Equipment Diagnosis Method and System Based on Power Grid Event Feature Matching and Event Source Tracing

Abstract

A device diagnosis method and system based on power grid event feature matching and event tracing includes: collecting and analyzing events occurring in the power grid to construct an event rule base; establishing a custom dictionary and building an inverted index library based on the custom dictionary; when querying device information or issuing early warnings of sudden device faults, using a feature extraction model to extract features from the device information and fault warning information to obtain query vectors, and using RAG technology and a vector index database to perform feature matching on the query vectors to obtain alarm events; matching the alarm events with the event rule base to derive the event characteristics and conditions under which the events occurred, and combining the inverted index library to infer the cause and location of the alarm events. This invention utilizes a rule base for matching, eliminating the need for complex data analysis and calculations, and uses a Transformer model with contextual information, thus improving the accuracy of fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of equipment diagnostic technology, and more specifically, relates to an equipment diagnostic method and system based on power grid event feature matching and event tracing. Background Technology

[0002] With the rapid development of power grid systems, their scale and complexity are constantly increasing, placing higher demands on the monitoring, diagnosis, and maintenance of power grid equipment. In power systems, equipment failures often lead to abnormal grid operation and even serious consequences such as large-scale power outages. Therefore, timely and accurate fault diagnosis of power grid equipment is of great significance for ensuring the safe and stable operation of the power system.

[0003] Fault diagnosis technology primarily involves monitoring and analyzing the operating status of equipment to promptly detect abnormalities and predict potential faults, thereby preventing losses caused by equipment failures. Meanwhile, with the continuous development of artificial intelligence technology, Natural Language Processing (NLP) technology is also increasingly being applied to power system management. NLP technology is capable of understanding and processing human natural language, helping people to more easily acquire, process, and utilize information. In power systems, NLP technology can be used to analyze reports and records of power grid events, thereby enabling a more accurate understanding and handling of these events.

[0004] Existing fault diagnosis technologies primarily rely on monitoring equipment operating status, collecting relevant data, and using data analysis and logical reasoning to infer the cause and location of equipment failures. This approach requires large amounts of data and complex calculations, has numerous limitations, and demands high data accuracy and completeness. Furthermore, existing power system management matching mechanisms are not flexible enough, requiring matching based on manually set rules, which is insufficient for the increasingly complex power grid operating environment.

[0005] Fault diagnosis techniques related to natural language processing include:

[0006] Patent CN118410360B discloses a method and system for judging power line fault events based on natural language processing. This includes acquiring switch remote signaling change data and protection device action information data, converting them into uniformly formatted data sets; analyzing and processing the data sets using natural language processing technology to extract key information and establish an event rule model; automatically generating power line outage events according to predetermined rules and judging the fault type of the power line outage events; comparing the event judgment based on the fault model with the results verified on-site, updating and correcting the power line outage events in real time, and optimizing the event rule model. However, this method relies excessively on natural language processing technology and the reliability of the data source. If the technology malfunctions or if there are errors or omissions in the switch remote signaling change data and protection device action information, it will directly affect the accuracy of subsequent analysis results, directly impacting the normal operation of the system and the judgment results. Furthermore, this method mainly focuses on judging and correcting power line fault events that have already occurred, lacking the ability to trace the source of fault events and locate the cause of equipment failures. It mainly relies on manual on-site verification for correction, increasing the dependence on manual intervention and costs. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a device diagnosis method and system based on power grid event feature matching and event tracing.

[0008] The present invention adopts the following technical solution.

[0009] The first aspect of this invention proposes a device diagnosis method based on power grid event feature matching and event tracing, specifically:

[0010] By collecting and analyzing events occurring in the power grid, an event rule base is constructed. The event rule base includes event feature rule bases for different devices and event-based model rule bases. The event feature rule base contains event features corresponding to faults or anomalies in different devices, and the event-based model base contains event attributes and conditions for event occurrence. The conditions for an event to occur are that several event features appear and these event features satisfy a certain event logic, at which point the event is determined to have occurred.

[0011] A custom dictionary is established, and an inverted index library is built based on the custom dictionary. The custom dictionary is a custom dictionary containing power grid terminology, and the inverted index library is an index library that can find the document or document fragment containing the words in the power grid monitoring information. The monitoring information includes equipment status, operation records, and fault alarms.

[0012] A feature extraction model is established to collect text data related to power grid equipment faults. The data is then segmented into words. After word segmentation, the feature extraction model is used to extract features to obtain multi-dimensional vectors. All extracted multi-dimensional vectors are stored together to form a vector index database.

[0013] When performing equipment information queries or issuing sudden equipment failure warnings, feature extraction models are used to extract features from equipment information and failure warning information to obtain query vectors. RAG technology and vector index database are then used to match the features of the query vectors to obtain alarm events.

[0014] The alarm event is matched with the event feature rule base to obtain the event features. The alarm event is matched with the event-based model feature rule base to obtain the conditions under which the event occurs. Combined with the inverted index library, the cause and location of the alarm event are inferred. An alarm is issued and the alarm event, its event features, cause and location are displayed.

[0015] Preferably, the event feature rule base includes event features such as alarm type, alarm status, objectification information, equipment voltage level, circuit breaker connected equipment information, measurement data configuration, preconditions, and associated sub-events; the event feature rule base is divided into different categories according to equipment, including main transformer fault event rule base, busbar fault event rule base, line fault event rule base, automatic transfer switch fault event rule base, other fault event rule base, main transformer abnormal event rule base, busbar abnormal event rule base, line abnormal event rule base, automatic transfer switch abnormal event rule base, other abnormal event rule base, auxiliary equipment event rule base, and main and auxiliary equipment linkage rule base.

[0016] The event attributes in the event-based model rule base include the scope of operation, event priority, event type, event level, event category, and event device type. Specifically, the scope of operation includes the same substation, same bay, same main transformer, same equipment, circuit breaker connected equipment, and same line, etc.; the event priority includes high, medium, and low; the event type includes accident, abnormality, over-limit, change, and notification; the event level includes critical, severe, serious, and general; the event category includes comprehensive events and combined events; and the event device type includes line, busbar, transformer, and switch, etc.

[0017] Preferably, the step of establishing a custom dictionary specifically involves:

[0018] Monitoring information of the power grid is collected in advance. Based on word segmentation tools, all the collected monitoring information is divided into terms to form a custom dictionary.

[0019] Preferably, the establishment of the inverted index database specifically includes:

[0020] Based on the Lucene full-text search framework, each word obtained by segmenting all monitoring information is assigned one or more index terms. These index terms point to documents or document fragments containing the word. Each term in the custom dictionary is also added to the index, ultimately forming an inverted index library.

[0021] Preferably, the text data related to power grid equipment faults includes power grid fault reports, maintenance records, user feedback, and equipment logs;

[0022] The feature extraction model uses a bidirectional encoder representation model from Transformer. It uses the same word segmentation tools used to build the custom dictionary and inverted index library to segment the text data related to power grid equipment faults, and then converts it into a format that the feature extraction model can process. This format includes NumPy arrays, Pandas DataFrames, TensorFlow / Keras Datasets, or TFRecords. The feature extraction model is pre-trained using these formats. Then, the text data related to power grid equipment faults is labeled according to event classification, event fault causes, and fault analysis to obtain a labeled power grid equipment fault event dataset. This dataset is used to further train the model.

[0023] Preferably, the Transformer model is an end-to-end Seq2Seq structure with a 6-layer stacked structure, stacking 6 Encoders and Decoders for text input and output. The output multidimensional vector has a dimension of 6, and each vector represents the position of a word or phrase in the embedding space, containing semantic information of the words, the association between words and contextual information.

[0024] Preferably, RAG technology is used for fast retrieval and matching, specifically:

[0025] Similarity is ranked by calculating the cosine similarity of the vectors:

[0026] The formula for calculating cosine similarity is as follows:

[0027]

[0028] Here, A and B are two vectors whose similarity needs to be compared. i and B i These are the components of vectors A and B, respectively, where n is the dimension of the vector;

[0029] Calculate the cosine similarity between all vectors in the vector index database and the query vector, sort them by similarity, select the knowledge text or historical records with the highest similarity, select the m results with the highest similarity as candidate answers based on the similarity sorting results, combine the candidate answers with the device information or fault warning information of the query, and extract the alarm event.

[0030] Preferably, the step of matching the alarm event with the event feature rule base to obtain the event features, matching the alarm event with the event-based model feature rule base to obtain the conditions for the event to occur, and combining the custom dictionary and the inverted index library to deduce the cause and location of the alarm event, specifically: querying the inverted index library to find the event features of alarm events contained in the information of each device and the number of times these features occur, deducing the complete event logic from the conditions for the event to obtain the specific device where the event occurred, and obtaining the location of the device from the device information.

[0031] Preferably, when extracting an alarm event, determining its fault type, issuing an alarm, and displaying it, event priority management is performed. Specifically, when a new event is extracted, it is analyzed whether there is a business logic relationship between all existing events and the new event. The business logic is that the two events occur on the same device or that they are precursors or prerequisites. If there is a business logic, the new event is defined as the parent event, and all existing events with business logic related to the new event are combined to define the child event. The event priorities of the parent event and the child event are extracted from the event-based model rule base. If the set priority of the parent event is higher than that of the child event, only the alarm result of the parent event is displayed; otherwise, the alarm results of both the parent and child events are displayed.

[0032] A second aspect of the present invention provides a device diagnostic system based on power grid event feature matching and event tracing using the power grid fault diagnosis method described in the first aspect of the present invention, specifically as follows:

[0033] Rule base building module: Collects and analyzes events occurring in the power grid, and builds an event feature rule base and an event-based model event rule base;

[0034] Inverted index database building system: Collected typical monitoring information is divided according to terms, and based on the Lucene full-text search framework, one or more index items are assigned to each word and term of typical monitoring information to form an inverted index database;

[0035] Feature extraction module: The collected text data related to power grid equipment faults is segmented into words. After segmentation, the bidirectional encoder representation model from Teasformer is used to transform it into a multidimensional vector and store it in the vector index database. When performing equipment information queries or sudden equipment fault warnings, the equipment information and the fault warning information are transformed into query vectors through the bidirectional encoder representation model from Teasformer.

[0036] Matching module: Uses RAG technology to search and query knowledge text or historical records with the highest vector similarity in the vector index database;

[0037] Fault Analysis Module: Matches alarm events with the event feature rule base to obtain the event features, matches alarm events with the event-based model feature rule base to obtain the conditions under which the event occurred, and combines the inverted index library to deduce the cause and location of the alarm event.

[0038] The beneficial effects of this invention are that, compared with the prior art,

[0039] (1) It solves the problem that existing fault diagnosis technologies require a large amount of data and complex calculations.

[0040] By establishing an event rule base, historical events, their causes, and types accumulated in the power grid operation system are collected and integrated, covering the correspondence between common fault modes, fault phenomena, and fault causes. During fault diagnosis, the system can directly call rules from the event rule base for matching and reasoning, without needing to perform complex data analysis and calculations again. This reduces the reliance on large amounts of real-time data and the need for complex calculations, thereby lowering the implementation and maintenance costs of fault diagnosis technology.

[0041] (2) This solves the problem of insufficient flexibility in matching existing power grid equipment fault diagnosis technologies.

[0042] Applying the Transformer model and RAG fast indexing matching technology to parse power grid event text enables more accurate extraction of key information from textual information such as fault reports and maintenance records, thereby improving the accuracy of fault diagnosis. Furthermore, the applied Transformer model is not limited to keyword matching but can also understand the semantics and contextual relationships of the text. This allows for more flexible identification and processing of various fault information in the complex and ever-changing power grid operating environment.

[0043] (3) It solves the problems and limitations of existing technologies in fault diagnosis of power grid equipment.

[0044] By introducing an event rule base and multi-dimensional information such as the physical characteristics and operational patterns of the equipment, the reliance on data integrity and accuracy is reduced. Physical characteristics and operational patterns are used to determine the behavior patterns of the equipment under normal and abnormal conditions. Through rapid matching and reasoning using the event rule base, the scope of the fault is narrowed down and the fault point is located, improving diagnostic efficiency and overcoming the limitations of existing technologies. Attached Figure Description

[0045] Figure 1 The Transformer model structure;

[0046] Figure 2 The encoding and decoding process for the Transformer model;

[0047] Figure 3 This refers to the vector transformation process of the Transformer model.

[0048] Figure 4 For RAG fast index matching process;

[0049] Figure 5 This is a flowchart of the present invention;

[0050] Figure 6 Logic diagram for handling abnormal situations;

[0051] Figure 7 Logic diagram for judging abnormal three-phase current conditions;

[0052] Figure 8 Logic diagram for judging abnormal opening and closing conditions;

[0053] Figure 9 This is a logic diagram for judging abnormal spring energy storage conditions after the circuit breaker operates. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0055] like Figure 5 As shown, Embodiment 1 of the present invention provides a device diagnosis method based on power grid event feature matching and event tracing, specifically as follows:

[0056] By collecting and analyzing events occurring in the power grid, an event rule base is constructed. This event rule base includes an event feature rule base and an event-based model rule base. The event feature rule base contains the event characteristics of each event, including alarm type, alarm status, object-specific information, equipment voltage level, circuit breaker connection equipment information, measurement data configuration, preconditions, and associated sub-events. The event-based model rule base contains various attributes of events and the conditions for their occurrence. Attributes include calculation scope, event priority, event type, event level, event category, and event equipment type. The condition for an event to occur is the presence of several event characteristics, and when these characteristics satisfy a certain event logic, the event is determined to have occurred.

[0057] For example, the conditions for the occurrence of events entered into the event-based model rule base in this embodiment include: circuit breaker energy storage mechanism abnormality rules, three-phase current abnormality rules, opening and closing abnormality rules, and spring energy storage abnormality rules after operation. Specifically:

[0058] like Figure 6 As shown, the abnormal rules for the circuit breaker energy storage mechanism are as follows:

[0059] The number of alarms triggered by suppression actions within the statistical period is counted. If the number exceeds the frequent suppression warning or alarm limit, it is judged as frequent suppression.

[0060] Calculate the difference between the number of suppressions and the previous suppression count. If the difference is greater than the suppression trend warning or alarm limit, it is determined that the suppression count trend has exceeded the limit.

[0061] If the timing of the pressure start action is not reset and exceeds the abnormal pressure duration warning or alarm limit, it is judged as a pressure timeout.

[0062] The trend of the change is calculated by comparing the average duration of the previous three suppressions. If it is greater than the suppression trend warning or alarm limit, it is judged that the suppression duration trend has exceeded the limit.

[0063] If any of the above values ​​exceeds its corresponding limit, the circuit breaker energy storage mechanism is deemed abnormal. If any of the above values ​​exceeds its corresponding warning limit, a warning message is sent. If any of the above values ​​exceeds its corresponding alarm limit, an alarm message is sent.

[0064] Specifically, the frequent suppression warning limit is set to 4 times, and the frequent suppression alarm limit is set to 10 times;

[0065] The suppression trend warning limit is set to 100%, and the suppression trend alert limit is set to 200%.

[0066] The abnormal suppression duration warning limit is set to 180 seconds, and the abnormal suppression duration alarm limit is set to 360 seconds.

[0067] like Figure 7 As shown, the three-phase current anomaly rules are as follows:

[0068] Calculate the deviation of the currents in phases A, B, and C and the balance of the three-phase currents. If any one of the current deviations exceeds the deviation limit and the balance of the three-phase currents exceeds the balance warning limit, it is determined that the three-phase currents of the circuit breaker are seriously unbalanced and a warning signal is sent.

[0069] Specifically, the circuit deviation limit is set at 10%; the balance warning limit is set at 10%.

[0070] like Figure 8 As shown, the abnormal opening and closing rules are as follows:

[0071] Calculate the maximum time difference and transmission interval of the three-phase remote signaling SOE signal. If the maximum time difference is greater than the time limit or the transmission interval is greater than the delay time threshold, it is judged that the opening and closing are seriously out of sync, and an early warning signal is sent.

[0072] Specifically, the time limit is set to 10ms; the delay time threshold is set to 3 seconds.

[0073] like Figure 9 As shown, the abnormal rule for spring energy storage after action is as follows:

[0074] Record the time for waiting for the energy storage signal after the circuit breaker operates. If the time exceeds the energy storage signal time limit, it is determined that no energy has been stored.

[0075] If the timeout period for the reserve start-up action signal exceeds the timeout limit, it is judged as an energy storage timeout.

[0076] The trend of the change is calculated by comparing the average energy storage duration with the previous three values. If the trend exceeds the energy storage trend limit, it is judged that the energy storage duration trend exceeds the limit.

[0077] If any of the above values ​​exceeds its corresponding limit, it is determined that the spring energy storage is abnormal after the circuit breaker trips, and an early warning message is sent.

[0078] Specifically, the energy storage signal time limit is set to 2s, the non-reset time limit is set to 60s, and the energy storage trend limit is set to 30%.

[0079] Preferably, the event feature rule base includes a main transformer fault event rule base, a busbar fault event rule base, a line fault event rule base, a backup automatic transfer fault event rule base, other fault event rule bases, a main transformer abnormal event rule base, a busbar abnormal event rule base, a line abnormal event rule base, a backup automatic transfer abnormal event rule base, other abnormal event rule bases, an auxiliary equipment event rule base, and a main and auxiliary equipment linkage rule base. The various attributes of events in the event-based model rule base include the calculation scope, event priority, event type, event level, event category, and event equipment type. Specifically: the calculation scope includes the same substation, the same bay, the same main transformer, the same equipment, circuit breaker connected equipment, the same line, etc.; the event priority includes high, medium, and low; the event type includes accident, abnormal, over-limit, change, and notification; the event level includes critical, severe, serious, and general; the event category includes comprehensive events and combined events; and the event equipment type includes lines, buses, transformers, switches, etc.

[0080] A custom dictionary is established, and an inverted index library is built based on the custom dictionary. The custom dictionary is a custom dictionary containing power grid terminology, and the inverted index library is an index library that can find the document or document fragment containing the words in the power grid monitoring information. The monitoring information includes equipment status, operation records, and fault alarms.

[0081] Preferably, the step of establishing a custom dictionary specifically involves:

[0082] Monitoring information of the power grid is collected in advance. Based on word segmentation tools, all the collected monitoring information is divided into terms to form a custom dictionary.

[0083] Preferably, the establishment of the inverted index database specifically includes:

[0084] Based on the Lucene full-text search framework, each word obtained by segmenting all monitoring information is assigned one or more index terms. These index terms point to documents or document fragments containing the word. Each term in the custom dictionary is also added to the index, ultimately forming an inverted index library.

[0085] Collect text data related to power grid equipment faults, segment the data into words, and then use a feature extraction model to extract features to obtain multi-dimensional vectors. Store all the extracted multi-dimensional vectors together to form a vector index database.

[0086] When performing equipment information queries or issuing sudden equipment failure warnings, feature extraction models are used to extract features from equipment information and failure warning information to obtain query vectors. RAG technology and vector index database are then used to match the features of the query vectors to obtain alarm events.

[0087] Preferably, the text data related to power grid equipment faults includes power grid fault reports, maintenance records, user feedback, and equipment logs;

[0088] like Figure 1 As shown, the feature extraction model uses a bidirectional encoder representation model from Transformer. It uses the same word segmentation tools used to build the custom dictionary and inverted index library to segment the text data related to power grid equipment faults, then converts it into a format that the feature extraction model can process. This format includes NumPy arrays, PandasDataFrames, TensorFlow / Keras Datasets, or TFRecords. This is used to pre-train the feature extraction model. Then, the text data related to power grid equipment faults is labeled according to event classification, fault causes, and fault analysis to obtain a labeled power grid equipment fault event dataset. This dataset is then used to further train the model.

[0089] Preferably, such as Figure 2 As shown, the Transformer model is an end-to-end Seq2Seq structure with a 6-layer stacked structure. It stacks 6 Encoders and Decoders for text input and output. The output multidimensional vector has a dimension of 6. Each vector represents the position of a word or phrase in the embedding space and contains semantic information of the words, the association between words and contextual information.

[0090] It should be noted that the contextual information contained in the vector is obtained through the self-attention mechanism in the Transformer model, specifically:

[0091] like Figure 3 As shown, the Transformer model first converts all text into a set of vectors. Each vector is then transformed linearly (i.e., multiplied by a weight matrix and a bias term is added) to obtain the query vector Q, key vector K, and value vector V. For each query vector, its similarity to all other key vectors is calculated. The similarity calculation formula is:

[0092]

[0093] Where Q is the query vector, K T It is the transpose of the key vector, d k This is the dimension of the key vector; based on the similarity calculation results, the softmax function is used to normalize the similarity scores to the [0,1] interval, ensuring that the sum of all normalized similarity scores is 1; each query vector then obtains a weighted vector.

[0094] Multiply the value vector corresponding to each query vector by its corresponding weight, and sum all the weighted value vectors; the result of this weighted summation is the information after taking into account the context.

[0095] Preferably, such as Figure 4 As shown, RAG technology is used for fast retrieval and matching, specifically as follows:

[0096] Similarity is ranked by calculating the cosine similarity of the vectors:

[0097] The formula for calculating cosine similarity is as follows:

[0098]

[0099] Here, A and B are two vectors whose similarity needs to be compared. i and B i These are the components of vectors A and B, respectively, where n is the dimension of the vector;

[0100] Calculate the cosine similarity between all vectors in the vector index database and the query vector, sort them by similarity, select the knowledge text or historical records with the highest similarity, select the m results with the highest similarity as candidate answers based on the similarity sorting results, combine the candidate answers with the device information or fault warning information of the query, and extract the alarm event.

[0101] The alarm event is matched with the event feature rule base to obtain the event features. The alarm event is matched with the event-based model feature rule base to obtain the conditions under which the event occurs. Combined with the inverted index library, the cause and location of the alarm event are inferred. An alarm is issued and the alarm event, its event features, cause and location are displayed.

[0102] Preferably, the step of matching the alarm event with the event feature rule base to obtain the event features, matching the alarm event with the event-based model feature rule base to obtain the conditions for the event to occur, and combining the custom dictionary and the inverted index library to deduce the cause and location of the alarm event, specifically: querying the inverted index library to find the event features of alarm events contained in the information of each device and the number of times these features occur, deducing the complete event logic from the conditions for the event to obtain the specific device where the event occurred, and obtaining the location of the device from the device information.

[0103] Preferably, when extracting an alarm event, determining its fault type, issuing an alarm, and displaying it, event priority management is performed. Specifically, when a new event is extracted, it is analyzed whether there is a business logic relationship between all existing events and the new event. The business logic is that the two events occur on the same device or that they are precursors or prerequisites. If there is a business logic, the new event is defined as the parent event, and all existing events with business logic related to the new event are combined to define the child event. The event priorities of the parent event and the child event are extracted from the event-based model rule base. If the set priority of the parent event is higher than that of the child event, only the alarm result of the parent event is displayed; otherwise, the alarm results of both the parent and child events are displayed.

[0104] In actual testing, the feature retrieval matching accuracy of this embodiment reached over 98%.

[0105] Embodiment 2 of the present invention proposes a device diagnosis system based on power grid event feature matching and event tracing using the power grid fault diagnosis method described in Embodiment 1 of the present invention, specifically as follows:

[0106] Rule base building module: Collects and analyzes events occurring in the power grid, and builds an event feature rule base and an event-based model event rule base;

[0107] Inverted index database building system: Collected typical monitoring information is divided according to terms, and based on the Lucene full-text search framework, one or more index items are assigned to each word and term of typical monitoring information to form an inverted index database;

[0108] Feature extraction module: The collected text data related to power grid equipment faults is segmented into words. After segmentation, the bidirectional encoder representation model from Teasformer is used to transform it into a multidimensional vector and store it in the vector index database. When performing equipment information queries or sudden equipment fault warnings, the equipment information and the fault warning information are transformed into query vectors through the bidirectional encoder representation model from Teasformer.

[0109] Matching module: Uses RAG technology to search and query knowledge text or historical records with the highest vector similarity in the vector index database;

[0110] Fault Analysis Module: Matches alarm events with the event feature rule base to obtain the event features, matches alarm events with the event-based model feature rule base to obtain the conditions under which the event occurred, and combines the inverted index library to deduce the cause and location of the alarm event.

[0111] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0112] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0113] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0114] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A device diagnostic method based on power grid event feature matching and event tracing, characterized in that, include: By collecting and analyzing events occurring in the power grid, an event rule base is constructed. The event rule base includes event feature rule bases for different devices and event-based model rule bases. The event feature rule base contains event features corresponding to faults or anomalies in different devices, and the event-based model base contains event attributes and conditions for event occurrence. The conditions for an event to occur are that several event features appear and these event features satisfy a certain event logic, at which point the event is determined to have occurred. A custom dictionary is established, and an inverted index library is built based on the custom dictionary. The custom dictionary is a custom dictionary containing power grid terminology, and the inverted index library is an index library that can find the document or document fragment containing the words in the power grid monitoring information. The monitoring information includes equipment status, operation records, and fault alarms. A feature extraction model is established to collect text data related to power grid equipment faults. The data is then segmented into words. After word segmentation, the feature extraction model is used to extract features to obtain multi-dimensional vectors. All extracted multi-dimensional vectors are stored together to form a vector index database. When performing equipment information queries or issuing sudden equipment failure warnings, feature extraction models are used to extract features from equipment information and failure warning information to obtain query vectors. RAG technology and vector index database are then used to match the features of the query vectors to obtain alarm events. The alarm event is matched with the event feature rule base to obtain the event features. The alarm event is matched with the event-based model feature rule base to obtain the conditions under which the event occurs. Combined with the inverted index library, the cause and location of the alarm event are inferred. An alarm is issued and the alarm event, its event features, cause and location are displayed.

2. The equipment diagnosis method based on power grid event feature matching and event tracing according to claim 1, characterized in that: The event feature rule base includes event features such as alarm type, alarm status, objectification information, equipment voltage level, circuit breaker connected equipment information, measurement data configuration, preconditions, and associated sub-events. The event feature rule base is divided into different categories based on equipment: main transformer fault event rule base, busbar fault event rule base, line fault event rule base, automatic transfer switch fault event rule base, other fault event rule base, main transformer abnormal event rule base, busbar abnormal event rule base, line abnormal event rule base, automatic transfer switch abnormal event rule base, other abnormal event rule base, auxiliary equipment event rule base, and main / auxiliary equipment linkage rule base. The event attributes in the event-based model rule base include the scope of operation, event priority, event type, event level, event category, and event device type. Specifically: the scope of operation includes the same substation, same bay, same main transformer, same equipment, circuit breaker connected equipment, and same line; the event priority includes high, medium, and low; the event type includes accident, abnormality, over-limit, change, and notification; the event level includes critical, severe, serious, and general; the event category includes comprehensive events and combined events; and the event device type includes line, busbar, transformer, and switch.

3. The equipment diagnosis method based on power grid event feature matching and event tracing according to claim 1, characterized in that: The creation of the custom dictionary specifically involves: Monitoring information of the power grid is collected in advance. Based on word segmentation tools, all the collected monitoring information is divided into terms to form a custom dictionary.

4. The equipment diagnosis method based on power grid event feature matching and event tracing according to claim 3, characterized in that: The establishment of the inverted index database specifically involves: Based on the Lucene full-text search framework, each word obtained by segmenting all monitoring information is assigned one or more index terms. These index terms point to documents or document fragments containing the word. Each term in the custom dictionary is also added to the index, ultimately forming an inverted index library.

5. The equipment diagnosis method based on power grid event feature matching and event tracing according to claim 4, characterized in that: The text data related to power grid equipment faults includes power grid fault reports, maintenance records, user feedback, and equipment logs; The feature extraction model uses a bidirectional encoder representation model from Transformer. It segments the text data related to power grid equipment faults using the same word segmentation tools used to build the custom dictionary and inverted index library, then converts it to a format that the feature extraction model can process. This format includes NumPy arrays, Pandas DataFrames, TensorFlow / Keras Datasets, or TFRecords. The processed text data related to power grid equipment faults is used to pre-train the feature extraction model. The text data is then labeled according to event classification, fault causes, and fault analysis to obtain a labeled power grid equipment fault event dataset, which is used to further train the model.

6. The equipment diagnosis method based on power grid event feature matching and event tracing according to claim 5, characterized in that: The Transformer model is an end-to-end Seq2Seq structure with a 6-layer stacked structure. It stacks 6 Encoders and Decoders for text input and output. The output multidimensional vector has a dimension of 6. Each vector represents the position of a word or phrase in the embedding space and contains semantic information of the words, the association between words and contextual information.

7. The equipment diagnosis method based on power grid event feature matching and event tracing according to claim 1, characterized in that: Using RAG technology for fast retrieval and matching, specifically: Similarity is ranked by calculating the cosine similarity of the vectors: The formula for calculating cosine similarity is as follows: Here, A and B are two vectors whose similarity needs to be compared. i and B i These are the components of vectors A and B, respectively, where n is the dimension of the vector; Calculate the cosine similarity between all vectors in the vector index database and the query vector, sort them by similarity, select the knowledge text or historical records with the highest similarity, select the m results with the highest similarity as candidate answers based on the similarity sorting results, combine the candidate answers with the device information or fault warning information of the query, and extract the alarm event.

8. The equipment diagnosis method based on power grid event feature matching and event tracing according to claim 2, characterized in that: The process involves matching alarm events with an event feature rule base to obtain the event features, matching alarm events with an event-based model feature rule base to obtain the conditions under which the event occurred, and combining a custom dictionary and an inverted index library to deduce the cause and location of the alarm event. Specifically, this involves querying the inverted index library to find the event features of alarm events and the frequency of these features in the information of each device, deducing the complete event logic from the conditions under which the event occurred, obtaining the specific device where the event occurred, and obtaining the location of the device from its device information.

9. The equipment diagnosis method based on power grid event feature matching and event tracing according to claim 2, characterized in that: When an alarm event is extracted, its fault type is determined, and an alarm is issued and displayed, event priority management is performed. Specifically, when a new event is extracted, it is analyzed whether there is a business logic relationship between all existing events and the new event. The business logic is that the two events occur on the same device or that they are precursors or prerequisites. If a business logic exists, the new event is defined as the parent event, and all existing events with business logic related to the new event are combined to define the child event. The event priorities of the parent and child events are extracted from the event-based model rule base. If the set priority of the parent event is higher than that of the child event, only the alarm result of the parent event is displayed; otherwise, the alarm results of both the parent and child events are displayed.

10. A device diagnostic system based on power grid event feature matching and event tracing using the power grid fault diagnosis method according to any one of claims 1-9, characterized in that: Rule base building module: Collects and analyzes events occurring in the power grid, and builds an event feature rule base and an event-based model event rule base; Inverted index database building system: Collected typical monitoring information is divided according to terms, and based on the Lucene full-text search framework, one or more index items are assigned to each word and term of typical monitoring information to form an inverted index database; Feature extraction module: The collected text data related to power grid equipment faults is segmented into words. After segmentation, the bidirectional encoder representation model from Teasformer is used to transform it into a multidimensional vector and store it in the vector index database. When performing equipment information queries or sudden equipment fault warnings, the equipment information and the fault warning information are transformed into query vectors through the bidirectional encoder representation model from Teasformer. Matching module: Uses RAG technology to search and query knowledge text or historical records with the highest vector similarity in the vector index database; Fault Analysis Module: Matches alarm events with the event feature rule base to obtain the event features, matches alarm events with the event-based model feature rule base to obtain the conditions under which the event occurred, and combines the inverted index library to deduce the cause and location of the alarm event.

Images