risk units in the path to the path length. The number of high-risk units is based on the risk level statistics determined by the risk assessment module, and the path length is calculated by the shortest path between nodes in the topology graph. Weight normalization is performed to ensure comparability between different paths. The test sequences are arranged in descending order of weight value to ensure that high-risk paths are tested first. Each test sequence unit contains specific test excitation parameters. For example, the voltage amplitude increases in increments of 5% within the range of 80% to 120% of the rated value, the phase shifts in increments of 1 degree within the range of ±10 degrees, and the frequency is adjusted in increments of 0.1 Hz within the range of 45 Hz to 55 Hz. Digital twin-guided test optimization: A digital twin model of the relay protection device is established, which includes an electromagnetic transient calculation module, a logic judgment module, and a physical characteristic simulation module. Before each test excitation output, the digital twin model predicts the expected response waveform of the relay protection device based on the current test parameters. During actual testing, the difference between the measured response and the predicted response is compared. The difference is calculated using the root mean square error, with the formula being: the square root of the sum of the squares of the differences between the measured and predicted values ​​at each sampling point divided by the total number of sampling points. When the difference exceeds 5%, the gradient descent algorithm is activated to adjust the test parameters: the new voltage amplitude is equal to the original amplitude minus the learning rate (set to 0.01) multiplied by the partial derivative of the difference with respect to the amplitude. The phase and frequency are iteratively updated according to the same logic until the difference is less than 2% or the maximum number of iterations is reached. Composite Test Waveform Design and Real-Time Adjustment: Composite test waveforms are designed for associated nodes, superimposing the basic power frequency signal with specific harmonic components. The amplitude of the second harmonic component is set to 5% of the basic waveform, and the amplitude of the third harmonic component is set to 3% of the basic waveform, with each harmonic phase maintaining a fixed offset from the basic waveform. Key parameters are monitored in real time during the test: when the chip temperature exceeds 85 degrees Celsius or the power supply ripple factor exceeds 3%, the test is immediately paused, and the cooling system operating parameters are adjusted according to the preset cooling sequence. Based on the matching degree between the real-time response and the expected response, a strategy is dynamically selected to continue the current test path, switch to a backup test path, or terminate the test.

[0045] The explanation is as follows: The functional unit association topology diagram is established based on the wiring diagram and datasheet provided by the relay protection device manufacturer. Nodes represent functional units such as ADC sampling modules, DSP processing units, and output relays. Edges represent electrical connections or data flow relationships, and the weights of the edges are quantized by impedance values ​​or signal transmission delays. The digital twin model is trained using historical test data. The training set contains one thousand sets of relay protection device response data under different test stimuli. Mean square error is used as the loss function, and the Adam optimizer is used for parameter updates. The adjustment range of the test stimulus parameters is strictly limited to the safe operating range of the relay protection device.

[0046] Furthermore, the specific implementation process corresponding to the predictive maintenance module includes: Remaining service life (RUL) calculation method: The remaining service life of key components is determined by combining physical failure mechanism analysis with data-driven trend prediction. Based on the electrothermal stress accelerated aging model, the theoretical life loss of core components of the relay protection device (such as electrolytic capacitors and optocouplers) is calculated using the Arrhenius equation. A Long Short-Term Memory (LSTM) network model is adopted, with input features including the monthly average change trajectory of the digital signal processor's computing cycle utilization rate and the weekly peak sequence of the power bus ripple coefficient, among other time series data. The LSTM model structure contains two hidden layers (50 units per layer), using mean squared error as the loss function and trained through an adaptive moment estimation optimizer. Finally, the RUL is taken as the smaller value between the physical model and the data-driven prediction result. Maintenance Recommendation Decision-Making Methodology: A multi-objective decision-making method is employed to generate maintenance recommendations, simultaneously evaluating the effectiveness of equipment reliability improvements, estimated maintenance costs, and potential power outage losses. Based on the aforementioned Reliability Limit (RUL) calculation results, a multi-objective weighted scoring system is constructed, covering reliability improvement effectiveness (quantified by the deviation between predicted RUL and the threshold), maintenance costs (spare parts costs and labor hours), and power outage losses (calculated based on load level and downtime). The weights of each objective are determined using the Analytic Hierarchy Process (AHP) (consistency ratio CR < 0.1), and different maintenance schemes (such as preventative replacement and condition-based maintenance) are ranked based on their overall benefits. Diagnostic report generation mechanism: Predictive diagnostic reports are automatically generated through predefined structured templates. The templates include core content such as equipment health status summary, multi-dimensional performance analysis data, recommended maintenance solutions and their basis, and comparative analysis of alternative solutions. The reports are linked to a historical similar case library and provide auxiliary decision support through knowledge graph matching.

[0047] Example 2 addresses the problem that existing relay protection device testing methods, when faced with discrepancies between actual and expected responses, cannot transform observed phenomena (such as action delays and waveform distortion) into a quantifiable, interpretable, and directly guiding comprehensive risk assessment, thus hindering the intelligent leap from passive fault detection to proactive risk warning and predictive maintenance. Example 2 further includes a module that dynamically adjusts the testing process based on the actual action response of the relay protection device, comprising: When a difference is identified between the actual response and the expected response, the logical path consistency index, which characterizes the deviation of logical behavior, and the transient feature fidelity index, which characterizes the distortion of transient waveforms, are calculated based on the current state and transition rules of the event-driven state machine. The explanation explains that the current state of an event-driven state machine is an abstract logical link defined by a unique state identifier (format: protection type code-voltage level-function number), state attributes (divided into initial / intermediate / terminal state types, and core / important / auxiliary levels), and state data (inherited from three sets of parameters: electrical quantity characteristics, state judgment, and timing constraints of the corresponding test node). The transition rules explicitly define the logical and temporal conditions that must be met to transition from one current state to another target state. The core of these rules is the transition trigger condition, which requires at least one electrical quantity trigger condition (which must simultaneously satisfy amplitude thresholds and phase constraints, and this condition must be continuously valid for at least two consecutive power frequency cycles) and / or one time trigger condition (reaching a preset delay) to be simultaneously satisfied before a state transition will occur. Specifically, the unique status identifier is configured as a three-segment encoding structure of protection principle feature code - grid voltage level - logic timing function number, serving as a logical addressing index for retrieving reference data and calculating differences during logic verification. The protection principle feature code corresponds to a specific algorithm set to which the current logic belongs (such as longitudinal differential, distance I-segment, and zero-sequence overcurrent algorithm). Based on this feature code, the corresponding theoretical action equations and criterion sets are indexed and loaded from the database of the digital twin model to establish the calculation benchmark. The grid voltage level defines the parameter sensitivity range and procedural constraints applicable to the current logic (such as differentiating the time constraints of 220kV and 110kV systems in reclosing logic), used to limit the expected timing threshold within the physical constraints of a specific grid level when calculating the logic path consistency index. The logic timing function number is used to mark discrete nodes in the protection action chain (specifically including 01-starting element action, 02-fault quantity measurement and holding, 03-output relay drive, and 04-reclosing charging window). This function number constitutes the basic comparison unit when calculating the normalized editing distance. During execution, the continuous action response of the relay protection device is discretized into an actual state sequence composed of the aforementioned unique state identifiers. When the actual response sequence has missing nodes or disordered timing, the specific missing or abnormal logical link is located using the unique state identifiers. The discretized state identifier sequence is then input into the information entropy calculation model to generate a logical path consistency index that characterizes the disorder of logical execution, thereby realizing the location of the missing logical link and the quantitative characterization of the degree of logical disorder.

[0048] The above index is fused with the abnormal feature vectors associated with the degradation patterns matched from the latent fault feature library through a knowledge embedding transfer learning strategy (feature-level fusion achieved through vector concatenation), and then input into a test risk quantification model based on support vector machine regression. The test risk quantification model outputs a degradation risk coefficient. Based on the coupling relationship between the preset threshold range to which the degradation risk coefficient belongs and the propagation probability, the built-in decision rule engine adaptively selects and performs one of the following operations: generating composite test waveform instructions for in-depth fault location, switching to a targeted backup test branch to isolate risks, or terminating the test and generating a risk warning report containing quantified risk entropy values ​​and impact range. The explanation is as follows: The degradation risk coefficient integrates the logic path consistency index, which characterizes the deviation of logical behavior, and the transient feature fidelity index, which characterizes the distortion of transient waveforms, with the abnormal feature vector matched from the latent fault feature library through knowledge embedding transfer learning to form a comprehensive feature vector. This vector is then input into a support vector machine regression model with radial basis functions as the kernel and trained on historical fault case data. Through an internal nonlinear mapping function, the input features are comprehensively mapped into a specific risk value. This directly quantifies the comprehensive risk probability and severity of the latent component degradation in the relay protection device revealed by the current test, which may lead to the failure of the protection function. The higher the value, the more urgent the risk and the more serious the consequences. It serves as a quantitative basis for subsequent intelligent decision-making (such as in-depth testing, branch switching, or termination warning).

[0049] The propagation probability is obtained through logical and statistical analysis based on the functional unit association topology diagram of the relay protection device. The calculation logic is as follows: First, based on the output of the test risk quantification model or the abnormal feature vector, the initial suspected deterioration unit node is located in the topology diagram; then, based on the connection edges and dependencies defined in the topology diagram, all possible paths for the fault or deterioration effect to propagate from the initial node to other associated nodes (such as downstream logic units, exit units, etc.) are analyzed; finally, a probability value is assigned to each possible path. This probability value can be obtained by analyzing the actual propagation statistical frequency of similar deterioration in historical fault cases, or calculated by an expert rule model based on functional dependency strength and signal path reliability, thereby quantifying the possibility of the deterioration effect spreading inside the device.

[0050] The intelligent decision-making process automatically triggers corresponding operations based on two core parameters: the degradation risk coefficient and the propagation probability, through preset decision rules. The logic is as follows: The calculated degradation risk coefficient is compared with multiple preset risk threshold ranges to initially determine the risk level (e.g., low, medium, high). Simultaneously, the propagation probability obtained from topology analysis is used to assess the potential scope and speed of risk impact spread. Then, a comprehensive judgment is made based on a built-in decision matrix or rule engine: For example, when the risk coefficient is in the medium range and the propagation probability is low, the decision generates an optimization instruction to trigger a composite test waveform for in-depth positioning; if the risk coefficient is high and the propagation probability points to a critical redundant unit, the decision switches to a backup test branch targeting the suspected degradation unit to isolate the risk and continue testing; when the risk coefficient reaches the highest critical value and the propagation probability indicates that the impact may rapidly affect the core protection function, the decision immediately terminates the test and automatically generates a risk warning report integrating a quantitative assessment of risk entropy and an analysis of the expected impact range to ensure safety and prompt emergency intervention.

[0051] Furthermore, the test risk quantification model integrates an adaptive learning module to achieve predictive maintenance support: the adaptive learning module continuously collects degradation risk coefficients, corresponding abnormal feature vectors, final verified failure modes, and component life impact data from historical tests to form a training dataset; Based on the training dataset, the model parameters are dynamically updated through an online learning algorithm, driving the digital twin model to optimize the simulation accuracy of the performance degradation of key components of the relay protection device. By performing regression analysis on the degradation trajectories of the logic path consistency index and transient feature fidelity index in each test, and combining the risk development trajectory obtained from the failure propagation analysis, a performance degradation knowledge graph of key components is constructed. Based on this knowledge graph, the remaining service life of key components is calculated and predicted, and test sequences for high-degradation-risk units are prioritized in subsequent tests, realizing a closed loop from logic verification to predictive maintenance.

[0052] The explanation explains that the logical path consistency index compares the actual and expected state transition sequences. First, it calculates the normalized edit distance (reflecting differences in logical order) and the normalized average time difference (reflecting temporal deviations). Then, based on these two values, it estimates a two-dimensional discrete distribution and calculates its information entropy (i.e., calculates information entropy using an information entropy calculation model). This entropy value is the index; a higher value indicates more chaotic and disordered logical behavior. The transient characteristic fidelity index extracts the decaying DC component time constant and the high-frequency transient component propagation rate from the actual response, calculates their relative errors compared to theoretical expectations, and... The two errors are weighted geometrically and averaged according to their importance in the project (usually the time constant has a higher weight, set to 0.6-0.8). The lower the value, the more severe the transient waveform distortion. The degradation risk coefficient is formed by fusing the two indices with the abnormal feature vectors matched from the latent fault feature library. This comprehensive feature vector is then input into a support vector machine regression model trained on historical data with radial basis functions as the kernel. The continuous scalar value output by the model is the coefficient. The higher the value, the greater the risk of latent degradation in the relay protection device and the greater the potential for failure.

[0053] Rationale for this setup: This setup constructs a multi-layered, quantifiable, comprehensive risk assessment system that integrates microscopic electrical characteristics, macroscopic logical behavior, and historical experience. The logic path consistency index (information entropy) can sensitively capture any disordered sequence or timing anomalies in the execution of protection logic, reflecting the degree of confusion in the relay protection device's thinking. The transient characteristic fidelity index (weighted geometric mean) focuses on the fidelity of key transient components, and its design amplifies the deviation of parameters (such as attenuated DC) that have a decisive impact on protection criteria, directly related to the sensory accuracy of the relay protection device. Both provide objective and complementary anomaly measurements from the two physical dimensions of discrete logic and continuous signals, respectively. Finally, the degradation risk coefficient (support vector machine regression) is not simply a superposition of the previous two, but rather intelligently correlates and regresses them with a vast historical fault knowledge base (latent fault characteristics), enabling the system to quantify and locate the currently observed anomaly pattern to a specific risk level based on rich prior knowledge. This provides accurate and reliable numerical basis for subsequent intelligent decisions (such as in-depth testing, branch switching, or termination warning), achieving a leap from phenomenon description to risk prediction.

[0054] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A closed-loop logic verification system for relay protection devices based on direct data acquisition, characterized in that, include: Synchronous acquisition module: When executing the test sequence, it acquires the action logic messages generated by the relay protection device, as well as multi-modal physical quantity data including the digital signal processor's operation cycle occupancy rate and the harmonic spectrum components of the analog sampling channel; Feature recognition module: Extracts features from multimodal physical quantity data, and identifies abnormal feature vectors associated with degradation modes by matching them with a latent fault feature library through a transfer learning strategy that combines the topological prior knowledge of the relay protection device. Risk assessment module: Spatiotemporally aligns the abnormal feature vector with the action logic message, and obtains the risk entropy value based on risk intensity, trend instability and correlation coupling quantification assessment; The risk level is determined by comparing the risk entropy value with the adjusted dynamic threshold. Failure propagation analysis is performed based on the topology of the relay protection device. The risk development trajectory is predicted by combining the time gradient of the risk entropy value, and a comprehensive assessment report including the current risk level and development expectations is generated. Strategy optimization module: Constructs a topology diagram of the functional units of the relay protection device, dynamically generates test sequence priorities based on critical path analysis, optimizes test excitation parameters in real time through a digital twin model, uses composite test waveforms to simultaneously excite the response characteristics of multiple associated nodes, and dynamically adjusts the test strategy according to the difference between the actual response and the expected response. Predictive maintenance module: Based on real-time test data and historical operation records, it calculates the remaining service life of key components by integrating physical failure mechanism analysis and data-driven trend prediction, and generates predictive diagnostic reports.

2. The closed-loop logic verification system for relay protection devices based on direct data acquisition according to claim 1, characterized in that, The synchronous acquisition includes: generating a hardware synchronization pulse at the start of the test excitation signal output by the tester; obtaining the digital signal processor's operation cycle utilization rate by reading the internal status register of the relay protection device; obtaining the harmonic spectrum components by accessing the fault recording file of the relay protection device; obtaining the ripple coefficient of the input and output circuit power bus by measuring it with the high impedance probe of the tester; and acquiring real-time junction temperature data by acquiring it with the built-in temperature sensor of the relay protection device. All acquired data are embedded with a unified timestamp to form a spatiotemporal correlated dataset.

3. The closed-loop logic verification system for relay protection devices based on direct data acquisition according to claim 1, characterized in that, The construction and matching process of the latent fault feature library adopts a knowledge embedding transfer learning strategy, including: A graph-structured prior knowledge base is constructed based on historical failure cases and domain expert experience. The prior knowledge base defines the association between degradation modes and topological relationship nodes of functional units of relay protection devices. In the process of cross-domain feature distribution alignment, the prior knowledge of the relay protection device topology is transformed into structured constraint terms that characterize the physical connection logic, and a topology constraint loss function is constructed. The topology constraint loss function is used to quantify the degree of deviation between the feature matching result and the physical topology logic of the relay protection device. An alternating optimization strategy is adopted. First, the parameters of the matching network are fixed, and the topological embedding representation of the feature space is optimized by minimizing the deviation. Then, the topological embedding representation is fixed, and the main loss function of the matching network is optimized. The topology constraint loss function and the main loss function of the deep matching network jointly participate in adversarial training so that the matching results output by the matching network conform to the physical connection and signal flow logic inside the relay protection device.

4. The closed-loop logic verification system for relay protection devices based on direct data acquisition according to claim 3, characterized in that, The risk level classification adopts a dynamic clustering method based on Mahalanobis distance, which includes: constructing a feature space based on risk intensity, trend instability and correlation coupling degree, and determining the boundary conditions of risk level through iterative optimization; wherein the number of clusters is adaptively determined according to the distribution characteristics of historical data, the cluster centers are dynamically updated through the mean shift algorithm, and the level boundary is described by a confidence ellipse.

5. The closed-loop logic verification system for relay protection devices based on direct data acquisition according to claim 1, characterized in that, The failure propagation analysis adopts a probabilistic prediction method based on a random walk model. In the constructed functional unit association topology of the relay protection device, the node corresponding to the abnormal feature vector is taken as the starting point, and the risk propagation process is simulated by a random walk algorithm with restart. The walk transition probability is dynamically adjusted according to the edge set weight and node state. The hit probability of the walk path is used as a probability estimate of the propagation risk. Combined with the time gradient of the risk entropy value, the expected time window of each propagation path is predicted.

6. The closed-loop logic verification system for relay protection devices based on direct data acquisition according to claim 5, characterized in that, The test sequence priority generation includes: constructing a topology diagram of the relay protection device functional units based on the risk propagation path; calculating the test priority weight of each path through critical path analysis; the test priority weight is determined by the ratio of the number of high-risk units in the path to the path length; and generating the test sequence order by arranging the test priority weights in descending order. The test optimization includes: predicting the expected response waveform using a digital twin model, calculating the root mean square error between the measured response and the predicted response, iteratively updating the response using a gradient descent algorithm based on real-time difference feedback, and introducing a momentum term to accelerate convergence and avoid local optima.

7. The closed-loop logic verification system for relay protection devices based on direct data acquisition according to claim 6, characterized in that, The dynamic threshold used to determine the risk level establishes a closed-loop feedback link between its adjustment process and the digital twin model that optimizes the test stimulus parameters; the parameter adjustment trajectory generated by the digital twin model based on the difference between the measured response and the expected response is fed back to the regression model in real time as an additional feature characterizing the sensitivity of the relay protection device to the test stimulus under the current specific operating conditions.

8. The closed-loop logic verification system for relay protection devices based on direct data acquisition according to claim 7, characterized in that, When calculating the remaining service life of key components, the set of affected relay protection device functional units, the trigger probability of each propagation path, and the expected time window information contained in the risk development trajectory are encoded into a weighted time decay graph network. The time decay graph network is used as a priori structural constraint and trained together with the time series data of historical performance parameters into a graph convolutional recurrent neural network. In the graph convolutional recurrent neural network, the topological structural constraints are used to guide feature propagation.

9. The closed-loop logic verification system for relay protection devices based on direct data acquisition according to claim 1, characterized in that, It also includes a module that dynamically adjusts the test process based on the actual action response of the relay protection device, including: When a difference is identified between the actual response and the expected response, the logical path consistency index, which characterizes the deviation of logical behavior, and the transient feature fidelity index, which characterizes the distortion of transient waveforms, are calculated based on the current state and transition rules of the event-driven state machine. The above indexes are fused with abnormal feature vectors associated with degradation patterns matched from a latent fault feature library through a knowledge embedding transfer learning strategy, and then input into a test risk quantification model based on support vector machine regression. The test risk quantification model outputs a degradation risk coefficient. Based on the coupling relationship between the preset threshold range to which the degradation risk coefficient belongs and the propagation probability, the built-in decision rule engine adaptively selects and performs one of the following operations: generates a composite test waveform command for in-depth fault location, switches to a targeted backup test branch to isolate the risk, or terminates the test and generates a risk warning report containing a quantified risk entropy value and the scope of impact.

10. The closed-loop logic verification system for relay protection devices based on direct data acquisition according to claim 9, characterized in that, The test risk quantification model integrates an adaptive learning module to support predictive maintenance: the adaptive learning module continuously collects degradation risk coefficients, corresponding abnormal feature vectors, final verified failure modes, and component life impact data from historical tests to form a training dataset; Based on the training dataset, the model parameters are dynamically updated through an online learning algorithm, driving the digital twin model to optimize the simulation accuracy of the performance degradation of key components of the relay protection device. By performing regression analysis on the degradation trajectories of the logic path consistency index and transient feature fidelity index in each test, and combining the risk development trajectory obtained from the failure propagation analysis, a performance degradation knowledge graph of the key components is constructed. Based on this knowledge graph, the remaining service life of the key components is calculated and predicted, and test sequences for high-degradation-risk units are prioritized in subsequent tests.

Images