A power equipment fault analysis method based on a RAG knowledge graph

Abstract

The application discloses a power equipment fault analysis method based on RAG knowledge graph, comprising the following steps: (1) inputting the obtained power equipment multi-modal data into the power equipment exclusive knowledge graph, carrying out correlation reasoning through the power equipment exclusive knowledge graph, and constructing a device state evaluation feature set; (2) based on the device state evaluation feature set, searching through the RAG knowledge base, when it is detected that the obtained power equipment multi-modal data is abnormal or exceeds the allowed range, extracting and analyzing the fault data of the power equipment multi-modal data; (3) fusing the reasoning result of the power equipment exclusive knowledge graph and the RAG knowledge base. The application fuses the mechanism reasoning of the knowledge graph and the RAG search enhancement technology, complementarily improves the fault research and judgment accuracy in two ways, relies on the power exclusive knowledge graph to sort out the device topological relationship and the fault mechanism logic, and makes up for the defects of the traditional diagnosis without mechanism support and high subjective misjudgment rate.

Description

Technical Field

[0001] This invention belongs to the field of equipment fault analysis, specifically a power equipment fault analysis method based on RAG knowledge graph. Background Technology

[0002] Equipment fault analysis and diagnosis are of great significance to power system construction, serving as a crucial foundation for ensuring the stability and reliability of power system operation. For a long time, the power industry's fault handling has heavily relied on senior technical experts. After a fault occurs, personnel must be dispatched to the site to investigate the fault type and pinpoint its root cause through methods such as visual inspection, electrical testing, equipment disassembly, and manual data comparison. This purely manual diagnostic model is constrained by personnel capabilities, working methods, and objective conditions. Specifically: 1. Diagnostic ability is highly dependent on personal experience, with a high skill threshold and insufficient coverage.

[0003] Power equipment is diverse, encompassing transformers, switchgear, cables, instrument transformers, relay protection devices, and more. Fault causes can be categorized into dozens of types, including mechanical defects, electrical aging, insulation damage, external force damage, and environmental influences. The same symptom often corresponds to multiple root causes. Traditional diagnosis relies entirely on the experience, memory, and subjective judgment of experts, resulting in inconsistent diagnostic effectiveness: senior experts can quickly pinpoint problems, while ordinary maintenance and repair personnel are prone to misjudging complex faults. Furthermore, the number of senior experts in the industry is limited. In the event of a large-scale, multi-location concentrated fault, expert resources are dispersed, making it difficult to provide synchronized support across the entire region. Remote sites and grassroots teams often lack high-end technical support, leading to significant regional and hierarchical disparities in fault diagnosis capabilities.

[0004] 2. The fault analysis has only one dimension, making it difficult to achieve multi-source information linkage and judgment.

[0005] In the traditional model, staff primarily analyze currently faulty equipment and observable phenomena on-site, only able to retrieve a limited amount of historical data for that equipment. This makes it difficult to quickly correlate historical fault cases and experience in handling similar defects across the entire network for the same model and operating conditions. Power equipment faults often exhibit correlation and latency. Existing power equipment fault analysis methods (CN2024113199642) can only predict current faults, primarily describing numerical analysis methods. They do not utilize RAG knowledge graphs to deduce the component-level fault development mechanism of power equipment. Some latent defects do not immediately manifest as obvious faults, and manual analysis struggles to integrate data and cases across time periods, equipment, and regions. This approach only addresses the immediate issue, failing to delve into common fault patterns and potential hazards, and making it difficult to distinguish between sporadic faults and batch equipment defects.

[0006] 3. Diagnostic methods are greatly constrained by on-site conditions, and cannot be implemented in some scenarios.

[0007] On-site inspection, equipment disassembly, and offline testing are the main methods of traditional diagnosis, but they have many practical limitations: First, high-voltage, enclosed, and live equipment cannot be disassembled for inspection. Forcibly shutting down and disassembling the equipment is not only cumbersome but also expands the scope of the power outage. Second, severe weather, high altitudes, underground mines, cable trenches, and other dangerous working environments limit manual inspection and testing operations, making the work difficult and prolonging the diagnosis cycle. Third, some minor defects and early latent faults have no obvious external characteristics and are difficult to identify by visual inspection and conventional testing. They are often only discovered after they have developed into serious faults, missing the opportunity for early intervention.

[0008] 4. Fault diagnosis criteria rely on manual review, resulting in low efficiency in standard matching.

[0009] Fault characterization, cause identification, and response plan formulation all require adherence to national standards, industry specifications, and on-site operation instructions. Traditionally, staff must manually sift through paper documents and scattered electronic files, checking each clause meticulously. This process is time-consuming and prone to omissions and misapplication of regulations. Especially when dealing with complex faults, which require cross-referencing multiple procedures, manually unraveling the logic is difficult and can easily lead to non-compliance with operational and diagnostic requirements. Summary of the Invention

[0010] The purpose of this invention is to provide a power equipment fault analysis method based on RAG knowledge graph. The method constructs a power equipment-specific knowledge graph by inputting multimodal data from the power equipment. Through bidirectional enhancement of RAG and knowledge graph, a logical form guidance mechanism is introduced in the retrieval stage to improve the accuracy and logic of professional domain question answering.

[0011] The present invention is achieved by at least one of the following technical solutions. A power equipment fault analysis method based on RAG knowledge graph includes the following steps: (1) Input the acquired multimodal data of power equipment into the power equipment-specific knowledge graph, and construct the equipment status assessment feature set through association reasoning using the power equipment-specific knowledge graph; (2) Based on the equipment condition assessment feature set, the RAG knowledge base is used for retrieval. When the acquired power equipment multimodal data is found to be abnormal or exceeds the allowable range, the fault data of the power equipment multimodal data is extracted and analyzed. (3) Integrate the reasoning results of the power equipment-specific knowledge graph with the extraction results of the RAG knowledge base to conduct equipment status assessment, classify hidden danger and defect levels, and generate professional power operation and maintenance management guidance suggestions.

[0012] Furthermore, the ontology architecture of the power sector-specific knowledge graph includes device dimensions, fault mechanism dimensions, and professional theory dimensions.

[0013] Furthermore, the acquired multimodal data of the power equipment includes raw data, operational data, maintenance and testing data, and factors affecting the safe and stable operation of the equipment.

[0014] Furthermore, the RAG knowledge base implements enhanced retrieval based on entity-relation-entity triples.

[0015] Furthermore, during the retrieval process in step (2), if the device is detected to be performing an action, the device status is analyzed to determine whether it is caused by a dispatch order or an abnormal situation. If it is an abnormal situation, the fault data of the power equipment multimodal data is extracted and analyzed.

[0016] Furthermore, in step (2), a 3-based approach is adopted. s The analysis is performed using the rules and statistical range of the equipment's secondary time parameters in both horizontal and vertical analysis algorithms. Step 1: Summarize the historical operation data of equipment of the same model from the same manufacturer and record it as the initial sample; Step 2: Preprocess the multimodal data of power equipment, remove abnormal data, and form the first data sample; Step 3: Calculate the sample mean and standard deviation of the first data sample; Step 4: Analyze the frequency distribution pattern of the first data sample and draw a frequency distribution graph; Step 5: Based on the abnormal data judgment and the degree of deviation, filter out severely biased data from the first data sample to form the second data sample; repeat steps 3 to 4. Step 6: Based on the outlier determination and deviation level standards, the initial sample data is divided into four levels: normal, attention, abnormal, and severe, according to the mean and standard deviation of the second data sample.

[0017] Furthermore, according to the statistical principle of 3- s The rules categorize the degree of deviation into levels.

[0018] Furthermore, in step (3), the equipment status is evaluated based on the reliability of the faulty component. The reliability of the faulty component is determined by the degree of degradation of the status quantity and its weight. The degree of degradation of the status quantity corresponds to the deviation level, which are normal, attention, abnormal and severe levels, respectively. The more severe the level, the greater the weight.

[0019] Furthermore, the system utilizes a large language model to generate specialized power operation and maintenance management guidelines, which are then processed into natural language and made available to relevant personnel for viewing.

[0020] A computer device according to the present invention includes a memory and a processor, the memory being electrically connected to the processor, the memory storing a computer program, which, when executed by the processor, causes the processor to implement the method described herein.

[0021] The present invention provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor implements the method described herein.

[0022] Compared with existing technologies, the beneficial effects of the present invention are as follows: This invention integrates knowledge graph mechanism reasoning with RAG retrieval enhancement technology, achieving a two-way complementary improvement in fault assessment accuracy. It leverages a power-specific knowledge graph to organize equipment topology relationships and fault mechanism logic, overcoming the shortcomings of traditional diagnosis, such as lack of mechanistic support and high subjective misjudgment rates. By standardizing national standards, operating instructions, and historical maintenance cases to build a dedicated vector knowledge base, it overcomes the pain points of fragmented power industry data, inefficient retrieval, and difficulty in transferring experience. Based on a multimodal equipment input linkage dual knowledge system, it considers real-time equipment status, professional procedures and standards, and similar historical fault experience, automatically completing equipment status assessment and defect level classification. This reduces reliance on on-site inspections and disassembly tests by senior power experts, shortens fault assessment time, and lowers the costs of manual maintenance and cross-regional repairs. Simultaneously, the assessment results are highly compliant, automatically outputting standardized maintenance and rectification suggestions to avoid the risks of illegal maintenance, helping the power grid shift from post-fault emergency repair to pre-fault hazard management, effectively improving the efficiency of power equipment maintenance and the overall stability of the power grid. Attached Figure Description

[0023] Figure 1 This is a flowchart of a power equipment fault analysis method based on RAG knowledge graph, as an example.

[0024] Figure 2 This is a schematic diagram of a knowledge graph specific to power equipment, used as an example.

[0025] Figure 3 The frequency distribution of sample 1 before removing severely biased data is shown in the example.

[0026] Figure 4 The frequency distribution of sample 2 in the example is shown after removing severely biased data.

[0027] Figure 5 The original data deviation level classification diagram for Sample 1 is shown in the example.

[0028] Figure 6 The circuit breaker parallel capacitor fault mechanism diagram is shown in the example. Detailed Implementation

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] like Figure 1 As shown in this embodiment, a power equipment fault analysis method based on RAG knowledge graph includes the following steps: S1. Construct a knowledge graph specific to power equipment. The ontology architecture of this power-domain-specific knowledge graph includes equipment dimensions, fault mechanism dimensions, and professional theoretical dimensions. Its purpose is to solve problems related to entity association and relational reasoning, such as... Figure 2 As shown.

[0031] The equipment dimension integrates manufacturer equipment models, structures, and component topology parameters. The fault mechanism dimension includes fault causes, fault phenomena, fault development patterns, and impact range. The professional theory dimension integrates high-voltage technology, relay protection principles, and common knowledge of power operation. Through structured organization and semi-automated knowledge extraction, it completes the modeling and storage of power entities, attributes, and relationships.

[0032] This embodiment constructs a dedicated knowledge graph for power equipment using circuit breakers in operation as the state assessment object. The equipment dimension includes the equipment model, from which the voltage level, manufacturer, porcelain insulator material, and mechanism type of the circuit breaker can be determined. From the structure, it can be determined whether the circuit breaker is a tank type, porcelain column type, or combined electrical appliance type. Furthermore, the topology association parameters of the components can be determined. For example, for an LTB type circuit breaker from a joint venture manufacturer, its arc-extinguishing chamber is a vertically arranged porcelain column, and its mechanism is three-phase linkage. The components include crank arms, opening and closing coils, opening stop, closing stop, and insulating pull rods. These components are interlocked and linked together.

[0033] The fault mechanism dimensions include insulation performance, mechanical performance, and environmental tolerance performance. In terms of insulation performance, the fault cause may be overvoltage, which includes internal overvoltage and external overvoltage. Internal overvoltage includes switching overvoltage and power frequency overvoltage, while external overvoltage includes lightning intrusion overvoltage and line fault grounding overvoltage. Fault causes corresponding to mechanical performance include internal jamming of the mechanism, loose and worn parts, fatigue deformation of the transmission mechanism, incomplete mechanical closing and opening, and aging and falling off of fasteners. Fault causes corresponding to environmental tolerance performance include sudden changes in temperature and humidity, condensation and water accumulation, and dust and dirt accumulation.

[0034] The professional theoretical dimensions include high-voltage technology, relay protection principles, and basic knowledge of power operation. In high-voltage technology, it covers the breakdown mechanism of insulation media in power equipment, the laws governing insulation aging and deterioration, and the core theories of lightning protection grounding and insulation coordination. This supports the analysis of the causes of insulation-related faults, the assessment of insulation status, and the evaluation of withstand voltage capabilities. Relay protection principles include the operating logic of various power protection devices, fault discrimination criteria, protection setting coordination principles, tripping mechanisms under abnormal operating conditions, and secondary circuit fault mechanisms. This is used to analyze protection action behavior during fault occurrence, locate fault zones, and identify abnormal issues such as maloperation and failure to operate. Basic knowledge of power operation covers the constraints of normal grid operation, load fluctuation characteristics, equipment start-up and shutdown operation specifications, system operation mode switching rules, and on-site operation and maintenance safety and operating condition evaluation standards. This provides fundamental theoretical support for fault scenario matching, screening of operational hazards, and the formulation of operation and maintenance plans.

[0035] S2. Acquire multimodal data of power equipment. After acquiring the multimodal data, use the fault characteristics perceived by multimodal sensing as the input to the knowledge graph for inference. Call the knowledge graph to perform association retrieval, match the corresponding structural parameters, historical fault records, and relay protection action mechanisms of the equipment, and identify suspected fault causes and fault propagation paths. After the multimodal data is input into the knowledge graph, the knowledge graph begins to retrieve data such as factory parameters, operating years, load operation data, fault frequency, insulation degradation data, environmental meteorological data, power standards and specifications, and business guidelines to construct a feature set for equipment condition assessment.

[0036] The multimodal data in this embodiment includes multiple datasets: (1) Collection of original data. The original data of the equipment mainly includes: key design / material / production process information that affects the equipment status, ordering technical specifications, nameplate, type test report, equipment supervision report, factory test report, transportation and installation records, handover and acceptance report, installation and operation manual, and other original data information before commissioning.

[0037] (2) Collection of operational data. The operational data of the equipment mainly includes: the number of circuit breaker operations, circuit breaker fault trip records (number of fault trips, waveform, phase, amplitude, duration, etc. of fault current provided by relay protection and automatic devices), equipment inspection / maintenance records, historical defect / abnormal / fault information, infrared temperature measurement records, real-time operation monitoring data, intelligent sensor data, digital twin model prediction results, and other information obtained from cutting-edge technologies.

[0038] (3) Collection of maintenance and testing data. The maintenance and testing data of the equipment mainly include: maintenance reports, historical power outage test reports, SF6 gas test reports, live-line testing / online monitoring data, special test reports, special inspections / countermeasures and correlation of similar equipment fault hazards, equipment technical upgrades and replacement of major components, overhaul / technical upgrade information and test information after overhaul / modification, etc.

[0039] (4) Other data collection. Other data for the equipment mainly include: the operation, repair and testing, defects and failures of similar equipment; changes in the equipment operating environment such as meteorological and geological disasters, and changes in system operation mode; short-circuit current calculation report at the installation location; and other factors affecting the safe and stable operation of the equipment.

[0040] The aforementioned dataset is input as a multimodal input into the RAG knowledge base of the knowledge graph. The device structure tree of the RAG knowledge base can classify, collect, and accurately correlate maintenance data, status data, and history data, providing data support for status assessment, maintenance decision-making, and defect tracing.

[0041] Taking the knowledge graph of circuit breaker equipment as an example, its core graph is: Primary Unit (Equipment Body): The entire SF6 circuit breaker, which collects basic data, comprehensive status data, weather status, operating voltage and current data, and other general data throughout the equipment's entire life cycle.

[0042] Secondary unit (core functional module): includes five major modules: main body, operating mechanism, voltage equalizing capacitor, closing resistor, and online monitoring device. It collects specific operating data, test data, and maintenance data for each module. For example, the abnormal action time data analyzed in this embodiment reflects the internal problem of the operating mechanism. It will not be associated with or mapped to unrelated components such as the main body and voltage equalizing capacitor, and will not search more knowledge bases, thus avoiding misjudgment or illusion.

[0043] Level 3 Unit (Minimum Maintenance and Replacement Unit): This further breaks down the Level 2 unit to match the data traceability and control requirements for single-point maintenance and component replacement.

[0044] In one embodiment, multimodal data may also include two-dimensional time-domain signals such as voltage and current waveforms during operation, three-dimensional frequency-domain detection signals such as partial discharge fault location results, equipment image defects and photos generated during inspections, video recognition results captured by monitoring, equipment ledgers, operation and maintenance data, repair and testing data, and power equipment fault recordings. Multimodal information is typically propagated in the fiber optic channel of the power system. The power equipment fault recordings mainly include information such as voltage, current, zero-sequence current, circuit breaker position signals, and tripping signals on the line.

[0045] (3) Based on the equipment condition assessment feature set, retrieve and analyze the faults of the multimodal data of power equipment by using retrieval tools; During the retrieval process, a retrieval tool is used, which in turn searches the RAG knowledge base. The RAG knowledge base uses a semantic retrieval enhanced reasoning framework to achieve enhanced retrieval through "entity-relationship-entity" triples, such as "circuit breaker-lower-component-mechanism," to accurately capture the semantic relationships implicit in multimodal data. When multimodal inputs exhibit signals such as voltage signal disturbances, current signal disturbances, or circuit breaker displacement, the retrieval tool, following the prompts, progressively determines whether the disturbance signal conforms to standards and technical specifications.

[0046] The judgment is as follows: When the operating object is a circuit breaker, the basic information mainly includes voltage, current frequency, and position change information. During normal operation of the power system, voltage and current signal disturbances are within reasonable ranges. The allowable deviation of the supply voltage is set to ±5% of the rated voltage, which is taken from GB / T 12325 "Power Quality - Supply Voltage Deviation". The allowable deviation of the normal system current frequency is 50Hz ± 0.2Hz, which is taken from the national standard GB / 15945 "Power Quality - Permissible Frequency Deviation of Power Systems". When the voltage and current exceed the allowable deviation range, the retrieval tool is triggered. In addition, position change signals (0 and 1) correspond to the open and closed states. When the circuit breaker operates, the retrieval tool is triggered to analyze the circuit breaker status and rule out whether the tripping was caused by dispatch order or abnormal conditions.

[0047] When the search tool is triggered, it signifies more advanced search criteria and more comprehensive multimodal information input and analysis. For example, when a circuit breaker trips due to an external fault, both voltage and current signals are disturbed, triggering an abnormal event. The search tool then extracts and analyzes the required multimodal data according to the analysis rules. The analysis rules are as follows: Based on the positive linear relationship between circuit breaker opening / closing time and the time difference between the operation command and the switch position change, the negative correlation between opening / closing speed and the time difference between opening / closing position, and the normal distribution law of secondary time parameters of the same type of circuit breaker, the feasibility of using secondary parameters to warn of circuit breaker time and speed anomalies is theoretically demonstrated. Therefore, based on 3 s The algorithm for horizontal and vertical analysis of secondary time parameters of circuit breakers based on rules and statistical ranges adopts a circuit breaker mechanism state diagnosis strategy of secondary early warning and primary diagnosis. The diagnosis strategy involves four parameters, which are expressed as follows: t 11 , t 13 , t 21 , t 23 ,in, t 11 This is the time difference between the opening and closing positions when the circuit breaker is tripped. t 21It is the time difference between the opening and closing positions when the circuit is closed, reflecting the change process of the auxiliary switch during the opening (closing) process; t 13 The time difference between the tripping command and the switch closing position is 0. t 23 It represents the time difference between the closing command and the switch opening position 0, reflecting the time from the output of the opening (closing) command to the change of the closing (opening) position contact of the auxiliary switch.

[0048] The sample mean is denoted as x ave The standard deviation is denoted as s According to the normal distribution law, the sample falls within [ x ave - s , x ave + s The probability of a sample falling outside the range is 31.7%; x ave -2 s , x ave +2 s The probability of a sample falling outside the range is 5%; x ave -3 s , x ave +3 s The probability of being outside the specified range is 0.3%. According to statistical principles 3- s The rule categorizes data status levels according to the degree to which the sample deviates from the mean, as follows: 1) Normal: Sample x satisfy: x [ x ave - s , x ave + s ]; 2) Note: Sample x satisfy: x [ x ave -2 s , x ave - s )or x ( x ave + s , x ave +2 s ]; 3) Anomalies: Samples x satisfy: x [ x ave -3 s , x ave -2 s )or x ( x ave +2 s , x ave +3 s ]; 4) Severe: Sample x satisfy: x [-∞, x ave -3 s )or x ( x ave +3 s , +∞).

[0049] For switching equipment, abnormal operating characteristics usually manifest as positive deviations in characteristic parameters. Therefore, the deviation levels of switching operating parameters are classified according to the following rules: 1) Normal: Sample x satisfy: x [0, x ave + s ]; 2) Note: Sample x satisfy: x ( x ave + s , x ave +2 s ]; 3) Anomalies: Samples x satisfy: x ( x ave +2 s , x ave +3 s ]; 4) Severe: Sample x satisfy: x ( x ave+3 s , +∞).

[0050] For example, if a circuit breaker is detected to have triggered a change signal, after verification by staff, it is confirmed that the currents of all three phases (A, B, and C) of the circuit breaker normally passed through the zero-point arc extinguishing time within 28ms-33ms-85ms. Although the arc extinguished normally, the phase-to-phase deviation of the C-phase current breaking at 85ms falls within the scope of the warning value. Simultaneously, extracting the SOE message information from the multi-mode data reveals that the C-phase switch relay protection action time is normal, but the tripping time is delayed, also falling within the scope of the warning value. Inputting multi-mode data shows that equipment preventative testing, maintenance records, operation and maintenance data, and drone inspection data are all normal; however, the lightning location system and dispatch system did not report any warnings.

[0051] To eliminate interference from obviously abnormal data, this embodiment adopts a method for removing abnormal data, and the specific steps are as follows.

[0052] Step 1: Summarize the four parameters of the same model of switchgear from the same manufacturer ( t 11 , t 13 , t 21 , t 23 The historical action data of ) is denoted as sample 0.

[0053] Step 2: Remove obviously abnormal data, resulting in 4 data samples, denoted as Sample 1. Obviously abnormal data is defined as closing-related parameters exceeding 150ms and opening-related data exceeding 100ms.

[0054] Step 3: Calculate the sample average of each of the four parameters. x ave Standard deviation s .

[0055] Step 4: Calculate the frequency distribution patterns of the four parameters and draw a frequency distribution graph.

[0056] Step 5: Based on the above abnormal data judgment and deviation level standards, for each of the four samples 1, filter out severely biased data (3... s Other data points), such as Figure 3 As shown.

[0057] Step 6: Remove severely biased data from each of the four samples to form Sample 2. Repeat steps 3 and 4 again. The frequency distribution of Sample 2 after removing severely biased data is as follows: Figure 4 As shown.

[0058] Step 7: Based on the outlier determination and deviation level standards, according to the mean and standard deviation of sample 2, divide the data of sample 0 into four levels: normal, attention, abnormal, and severe. Figure 5 As shown.

[0059] The equipment status assessment feature set data undergoes multi-dimensional semantic similarity matching using the RAG knowledge base. This matching starts with surface-level phenomena and then uses a knowledge graph to infer potentially related equipment, components, and causes of failures. The system outputs the industry standard clauses, fault handling specifications, and historical maintenance cases with the highest matching degree. After retrieval using the RAG knowledge base, the attention value is determined to belong to abnormal equipment. Natural language processing of this output can be performed using a large language model, making it available for maintenance and technical personnel to view.

[0060] S3. By integrating the reasoning results of the power equipment-specific knowledge graph with the retrieval results of the RAG knowledge base, a weighted fusion decision algorithm is used to complete equipment status assessment, classification of hidden danger and defect levels, and generate professional power operation and maintenance management guidance suggestions.

[0061] The weighted fusion decision algorithm completes the equipment status assessment and hazard / defect level classification as follows: Based on the degree of influence of equipment status variables on the probability of equipment failure, the confidence levels are divided into four levels from low to high: 1, 2, 3, and 4, each corresponding to a different weight coefficient. Weights 1 and 2 correspond to general confidence levels, while weights 3 and 4 correspond to high confidence levels.

[0062] The degree of degradation of state variables is divided into four levels from mild to severe: normal, warning, abnormal, and severe. The above deviation levels are corresponding. Except for the tripping time, other parameters such as the gas composition of the main body, the capacitance value of the equalizing capacitor, and the dielectric loss value are all based on the deviation level.

[0063] The corresponding basic deduction values ​​are 1, 2, 3, and 4 points. The degree of degradation of the state variables is the deviation of the multimodal input data from the standard value; the greater the deviation, the greater the degree of degradation.

[0064] The reliability of a faulty component is determined by both the degree of degradation of the state variable and its weight. Specifically, the reliability of a faulty component equals the actual deduction value of that state variable multiplied by the weighting coefficient. No points are deducted when the state variable is normal, meaning its influence can be eliminated. When the state variable is abnormal, the deduction is more severe depending on the nature of the abnormality.

[0065] Table 1. Weights, Deterioration Levels, and Corresponding Scores of Faulty Components

[0066] Taking the arcing time of circuit breaker breaking as an example, the past arcing time data for this circuit breaker breaking are: Sample 1: 27ms-29ms-34ms, Sample 2: 29ms-30ms-33ms..., and the average value of the operating database of similar circuit breakers: 30.3ms-31.2ms-32.9ms. Therefore, according to 3... s The rule can determine that the phase arcing time of the circuit breaker is abnormal. Judging from the degree of deviation, it is abnormal and the basic score is 4 points.

[0067] Abnormal arcing time is a symptom, and the equipment-fault mapping relationships involved include: jamming of the arc-extinguishing chamber mechanism of the circuit breaker. This has a very rare history, and once jamming occurs, it may fail to operate, and the symptoms are different from the current ones. Therefore, the weight coefficient is set to 1; abnormal relay protection operating circuit. Based on the analysis and statistics of historical hidden dangers, defects and faults, the weight coefficient is set to 2; external overvoltage influence, etc. Based on the analysis and statistics of historical hidden dangers, defects and faults, the weight coefficient is set to 3; mechanical traditional components have the highest correlation, and the weight coefficient is set to 4.

[0068] After the search is completed, i.e., the mechanically related components are identified, further searching is performed on the equipment-component dependency relationships. In a specific embodiment, firstly, based on the manufacturer's product manual, high-voltage technology, etc., equipment-component dependency relationships and equipment-fault mapping relationships are established. This allows for the chaining of mechanically related equipment, including: time relays, secondary control circuits, action relays, tripping coils, closing coils, tripping crank arms, closing crank arms, tripping stoppers, closing stoppers, insulating pull rods, and other components. From the components in the mechanically related components, the search begins with the factory inspection reports, maintenance records, repair records, and test record forms for each component. After the search is completed, it is found that the saddle-shaped current of the tripping coil has shown a certain distortion in the preventative testing over the past six years. If there is a 10% lag in the time node of the tripping coil's action current curve, it indicates a certain abnormality in the tripping circuit. Accordingly, if... Figure 6 As shown, it can be deduced that the tripping coil, tripping stop, and tripping crank arm of the tripping circuit are key inspection targets. Then, by integrating the mechanistic reasoning results with the standard retrieval results, a weighted fusion decision algorithm is used to classify the root cause of the fault into major hidden dangers, general hidden dangers, and minor hidden dangers based on health and importance. Given that this accident may lead to the failure of a critical substation to operate, considering the cause of the fault and the importance of the equipment, this is classified as a major hidden danger. After the analysis is completed, a large language model is used to process the output into natural language, providing it to maintenance and technical personnel for review.

[0069] The power industry-specific knowledge graph involved in this invention supports dynamic updates. It can iteratively supplement entities, attributes, and relationships in real time based on newly added equipment models, new fault cases, and updated industry standards and specifications, adapting to the fault diagnosis and assessment needs under different operating scenarios. The knowledge graph library constructed in this invention primarily aims to enhance reasoning capability and robustness, preventing over-reasoning or illusions. Through the method of this invention, the problem of retrieval bias caused by low semantic similarity of power industry texts and strong specificity of professional terminology is solved, improving the matching accuracy between fault scenarios and standards and specifications.

[0070] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A power equipment fault analysis method based on RAG knowledge graph, characterized in that, Includes the following steps: (1) Input the acquired multimodal data of power equipment into the power equipment-specific knowledge graph, and construct the equipment status assessment feature set through association reasoning using the power equipment-specific knowledge graph; (2) Based on the equipment condition assessment feature set, the RAG knowledge base is used for retrieval. When the acquired power equipment multimodal data is found to be abnormal or exceeds the allowable range, the fault data of the power equipment multimodal data is extracted and analyzed. (3) By integrating the reasoning results of the power equipment-specific knowledge graph with the extraction results of the RAG knowledge base, we can conduct equipment status assessment, classify hidden dangers and defects, and generate professional power operation and maintenance management guidance suggestions.

2. The power equipment fault analysis method based on RAG knowledge graph according to claim 1, characterized in that, The ontology architecture of the power industry-specific knowledge graph includes equipment dimensions, fault mechanism dimensions, and professional theory dimensions.

3. The power equipment fault analysis method based on RAG knowledge graph according to claim 2, characterized in that, The acquired multimodal data of the power equipment includes raw data, operational data, maintenance and testing data, and factors affecting the safe and stable operation of the equipment.

4. The power equipment fault analysis method based on RAG knowledge graph according to claim 4, characterized in that, The RAG knowledge base enables enhanced retrieval based on entity-relation-entity triples.

5. The power equipment fault analysis method based on RAG knowledge graph according to claim 4, characterized in that, During the retrieval process in step (2), if the device is detected to be performing an action, the device status is analyzed to determine whether it is caused by a dispatch order or an abnormal situation. If it is an abnormal situation, the fault data of the power equipment multimodal data is extracted and analyzed.

6. The power equipment fault analysis method based on RAG knowledge graph according to claim 1, characterized in that, In step (2), a 3D-based approach is adopted. σ The analysis is performed using the rules and statistical range of the equipment's secondary time parameters in both horizontal and vertical analysis algorithms. Step 1: Summarize the historical operation data of equipment of the same model from the same manufacturer and record it as the initial sample; Step 2: Preprocess the multimodal data of power equipment, remove abnormal data, and form the first data sample; Step 3: Calculate the sample mean and standard deviation of the first data sample; Step 4: Analyze the frequency distribution pattern of the first data sample and draw a frequency distribution graph; Step 5: Based on the abnormal data judgment and the degree of deviation, filter out severely biased data from the first data sample to form the second data sample; repeat steps 3 to 4. Step 6: Based on the outlier determination and deviation level standards, the initial sample data is divided into four levels: normal, attention, abnormal, and severe, according to the mean and standard deviation of the second data sample.

7. The power equipment fault analysis method based on RAG knowledge graph according to claim 6, characterized in that, According to statistical principles 3- σ The rules categorize the degree of deviation into levels.

8. The power equipment fault analysis method based on RAG knowledge graph according to claim 6, characterized in that, In step (3), the equipment status is evaluated based on the reliability of the faulty component. The reliability of the faulty component is determined by the degree of degradation of the status quantity and its weight. The degree of degradation of the status quantity corresponds to the deviation level, which is normal, attention, abnormal and severe. The more severe the level, the greater the weight.

9. A computer device comprising a memory and a processor, the memory being electrically connected to the processor, the memory storing a computer program, characterized in that: When the computer program is executed by the processor, it causes the processor to implement the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the processor implements the method as described in any one of claims 1 to 8.

Images