A method and system for analyzing and diagnosing faults of PSCADA messages based on a Bert large language model

Abstract

The application provides a PSCADA message analysis and fault diagnosis method and system based on a Bert large language model, belongs to the field of intelligent operation and maintenance of power systems, and solves the technical defects of weak semantic understanding, low fault diagnosis precision, low operation and maintenance efficiency and difficult knowledge reuse in the existing PSCADA message analysis technology. The application provides a PSCADA message analysis and fault diagnosis method and system, collects and standardizes power messages, extracts semantics by using an improved BERT, completes Bayesian fault reasoning by combining a probability graph model, automatically generates an operation and maintenance report, and realizes intelligent diagnosis of power system faults. The application effectively improves message semantic recognition and fault diagnosis precision, reduces missed diagnosis and misdiagnosis problems. Full-process automation significantly improves operation and maintenance efficiency, realizes efficient reuse of fault knowledge, is compatible with various power communication protocols, has strong system expansibility, and is suitable for various power operation and maintenance scenes.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for power systems, and in particular to a method and system for PSCADA message analysis and fault diagnosis based on the Bert large language model. Background Technology

[0002] PSCADA (Power Monitoring and Data Acquisition) system is the core system for power system operation monitoring, responsible for real-time acquisition, transmission, and processing of operational status data from power equipment such as substations, overhead contact lines, and switchgear. During system operation, PSCADA generates a large amount of message data containing equipment status, alarm information, and event logs, which are crucial for fault diagnosis.

[0003] The existing PSCADA message analysis technology has the following main drawbacks: (1) Weak semantic comprehension ability Existing methods primarily rely on regular expression matching and keyword retrieval for message analysis, lacking the ability to understand the deep semantics of messages. For messages with diverse expressions but identical meanings, existing methods struggle to accurately identify their semantic equivalence, resulting in a semantic extraction accuracy rate of less than 75%, and failing to effectively handle non-standardized, colloquial message descriptions.

[0004] (2) Low fault diagnosis accuracy Traditional methods rely on manually set fixed rules for fault diagnosis, lacking the ability to effectively process uncertain information. When faced with complex or latent faults, rule-matching methods are prone to misdiagnosis and missed diagnosis, with a missed diagnosis rate of no less than 15% and a misdiagnosis rate of no less than 10%, making it difficult to meet the requirements of high reliability operation of power systems.

[0005] (3) Low operation and maintenance efficiency Currently, PSCADA message analysis mainly relies on manual methods. Maintenance personnel need to review each message, compare it with the rule manual, and analyze the cause of the fault. A single fault analysis takes no less than 30 minutes. With the expansion of the power system and the surge in the number of messages, manual analysis methods can no longer meet the needs of rapid response.

[0006] (4) Difficulty in knowledge reuse Fault diagnosis knowledge mainly relies on expert experience and is stored in a scattered manner in the form of rule documents, lacking a unified knowledge representation and reuse mechanism. When experts leave or rules are updated, the performance of the diagnostic system deteriorates significantly, and knowledge transfer and reuse face serious challenges.

[0007] (5) Poor system scalability Adding new fault types requires manually writing new diagnostic rules, with an adaptation period of no less than two weeks. Furthermore, there are conflicts and redundancies among the rules, making it difficult to quickly adapt to the needs of power system equipment updates and fault mode evolution.

[0008] In recent years, with the development of deep learning technology, neural network-based natural language processing methods have made significant progress in tasks such as text classification, named entity recognition, and semantic matching. The BERT model, as a bidirectional Transformer pre-trained language model, can effectively capture the deep semantic features of text through pre-training on large-scale corpora and fine-tuning on downstream tasks. However, directly applying the general BERT model to PSCADA message analysis still has the following problems: the semantic specificity of power industry terms is not fully modeled; the general pre-trained model has insufficient sensitivity to power equipment status descriptions; and there is a lack of effective integration with fault diagnosis reasoning mechanisms.

[0009] Therefore, there is an urgent need for a PSCADA message analysis and fault diagnosis method that integrates deep semantic understanding and uncertainty reasoning to improve the accuracy, efficiency and scalability of fault diagnosis. Summary of the Invention

[0010] This invention proposes a PSCADA message analysis and fault diagnosis method and system based on the BERT large language model. By constructing an improved BERT model to extract deep semantic vectors of messages and integrating a probabilistic graphical model for Bayesian inference, it realizes intelligent processing from message understanding to fault diagnosis, thereby solving the technical defects of existing PSCADA message analysis technology, such as weak semantic understanding, low fault diagnosis accuracy, low operation and maintenance efficiency, and difficulty in knowledge reuse.

[0011] A method for PSCADA message analysis and fault diagnosis based on the BERT large language model includes the following steps: S1. Message Data Acquisition: Through the PSCADA system API interface, communication protocol parser and log acquisition module, the PSCADA system operation messages are collected in real time according to the preset acquisition period. The operation messages include switch messages, analog messages, event sequence record messages and alarm messages. The acquisition period is no more than 30 seconds. S2. Message Preprocessing: The collected raw messages are processed sequentially by regular expression segmentation, noise reduction algorithm to remove redundant information, numerical normalization and entity alignment to convert unstructured message text into a standardized token sequence. The preprocessing accuracy is no less than 98%. S3. BERT Semantic Understanding: The preprocessed standardized token sequence is input into the improved BERT model. This improved BERT model introduces a domain adaptation layer and a dynamic masking strategy, integrates power domain knowledge vectors for deep semantic encoding, and outputs a message semantic vector. The semantic vector extraction formula is as follows:

[0012] In the formula, The output semantic vector, Input a sequence of tokens for the segmented message. For Bert encoder functions, For knowledge vectors in the power sector, To adapt the weight matrix to the domain, This represents matrix multiplication. S4. Fault Diagnosis: The semantic vector is used as evidence input into the probabilistic graphical model. Bayesian inference is performed by combining the prior probability and conditional probability of the fault to calculate the posterior probability of each fault type. The fault type corresponding to the highest posterior probability is taken as the diagnosis result. The fault probability calculation formula is:

[0013] In the formula, In the evidence Fault types under conditions The posterior probability, The evidence set constituted by the semantic vectors. Fault type The prior probability, Fault type Provide evidence The conditional probability (likelihood). This represents the total number of fault types. For the first Types of faults; S5. Result Output and Operation and Maintenance Assistance: Generate a structured fault report based on the fault diagnosis results. The fault report includes the fault type, fault cause, scope of impact, handling suggestions and historical similar fault cases, and feeds back the diagnosis results and handling solutions to the operation and maintenance personnel's terminal.

[0014] Furthermore, the message data acquisition in S1 also includes: synchronizing the timestamps and marking the source of the acquired raw messages, using a unified clock source to time-align the messages from multiple sources, with a time synchronization error of no more than 100 milliseconds, and marking the acquisition source, protocol type, and device identifier in the message metadata.

[0015] Furthermore, message preprocessing in S2 includes: segmenting the message text using a power industry-specific word segmenter based on regular expressions, with a segmentation granularity of mixed characters and words; removing redundant information from the message using a rule-based and statistical denoising algorithm, including communication protocol headers, duplicate timestamp fields, null value fields, and transmission noise; and normalizing the analog numerical values ​​using the Min-Max normalization method, mapping the values ​​to... For each interval, an entity alignment method based on the power equipment itself is used to match and align the equipment name, location information, and status description in the message with the standard power equipment knowledge base.

[0016] Furthermore, the improved BERT model in S3 also includes: adding a domain adaptation layer after the output layer of the BERT model. This domain adaptation layer is a fully connected neural network layer used to map the general semantic space to the power domain semantic space; employing a dynamic masking strategy, during the pre-training stage, the masking probability is dynamically adjusted based on the word frequency and importance of power domain terms, with the masking probability of key terms being no less than twice that of ordinary words. The power domain knowledge vector... It is obtained through embedding learning based on the knowledge graph of power equipment ontology.

[0017] Furthermore, fault diagnosis in S4 also includes: constructing a conditional probability table between fault types and semantic features. This conditional probability table is obtained through statistical learning from a historical fault case database. A multi-evidence fusion mechanism is employed; when multiple related messages exist, the semantic vectors of each message are used as independent evidence for joint inference. The joint fault probability calculation formula is as follows:

[0018] In the formula, for The semantic vectors corresponding to each independent piece of evidence are used to set a fault probability threshold. ,when Output the fault diagnosis conclusion in real time; otherwise, mark it as pending manual review.

[0019] Furthermore, the operation and maintenance assistance in S5 also includes: automatically searching the historical fault knowledge base according to the fault type, pushing similar historical fault cases and handling solutions, assessing the cascading impact of faults on various subsystems of the power supply system based on the fault impact range analysis model; generating a fault handling priority ranking, and pushing emergency faults to the mobile terminals of operation and maintenance personnel first.

[0020] Furthermore, the improved BERT model is characterized by the following training loss function:

[0021] In the formula, Train the loss function for the model. The total number of training samples, For the first The true label of each sample For the first The predicted probabilities of each sample are used for model training, employing the AdamW optimizer with an initial learning rate of [value missing]. The batch size is 32, the number of training rounds is no less than 10, and an early stopping strategy is adopted to prevent overfitting.

[0022] Furthermore, the PSCADA system operation messages include IEC 60870-5-104 protocol messages, IEC 61850MMS messages, Modbus TCP messages, and DL / T 645 protocol messages; the communication protocol parser supports multi-protocol adaptive identification and parsing, with a protocol identification accuracy of no less than 99%.

[0023] A PSCADA message analysis and fault diagnosis system based on the Bert large language model, applied to the aforementioned PSCADA message analysis and fault diagnosis method based on the Bert large language model, includes: a message data acquisition module, used to acquire running messages in real time through the PSCADA system API interface, communication protocol parser and log acquisition module; The message preprocessing module is used to perform word segmentation, noise reduction, normalization, and entity alignment on the original message. The Bert semantic understanding module is used to extract message semantic vectors using an improved Bert model; the fault diagnosis module is used for Bayesian inference and fault diagnosis based on a probabilistic graphical model. The results output and operation and maintenance assistance module is used to generate structured fault reports and provide operation and maintenance assistance decision support.

[0024] Furthermore, the system also includes: The fault knowledge base module is used to store and manage historical fault cases, fault handling solutions, and domain knowledge ontology; The model management module is used for version management, online updates, and performance monitoring of the improved BERT model. The visualization module is used to display fault diagnosis results, system operating status, and trend analysis in the form of charts.

[0025] Compared with the prior art, the present invention achieves significant beneficial effects through the above technical solution: (1) High accuracy in semantic understanding By leveraging the deep semantic encoding capabilities of the improved BERT model and combining it with a knowledge vector fusion mechanism in the power sector, this invention can accurately understand the deep semantics of messages, maintaining high accuracy even when faced with non-standardized messages with diverse expressions. Experimental results show that the semantic extraction accuracy reaches over 95%, an improvement of more than 20 percentage points compared to existing regular expression-based methods.

[0026] (2) High accuracy in fault diagnosis By utilizing the Bayesian inference mechanism of a probabilistic graphical model, this invention can effectively integrate multi-source evidence and address uncertainties in fault diagnosis. Testing shows that the fault diagnosis accuracy exceeds 94%, with a false negative rate of no more than 3% and a false positive rate of no more than 2%, representing reductions of 12 percentage points and 8 percentage points respectively compared to existing rule-matching methods.

[0027] (3) Operation and maintenance efficiency has been significantly improved This invention automates the entire process from message acquisition to fault report generation, eliminating the need for manual review of each message and analysis of fault causes. Fault report generation takes no more than 5 minutes, improving efficiency by more than 6 times compared to existing manual analysis methods and significantly shortening fault response time.

[0028] (4) Convenient knowledge reuse This invention achieves structured storage and automatic accumulation of fault knowledge by constructing a knowledge graph and fault knowledge base in the power field. Newly confirmed fault cases are automatically added to the database, eliminating the need for verbal instruction of expert experience, significantly improving knowledge reuse rate, and shortening the onboarding cycle for new maintenance personnel by more than 70%.

[0029] (5) The system has strong scalability When adding new fault types, only a small number of labeled samples are needed to incrementally fine-tune the model, with an adaptation time of no more than 3 days, which is more than 80% more efficient than the existing rule writing method. At the same time, the network structure of the probabilistic graphical model supports dynamic expansion, and new fault types can be seamlessly integrated into the existing inference framework.

[0030] (6) Good multi-protocol compatibility The message data acquisition module of this invention supports adaptive recognition and parsing of multiple mainstream power communication protocols such as IEC 60870-5-104, IEC 61850 MMS, Modbus TCP, and DL / T 645. The protocol recognition accuracy reaches over 99%, and it has good versatility and compatibility. Attached Figure Description

[0031] Figure 1 This is a flowchart of a PSCADA message analysis and fault diagnosis method based on the Bert large language model according to the present invention; Figure 2This is a schematic diagram of the architecture of the message data acquisition module; Figure 3 A detailed schematic diagram of the message preprocessing flow; Figure 4 This is a schematic diagram of the network structure of the improved Bert model; Figure 5 This is a schematic diagram of the Bayesian network structure of a probabilistic graphical model. Figure 6 This is the overall architecture diagram of the fault diagnosis system. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Reference Figures 1-6 As shown, a PSCADA message analysis and fault diagnosis method based on the BERT large language model includes the following steps: S1. Message Data Acquisition: Through the PSCADA system API interface, communication protocol parser and log acquisition module, the PSCADA system operation messages are collected in real time according to the preset acquisition period. The operation messages include switch messages, analog messages, event sequence record messages and alarm messages. The acquisition period is no more than 30 seconds. S2. Message Preprocessing: The collected raw messages are processed sequentially by regular expression segmentation, noise reduction algorithm to remove redundant information, numerical normalization and entity alignment to convert unstructured message text into a standardized token sequence. The preprocessing accuracy is no less than 98%. S3. BERT Semantic Understanding: The preprocessed standardized token sequence is input into the improved BERT model. This improved BERT model introduces a domain adaptation layer and a dynamic masking strategy, integrates power domain knowledge vectors for deep semantic encoding, and outputs a message semantic vector. The semantic vector extraction formula is as follows:

[0034] In the formula, The output semantic vector, Input a sequence of tokens for the segmented message. For Bert encoder functions, For knowledge vectors in the power sector, To adapt the weight matrix to the domain, This represents matrix multiplication. S4. Fault Diagnosis: The semantic vector is used as evidence input into the probabilistic graphical model. Bayesian inference is performed by combining the prior probability and conditional probability of the fault to calculate the posterior probability of each fault type. The fault type corresponding to the highest posterior probability is taken as the diagnosis result. The fault probability calculation formula is:

[0035] In the formula, In the evidence Fault types under conditions The posterior probability, The evidence set constituted by the semantic vectors. Fault type The prior probability, Fault type Provide evidence The conditional probability (likelihood). This represents the total number of fault types. For the first Types of faults; S5. Result Output and Operation and Maintenance Assistance: Generate a structured fault report based on the fault diagnosis results. The fault report includes the fault type, fault cause, scope of impact, handling suggestions and historical similar fault cases, and feeds back the diagnosis results and handling solutions to the operation and maintenance personnel's terminal.

[0036] Specifically, this invention establishes a complete technical solution from PSCADA system operation message collection, message standardization preprocessing, deep semantic extraction of messages, to intelligent fault diagnosis, fault report generation, and operation and maintenance assistance. It relies on multiple interfaces to collect various operation messages such as switch quantities, analog quantities, event sequence records, and alarms in real time. Through word segmentation, noise reduction, numerical normalization, and entity alignment processing, the unstructured raw messages are transformed into standardized sequences. On this basis, an improved BERT model with added domain adaptation layer and dynamic masking strategy is adopted. Combined with power domain knowledge vectors, deep semantic encoding is completed and semantic vectors are output. With the help of probabilistic graphical models and Bayesian inference, fault type determination is completed. Finally, a structured fault report containing fault type, fault cause, impact range, handling suggestions, and historical similar fault cases is output and pushed to the operation and maintenance terminal. This solution changes the traditional PSCADA message analysis model, which mainly relies on regular expression matching and keyword retrieval. It effectively enhances the understanding of deep semantics in messages, maintaining good recognition even with non-standardized messages of diverse expressions. Simultaneously, leveraging the uncertainty reasoning capabilities of probabilistic graphical models, it can effectively handle complex operating conditions such as compound faults and latent faults, effectively reducing missed and misdiagnosed faults and improving overall fault diagnosis accuracy. The entire process is automated, eliminating the need for maintenance personnel to manually review messages and analyze faults against rule manuals, significantly reducing fault analysis time and effectively improving the efficiency of power system operation and maintenance. This invention combines a power equipment ontology knowledge graph and a fault knowledge base to achieve structured management and automatic accumulation of diagnostic knowledge. It overcomes the previous limitations of over-reliance on expert experience and fragmented storage that hindered knowledge transfer and reuse, making the reuse of domain knowledge more convenient. In terms of system expansion, facing power equipment updates and the continuous evolution of fault modes, this invention eliminates the need for manual batch writing of diagnostic rules. New fault types can be added simply by incrementally adapting the model, demonstrating excellent scalability. Meanwhile, the message acquisition process can support the parsing of various mainstream power communication protocols such as IEC60870-5-104, IEC61850MMS, ModbusTCP, and DL / T645. It has outstanding protocol recognition capabilities and can be adapted to various power operation and maintenance scenarios such as urban rail power supply systems, substations, and distribution networks. It is flexible in deployment and highly versatile.

[0037] Furthermore, the message data acquisition in S1 also includes: synchronizing the timestamps and marking the source of the acquired raw messages, using a unified clock source to time-align the messages from multiple sources, with a time synchronization error of no more than 100 milliseconds, and marking the acquisition source, protocol type, and device identifier in the message metadata.

[0038] Specifically, during the collection of PSCADA system operation messages, this invention performs timestamp synchronization and source marking processing on the acquired raw messages. It utilizes a unified clock source to achieve time alignment of multi-source messages, strictly controlling time synchronization errors. This effectively solves the problem of inconsistent timing in messages collected from multiple channels and protocols, ensuring the authenticity and accuracy of the time sequence of all messages. It avoids errors in determining the correlation between different messages due to time discrepancies, mitigating the adverse effects of timing issues on subsequent message semantic analysis and fault probability reasoning from the data source. Simultaneously, this invention fully annotates the collection source, protocol type, and device identifier in the message metadata, clearly distinguishing the basic information of each message. This facilitates the classification and rapid tracing of massive amounts of messages according to collection channel, communication protocol, and associated device. It also allows subsequent processing stages to accurately match the corresponding power equipment and communication links, ensuring that message analysis and fault diagnosis accurately target the intended object. Furthermore, it helps maintenance personnel quickly locate the corresponding collection node and field equipment during fault handling.

[0039] Furthermore, message preprocessing in S2 includes: segmenting the message text using a power industry-specific word segmenter based on regular expressions, with a segmentation granularity of mixed characters and words; removing redundant information from the message using a rule-based and statistical denoising algorithm, including communication protocol headers, duplicate timestamp fields, null value fields, and transmission noise; and normalizing the analog numerical values ​​using the Min-Max normalization method, mapping the values ​​to... For each interval, an entity alignment method based on the power equipment itself is used to match and align the equipment name, location information, and status description in the message with the standard power equipment knowledge base.

[0040] Specifically, this invention performs multi-dimensional standardization processing on unstructured raw PSCADA messages. It employs a power industry-specific regular expression-based word segmenter adapted to industry scenarios, and completes message segmentation at a mixed word-to-speech granularity. This accurately identifies various power industry terms, avoiding semantic breaks caused by the inability of general word segmentation methods to adapt to domain-specific vocabulary. Simultaneously, through a noise reduction method combining rules and statistics, redundant content such as communication protocol headers, duplicate timestamps, null fields, and transmission noise in the messages is cleaned up, minimizing the interference of invalid information on subsequent analysis processes. Min-Max algorithm is applied to analog numerical values ​​within the messages. The normalization method maps data uniformly to the [0,1] interval, eliminating the problem of inconsistent dimensions between different types of numerical values ​​and making the data format more suitable for model operation requirements. It also relies on the power equipment itself to complete entity alignment, matching the equipment name, location information, and status description in the message with the standard power equipment knowledge base, effectively solving the semantic ambiguity problem caused by multiple names for power equipment. After the entire preprocessing process, the messy original message is transformed into standardized content that is orderly and can fully guarantee the extraction effect of the deep semantics of the message by the subsequent improved BERT model, thereby improving the accuracy of semantic recognition from the data level.

[0041] Furthermore, the improved BERT model in S3 also includes: adding a domain adaptation layer after the output layer of the BERT model. This domain adaptation layer is a fully connected neural network layer used to map the general semantic space to the power domain semantic space; employing a dynamic masking strategy, during the pre-training stage, the masking probability is dynamically adjusted based on the word frequency and importance of power domain terms, with the masking probability of key terms being no less than twice that of ordinary words. The power domain knowledge vector... It is obtained through embedding learning based on the knowledge graph of power equipment ontology.

[0042] Specifically, this invention makes targeted optimizations to the BERT model used for PSCADA message semantic encoding. A domain adaptation layer in the form of a fully connected neural network is added after the model output layer, which can accurately map the general semantic space to the power domain semantic space, effectively improving the problem of insufficient sensitivity of general pre-trained models to power-related content. Simultaneously, a dynamic masking strategy is adopted, setting mask probabilities based on the frequency and importance of power-related terms during pre-training, allowing the model to focus on learning key terms such as equipment names and fault types, strengthening its ability to capture core message information. Power domain knowledge vectors are obtained through embedding learning based on the power equipment ontology knowledge graph, and power industry ontology knowledge is integrated into the model system, supplementing the lack of power domain professional knowledge in general models. The synergistic effect of multiple optimization methods solves the problem that the general BERT model cannot fully model the semantic specificity of power domain terms, significantly enhancing the model's ability to parse the deep semantics of PSCADA messages, so that the final output semantic vector can completely and accurately carry effective information such as equipment status and alarm prompts within the message.

[0043] Furthermore, fault diagnosis in S4 also includes: constructing a conditional probability table between fault types and semantic features. This conditional probability table is obtained through statistical learning from a historical fault case database. A multi-evidence fusion mechanism is employed; when multiple related messages exist, the semantic vectors of each message are used as independent evidence for joint inference. The joint fault probability calculation formula is as follows:

[0044] In the formula, for The semantic vectors corresponding to each independent piece of evidence are used to set a fault probability threshold. ,when Output the fault diagnosis conclusion in real time; otherwise, mark it as pending manual review.

[0045] Specifically, this invention combines statistical learning from a historical fault case database to construct a conditional probability table between fault types and semantic features. This ensures that the probability basis for fault reasoning comes from real cases accumulated during power operation and maintenance, guaranteeing that the probability judgment rules align with actual on-site operating conditions. In the common scenario of multiple related messages appearing simultaneously, this invention employs a multi-evidence fusion mechanism, using the semantic vectors corresponding to each message as independent evidence for joint reasoning. This allows for comprehensive analysis of information conveyed by multiple messages, overcoming the limitations of single-evidence analysis and better adapting to complex diagnostic scenarios involving compound faults and concurrent alarm information in power systems. Simultaneously, a fault probability threshold is set as the judgment criterion. When the calculated posterior probability of the fault meets the threshold requirement, the fault diagnosis conclusion is directly output. Cases that do not meet the threshold are marked as requiring manual review. This hierarchical judgment mode fully leverages the efficiency and speed of automated reasoning while retaining a manual review step for uncertain diagnostic results. This effectively compensates for the shortcomings of automated algorithms in complex and uncertain scenarios, reducing the occurrence of missed and misdiagnosed faults and making the fault diagnosis process more rigorous and standardized.

[0046] Furthermore, the operation and maintenance assistance in S5 also includes: automatically searching the historical fault knowledge base according to the fault type, pushing similar historical fault cases and handling solutions, assessing the cascading impact of faults on various subsystems of the power supply system based on the fault impact range analysis model; generating a fault handling priority ranking, and pushing emergency faults to the mobile terminals of operation and maintenance personnel first.

[0047] Specifically, this invention can automatically retrieve content from the historical fault knowledge base based on the identified fault type, and push matching similar historical fault cases and corresponding handling solutions to maintenance personnel. By leveraging long-term accumulated maintenance knowledge, it provides intuitive reference for on-site fault handling, changing the previous situation where fault handling relied excessively on the personal experience of senior maintenance personnel. This allows the accumulated fault handling knowledge to be efficiently transferred and reused. At the same time, relying on the fault impact range analysis model, it accurately judges the cascading impact of the fault on various subsystems of the power supply system, helping maintenance personnel to fully grasp the fault's scope and potential derivative risks, facilitating early planning and risk prevention and control. On this basis, it also prioritizes handling based on the actual fault situation, prioritizing the push of emergency fault information to the maintenance personnel's mobile terminals, ensuring that high-risk faults are received and responded to in a timely manner. The entire set of maintenance assistance functions connects fault diagnosis and on-site handling, further simplifying the maintenance workflow, shortening the overall fault response time, and comprehensively improving the intelligence level and actual handling efficiency of power system fault maintenance.

[0048] Furthermore, the improved BERT model is characterized by the following training loss function:

[0049] In the formula, Train the loss function for the model. The total number of training samples, For the first The true label of each sample For the first The predicted probabilities of each sample are used for model training, employing the AdamW optimizer with an initial learning rate of [value missing]. The batch size is 32, the number of training rounds is no less than 10, and an early stopping strategy is adopted to prevent overfitting.

[0050] Specifically, this invention employs a corresponding loss function to train the improved BERT model. This function accurately calculates the deviation between the model's predicted probability and the true label of the sample, serving as a crucial basis for iterative optimization of model parameters. It guides the model to continuously learn various power semantic features contained in PSCADA messages, ensuring that the model's predictions increasingly align with the true labels. During training, the AdamW optimizer is used, along with a predetermined initial learning rate, batch size, and number of training rounds. This reasonable parameter combination allows for a smoother model parameter update process, improving overall training efficiency while ensuring the model's learning quality of professional semantic features in the power field. The early stopping strategy implemented during training effectively terminates ineffective training processes, mitigating overfitting and preventing the model from performing well only in training samples but failing to adapt to actual field operation messages. This allows the improved BERT model after training to possess superior generalization ability, stably and accurately extracting semantic vectors from various PSCADA messages. By strengthening the foundation of semantic parsing during model training, this further ensures the stable operation of the entire message analysis and fault diagnosis solution in real-world power operation and maintenance scenarios, maintaining reliable identification and diagnostic accuracy.

[0051] Furthermore, the PSCADA system operation messages include IEC 60870-5-104 protocol messages, IEC 61850MMS messages, Modbus TCP messages, and DL / T 645 protocol messages; the communication protocol parser supports multi-protocol adaptive identification and parsing, with a protocol identification accuracy of no less than 99%.

[0052] Specifically, this invention is compatible with PSCADA messages corresponding to various mainstream power communication protocols such as IEC 60870-5-104, IEC 61850 MMS, Modbus TCP, and DL / T 645. Relying on a communication protocol parser with adaptive recognition capabilities, it can automatically complete the identification and parsing of messages of different protocols without the need for separate configuration or secondary development for a single protocol, effectively broadening the scope of application of the solution. The high protocol recognition accuracy ensures that various messages are completely and accurately parsed, avoiding problems such as missing message content and parsing errors caused by protocol recognition errors. It ensures the effectiveness of subsequent word segmentation, normalization, semantic extraction, and fault reasoning from the source. This also allows the entire message analysis and fault diagnosis solution to be flexibly deployed in various power scenarios such as urban rail power supply systems, substations, and distribution networks, reducing the difficulty of system adaptation under different equipment and scenarios.

[0053] A PSCADA message analysis and fault diagnosis system based on the Bert large language model, applied to the aforementioned PSCADA message analysis and fault diagnosis method based on the Bert large language model, includes: a message data acquisition module, used to acquire running messages in real time through the PSCADA system API interface, communication protocol parser and log acquisition module; The message preprocessing module is used to perform word segmentation, noise reduction, normalization, and entity alignment on the original message. The Bert semantic understanding module is used to extract message semantic vectors using an improved Bert model; the fault diagnosis module is used for Bayesian inference and fault diagnosis based on a probabilistic graphical model. The results output and operation and maintenance assistance module is used to generate structured fault reports and provide operation and maintenance assistance decision support.

[0054] Specifically, this invention constructs a deployable hardware and software system through a modular architecture, engineering the entire PSCADA message analysis and fault diagnosis method. Each functional module has a clear division of labor and is sequentially connected, enabling the orderly completion of the entire process from message acquisition to final output of maintenance results. The message data acquisition module can stably acquire various operational messages generated on-site, providing reliable raw data for subsequent analysis. The message preprocessing module standardizes the raw messages, removing invalid information and unifying the data format to ensure the overall quality of the input data. The BERT semantic understanding module, based on an improved model, performs deep semantic extraction of the messages, outputting semantic vectors that represent the message information. The fault diagnosis module is based on a probabilistic graphical model. By employing Bayesian inference, the system automates fault type determination. The output and maintenance support modules generate corresponding reports and provide decision support based on the diagnostic conclusions. The entire system relies on the collaborative operation of each module to fully implement intelligent analysis and diagnostic logic, completely changing the previous work mode of manually checking messages one by one and manually judging faults. It achieves full automation of power message analysis and fault diagnosis. The clear module division also facilitates debugging, inspection, and functional maintenance of the system during actual deployment. It can stably connect to on-site PSCADA equipment and is compatible with various power operation and maintenance sites such as urban rail power supply systems and substations, allowing intelligent fault diagnosis solutions to be truly implemented and continuously improving the intelligence level and work efficiency of power operation and maintenance.

[0055] Furthermore, the system also includes: The fault knowledge base module is used to store and manage historical fault cases, fault handling solutions, and domain knowledge ontology; The model management module is used for version management, online updates, and performance monitoring of the improved BERT model. The visualization module is used to display fault diagnosis results, system operating status, and trend analysis in the form of charts.

[0056] Specifically, this invention adds supporting functional modules to the original system architecture. The fault knowledge base module can centrally store and manage historical fault cases, fault handling solutions, and domain knowledge ontology, realizing the unified collection and continuous accumulation of power operation and maintenance related knowledge. This effectively solves the problems of scattered storage of fault knowledge and over-reliance on personal experience in the traditional model, allowing various handling experiences and professional knowledge to be inherited and repeatedly used for a long time, continuously providing reference support for fault diagnosis and on-site operations. The model management module can complete the version management, online updates, and operational performance monitoring of the improved BERT model. It can iteratively optimize the model based on newly added fault data and equipment operation status on-site, timely control the model's operating status, and ensure... The fault model maintains stable semantic extraction capabilities throughout long-term use, allowing the entire system to better adapt to the continuous changes in power equipment and fault modes. The visualization module presents fault diagnosis results, system operating status, and operating trends in chart form, intuitively transforming complex message data and diagnostic information. This facilitates maintenance personnel in quickly grasping the overall operating status, fault distribution, and situational changes, assisting in global control and comprehensive decision-making. These auxiliary modules work in conjunction with the original functional modules of the system to further improve the functional system of the entire fault diagnosis system, enhance the convenience of system operation and maintenance, knowledge management capabilities, and human-computer interaction experience, and better support the routine intelligent operation and maintenance work of the power system.

[0057] The following are specific embodiments of the present invention: Example 1: Fault Diagnosis of Low SF6 Gas Pressure in Switchgear of Urban Rail Transit Main Substation 1. Message Data Acquisition In a main substation of a certain urban rail transit system, the PSCADA system collects real-time operating status messages of the switchgear using the IEC 61850 MMS protocol. When the SF6 gas pressure sensor in the switchgear detects that the gas pressure is lower than a set threshold, the system generates an alarm message. The acquisition module collects messages in real time at a 10-second interval. An example of the original messages collected is shown below: [2024-01-15 09:23:45.237] MMS_REPORT | Device: CB_110kV_01 | Signal: SF6_Pressure_Low_Alarm | Value: 0.32 MPa | Threshold: 0.35 MPa | Status: ALARM | Location: MainSub_110kV_RoomA 2. Message Preprocessing The preprocessing module processes the above raw messages: (a) Regular expression tokenization: The message is decomposed into a token sequence to identify key entities: device name "CB_110kV_01" (110kV circuit breaker No. 1), signal type "SF6_Pressure_Low_Alarm" (SF6 low air pressure alarm), value "0.32 MPa", threshold "0.35 MPa", status "ALARM", and location "MainSub_110kV_RoomA" (main transformer 110kV room A). (b) Noise reduction: Remove duplicate fields of the "MMS_REPORT" header and timestamp from the communication protocol, and retain valid information content; (c) Numerical normalization: The SF6 pressure values ​​were normalized. The minimum value in the historical data was 0.20 MPa, and the maximum value was 0.50 MPa. The normalized values ​​are as follows:

[0058] (d) Entity alignment: Align the device name "CB_110kV_01" to "110kV main transformer incoming line circuit breaker #1" in the standard device body, and align the location information to "Main substation 110kV high voltage room A zone".

[0059] The preprocessed normalized token sequence is as follows: [CLS] 110kV Main Transformer Incoming Line Circuit Breaker #1 SF6 Low Gas Pressure Alarm 0.40 [SEP] 3. BERT semantic understanding The preprocessed token sequence is input into the improved BERT model. The model parameters are set as follows: the base BERT model uses bert-base-chinese (12 Transformer layers, hidden layer dimension...). (Attention head count 12); the domain adaptation layer is a fully connected layer with an input dimension of 768 and an output dimension of 768; the knowledge vector dimension for the power domain is... In the dynamic masking strategy, the masking probability for key terms is 0.30, and the masking probability for ordinary words is 0.15.

[0060] The model outputs semantic vectors. Domain knowledge vector Obtained through TransE embedding of the power equipment knowledge graph, the embedding vector corresponding to the entity "low SF6 gas pressure" can effectively represent "low SF6 gas pressure". May lead to The semantic association of "the insulation performance of the circuit breaker has deteriorated".

[0061] The semantic vector calculation process is as follows:

[0062] in For general semantic representation, The two are domain knowledge enhancement terms, and their addition yields a semantic vector that integrates domain knowledge.

[0063] 4. Fault Diagnosis Using semantic vectors as evidence The input is used for inference via a Bayesian network. In this embodiment, the fault knowledge base includes... Common switchgear fault types. The relevant prior probabilities and conditional probabilities are set as follows: • (Prior probability of SF6 low pressure fault, based on historical statistics) • (Prior probability of normal operation) • (Prior probability of sensor failure) • (Conditional probability of observing this evidence under SF6 low pressure fault) • (Probability of false alarms under normal operation) • (The probability of a similar alarm being generated due to sensor failure) Calculate the posterior probability using Bayes' theorem:

[0064] The calculation results show that the posterior probability of a low SF6 pressure fault is 87.6%, which is greater than the threshold. The system directly outputs the diagnostic conclusion: the fault type is "low SF6 gas pressure fault in switchgear", with a confidence level of 87.6%.

[0065] 5. Results Output and Operation and Maintenance Assistance The system generates structured fault reports: • Fault type: Low SF6 gas pressure fault in switchgear (confidence level: 87.6%) • Cause of the fault: The SF6 gas pressure dropped to 0.32 MPa, below the alarm threshold of 0.35 MPa, which may lead to a decrease in the insulation performance of the circuit breaker. • Affected area: The affected equipment is the 110kV main transformer incoming line circuit breaker #1, located in Zone A of the 110kV high-voltage room of the main substation. This may cause the circuit breaker to malfunction in normal opening and closing operations. • Recommendations: (1) Immediately arrange for maintenance personnel to inspect the SF6 gas pipeline for leaks; (2) Use an SF6 gas leak detector to check the circuit breaker body, valves, and pipeline connections for leaks; (3) After confirming the leak, seal it or replace the seals; (4) Refill the gas to the rated pressure of 0.45 MPa and then retest to confirm; (5) After the gas replenishment is completed, conduct a mechanical characteristic test on the circuit breaker to verify its performance. • Historical Cases: Three similar historical fault cases were found. The most recent case was an SF6 leakage fault in the main transformer outgoing circuit breaker, which was handled in November 2023. The result was that the circuit was restored to normal after the sealing ring was replaced. In this embodiment, the time from message acquisition to report generation is 3 minutes, and the fault diagnosis accuracy rate is 94.5%.

[0066] Example 2: Fault Diagnosis of Abnormal Voltage Fluctuations in Urban Rail Contact Network 1. Message Data Acquisition In a certain urban rail transit traction power supply system, the PSCADA system collects overhead contact line voltage data via the IEC 60870-5-104 protocol. When abnormal fluctuations occur in the overhead contact line voltage, the system generates multiple sets of related messages. Typical collected messages are as follows: Message A: [2024-02-20 14:15:32.512] IEC104_M_ME_TF_1 | IOA: 20101 | Value: 1850 V | QDS: GOOD | Timestamp: 14:15:32.500 Message B: [2024-02-20 14:15:35.108] IEC104_M_SP_NA_1 | IOA: 30015 | Value: ON | Description: Overvoltage protection action of traction converter | Timestamp: 14:15:35.000 Message C: [2024-02-20 14:15:38.245] IEC104_M_ME_TF_1 | IOA: 20102 | Value: 2200 V | QDS: QUESTIONABLE | Timestamp: 14:15:38.200 2. Message Preprocessing The three messages are preprocessed and entity aligned respectively: • Message A: Contact line voltage 1850V (lower than the lower limit of the normal range of 1500V-2000V), data quality is good. • Message B: Traction converter overvoltage protection activated, switch status is "closed". • Message C: Overhead contact line voltage 2200V (above the upper limit of the normal range), data quality questionable. Normalization: The rated voltage of the overhead contact line is DC 1500V, with an allowable fluctuation range of 1000V-2000V. After normalization:

[0067] 3. BERT semantic understanding The three messages are input into the improved BERT model respectively, resulting in three semantic vectors. , , All three semantic vectors indicate the semantic topic of "abnormal contact network voltage", but with different emphases: message A emphasizes low voltage, message B emphasizes protection action events, and message C emphasizes high voltage and questionable data quality.

[0068] 4. Fault Diagnosis (Multi-Evidence Fusion) This embodiment involves multiple related messages and employs a multi-evidence joint reasoning mechanism. It assumes three fault types: abnormal fluctuations in overhead contact line voltage (…). ), traction converter fault ( Voltage measurement sensor failure () ).

[0069] The conditional probabilities of each piece of evidence are shown in the table below:

[0070] Prior probability: , , .

[0071] Calculated using a multi-evidence fusion formula: for (Abnormal fluctuations in overhead contact line voltage):

[0072] for (Traction converter failure):

[0073] for (Voltage measurement sensor malfunction):

[0074] After normalization, the posterior probability is obtained:

[0075] The diagnosis was "abnormal fluctuation fault in overhead contact line voltage," with a confidence level of 82.3%, falling within the range where manual review is recommended. The system outputs diagnostic conclusions and suggests manual review.

[0076] 5. Results Output and Operation and Maintenance Assistance The system generates a structured fault report, which includes: fault type: abnormal fluctuation of contact network voltage (confidence level 82.3%, review recommended); fault cause: contact network voltage jumps from 1850V to 2200V in a short period of time, accompanied by overvoltage protection activation of the traction converter, possibly indicating abnormal regenerative braking energy feedback or sudden load changes in adjacent power supply zones; affected area: 3 stations and 2 trains in operation within the power supply zone; recommended actions include checking the operating status of the traction substation rectifier units, checking the status of adjacent zone switches, and checking the operation of the regenerative braking device. In this embodiment, the report generation time is 4 minutes, and the fault diagnosis accuracy rate is 93.8%.

[0077] Example 3: Troubleshooting Overheating of Urban Rail Traction Transformer Oil 1. Message Data Acquisition At a traction substation of a certain urban rail transit system, the PSCADA system collects traction transformer operating data via the Modbus TCP protocol. The collected alarm messages are as follows: [2024-03-10 11:42:18.905] MODBUS_TCP | SlaveID: 5 | Function: 04 | Register: 40125 | Value: 95 | Unit: °C | Alarm: Transformer_Temp_High | Threshold: 85°C | Device: TractionTransformer_02 2. Message Preprocessing After word segmentation and entity alignment, key information was identified: the device is traction transformer No. 2 (TractionTransformer_02), the measured value is transformer oil temperature of 95°C, the alarm threshold is 85°C, and it is currently in a high-temperature alarm state. Value normalization: the historical temperature range is 20°C-120°C, after normalization:

[0078] Preprocessed token sequence: [CLS] High oil temperature of 95 degrees Celsius for traction transformer No. 2 (alarm 0.75). [SEP] 3. BERT semantic understanding The token sequence is input into the improved BERT model to extract semantic vectors. The model correctly identifies the core semantic feature "excessive oil temperature" and combines it with the domain knowledge vector "high transformer oil temperature". May lead to The semantic association of "accelerated insulation aging" enhances the understanding of the severity of the fault.

[0079] 4. Fault Diagnosis Consider four possible fault types: transformer overload ( Cooling system malfunction ( Oil temperature sensor malfunction ( Transformer internal faults ( ).

[0080] Prior probabilities of each fault: , , , .

[0081] The conditional probability setting for the evidence that the oil temperature is 95°C and the alarm threshold is 85°C: - (The probability of oil temperature exceeding the threshold is higher under overload conditions) (Cooling system malfunction directly caused the oil temperature to rise) (Sensor malfunction may produce false alarms) (Internal malfunctions may be accompanied by temperature rise) Calculate the posterior probability:

[0082] because The posterior probability was 40.1%, which did not reach the direct diagnosis threshold. The system further searched for related messages from the same time period. A message from 10 minutes prior was found showing a transformer load rate of 115% (exceeding the rated load), and this message was used as additional evidence. Perform joint reasoning.

[0083] Updated joint reasoning:

[0084]

[0085] After normalization The diagnosis result was "transformer overload causing oil temperature rise", with a confidence level of 73.4%, and manual review was recommended.

[0086] 5. Results Output and Operation and Maintenance Assistance The system generates a fault report, which includes: the fault type is transformer overload with excessive oil temperature (confidence level 73.4%, review recommended); the fault cause is that the transformer load rate continues to exceed 115%, resulting in increased winding and core losses, and the oil temperature rises to 95°C, exceeding the alarm threshold of 85°C; the affected area is traction transformer No. 2 and its two power supply zones; the handling recommendations include: (1) immediately verify the actual load situation and coordinate the scheduling to reduce the load of the transformer; (2) check the operating status of the cooling fan and confirm that the cooling system is working normally; (3) continuously monitor the oil temperature change trend, and if the oil temperature continues to rise to 105°C, it is necessary to consider cutting off part of the load; (4) arrange for night maintenance to check the DC resistance and insulation status of the transformer winding. The report generation time in this embodiment is 2 minutes, and the fault diagnosis accuracy rate is 95.2%.

[0087] Model training implementation details The training process of the improved BERT model includes the following steps: (a) Pre-training phase: Domain-adaptive pre-training is conducted based on a corpus of texts in the power industry (including power equipment manuals, operation and maintenance procedures, fault case reports, etc., totaling approximately 5 million words). A dynamic masking strategy is adopted, with key term masking probabilities... Probability of ordinary word mask The pre-training tasks included masked language modeling (MLM) and next-sentence prediction (NSP), with a batch size of 64 and a learning rate of [missing information]. The pre-training rounds are 100.

[0088] (b) Fine-tuning phase: Downstream task fine-tuning is performed based on the labeled PSCADA message-fault type pairing dataset (containing 10,000 labeled samples covering 50 fault types). The training loss function uses binary cross-entropy loss:

[0089] In the formula, The total number of training samples, For the first The true labels of each sample (one-hot encoded) For the first The predicted probability of a sample after Softmax normalization.

[0090] Training hyperparameter settings: Optimizer is AdamW, initial learning rate... Weight decay coefficient Batch size Number of training rounds An early stopping strategy is adopted, and training stops if the validation set loss does not decrease for three consecutive rounds. The learning rate scheduling uses a Warmup-LinearDecay strategy, with a Warmup step size of 500 steps.

[0091] (c) Domain adaptation layer training: After the BERT model is fine-tuned, the domain adaptation weight matrix is ​​trained separately. Using entity vectors embedded in knowledge graphs as the target, the mean squared error loss is optimized. , making The BERT semantic representation should be as close as possible to the corresponding entity.

[0092] Performance evaluation The performance of the method of the present invention was evaluated on a test set containing 2000 test samples, and the results are as follows:

[0093] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0094] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A PSCADA message analysis and fault diagnosis method based on the BERT large language model, characterized in that, Includes the following steps: S1. Message Data Acquisition: Through the PSCADA system API interface, communication protocol parser and log acquisition module, the PSCADA system operation messages are collected in real time according to the preset acquisition period. The operation messages include switch messages, analog messages, event sequence record messages and alarm messages. The acquisition period is no more than 30 seconds. S2. Message Preprocessing: The collected raw messages are processed sequentially by regular expression segmentation, noise reduction algorithm to remove redundant information, numerical normalization and entity alignment to convert unstructured message text into a standardized token sequence. The preprocessing accuracy is no less than 98%. S3. BERT Semantic Understanding: The preprocessed standardized token sequence is input into the improved BERT model. This improved BERT model introduces a domain adaptation layer and a dynamic masking strategy, integrates power domain knowledge vectors for deep semantic encoding, and outputs a message semantic vector. The semantic vector extraction formula is as follows: In the formula, The output semantic vector, Input a sequence of tokens for the segmented message. For Bert encoder functions, For knowledge vectors in the power sector, To adapt the weight matrix to the domain, This represents matrix multiplication. S4. Fault diagnosis: The semantic vector is used as evidence input to the probabilistic graphical model. Bayesian inference is performed by combining the prior probability and conditional probability of the fault. The posterior probability of each fault type is calculated, and the fault type corresponding to the maximum posterior probability is taken as the diagnosis result. The formula for calculating the failure probability is: In the formula, In the evidence Fault types under conditions The posterior probability, The evidence set constituted by the semantic vectors. Fault type The prior probability, Fault type Provide evidence The conditional probability, This represents the total number of fault types. For the first Types of faults; S5. Result Output and Operation and Maintenance Assistance: Generate a structured fault report based on the fault diagnosis results. The fault report includes the fault type, fault cause, scope of impact, handling suggestions and historical similar fault cases, and feeds back the diagnosis results and handling solutions to the operation and maintenance personnel's terminal.

2. The PSCADA message analysis and fault diagnosis method based on the BERT large language model according to claim 1, characterized in that, The message data acquisition in S1 also includes: synchronizing the timestamps and marking the source of the acquired raw messages, using a unified clock source to align the time of multi-source messages, with a time synchronization error of no more than 100 milliseconds, and marking the acquisition source, protocol type and device identifier in the message metadata.

3. The PSCADA message analysis and fault diagnosis method based on the BERT large language model according to claim 1, characterized in that, The message preprocessing in S2 includes: segmenting the message text using a power industry-specific word segmenter based on regular expressions, with a segmentation granularity of mixed characters and words; removing redundant information from the message using a denoising algorithm combining rules and statistics, including communication protocol headers, duplicate timestamp fields, null value fields, and transmission noise; and normalizing the analog numerical values ​​using the Min-Max normalization method, mapping the values ​​to... For each interval, an entity alignment method based on the power equipment itself is used to match and align the equipment name, location information, and status description in the message with the standard power equipment knowledge base.

4. The PSCADA message analysis and fault diagnosis method based on the BERT large language model according to claim 1, characterized in that, The improved BERT model in S3 further includes: adding a domain adaptation layer after the output layer of the BERT model. This domain adaptation layer is a fully connected neural network layer used to map the general semantic space to the power domain semantic space. A dynamic masking strategy is employed, dynamically adjusting the masking probability based on the word frequency and importance of power domain terms during the pre-training stage. The masking probability of key terms is no less than twice that of ordinary words. The power domain knowledge vector... It is obtained through embedding learning based on the knowledge graph of power equipment ontology.

5. The PSCADA message analysis and fault diagnosis method based on the BERT large language model according to claim 1, characterized in that, Fault diagnosis in S4 also includes: constructing a conditional probability table between fault types and semantic features. This conditional probability table is obtained through statistical learning from a historical fault case database. A multi-evidence fusion mechanism is used; when multiple related messages exist, the semantic vectors of each message are treated as independent evidence for joint inference. The joint fault probability calculation formula is as follows: In the formula, for The semantic vectors corresponding to each independent piece of evidence are used to set a fault probability threshold. ,when Output the fault diagnosis conclusion in real time; otherwise, mark it as pending manual review.

6. The PSCADA message analysis and fault diagnosis method based on the BERT large language model according to claim 1, characterized in that, The operation and maintenance assistance in S5 also includes: automatically searching the historical fault knowledge base according to the fault type, pushing similar historical fault cases and handling solutions, assessing the cascading impact of faults on various subsystems of the power supply system based on the fault impact range analysis model, generating fault handling priority ranking, and pushing emergency faults to the mobile terminals of operation and maintenance personnel first.

7. The PSCADA message analysis and fault diagnosis method based on the BERT large language model according to any one of claims 1 or 4, characterized in that, The training loss function of the improved BERT model is: In the formula, Train the loss function for the model. The total number of training samples, For the first The true label of each sample For the first The predicted probabilities of each sample are used for model training, employing the AdamW optimizer with an initial learning rate of [value missing]. The batch size is 32, the number of training rounds is no less than 10, and an early stopping strategy is adopted to prevent overfitting.

8. The PSCADA message analysis and fault diagnosis method based on the BERT large language model according to claim 1, characterized in that, The PSCADA system operation messages include IEC 60870-5-104 protocol messages, IEC 61850 MMS messages, Modbus TCP messages, and DL / T 645 protocol messages; the communication protocol parser supports multi-protocol adaptive identification and parsing, with a protocol identification accuracy of no less than 99%.

9. A PSCADA message analysis and fault diagnosis system based on the BERT large language model, applied to the PSCADA message analysis and fault diagnosis method based on the BERT large language model as described in any one of claims 1-8, characterized in that, include: The message data acquisition module is used to collect runtime messages in real time through the PSCADA system API interface, communication protocol parser, and log acquisition module. The message preprocessing module is used to perform word segmentation, noise reduction, normalization, and entity alignment on the original message. The Bert semantic understanding module is used to extract message semantic vectors using an improved Bert model; the fault diagnosis module is used for Bayesian inference and fault diagnosis based on a probabilistic graphical model. The results output and operation and maintenance assistance module is used to generate structured fault reports and provide operation and maintenance assistance decision support.

10. The PSCADA message analysis and fault diagnosis system based on the BERT large language model according to claim 9, characterized in that, The system also includes: The fault knowledge base module is used to store and manage historical fault cases, fault handling solutions, and domain knowledge ontology; The model management module is used for version management, online updates, and performance monitoring of the improved BERT model. The visualization module is used to display fault diagnosis results, system operating status, and trend analysis in the form of charts.

Images