An event knowledge graph-based automated device fault-tolerant test method and system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- STATE GRID HUBEI ELECTRIC POWER RES INST
- Filing Date
- 2026-02-06
- Publication Date
- 2026-06-02
Smart Images
Figure CN122132992A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system automation testing technology, specifically to an automated equipment fault-tolerant testing method and system based on event knowledge graphs. Background Technology
[0002] As the core hub of the smart grid, digital substations have achieved standardized information modeling and networked communication in accordance with the IEC 61850 standard. Their operation is manifested as a series of discrete but strictly sequential and logical communication event flows. To ensure the reliability of automated equipment under complex fault scenarios, fault-tolerant testing has become a key step before equipment commissioning. Knowledge graph technology, with its powerful semantic expression and relational reasoning capabilities, provides an ideal tool for describing complex fault event logic. Deeply integrating event-based thinking with knowledge graph technology to construct a fault-tolerant testing method that can automatically generate, execute, and evaluate complex event scenarios is of great significance for improving the reliability of digital substations.
[0003] However, traditional test case generation relies on human experience, making it difficult to extract related event chains from massive amounts of historical fault data and construct test scenarios with rich contextual relationships, resulting in incomplete test coverage. On the other hand, fault injection lacks temporal logic and physical model constraints, and the generated abnormal message sequences are disconnected from real fault characteristics, making it difficult to effectively verify the overall logic of the device when facing the impact of related events. In addition, the judgment of test results often relies on simple threshold comparisons or status checks, lacking the ability to comprehensively analyze all response signals of the device within the entire event time window, failing to identify abnormal patterns from a macro perspective of "event handling," and potentially missing hidden defects that do not trigger hard thresholds but exhibit abnormal patterns. Summary of the Invention
[0004] The purpose of this invention is to address the problems in existing fault-tolerant testing methods, such as test case generation relying on human experience, difficulty in automatically extracting related event chains from multi-source heterogeneous data to construct scenario-based test cases, lack of temporal logic and physical model constraints in fault injection leading to a disconnect between abnormal message sequences and real fault characteristics, and lack of comprehensive intelligent analysis capabilities for multi-dimensional response signals and a self-evolution mechanism for test knowledge in test result analysis, thus failing to form a continuously learning adaptive intelligent testing system. Therefore, this invention proposes an automated equipment fault-tolerant testing method and system based on event knowledge graphs.
[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:
[0006] An automated device fault-tolerant testing method based on event knowledge graph includes the following steps:
[0007] S1. Extract entities, relationships, and attributes of fault events from multi-source heterogeneous data, construct triple knowledge representations, eliminate entity ambiguity through semantic alignment technology, and form a unified fault event knowledge graph.
[0008] S2. Based on the fault event knowledge graph, a weighted subgraph traversal algorithm is used to calculate the path weight according to the relation type weight, historical frequency and severity. High-weight fault signal subgraphs are extracted from the fault event knowledge graph, and structured test cases containing test scenarios, fault sequences and judgment criteria are automatically generated based on the high-weight fault signal subgraphs.
[0009] S3. Establish a dynamic mapping rule base from fault signals to communication message parameters. Based on the temporal logic and causal relationship of the fault signal sub-graph, and combined with the physical model calculation, generate an abnormal message sequence that conforms to the power communication protocol, and perform scenario-based fault injection.
[0010] S4. Collect the multi-dimensional response signal of the device under test after performing the scenario-based fault injection in S3, construct a spatiotemporal graph data with response time series data as node features and high-weight fault signal sub-graph as spatial structure, perform feature extraction and pattern recognition through spatiotemporal graph convolutional network, output multi-label classification results, update the entities, attributes and relationships of the knowledge graph according to the multi-label classification results, and realize the closed-loop evolution of the knowledge graph.
[0011] Based on the above technical solution, the present invention can be further improved as follows.
[0012] Furthermore, the multi-source heterogeneous data in S1 includes the whole-station system configuration file SCD, production management system PMS defect records, fault waveform files, protection action reports, and equipment ledgers;
[0013] The entity types include faulty equipment, fault type, fault level, fault signal, and associated equipment;
[0014] The semantic alignment technique specifically involves: generating high-dimensional word vectors using a language model pre-trained on a corpus in the power industry; calculating the semantic similarity between entities using a cosine similarity algorithm enhanced by a power industry dictionary; and merging entities that exceed a preset threshold. The similarity calculation formula is as follows:
[0015]
[0016] in, and Entities and Word vectors, For dictionary matching functions, and These are the weighting coefficients.
[0017] Furthermore, the weighted subgraph traversal algorithm in S2 specifically includes: assigning basic weights to relation types in the knowledge graph, where the weight of "inevitably leading to" relations is 0.9-1.0, the weight of "possibly accompanying" relations is 0.6-0.8, and the weight of "precondition" relations is 0.7-0.9; and introducing dynamic correction factors, including historical frequency factors. Severity factor and timeliness factor The final weight value is calculated using the weighting formula:
[0018]
[0019] in, Based on the weights, To adjust the coefficients, a priority-based traversal strategy is adopted, starting from the target fault node, maintaining a dynamic priority queue, always selecting the path with the highest current weight for expansion, setting a dual termination condition of path weight accumulation threshold and search depth limit, determining key elements through topology analysis, and performing redundancy elimination and structural optimization.
[0020] Furthermore, the dynamic mapping rule base in S3 stores mapping rules from abstract fault signals to specific message parameters conforming to the IEC61850 standard, including IEC61850-9-2SV sampling values, GOOSE general object-oriented substation events, and MMS manufacturing message specifications. The scenario-based fault injection specifically includes: performing deep analysis and time-series planning on the fault signal sub-map, establishing an event sequence with strict time constraints by analyzing the causal relationships and weight information between nodes, and calling the electromagnetic transient model to calculate the waveform characteristics in the fault process in real time, dynamically generating the corresponding abnormal message sequence.
[0021] Furthermore, the abnormal message generation process in S3 includes: for current transformer CT saturation abnormalities, calling the electromagnetic transient model to calculate the waveform distortion characteristics during the saturation process:
[0022]
[0023] in, For primary current, For magnetic flux, The saturation flux is used; for the merging unit out-of-step scenario, the phase accumulation error caused by clock drift is simulated by adjusting the sampling counters of multiple merging units; for the voltage transformer PT open circuit fault, the single-phase grounding scenario on the load side is simulated by setting the voltage channel value to zero or an abnormally low value and keeping the current channel normal.
[0024] Furthermore, the response signal acquisition in S4 includes the hard contact action signal of the protection device, the GOOSE message of the smart terminal, the data sent by the measurement and control device, the sampling record of the fault recorder, and the communication message of the network analyzer. The acquisition system uses a working mechanism that combines intelligent caching and real-time processing to cache the raw data in a high-speed storage device, while simultaneously initiating a real-time preprocessing process, including data cleaning, format standardization, and outlier detection. A test timeline is established to precisely align the injected fault sequence with the device's response signal.
[0025] Furthermore, the spatiotemporal graph convolutional network ST-GCN model in S4 achieves intelligent analysis through a multi-level feature extraction mechanism. In the spatial dimension, the graph convolutional layer learns the mutual influence relationships between devices by aggregating neighbor node information.
[0026]
[0027] in, To add self-connected adjacency matrices, For degree matrix, For the first The feature matrix of the layer, These are trainable parameters; in the time dimension, the temporal convolutional layer captures the changing trends and dynamic characteristics of each signal through a sliding time window; the model output includes the overall judgment of whether the test passed, the type of anomaly detected, the severity level of the anomaly, the time period in which the anomaly occurred, and the confidence level assessment.
[0028] Furthermore, the knowledge graph closed-loop evolution in S4 includes three levels: the entity level, which automatically identifies newly discovered fault modes during the testing process, creates corresponding entity nodes, and establishes a device association network; the attribute level, which dynamically adjusts node weight coefficients based on test statistics to correct the evaluation indicators of fault occurrence frequency and impact; and the relationship level, which uses association rule mining algorithms to discover hidden device association characteristics and establish a multi-dimensional relationship network.
[0029] Furthermore, the closed-loop evolution of the knowledge graph adopts a progressive knowledge update strategy. New knowledge elements first enter the knowledge base to be verified, and can only be formally added to the database after multiple tests and verifications and a confidence level reaching a preset threshold. A version history management system is established to record every important change to the knowledge graph, supporting retrospective analysis and comparative verification of the knowledge structure. By constructing a complete closed loop of "test-analysis-learning," an adaptive intelligent testing system is formed, enabling testing capabilities to continuously improve with knowledge accumulation. An automated equipment fault-tolerant testing system based on event knowledge graphs is used to execute the method described above, the system comprising:
[0030] The knowledge graph construction module is used to extract entities, relations and attributes of fault events from multi-source heterogeneous data to form triples, and to construct a fault event knowledge graph by eliminating ambiguity through semantic alignment.
[0031] The test case generation module is used to calculate path weights based on relation type weights, historical frequency and severity using a weighted subgraph traversal algorithm, extract high-weight fault signal subgraphs from the fault event knowledge graph, and automatically generate structured test cases containing test scenarios, fault sequences and judgment criteria based on the high-weight fault signal subgraphs.
[0032] The fault injection execution module is used to establish a dynamic mapping rule base from fault signals to communication message parameters. Based on the temporal logic and causal relationship of the fault signal sub-graph, combined with physical model calculation, it generates an abnormal message sequence that conforms to the power communication protocol and performs scenario-based fault injection.
[0033] The intelligent analysis and evaluation module is used to collect multi-dimensional response signals of the device under test after performing scenario-based fault injection, construct spatiotemporal graph data with response time series data as node features and high-weight fault signal sub-graphs as spatial structures, perform feature extraction and pattern recognition through spatiotemporal graph convolutional networks, and output multi-label classification results.
[0034] The knowledge evolution and update module is used to update the entities, attributes, and relationships of the knowledge graph based on the multi-label classification results, thereby realizing the closed-loop evolution of the knowledge graph.
[0035] Compared with the prior art, the technical solution of this application has the following beneficial technical effects:
[0036] This invention employs a weighted subgraph traversal algorithm to calculate path weights by comprehensively considering multi-dimensional factors such as relation type weights, historical frequency, and severity. This enables intelligent identification of high-risk fault scenarios and automatic generation of structured test cases, eliminating the reliance on human experience in traditional test case generation. By establishing a dynamic mapping rule base and combining it with physical models such as electromagnetic transients for real-time calculations, it generates abnormal message sequences based on the temporal logic and causal relationships of the fault signal subgraph, ensuring that the fault injection process conforms to real physical laws and temporal constraints. This overcomes the shortcomings of traditional methods where abnormal messages are disconnected from real fault characteristics. Furthermore, by constructing spatiotemporal graph data and using a spatiotemporal graph convolutional network for feature extraction and pattern recognition, it achieves comprehensive intelligent analysis of multi-dimensional response signals from equipment in both time and space dimensions. This allows it to identify latent abnormal patterns that have not triggered hard thresholds, overcoming the limitations of traditional threshold comparison methods. Most importantly, by continuously updating the entities, attributes, and relationships of the knowledge graph based on test analysis results, it constructs a closed-loop evolution mechanism of test-analysis-learning, forming an intelligent testing system with adaptive learning capabilities, enabling testing capabilities to continuously improve with knowledge accumulation. Attached Figure Description
[0037] Figure 1 This is a flowchart of the overall testing method of the present invention;
[0038] Figure 2 This is a flowchart of the knowledge graph construction process for this invention;
[0039] Figure 3 This is a flowchart illustrating the test case generation and fault injection process of this invention.
[0040] Figure 4 This is a flowchart of the intelligent analysis and evaluation process of the present invention;
[0041] Figure 5 This is a flowchart illustrating the closed-loop evolution of the knowledge graph in this invention. Detailed Implementation
[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0043] like Figure 1 As shown, the present invention provides an automated device fault-tolerant testing method based on event knowledge graphs, comprising the following steps:
[0044] S1. Extract entities, relationships, and attributes of fault events from multi-source heterogeneous data, construct triple knowledge representations, eliminate entity ambiguity through semantic alignment technology, and form a unified fault event knowledge graph.
[0045] S2. Based on the fault event knowledge graph, a weighted subgraph traversal algorithm is used to calculate the path weight according to the relation type weight, historical frequency and severity. High-weight fault signal subgraphs are extracted from the fault event knowledge graph, and structured test cases containing test scenarios, fault sequences and judgment criteria are automatically generated based on the high-weight fault signal subgraphs.
[0046] S3. Establish a dynamic mapping rule base from fault signals to communication message parameters. Based on the temporal logic and causal relationship of the fault signal sub-graph, combined with physical model calculation, generate an abnormal message sequence that conforms to the power communication protocol and perform scenario-based fault injection.
[0047] S4. Collect the multi-dimensional response signal of the device under test after performing the scenario-based fault injection in S3, construct a spatiotemporal graph data with response time series data as node features and high-weight fault signal sub-graph as spatial structure, perform feature extraction and pattern recognition through spatiotemporal graph convolutional network, output multi-label classification results, update the entities, attributes and relationships of the knowledge graph according to the multi-label classification results, and realize the closed-loop evolution of the knowledge graph.
[0048] The multi-source heterogeneous data in S1 includes the whole station system configuration file SCD, production management system PMS defect records, fault waveform files, protection action reports and equipment ledgers;
[0049] Entity types include faulty equipment, fault type, fault level, fault signal, and associated equipment;
[0050] The semantic alignment technique specifically involves: generating high-dimensional word vectors using a language model pre-trained on a corpus in the power industry; calculating the semantic similarity between entities using a cosine similarity algorithm enhanced by a power industry-specific dictionary; and merging entities that exceed a preset threshold. The similarity calculation formula is as follows:
[0051]
[0052] in, and Entities and Word vectors, For dictionary matching functions, and Using weighted coefficients, the system automatically extracts fault event information from five types of data sources during the multi-source data acquisition and semantic alignment stages. The substation system configuration file (SCD) provides the substation's topology and equipment configuration information, including bay logical relationships and communication address mappings; the production management system (PMS) defect record stores detailed descriptions of historical faults and handling experience; fault waveform files contain electrical quantity waveform data at the time of the fault; protection action reports record the response behavior of protection devices; and the equipment ledger provides basic information such as equipment model and production batch.
[0053] The system employs natural language processing (NLP) technology to extract entities from unstructured text. Faulty equipment entities include primary and secondary equipment such as transformers, circuit breakers, and merging units; fault types cover abnormal situations such as short circuits, open circuits, and failure to operate; fault levels are categorized into equipment-level, bay-level, and station-level based on their impact range; fault signal entities include "SV packet loss" (meaning lost sampled data packets), "GOOSE interruption" (meaning interrupted object-oriented communication), and "overcurrent" (meaning current exceeding the set value); associated equipment records indicate the fault's affected area. Relationship extraction uses a dependency parsing-based method. For example, from the PMS defect record text "110kV bus short circuit caused main transformer differential protection maloperation," by identifying the verb "caused" and its subject and object, a triplet is automatically extracted: (Subject: 110kV bus short circuit, Relation: necessarily caused, Object: main transformer differential protection maloperation).
[0054] The triples are stored in the Neo4j database in RDF format:
[0055] <Faulty Entity ID, Relationship Type, Related Entity ID>
[0056] Simultaneously store attribute tuples:
[0057] <Entity ID, Attribute Name, Attribute Value>
[0058] To address the issue of inconsistent entity descriptions across different data sources—for example, "PT disconnection" and "voltage transformer disconnection" referring to the same fault—the system employs deep semantic alignment technology. First, an ALBERT model, trained on a large-scale corpus of power industry terminology (including regulatory documents, technical manuals, and maintenance records), maps entity texts into a 768-dimensional high-dimensional word vector space. This pre-training enables the model to deeply understand the semantics of power industry terminology.
[0059] In the similarity calculation formula, Cosine similarity represents the similarity between two entity word vectors, with a value ranging from [−1, 1]. The closer the value is to 1, the more semantically similar they are. This is a dictionary matching function that returns 1 if two entities are synonymous in an electrical engineering dictionary, and 0 otherwise. Weighting coefficients. and Based on experience, it is generally set up as follows: , The weighting coefficients α and β were determined through experiments on 500 sets of historical data. When α=0.7 and β=0.3, the semantic alignment accuracy reached a maximum of 94.5%, and therefore this was used as the default configuration. In practical applications, α∈[0.6,0.8] and β∈[0.2,0.4] can be adjusted according to the specific characteristics of the data source, indicating that word vector similarity is the primary metric, supplemented by dictionary matching. When the threshold is exceeded (usually set to 0.85), the system merges the two entities into a unified representation and selects the name used more frequently as the standard form.
[0060] Through this semantic alignment process, the system successfully eliminated entity ambiguity in multi-source data and constructed a fault event knowledge graph containing approximately 5,000 entity nodes and 15,000 relationship edges, laying a knowledge foundation for subsequent automated testing.
[0061] like Figure 2 and Figure 3 As shown, the weighted subgraph traversal algorithm in S2 specifically includes: assigning basic weights to relation types in the knowledge graph, where the weight of "inevitably leading to" relations is 0.9-1.0, the weight of "possibly accompanying" relations is 0.6-0.8, and the weight of "precondition" relations is 0.7-0.9; and introducing dynamic correction factors, including historical frequency factors. Severity factor and timeliness factor The final weight value is calculated using the weighting formula:
[0062]
[0063] in, Based on the weights, To adjust the coefficients, a priority-based traversal strategy is adopted. Starting from the target fault node, a dynamic priority queue is maintained, and the path with the highest current weight is always selected for expansion. A dual termination condition is set: a path weight accumulation threshold and a search depth limit. Key elements are identified through topology analysis, and redundancy elimination and structural optimization are performed. Based on the constructed knowledge graph, the system automatically generates test cases using a weighted subgraph traversal algorithm. The weight system design reflects the importance and determinism of different relationship types. "Inevitable Cause" relationships represent deterministic causal relationships, such as "bus short circuit" "inevitably causes" "bus differential protection activation," with weights set at 0.9-1.0 to ensure priority coverage in testing. "Possibly Accompanying" relationships represent probabilistic associations, such as "CT saturation" possibly accompanying "differential protection maloperation," with weights set at 0.6-0.8. "Precondition" relationships represent necessary conditions for the occurrence of an event, such as "GOOSE subscription configuration error" being a precondition for "GOOSE interruption," with weights set at 0.7-0.9.
[0064] To improve the targeting of tests, the system introduces three types of dynamic correction factors. Historical frequency factor. Calculated based on the number of times the fault occurs in historical records, the calculation method is as follows: ,in This represents the historical number of times this fault has occurred. For the total number of failures, high-frequency failures are given higher weight; severity factor The value range is [0, 1], with 1.0 for station-level faults, 0.6 for interval-level faults, and 0.3 for equipment-level faults; Timeliness factor More recent failures are given higher weight and calculated as follows: ,in For the present time, This is the attenuation coefficient, which is usually taken as 0.01.
[0065] The adjustment coefficient can be flexibly adjusted according to the testing focus; a common configuration is as follows. This indicates a testing strategy that prioritizes severity.
[0066] The traversal process employs a priority queue mechanism similar to Dijkstra's algorithm. Using "main transformer differential protection" as the target fault node, the system first adds this node to the queue, and then proceeds according to its final weight. The system expands adjacent nodes from high to low weight. For example, when expanding from the "Main Transformer Differential Protection Malfunction" node, it may encounter adjacent nodes such as "CT Secondary Disconnection" (weight 0.92), "CT Saturation" (weight 0.85), and "TA Ratio Configuration Error" (weight 0.78). The system prioritizes "CT Secondary Disconnection" for expansion.
[0067] A dual termination condition ensures traversal efficiency: the path weight accumulation threshold is set to 2.5, and the branch expansion stops when the sum of the weights from the starting point to the current node exceeds this value; the search depth is limited to 5 layers to avoid overly complex test scenarios. After traversal, the system identifies key signal nodes through topology analysis, performs redundancy elimination to remove duplicate paths, and finally extracts a subgraph containing 8-15 key signal nodes.
[0068] For the scenario of "external fault transfer to internal fault", the generated sub-map includes: initial fault state → external short circuit signal → first-stage protection start → delay judgment → fault nature change → internal short circuit characteristics → second-stage protection start → circuit breaker trip command → circuit breaker action confirmation. Each node is marked with a corresponding weight value, which provides clear priority guidance for subsequent test execution.
[0069] The dynamic mapping rule base in S3 stores mapping rules from abstract fault signals to specific message parameters conforming to the IEC 61850 standard, including IEC 61850-9-2SV sampled values, GOOSE general object-oriented substation events, and MMS manufacturing message specifications. Scenario-based fault injection specifically includes: deep analysis and timing planning of the fault signal sub-graph; establishing an event sequence with strict time constraints by analyzing the causal relationships and weight information between nodes; and calling the electromagnetic transient model to calculate waveform characteristics during the fault process in real time, dynamically generating corresponding abnormal message sequences. The dynamic mapping rule base serves as a bridge connecting abstract fault concepts and specific communication messages. This rule base is organized according to the three-layer communication protocol of the IEC 61850 standard: SV sampled value messages at the process layer, GOOSE messages at the bay layer, and MMS messages at the station control layer.
[0070] For the SV sampling value protocol, the rule base stores the mapping relationship from fault signals to the fields of the 80-byte / 100-byte SV message. For example, the "excessive phase current" signal is mapped to modifying the InstMag (instantaneous value) field in the ASDU (Application Service Data Unit) to set the current amplitude to a specific multiple exceeding the rated value; the "abnormal sampling synchronization" signal is mapped to modifying the SmpSynch (synchronization status) field, changing it from the normal value of 2 to the abnormal value of 0 or 1.
[0071] For the GOOSE protocol, the rule base defines the mapping between control signals and GOOSE datasets. "GOOSE status number error" is mapped to modifying the StNum field, setting it to a non-continuous incrementing value to simulate message loss or retransmission; "GOOSE timestamp error" is mapped to modifying the T field, setting an unreasonable timestamp or time reversal.
[0072] Scenario-based fault injection simulates complex scenarios through meticulous timing planning. Taking "external fault to internal fault transfer" as an example, the system execution flow is as follows:
[0073] At time T0, a short-circuit signal of phase B outside the zone is injected. The system calls the short-circuit calculation module to calculate the amplitude and phase of the short-circuit current flowing through each side of the main transformer at this time based on the system impedance and the location of the fault point, and generates the corresponding SV sampling value sequence. The current value is set to 1.5 times the rated current and lasts for 500ms to simulate the delay action time of the backup protection.
[0074] At time T0+500ms, the nature of the fault changes, and the fault point is moved from outside the zone to inside the main transformer. The short-circuit current distribution is recalculated. At this time, the current difference between the two sides of the main transformer changes abruptly, and the differential current jumps from near zero to 3 times the rated value, which meets the differential protection start-up criterion.
[0075] At time T0+520ms, based on the correlation in the graph, the differential protection start signal is simulated. By modifying the dataset of the corresponding GOOSE message, the value of the "DIF_START" data object is changed from FALSE to TRUE.
[0076] At time T0+550ms, a circuit breaker trip command is sent, and the "TRIP" data object is modified to TRUE;
[0077] At time T0+600ms, the circuit breaker is simulated to operate, and the current value in the GOOSE and SV messages of the circuit breaker position signal is changed synchronously. The current drops to zero, and the position signal changes from closed to open.
[0078] The entire event sequence is precisely synchronized on the timeline, and the signals maintain a strict temporal logical relationship, fully restoring the evolution of the real fault.
[0079] The abnormal message generation process in S3 includes: for current transformer CT saturation abnormalities, calling the electromagnetic transient model to calculate the waveform distortion characteristics during the saturation process:
[0080]
[0081] in, For primary current, For magnetic flux, For saturated magnetic flux; for the merging unit out-of-synchronization scenario, the phase accumulation error caused by clock drift is simulated by adjusting the sampling counters of multiple merging units; for the voltage transformer PT open circuit fault, the single-phase grounding scenario on the load side is simulated by setting the voltage channel value to zero or an abnormally low value and keeping the current channel normal. During the abnormal message generation process, the system calls the corresponding physical model for accurate calculation for different fault types.
[0082] For CT saturation anomalies, the system uses an electromagnetic transient model to simulate the core saturation process. In the formula, The fault current that actually flows through the primary side is usually a complex waveform containing both DC and AC components. The magnetic flux in the CT core is proportional to the rate of change of the secondary induced voltage. The core saturation flux is an intrinsic parameter of the CT, typically ranging from 1.0 to 1.5 Weber. Hyperbolic tangent function. The nonlinear magnetization characteristics of the iron core were simulated: when hour, This indicates that it is much smaller than, CT scans operate in the linear region; when Approaching or exceeding hour, When the function approaches saturation, the secondary current is clipped and cannot accurately reflect the primary current.
[0083] The system calculates the saturated current value based on each sampling point of the model (sampling rate 4000Hz or 12800Hz) and generates a distorted SV message sequence. For example, when the primary current peak value is 20 times the rated value, the secondary current after saturation may be limited to 8 times the rated value at the peak value, accompanied by obvious waveform distortion.
[0084] For scenarios where the merging unit loses synchronization, the system simulates clock drift by precisely controlling the SmpCnt field of the sampling counter. Assuming a standard sampling frequency of 4000Hz, there should be 80 sampling points within a single cycle, with SmpCnt cycling between 0 and 79. If clock drift is simulated, the system can gradually shift the sampling period of a merging unit from 250μs to 252μs, accumulating phase errors of 8 sampling points within 1 second, corresponding to a phase angle difference of approximately 36°. During this gradual shift, the differential protection device may generate false differential current due to sampling asynchrony, thus verifying the effectiveness of its synchronization monitoring function.
[0085] For PT disconnection faults, the system sets all sampled values of the voltage channel of the faulty phase (e.g., phase A) to 0 or a very small value (e.g., 0.5V), while keeping the voltages of phases B and C and the three-phase currents normal. This asymmetrical state may be misjudged as a single-phase ground fault. This test is used to check whether the PT disconnection discrimination logic of the protection device is accurate and whether it will falsely generate a ground fault signal due to PT disconnection.
[0086] The response signal acquisition in S4 includes hard contact action signals from protection devices, GOOSE messages from intelligent terminals, data transmitted from measurement and control devices, sampling records from fault recorders, and communication messages from network analyzers. The acquisition system uses a working mechanism that combines intelligent caching and real-time processing. It caches raw data in high-speed storage devices and simultaneously initiates real-time preprocessing processes, including data cleaning, format standardization, and outlier detection. A test timeline is established to precisely align the injected fault sequence with the device's response signals. The test response acquisition system constructs a comprehensive data acquisition network. Hard contact action signals are acquired through a high-speed digital acquisition card with a sampling rate of no less than 1kHz, recording the action time and duration of the output contacts of the protection device, such as tripping, starting, and alarming. GOOSE messages from intelligent terminals are captured by a network analysis device connected to the mirror port of the process layer switch, recording the transmission time, content changes, and StNum and SqNum sequences of all GOOSE messages. The data sent by the measurement and control device is acquired through the station control layer communication network, including information such as remote signaling changes, telemetry over-limits, and SOE (Sequence of Events). The sampling recorder of the fault waveform recorder provides high-precision electrical waveforms with a sampling rate of up to 12.8kHz or higher. The network analyzer monitors parameters such as message traffic, packet loss rate, and latency jitter of the communication network.
[0087] The acquisition system employs an intelligent caching mechanism to handle high-speed data streams. Raw data is first written to a high-speed solid-state drive buffer based on a PCIe interface, with a write bandwidth of 2GB / s, ensuring no test data is lost. Simultaneously, a real-time preprocessing thread performs data cleaning in parallel: removing irrelevant broadcast packets from the network and filtering communication noise; converting data from different devices and protocols into a unified standard format, such as representing hard contact signals, GOOSE Boolean values, and remote signaling changes as timestamp-state pairs; and using the 3σ criterion or sliding window-based statistical methods to detect and mark outliers, such as suddenly appearing over-range data or abrupt changes that violate physical laws.
[0088] Establishing a test timeline is crucial for response analysis. The system uses the fault injection time T0 as a baseline, aligning the timestamps of all acquired data to nanosecond precision. A time synchronization protocol (such as PTP or NTP) ensures that the clock deviation of each acquisition device is less than 1ms. For example, recording a "phase A ground" fault signal injected at time T0: the protection start contact closes at T0+15ms, a GOOSE trip message is sent at T0+35ms, the circuit breaker position changes from closed to open at T0+55ms, and the current drops to zero at T0+80ms. This precise time alignment allows subsequent analysis to accurately assess whether the protection action speed and equipment response timing meet expectations.
[0089] like Figure 4 As shown, the spatiotemporal graph convolutional network ST-GCN model in S4 achieves intelligent analysis through a multi-level feature extraction mechanism. The ST-GCN in this scheme is not a general-purpose model; its input representation and graph structure are specifically designed for power automation equipment testing scenarios. This sub-graph is defined as a graph...
[0090] G=(V,E,A)
[0091] In this context, the node set V corresponds to each fault signal and device status in the subgraph; the edge set E consists of causal relationships, temporal relationships, or logical connections defined in the subgraph; and the adjacency matrix A not only reflects the connectivity between nodes, but its elements A ij The value of v can be initialized based on the weights of the relationships in the knowledge graph, thus incorporating prior knowledge into the model. Each node v i The corresponding row in the node feature matrix X contains multi-dimensional time-series data of the signal represented by that node within the test time window. This data originates from multiple source response signals, including hard contact action signals, GOOSE messages, and uploaded data, collected in step S4, and has been precisely aligned using the test time axis. This ensures that each node feature simultaneously contains both spatial topological attribution and temporal evolution information.
[0092] The ST-GCN model of this invention consists of multiple cascaded spatiotemporal convolutional blocks (ST-ConvBlock), each block containing alternating convolution operations in the spatial and temporal dimensions to achieve hierarchical feature extraction.
[0093] In the spatial dimension, graph convolutional layers learn the inter-device relationships by aggregating neighbor node information: the spatial graph convolutional layer is responsible for aggregating the information of adjacent nodes in the graph within a single time slice to extract spatial dependencies. We employ graph convolution operations based on the first-order approximation of Chebyshev polynomials, and its single-layer forward propagation formula is:
[0094]
[0095] in, To add self-connected adjacency matrices, For degree matrix, For the first The feature matrix of the layer, These are trainable parameters; in the time dimension, the temporal convolutional layer captures the changing trends and dynamic characteristics of each signal through a sliding time window; the model output includes the overall judgment of whether the test passed, the type of anomaly detected, the severity level of the anomaly, the time period of the anomaly, and the confidence level assessment. The innovation of this scheme lies in applying this spatial convolution independently to each time step after time axis alignment, thereby accurately extracting the spatial correlation features determined by the fault propagation path at the current moment in each snapshot of the sequence.
[0096] After spatial convolution extracts features from each time slice, the temporal convolutional layer follows to capture the dynamic evolution pattern of each node along the time axis. This layer uses a one-dimensional standard convolutional kernel, performing sliding convolution on the feature sequence of each node along the time axis. The model output includes the overall judgment of whether the test passed, the type of detected anomaly, the severity level of the anomaly, the time period of the anomaly, and the confidence level assessment. The Spatiotemporal Graph Convolutional Network (ST-GCN) model is the core of the intelligent analysis in this invention. For example, using a width of K... t A one-dimensional convolutional kernel can learn the node state in its nearest K neighbors. t Local temporal patterns (such as signal mutations, trends, and periodicity) within a time step. The innovation of this scheme lies in the fact that by alternately stacking temporal convolution and spatial graph convolution, the model can first understand "which fault signals are related at a certain moment", and then understand "how these relationships evolve over time", thus achieving deep fusion and collaborative perception of spatiotemporal features.
[0097] The model input consists of two parts: the spatial structure is represented by the adjacency matrix of the fault signal subgraph extracted from the knowledge graph, if the subgraph contains... If there are 1 signal node, then the adjacency matrix is... for matrix, Represents a node and nodes There is a causal or correlational relationship between them; the time feature is the device response time series corresponding to each node, with a sampling window of 1 second and a sampling frequency of 100Hz, forming... Feature matrix Each row represents a signal node, and each column represents a time step.
[0098] In the calculation formula of the graph convolutional layer This means adding self-connections to the original adjacency matrix, so that each node not only aggregates neighbor information but also retains its own characteristics; for The degree matrix is a diagonal matrix. The adjacency matrix is symmetrically normalized to eliminate the influence of degree differences; For the first The node feature matrix of the layer has a dimension of , For the first The feature dimensions of the layer; This is the trainable weight matrix for this layer; For activation functions, ReLU is typically used.
[0099] The model constructs a dynamic spatiotemporal graph from the collected multi-device response signals related to specific fault scenarios. Through ST-GCN, the model automatically learns the spatial interactions of these signals (e.g., how a circuit breaker's failure to operate affects subsequent protection actions) and the temporal causal sequence (e.g., the strict timing sequence between protection activation, tripping signals, and circuit breaker position changes after fault injection). This powerful ability to perceive "spatiotemporally correlated fault modes" enables it to discover hidden defects in the multi-signal interaction process that cannot be identified by traditional threshold comparisons or single-signal analysis.
[0100] The physical meaning of this formula is that the new features of each node are obtained by weighted averaging of its own features and those of all its neighboring nodes, followed by a nonlinear transformation. For example, if the "CT saturation" node is adjacent to the "differential protection malfunction" node, then during feature propagation, the "differential protection malfunction" node will incorporate the abnormal features of the "CT saturation" node, and the model can learn the correlation pattern between the two.
[0101] Temporal convolutional layers use one-dimensional convolution to handle the time dimension, with kernel sizes typically of 3 or 5 and strides of 1. By using a sliding time window, the model extracts dynamic features such as signal trends, peak positions, and frequency characteristics. For example, a normal protection activation signal should show a jump within 10-30ms after a fault, while abnormal situations may manifest as excessive delays (exceeding 100ms) or no response; temporal convolution can capture these temporal patterns.
[0102] The model employs a stacked structure, consisting of three layers of graph convolutions and two layers of temporal convolutions, alternating between the two, totaling approximately 500,000 trainable parameters. The output layer uses a multi-label sigmoid classifier to predict multiple anomaly categories simultaneously. Outputs include: an overall judgment (pass / fail, threshold 0.5); anomaly type identification, such as "response timeout," "logic error," and "communication anomaly," with a probability value for each type; anomaly severity categorized into three levels: minor (0-0.3), moderate (0.3-0.7), and severe (0.7-1.0); anomaly occurrence timeframes pinpointed to specific time windows using an attention mechanism; and a confidence score reflecting the model's certainty regarding the judgment result. The ST-GCN model is trained using supervised learning. The training set contains 5000 sets of historical test data, each containing a complete fault injection sequence and device response record, labeled with test results (pass / fail) and anomaly types by experienced test engineers. The training process uses the Adam optimizer with a learning rate of 0.001 and a batch size of 32. After 50 epochs, it achieves 92% accuracy on the validation set. The model parameters are saved and used for real-time testing and analysis.
[0103] For example, in one test, the model output: overall failure (probability 0.85), "circuit breaker malfunction" anomaly detected (probability 0.92), severity 0.88 (severity level), the anomaly occurred in the time period from T0+50ms to T0+200ms, and confidence level 0.89. This detailed output provides test engineers with precise defect location information.
[0104] The closed-loop evolution of the knowledge graph in S4 includes three levels: the entity level, which automatically identifies newly discovered fault modes during testing, creates corresponding entity nodes, and establishes a device association network; the attribute level, which dynamically adjusts node weight coefficients based on test statistics to correct the evaluation indicators of fault occurrence frequency and impact; and the relationship level, which uses association rule mining algorithms to discover hidden device association characteristics and establish a multi-dimensional relationship network. The closed-loop evolution mechanism of the knowledge graph achieves knowledge updates through these three levels.
[0105] Entity-level updates: When a new fault mode is discovered during testing, such as a "sampling drift" anomaly occurring in a high-temperature environment for a specific model of merging unit, and this anomaly does not exist in the original knowledge graph, the system automatically creates a new "sampling drift" entity node, labels its attributes (faulty equipment model, environmental conditions, discovery time, etc.), and establishes a relationship with related equipment (this model of merging unit, and the protection device using this merging unit). By analyzing multiple test data, if the anomaly recurs, the system increases the weight of this node, giving it higher priority in subsequent test case generation.
[0106] Attribute-level updates: The weight coefficient of each node in the knowledge graph is dynamically adjusted based on test statistics. For example, initially, the historical frequency factor of the "GOOSE message lost" node... After 100 tests, it was found that the fault actually occurred 15 times, higher than expected, and the system recalculated. The weight of that node will be increased accordingly. The severity factor will also be adjusted based on the actual impact of the fault. If a fault occurs infrequently but causes protection failure to operate every time, its severity factor will be increased from the initial 0.5 to 0.9.
[0107] Relationship-level updates: The system employs the Apriori association rule mining algorithm to discover hidden device association characteristics. For example, analysis of test logs revealed that when "switch port packet loss rate > 5%" and "GOOSE subscription configuration error" occur simultaneously, it will 100% lead to "protection malfunction," but when they occur alone, the probability is only 20%. Based on this, the system establishes a ternary relationship in the knowledge graph: "(switch packet loss AND GOOSE configuration error) → protection malfunction," and assigns it a high weight of 0.95. The discovery of this composite relationship makes the test cases more closely resemble the complex scenarios of real-world faults.
[0108] The knowledge graph's closed-loop evolution employs a progressive knowledge update strategy. New knowledge elements first enter the unverified knowledge base, undergoing multiple tests and verifications until their confidence level reaches a preset threshold before being formally added to the database. A version history management system is established to record every significant change to the knowledge graph, supporting retrospective analysis and comparative verification of the knowledge structure. By constructing a complete "test-analysis-learning" closed loop, an adaptive intelligent testing system is formed, enabling testing capabilities to continuously improve with knowledge accumulation. The progressive knowledge update strategy ensures the stability and reliability of the knowledge graph's evolution. Newly discovered knowledge elements are not directly added to the main knowledge base but first enter the "unverified knowledge base." The unverified knowledge carries a temporary identifier and an initial confidence level (usually set to 0.3).
[0109] In subsequent tests, the system will consciously generate test cases containing the knowledge to be verified for repeated verification. Each verification update the confidence level based on the result: if verification is successful (consistent with expectations), the confidence level increases by 0.15; if verification fails, the confidence level decreases by 0.2. Only when the confidence level reaches a preset threshold (generally set to 0.8) and the number of verifications is no less than 3 will the knowledge to be verified be officially recognized and included in the main knowledge base. If the confidence level continues to drop below 0.1, the knowledge will be marked as "invalid" and removed.
[0110] The version history management system adopts a Git-like version control mechanism. Each significant change to the knowledge graph (such as adding more than 10 entities, deleting key nodes, or significantly adjusting weights) generates a version snapshot, recording the change time, content, and triggering reason. The system retains complete data for the most recent 20 versions, supporting backtracking and comparison of any version.
[0111] If a knowledge update leads to a decrease in test performance, engineers can use version comparison tools to examine the specific changes, analyze the cause of the problem, and, if necessary, roll back to a previous stable version. This version management mechanism provides a safeguard for knowledge evolution, preventing erroneous updates from damaging system performance.
[0112] Through a complete closed loop of "testing-analysis-learning," the system has achieved adaptive capabilities. Initially, the knowledge graph may have contained only 3,000 entities and 8,000 relationships. After one year of continuous operation and evolution, the scale has expanded to 8,000 entities and 25,000 relationships. Test case coverage has increased from 60% to 92%, and defect detection rate has increased from 75% to 89%, fully demonstrating the improved testing capabilities brought about by knowledge accumulation.
[0113] like Figure 5 As shown, this embodiment of the invention also provides an automated equipment fault-tolerant testing system based on an event knowledge graph, used to execute the method described above. The system includes:
[0114] The knowledge graph construction module is used to extract entities, relations and attributes of fault events from multi-source heterogeneous data to form triples, and to construct a fault event knowledge graph by eliminating ambiguity through semantic alignment.
[0115] The test case generation module is used to calculate path weights based on relation type weights, historical frequency and severity using a weighted subgraph traversal algorithm, extract high-weight fault signal subgraphs from the fault event knowledge graph, and automatically generate structured test cases containing test scenarios, fault sequences and judgment criteria based on the high-weight fault signal subgraphs.
[0116] The fault injection execution module is used to establish a dynamic mapping rule base from fault signals to communication message parameters. Based on the temporal logic and causal relationship of the fault signal sub-graph, combined with physical model calculation, it generates an abnormal message sequence that conforms to the power communication protocol and performs scenario-based fault injection.
[0117] The intelligent analysis and evaluation module is used to collect multi-dimensional response signals of the device under test after performing the S3 scenario-based fault injection, construct spatiotemporal graph data with response time series data as node features and high-weight fault signal sub-graph as spatial structure, perform feature extraction and pattern recognition through spatiotemporal graph convolutional network, and output multi-label classification results.
[0118] The knowledge evolution and update module is used to update the entities, attributes, and relationships of the knowledge graph based on the multi-label classification results, so as to realize the closed-loop evolution of the knowledge graph. The system implementation of this invention includes five functional modules, and the modules work together to form a complete testing system.
[0119] The knowledge graph construction module is deployed on a backend server, employing a distributed architecture to process large-scale data. The data acquisition submodule connects to SCD file parsers, PMS databases, and waveform file readers via interfaces to acquire multi-source data in real time. The triplet extraction submodule uses the BERT model to achieve named entity recognition and relation extraction with an accuracy exceeding 95%. The semantic alignment submodule loads a pre-trained ALBERT model (18 million parameters) and a professional dictionary containing 8,000 power-related terms, calculating entity similarity in real time and performing merging operations. The knowledge storage submodule uses the Neo4j graph database, supporting efficient graph query and traversal operations.
[0120] The test case generation module implements automated test case generation based on weighted subgraph traversal. The weight calculation engine calculates the final weight of each edge in real time according to configured coefficients and formulas. The traversal engine uses a priority queue data structure and is implemented in C++ to ensure computational performance, with a single traversal time controlled within 100ms. The subgraph extractor ensures the logical integrity of the extracted subgraph through topological sorting and connectivity analysis. The test case generator converts the subgraph into structured test cases according to predefined templates and outputs JSON format files for use by subsequent modules.
[0121] The fault injection execution module is the core execution unit of the system, deployed on a dedicated test workstation. The mapping rule library, stored in XML format, contains over 200 mapping rules and supports online editing and expansion. The timing planning engine parses the event sequences in test cases to generate an execution schedule accurate to the microsecond level. The physical model library integrates power system simulation models such as short-circuit calculation, electromagnetic transient simulation, and power flow calculation, performing complex calculations by calling the MATLAB engine or PSCAD interface. The message generator encodes SV, GOOSE, and MMS messages in real time according to the IEC61850 standard and sends them to the device under test via a gigabit network card.
[0122] The intelligent analysis and evaluation module is responsible for in-depth analysis of the test results. The data acquisition subsystem acquires response signals through a multi-channel synchronous acquisition card and network packet capture equipment, with an acquisition bandwidth supporting 1 million packets per second. The data preprocessing module performs real-time filtering, format conversion, and time alignment, outputting standardized spatiotemporal graph data. The ST-GCN analysis engine is implemented based on the PyTorch deep learning framework, with the model loaded onto the GPU for inference, and a single analysis taking approximately 200ms. The results output module generates detailed test reports, including judgment conclusions, anomaly details, confidence assessments, and improvement suggestions, supporting export in PDF and HTML formats.
[0123] The knowledge evolution and update module automatically optimizes the knowledge graph. The test result parser extracts key information from the analysis report, identifying newly discovered failure modes and relationships. The knowledge proposal generator converts the extracted information into knowledge graph update proposals and writes them into the knowledge base to be verified. The verification scheduler prioritizes the repeated verification of the knowledge to be verified in subsequent tests. The confidence evaluator updates the confidence score based on the verification results. The knowledge fusion unit formally incorporates the verified knowledge into the main knowledge base and triggers a version snapshot. The statistical analyzer periodically generates knowledge evolution reports, displaying key indicators such as changes in graph size and improvements in test performance.
[0124] The five modules communicate via message queues (such as RabbitMQ) and RESTful APIs, ensuring loose coupling and high scalability. The system as a whole adopts a microservice architecture, supporting independent deployment, upgrades, and horizontal scaling of modules, and can adapt to the testing needs of substations of different sizes.
[0125] To enhance the applicability of the system engineering, the present invention also integrates three auxiliary function modules.
[0126] The visualization and interaction module adopts a B / S architecture, using Vue.js and D3.js to implement the graphical display of the knowledge graph. The system employs a force-directed layout algorithm to automatically arrange nodes; node size reflects entity importance, and edge thickness represents relationship strength. It supports keyword search, multi-level node expansion, and filtering and aggregation display by fault type and time range. During test execution, a dynamic demonstration function is provided, showing the fault signal propagation path in animation, with nodes highlighted sequentially. After the test, frame-by-frame playback is supported for easy analysis of the device response process.
[0127] The test management module adopts a distributed architecture, supporting the parallel execution of up to eight test tasks. The task queue is sorted by priority, and a load balancing algorithm monitors the CPU, memory, and network resources of each node in real time, dynamically allocating tasks to ensure resource utilization remains within a reasonable range. It provides test progress monitoring, anomaly alerts, and resource coordination mechanisms to avoid multi-task conflicts.
[0128] The interface adapter module supports multiple power communication protocols such as IEC61850, IEC60870-5-104, Modbus, and DNP3. A built-in protocol parsing library automatically identifies message formats, and a protocol conversion engine enables data mapping between different protocols, such as converting IEC61850 remote signaling to IEC61850 GOOSE messages. It adopts a plug-in architecture, with new protocols loaded as dynamic link libraries, automatically scanned and registered during system startup, and supports hot updates without requiring a restart.
[0129] The visualization module employs a force-directed layout algorithm, treating the knowledge graph as a spring-mass system. It iterative calculations reach a state of minimum energy to form a clear topology. Node radii are determined based on a combination of degree and weight.
[0130]
[0131] in, Based on the radius, To normalize the node degree, To normalize the weights, and Adjustment coefficient (typical value) This makes important nodes visually stand out.
[0132] For new equipment or rare faults, when historical data is lacking in the knowledge graph, the system adopts a transfer learning method based on equipment type similarity, and initializes by referencing knowledge of similar equipment. The historical frequency factor fh is set to 0.1 (minimum value), and data is gradually accumulated through actual testing.
[0133] When the probability difference between multiple anomaly types output by the ST-GCN model is less than 0.1, it is judged as an "uncertain" state. The system marks the test as "manual verification" and includes the case in the model's retraining samples to optimize the classification boundary.
[0134] For knowledge that lingers in the knowledge base to be verified for a long time (failed to reach the threshold after more than 10 verifications), the system transfers it to the "low confidence knowledge base". It does not participate in the generation of test cases, but retains the data for manual analysis to avoid frequent failures and waste of test resources.
[0135] The multi-dimensional filtering function supports combined condition queries, extracting subgraphs that meet the conditions from the entire graph and rearranging them. Aggregated display collapses similar entities into super nodes; clicking expands them to view details. The dynamic demonstration of the test path uses a timeline control; nodes change color according to the actual time sequence, as shown in Table 1.
[0136] Table 1
[0137]
[0138] The test management module adopts a master-slave mode, with the master node running the task scheduler and the slave nodes running the test execution engine. The task queue is implemented based on Redis, supporting persistence and fault recovery.
[0139] The load balancing algorithm uses a weighted strategy to comprehensively calculate the load score of each node:
[0140]
[0141] in, CPU utilization For memory usage, This represents the current number of tasks. The maximum number of tasks, weight coefficient New tasks are prioritized for allocation to the node with the lowest load score, and allocation is paused when a node's load exceeds 85%.
[0142] The task queue supports priority scheduling (levels 1-10) and dependency definition, automatically handling the execution order. Task status includes five categories: pending, executing, completed, failed, and timed out. Failed tasks can be automatically retried. The system dynamically adjusts execution priorities based on device idle status, fault frequency, and time period characteristics to optimize resource utilization.
[0143] The interface adaptation module adopts a plug-in architecture, with each protocol encapsulated as an independent plug-in, providing a unified interface method: initialization, connection establishment, data transmission and reception, message parsing, and connection closure. New protocols are compiled into dynamic link libraries (.dll or .so) and placed in the plugins directory, automatically scanned and loaded during system startup.
[0144] Protocol conversion is achieved through a mapping table to convert data between different protocols. An example of the mapping relationship is shown in Table 2:
[0145] Table 2
[0146]
[0147] The data format mapper parses the source message to extract field values and queries the mapping table to populate the target message. The system currently supports 15 protocols, including IEC61850 (SV / GOOSE / MMS), IEC60870-5-104, ModbusTCP / RTU, DNP3, and MQTT, and the plugin library can be continuously expanded.
[0148] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0149] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. An automated equipment fault-tolerant testing method based on event knowledge graph, characterized in that, Includes the following steps: S1. Extract entities, relationships, and attributes of fault events from multi-source heterogeneous data, construct triple knowledge representations, eliminate entity ambiguity through semantic alignment technology, and form a unified fault event knowledge graph. S2. Based on the fault event knowledge graph, a weighted subgraph traversal algorithm is used to calculate the path weight according to the relation type weight, historical frequency and severity. High-weight fault signal subgraphs are extracted from the fault event knowledge graph, and structured test cases containing test scenarios, fault sequences and judgment criteria are automatically generated based on the high-weight fault signal subgraphs. S3. Establish a dynamic mapping rule base from fault signals to communication message parameters. Based on the temporal logic and causal relationship of the fault signal sub-graph, and combined with the physical model calculation, generate an abnormal message sequence that conforms to the power communication protocol, and perform scenario-based fault injection. S4. Collect the multi-dimensional response signal of the device under test after performing the scenario-based fault injection in S3, construct a spatiotemporal graph data with response time series data as node features and high-weight fault signal sub-graph as spatial structure, perform feature extraction and pattern recognition through spatiotemporal graph convolutional network, output multi-label classification results, update the entities, attributes and relationships of the knowledge graph according to the multi-label classification results, and realize the closed-loop evolution of the knowledge graph.
2. The automated equipment fault-tolerant testing method based on event knowledge graph according to claim 1, characterized in that, The multi-source heterogeneous data in S1 includes the whole station system configuration file SCD, production management system PMS defect records, fault waveform files, protection action reports and equipment ledgers; The entity types include faulty equipment, fault type, fault level, fault signal, and associated equipment; The semantic alignment technique specifically involves: generating high-dimensional word vectors using a language model pre-trained on a corpus in the power industry; calculating the semantic similarity between entities using a cosine similarity algorithm enhanced by a power industry dictionary; and merging entities that exceed a preset threshold. The similarity calculation formula is as follows: ； in, and Entities and Word vectors, For dictionary matching functions, and These are the weighting coefficients.
3. The automated equipment fault-tolerant testing method based on event knowledge graph according to claim 1, characterized in that, The weighted subgraph traversal algorithm in S2 specifically includes: assigning basic weights to relation types in the knowledge graph, where the weight of "inevitably leading to" relations is 0.9-1.0, the weight of "possibly accompanying" relations is 0.6-0.8, and the weight of "precondition" relations is 0.7-0.9; and introducing dynamic correction factors, including historical frequency factors. Severity factor and timeliness factor The final weight value is calculated using the weighting formula: ； in, Based on the weights, To adjust the coefficients, a priority-based traversal strategy is adopted, starting from the target fault node, maintaining a dynamic priority queue, always selecting the path with the highest current weight for expansion, setting a dual termination condition of path weight accumulation threshold and search depth limit, determining key elements through topology analysis, and performing redundancy elimination and structural optimization.
4. The automated equipment fault-tolerant testing method based on event knowledge graph according to claim 1, characterized in that, The dynamic mapping rule base in S3 stores mapping rules from abstract fault signals to specific message parameters conforming to the IEC61850 standard, including IEC61850-9-2SV sampled values, GOOSE general object-oriented substation events, and MMS manufacturing message specifications. The scenario-based fault injection specifically includes: performing in-depth analysis and time-series planning on the fault signal sub-graph, and establishing an event sequence with strict time constraints by analyzing the causal relationships and weight information between nodes; The electromagnetic transient model is invoked to calculate the waveform characteristics during the fault process in real time and dynamically generate the corresponding abnormal message sequence.
5. The automated equipment fault-tolerant testing method based on event knowledge graph according to claim 4, characterized in that, The abnormal message generation process in S3 includes: for current transformer CT saturation abnormalities, calling the electromagnetic transient model to calculate the waveform distortion characteristics during the saturation process: ； in, For primary current, For magnetic flux, The saturation flux is used; for the merging unit out-of-step scenario, the phase accumulation error caused by clock drift is simulated by adjusting the sampling counters of multiple merging units; for the voltage transformer PT open circuit fault, the single-phase grounding scenario on the load side is simulated by setting the voltage channel value to zero or an abnormally low value and keeping the current channel normal.
6. The automated equipment fault-tolerant testing method based on event knowledge graph according to claim 1, characterized in that, The response signal acquisition in S4 includes the hard contact action signal of the protection device, the GOOSE message of the smart terminal, the data sent by the measurement and control device, the sampling record of the fault recorder, and the communication message of the network analyzer. The acquisition system uses a working mechanism that combines intelligent caching and real-time processing. It caches the raw data in a high-speed storage device while simultaneously initiating a real-time preprocessing process, including data cleaning, format standardization, and outlier detection. Establish a test timeline to precisely align the injected fault sequence with the device's response signals.
7. The automated equipment fault-tolerant testing method based on event knowledge graph according to claim 1 or 6, characterized in that, The spatiotemporal graph convolutional network ST-GCN model in S4 achieves intelligent analysis through a multi-level feature extraction mechanism: In the spatial dimension, the graph convolutional layer learns the mutual influence relationships between devices by aggregating neighbor node information. ； in, To add self-connected adjacency matrices, For degree matrix, For the first The feature matrix of the layer, These are trainable parameters; in the time dimension, the temporal convolutional layer captures the changing trends and dynamic characteristics of each signal through a sliding time window; the model output includes the overall judgment of whether the test passed, the type of anomaly detected, the severity level of the anomaly, the time period in which the anomaly occurred, and the confidence level assessment.
8. The automated equipment fault-tolerant testing method based on event knowledge graph according to claim 1 or 7, characterized in that, The knowledge graph closed-loop evolution in S4 includes three levels: the entity level, which automatically identifies newly discovered fault modes during the testing process, creates corresponding entity nodes, and establishes a device association network; the attribute level, which dynamically adjusts node weight coefficients based on test statistics to correct the evaluation indicators of fault occurrence frequency and impact; and the relationship level, which uses association rule mining algorithms to discover hidden device association characteristics and establish a multi-dimensional relationship network.
9. The automated equipment fault-tolerant testing method based on event knowledge graph according to claim 8, characterized in that, The knowledge graph closed-loop evolution adopts a progressive knowledge update strategy. New knowledge elements first enter the knowledge base to be verified, and can only be formally added to the database after multiple tests and verifications and the confidence level reaches a preset threshold. Establish a version history management system to record every significant change to the knowledge graph, supporting retrospective analysis and comparative verification of the knowledge structure; by constructing a complete closed loop of "test-analysis-learning", an adaptive intelligent testing system is formed, enabling testing capabilities to continuously improve with the accumulation of knowledge.
10. An automated equipment fault-tolerant testing system based on event knowledge graph, characterized in that, The system for performing the method according to any one of claims 1 to 9, the system comprising: The knowledge graph construction module is used to extract entities, relations and attributes of fault events from multi-source heterogeneous data to form triples, and to construct a fault event knowledge graph by eliminating ambiguity through semantic alignment. The test case generation module is used to calculate path weights based on relation type weights, historical frequency and severity using a weighted subgraph traversal algorithm, extract high-weight fault signal subgraphs from the fault event knowledge graph, and automatically generate structured test cases containing test scenarios, fault sequences and judgment criteria based on the high-weight fault signal subgraphs. The fault injection execution module is used to establish a dynamic mapping rule base from fault signals to communication message parameters. Based on the temporal logic and causal relationship of the fault signal sub-graph, combined with physical model calculation, it generates an abnormal message sequence that conforms to the power communication protocol and performs scenario-based fault injection. The intelligent analysis and evaluation module is used to collect multi-dimensional response signals of the device under test after performing scenario-based fault injection, construct spatiotemporal graph data with response time series data as node features and high-weight fault signal sub-graphs as spatial structures, perform feature extraction and pattern recognition through spatiotemporal graph convolutional networks, and output multi-label classification results. The knowledge evolution and update module is used to update the entities, attributes, and relationships of the knowledge graph based on the multi-label classification results, thereby realizing the closed-loop evolution of the knowledge graph.