A method and system for generating substation inspection report

Abstract

The present invention relates to a method and system for generating a substation inspection report, and belongs to the technical field of intelligent substation inspection. The method comprises the following steps: based on a large model of a multimodal reasoning framework, the inspection data, equipment status data and operation and maintenance data of the substation are analyzed in real time by integrating image recognition technology and integrated time series prediction algorithm, the abnormal data obtained by the real-time analysis is deeply analyzed by the thinking chain prompt technology of the large model, and the in-depth analysis results are obtained by combining the historical operation and maintenance data, the knowledge base called by the large model and at least one of the predefined standards, and the substation inspection report is generated by integrating the real-time analysis results and the in-depth analysis results in combination with the inspection template. The large model of the multimodal reasoning framework based on this method can analyze and process the information provided by different data sources, realize the deep learning and understanding of massive data, and effectively improve the accuracy of the substation inspection report and the generation efficiency of the inspection report.

Description

Technical Field

[0001] The invention relates to a method and system for generating a substation inspection report, belonging to the technical field of substation intelligent inspection. Background Art

[0002] With the continuous development of the power system, substations play a vital role in the stable operation of the power grid. The traditional substation inspection method relies on manual inspection, which is inefficient and prone to omissions. Especially when dealing with complex equipment and data, manual operation has great limitations. At the same time, after manual inspection, the substation inspection report generated often has problems such as incomplete data analysis, long report generation time, and inaccurate content, which makes it difficult to provide timely and effective support for the maintenance and management of power equipment.

[0003] At present, the generation of substation inspection reports usually adopts simple rules and data collection methods. Although the inspection reports generated by some inspection systems have achieved preliminary automatic data processing, these systems still have great limitations in data analysis and report generation due to the wide variety of equipment, huge amount of data and uneven distribution. The existing automated inspection report generation methods are often limited to simple data integration and visual display, lack of in-depth intelligent analysis, and cannot achieve deep mining of inspection data.

[0004] Currently, some "knowledge base" technologies have been applied to the field of equipment maintenance and fault diagnosis. These inspection systems usually rely on the accumulation of historical data and simple rule matching, and they have the following shortcomings: First, the existing knowledge base often cannot be updated in real time, resulting in the inability to integrate new cases into the system in a timely manner, thereby reducing the accuracy of the analysis; second, these inspection systems have limited capabilities in processing complex relationships and multidimensional data, and it is difficult to provide comprehensive decision support; finally, the models of existing self-learning operation and maintenance knowledge bases usually lack self-optimization mechanisms, resulting in poor adaptability to different equipment or environments, and cannot effectively improve maintenance efficiency and accuracy.

[0005] At the same time, existing inspection systems generally have difficulty processing and analyzing massive inspection data in real time, especially when encountering sudden equipment failures or abnormal situations, and lack effective decision-making support. Traditional inspection report generation methods also fail to fully utilize modern artificial intelligence technology, resulting in insufficient accuracy and timeliness of inspection reports, which in turn affects the operating efficiency and safety of the power system. Summary of the invention

[0006] The object of the present invention is to provide a method and system for generating a substation inspection report, so as to solve the problem of low accuracy of the generated inspection report.

[0007] To achieve the above object, the solution of the present invention includes:

[0008] A method for generating a substation inspection report of the present invention comprises the following steps:

[0009] Based on the big model of the multimodal reasoning framework, the inspection data, equipment status data and operation and maintenance data of the substation are analyzed in real time through integrated image recognition technology and integrated time series prediction algorithm. The abnormal data obtained by the real-time analysis is deeply analyzed through the thinking chain prompt technology of the big model, and the historical operation and maintenance data, the knowledge base called by the big model and at least one of the predefined standards are combined to obtain the in-depth analysis results. The substation inspection report is generated by integrating the real-time analysis results and the in-depth analysis results in combination with the inspection template; the inspection data includes at least one of the images, videos and voiceprint information of the substation equipment.

[0010] Furthermore, during in-depth analysis, relevant contextual information used to provide a basis for the interpretation of abnormal data is obtained from the knowledge base through retrieval enhancement generation technology based on vector retrieval.

[0011] Furthermore, the knowledge base uses a self-learning operation and inspection knowledge base, which is constructed in the following way:

[0012] Utilize the natural language processing capabilities of the big model to analyze historical maintenance cases and equipment failure records, extract fault feature data, and use the trained machine learning model combined with operation and inspection data to classify and predict situations where there is existing equipment status data but the abnormal cause, diagnosis method, and repair method are unknown, forming a self-learning operation and inspection knowledge base;

[0013] The machine learning model takes fault feature data as input and the classification result of the fault feature data as output.

[0014] Furthermore, the self-learning operation and maintenance knowledge base is also updated based on the real-time feedback of the operation and maintenance personnel on the substation inspection reports.

[0015] Furthermore, the substation inspection report includes status assessment, abnormality analysis, risk prediction and maintenance optimization suggestions for substation equipment.

[0016] Furthermore, before real-time analysis, the patrol data, equipment status data, and operation and maintenance data are preliminarily processed through edge computing technology, and the results of the preliminary processing are used as the patrol data, equipment status data, and operation and maintenance data for the final real-time analysis.

[0017] Furthermore, the integrated image recognition technology includes convolutional neural network and YOLO algorithm; the integrated time series prediction algorithm includes long short-term memory network, recurrent neural network and Prophet algorithm.

[0018] Furthermore, the inspection template is obtained according to the inspection standard template and / or user requirements, and the inspection standard template selects an industry standard template.

[0019] Furthermore, the LangChain framework is used as the multimodal reasoning framework.

[0020] A substation inspection report generation system of the present invention comprises a processor, and the processor is used to execute a computer program to implement the steps of the substation inspection report generation method as described above.

[0021] Beneficial effects of the present invention:

[0022] The present invention is a pioneering invention, which provides a method for generating a substation inspection report. The method utilizes the multi-module fusion capability of a multimodal reasoning framework, and can flexibly integrate different types of data and large models. Specifically, based on the large model of the multimodal reasoning framework, the inspection data, equipment status data and operation and maintenance data of the substation are analyzed in real time by integrating image recognition technology and integrated time series prediction algorithm. The abnormal data obtained by real-time analysis is deeply analyzed by the thinking chain prompt technology of the large model, and the in-depth analysis results are obtained by combining the historical operation and maintenance data, the knowledge base called by the large model and at least one of the predefined standards. The substation inspection report is generated by integrating the real-time analysis results and the in-depth analysis results, and the inspection template is combined. The multimodal reasoning framework used in the method supports different types of data input (such as text, image, voiceprint data, etc.), and can realize information fusion and intelligent decision-making through a cross-modal reasoning process. The large model based on the multimodal reasoning framework can analyze and process information provided by different data sources, is suitable for multi-source data environments such as substations, and can realize deep learning and understanding of massive data, effectively improving the accuracy of substation inspection reports and the generation efficiency of inspection reports. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a flowchart of the substation inspection report generation system based on the multimodal reasoning framework;

[0024] Figure 2 This is the equipment architecture diagram of the substation inspection report generation system based on the multimodal reasoning framework;

[0025] Figure 3 This is the architecture diagram of the self-learning operation and inspection knowledge base;

[0026] Figure 4 This is a technical architecture diagram of a substation inspection report generation method based on a multimodal reasoning framework. DETAILED DESCRIPTION

[0027] In order to solve the problems in the background technology, the present invention provides a substation inspection report generation method. The large model based on the multimodal reasoning framework can analyze and process the information provided by different data sources, is suitable for multi-source data environments such as substations, and can realize deep learning and understanding of massive data, effectively improving the accuracy of substation inspection reports and the generation efficiency of inspection reports.

[0028] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0029] An embodiment of a substation inspection report generation system:

[0030] A substation inspection report generation system comprises a processor, and the processor is used for executing a computer program to realize the steps of a substation inspection report generation method.

[0031] Specifically, the substation inspection report generation system is also called the substation inspection report (referred to as report) generation system based on the multimodal reasoning framework. Figure 1 As shown, the system mainly includes a data acquisition module, an anomaly detection module, an intelligent analysis module and a report generation module.

[0032] Among them, the anomaly detection module, intelligent analysis module and report generation module serve as processors of the substation inspection report generation system.

[0033] As another implementation method, the data acquisition module, the anomaly detection module, the intelligent analysis module and the report generation module all serve as processors of the substation inspection report generation system.

[0034] Among them, the data acquisition module can obtain the multimodal data of the substation through direct acquisition or calling the substation intelligent patrol system interface and other methods. The multimodal data of the substation includes the substation patrol data, equipment status data, and operation and maintenance data.

[0035] The substation intelligent patrol system is deployed at the substation site to collect patrol data, equipment status data and operation and maintenance data of the substation in real time. The substation intelligent patrol system collects patrol data of substation equipment through voiceprint detection devices, cameras and robots. Among them, voiceprint detection devices, cameras and robots are used as collection devices to monitor the operating status of substation equipment.

[0036] The inspection data includes at least one of images, videos and voiceprint information of the substation equipment.

[0037] Specifically, the inspection data refers to equipment photos (images), videos, infrared images, voiceprint information, etc. of substation equipment.

[0038] Among them, equipment status data refers to the substation primary equipment status data, secondary equipment status data, substation auxiliary control data, or status data generated after analysis based on patrol data, etc., collected by sensors.

[0039] Among them, operation and maintenance data refers to maintenance records, fault types, operation logs and other data.

[0040] The data acquisition module can simultaneously acquire data from multiple data sources and ensure the continuity and real-time nature of data acquisition, thereby providing complete and accurate basic data for subsequent analysis and anomaly detection.

[0041] Specifically, the equipment architecture of the substation inspection report generation system based on the multimodal reasoning framework, such as Figure 2 As shown, it includes data collection equipment (voiceprint detection, camera, robot, etc.), regional patrol host, intelligent analysis host, edge node, switch and other equipment. The data collection equipment sends the collected data to the regional patrol host through the switch and edge node. The regional patrol host and the intelligent analysis host interact with each other. The two hosts are also used to send data to the abnormal detection module of the substation intelligent patrol system. The substation intelligent patrol system also uses edge computing technology to perform preliminary processing on the data it collects. The results of the preliminary processing can be used as the patrol data, equipment status data and operation and maintenance data for the final real-time analysis. It can also be transmitted to the regional patrol host and the intelligent analysis host through a high-speed network for further analysis. The results of the further analysis here are used as the patrol data, equipment status data and operation and maintenance data for the final real-time analysis.

[0042] Among them, the anomaly detection module is used to detect anomalies in the collected multimodal data of the substation. It is specifically based on the large model of the multimodal reasoning framework. It integrates image recognition technology and time series prediction algorithm to perform real-time analysis on the substation's patrol data, equipment status data and operation and maintenance data. Possible abnormal situations can be discovered in time through monitoring or prediction.

[0043] Integrated image recognition technology can use the YOLO algorithm that can simultaneously predict the categories and positions of multiple targets through a single convolutional network, or the convolutional neural network (R-CNN) that extracts candidate areas and uses a convolutional neural network algorithm for classification and regression in each area to achieve detection of multiple targets in the image, or single-shot multi-frame detection (SSD) that extracts candidate boxes of different sizes and classifiers through multi-scale convolutional layers to identify targets.

[0044] The integrated time series prediction algorithms include long short-term memory network (LSTM), recurrent neural network (GRU), Prophet algorithm, etc.

[0045] For example, the intelligent inspection system of a substation can detect problems such as meter anomalies, abnormal sounds, or defects in the appearance of equipment, thereby achieving accurate anomaly detection. For example, by predicting the oil temperature of the transformer, timely warnings can be given for early hidden dangers.

[0046] To further improve the accuracy of anomaly detection, the anomaly detection module also integrates a real-time feedback mechanism, which dynamically adjusts the anomaly data identified in the real-time analysis process based on historical data. That is, the anomaly detection module uses past data and patterns to continuously optimize the current anomaly assessment process. For example, if historical records show that a certain type of anomaly usually does not lead to serious consequences, the anomaly detection module will lower the priority of such anomalies. On the contrary, if historical data shows that a certain anomaly indicates a major failure or risk, the anomaly detection module will increase the level of the anomaly. The anomaly detection module also preprocesses the collected multimodal data, including data cleaning, noise reduction, and feature extraction, to ensure the consistency and accuracy of the data, providing a solid foundation for subsequent analysis, and also monitors the anomaly of multimodal data through data analysis or rule matching.

[0047] Among them, the intelligent analysis module is used to conduct in-depth analysis of the detected abnormal data, specifically through the Chain of Thought (CoT) prompt technology of the big model to conduct in-depth analysis of the abnormal data obtained by real-time analysis, gradually decompose complex problems, form a clear reasoning path, and combine historical operation and maintenance data, or the knowledge base called by the big model, or predefined standards, or historical operation and maintenance data and the knowledge base called by the big model, or historical operation and maintenance data and predefined standards, or the knowledge base called by the big model and predefined standards, or historical operation and maintenance data, the knowledge base called by the big model and predefined standards, to deeply analyze the causes of the abnormality and its possible impact. Among them, the predefined standards refer to the performance parameter thresholds, risk levels, etc. preset by humans.

[0048] The intelligent analysis module can also obtain relevant contextual information from the substation's operation and maintenance knowledge base (knowledge base) through the retrieval-augmented generation (RAG) technology based on vector retrieval, providing a basis for the interpretation of abnormal data.

[0049] The technical architecture of the substation inspection report generation method based on the multimodal reasoning framework is as follows Figure 4As shown in the figure, the substation patrol report generation system uses the multi-module fusion mechanism in the multimodal reasoning framework to perform intelligent analysis on multi-source data: First, for the detected abnormal data, the substation intelligent patrol system will generate preliminary equipment status data according to the established rules or machine learning models. For example, if the reading data of the equipment meter captured by the patrol camera exceeds the limit, the abnormal meter reading can be analyzed by the image recognition algorithm to generate the abnormal data of the equipment status. When the user asks a question or needs to generate a patrol report (intelligent patrol report or intelligent patrol report) intelligently, the intelligent analysis module will retrieve the relevant operation and inspection knowledge base through the RAG technology based on vector retrieval to obtain historical operation and maintenance experience data, and pass the detected abnormal data and tools to the large model LLM by combining the given prompt template and local function library (Function). After the large model gradually infers, it further analyzes the abnormal causes and possible impacts, and provides more detailed explanations and handling suggestions for abnormal situations.

[0050] The traditional expert database (expert knowledge base) can be used as an operation and maintenance knowledge base or knowledge base called by the large model. The construction of the traditional expert database usually follows the following steps: First, by classifying the substation signal types, the information features in the signal are extracted according to the preset classification rules. These information features are combined into multiple information feature groups according to certain event feature rules, and the correspondence between feature groups and events is established through the defined event feature rules. Then, using the preset event generation rules, according to the existing substation data, the event features are converted into specific events, and by verifying whether the various information points in the substation data information can successfully generate corresponding events, the validity of the rule configuration is further confirmed. Although this traditional expert knowledge base method can effectively help the system identify and process events, its limitations are also obvious.

[0051] First, the setting of rules often relies on the experience of experts, which makes it difficult to flexibly respond to new and uncertain situations; second, since the content of the knowledge base is statically updated, it has poor adaptability to changes in equipment status and environment, resulting in the system lacking sufficient intelligent response capabilities when facing new problems. Therefore, based on these problems, a method for constructing a self-learning operation and maintenance knowledge base is proposed, aiming to overcome the shortcomings of the traditional expert database through dynamic learning and automatic optimization, and to more flexibly and accurately respond to the ever-changing anomaly detection needs in the substation operation and maintenance process.

[0052] Among them, the self-learning operation and maintenance knowledge base is constructed and maintained through the following steps: Figure 3 As shown:

[0053] 1. Knowledge learning: Using the natural language processing capabilities of the big model, historical maintenance cases and equipment failure records are analyzed, and fault feature data such as environmental factors, equipment parameters, and failure modes are extracted from multi-source data (multimodal data) such as historical maintenance records, failure types, and operation logs. The mathematical expression for feature extraction is:

[0054] X={x 1 ,x 2 ,x 3 ,…,x i ,…,x n}

[0055] In the formula, X is the feature set, x i is the i-th fault feature, and n is the total number of fault features.

[0056] 2. Machine learning: Use machine learning models (such as BP neural network classification models or other existing classification models) combined with operation and inspection data to classify and predict situations where there is existing status data but the abnormal cause, diagnosis method, and repair method are unknown, forming a complete operation and inspection knowledge base for subsequent large models to search and improve the accuracy of fault prediction. The machine learning model takes fault feature data as input and the classification result of fault feature data as output.

[0057] Among them, BP neural network is a multi-layer feedforward neural network, and its main characteristics are: the signal is forward propagated, while the error is backward propagated. The process of BP neural network is mainly divided into two stages. The first stage is the forward propagation of the signal, from the input layer through the hidden layer, and finally to the output layer; the second stage is the backward propagation of the error, from the output layer to the hidden layer, and finally to the input layer, and the weights and biases from the hidden layer to the output layer, and the weights and biases from the input layer to the hidden layer are adjusted in turn.

[0058] The steps to construct a BP neural network are as follows:

[0059] Step 1: Initialize the network structure: set the number of network layers, the number of neurons in each layer, and the activation function. The three-layer BP neural network includes input layer, hidden layer, and output layer. The input layer has m neurons, corresponding to the feature data, the hidden layer has n neurons, and the output layer has k neurons, corresponding to the classification results.

[0060] Step 2: Weight initialization:

[0061] Randomly initialize the weight matrix:

[0062] W (1) : The weight matrix from the input layer to the hidden layer, with a dimension of n×m;

[0063] W (2): The weight matrix from the hidden layer to the output layer, with a dimension of k×n.

[0064] Step 3: Forward propagation:

[0065] Calculate the output of each layer: Input layer: x = [x 1 ,x 2 ,…,x m ]

[0066] Hidden layer: h = σ(W (1) x+b (1) ), where σ is the activation function (such as sigmoid or ReLU), b (1) is the bias term.

[0067] Output layer: Among them, softmax is used for multi-classification problems, b (2) is the bias term.

[0068] Step 4: Loss function:

[0069] The cross entropy loss function is used to measure the error of the model. The expression of the cross entropy loss function is as follows:

[0070]

[0071] Where L is the cross entropy loss function, k is the total number of true values ​​or labels, and y i is the ith true value or label, is the value or label predicted by the model.

[0072] Step 5: Back Propagation:

[0073] The gradient of the loss function for each weight is calculated using the chain rule, and the weight is updated:

[0074] Output layer error:

[0075] Hidden layer error: δ 1 =(W (2) ) T δ 2 ⊙σ′(W (1) x+b (1) )

[0076] Among them, ⊙ represents the bit-by-bit multiplication of elements, δ ′ is the derivative of the activation function.

[0077] Update weights:

[0078]

[0079] Where η is the learning rate.

[0080] Step 6: Iterative training:

[0081] Repeat steps 3 to 5 until the loss function converges or the preset number of iterations is reached.

[0082] Step 7: Generate behavior sequence:

[0083] The trained BP neural network is used to predict new inputs. The expression of the neural network output result is as follows:

[0084] output(n)=softmax(W (2) σ(W (1) x+b (1) )+b (2) )

[0085] In the formula, x is the input data, W (1) is the weight from the first layer to the hidden layer, W (2) is the weight from the hidden layer to the output layer, b is the bias term, σ is the hidden layer activation function, softmax is the output layer activation function, and output(n) is the n output results of the neural network. The result is a probability distribution, which indicates the probability that the input data belongs to different categories. This output value can further output the cause of the abnormality or the recommended repair method according to the category with the highest probability. The output results of the neural network include the predicted behavior method and the cause of the abnormality.

[0086] The status data is then combined with the predicted behavior method and then with the predicted cause of the abnormality to form a behavior sequence. The behavior sequence refers to a complete data strip from discovering the abnormality, analyzing the cause of the abnormality, providing a diagnostic method, and providing a repair method.

[0087] The training process of the BP neural network includes forward propagation to calculate output, calculate loss, and back propagation to update weights. By continuously iterating and optimizing model parameters, the classification of equipment fault feature data is ultimately achieved.

[0088] Among them, the construction of the self-learning operation and maintenance knowledge base automatically generates maintenance-related knowledge items through the self-learning mechanism. Knowledge items include abnormal phenomena, abnormal causes, diagnostic methods, and repair methods. Whenever new data is collected and analyzed, the self-learning operation and maintenance knowledge base will convert it into knowledge and establish a knowledge model. When a new maintenance case is entered, the self-learning operation and maintenance knowledge base automatically evaluates its impact on the knowledge base. The update formula of the self-learning operation and maintenance knowledge base is:

[0089] K new =K old +α(D new -D old )

[0090] Among them, K new D is the updated self-learning operation and inspection knowledge base. new is the data of the new case, α is the learning rate, K old D is the self-learning operation and inspection knowledge base before updating. old This is data from old cases.

[0091] Among them, the report generation module is used to automatically generate inspection reports based on the analysis results. Specifically, it generates substation inspection reports by integrating the real-time analysis results and the in-depth analysis results and combining the inspection template. That is, the report generation module integrates the output of the anomaly detection module and the output of the intelligent analysis module, and generates a substation inspection report that meets the specifications in combination with the inspection template.

[0092] The inspection template is obtained according to the inspection standard template, or user requirements, or the inspection standard template and user requirements. Among them, the inspection standard template can adopt an industry standard template or a custom standard template.

[0093] Specifically, the inspection template is obtained based on the State Grid inspection standard template (as a kind of industry standard template) and user-defined requirements.

[0094] Among them, the content of the substation inspection report may include detected abnormal conditions, detailed analysis results and corresponding maintenance suggestions. The substation inspection report also includes status assessment, abnormality analysis, risk prediction and maintenance optimization suggestions for substation equipment.

[0095] Equipment status assessment report: Equipment status assessment is one of the core functions of the system. In the prompt template, the big model is prompted with what data to collect and how to output it, so as to trigger the thinking chain technology of the big model. By collecting and integrating multi-source data such as equipment status data (such as temperature, appearance, etc.), maintenance logs, and operation logs, the system can conduct real-time assessment of the current health status of the equipment. Specifically, the system uses big data analysis technology to model historical data and compare it with current data to analyze whether the equipment is in normal operation.

[0096] In order to improve the accuracy of equipment status assessment, multiple machine learning models can be introduced through Function based on the multimodal reasoning framework, such as time series-based prediction models, anomaly detection models, etc. These models can capture subtle changes in the operation of equipment, identify early signals that may cause failures, and output signals and other data to the big model. With the help of the big model, detailed status reports are generated to help maintenance personnel find problems in a timely manner.

[0097] Abnormal analysis report: When the equipment runs abnormally, the system can detect and generate abnormal analysis reports in real time. The abnormal analysis module combines the natural language processing function of the multimodal reasoning framework to analyze unstructured data (such as maintenance personnel's logs and reports) and automatically extract key information related to the abnormality. In addition, the system will also combine the equipment status data to conduct in-depth analysis of the abnormality to determine which component or parameter is out of the normal range and may have caused the abnormality.

[0098] The abnormality analysis report includes: (1) Abnormality description: The system generates a detailed description of the current abnormality by analyzing text and data, including the time when the abnormality occurred, the abnormality type, the scope of impact, etc. (2) Possible causes: Combining historical data and empirical models, the system can infer the possible causes of the abnormality, such as aging of equipment components, overload operation, external environmental impact, etc. (3) Handling suggestions: Based on the type and cause of the abnormality, the system automatically generates maintenance suggestions, such as the need to replace components, adjust operating parameters, or strengthen monitoring.

[0099] Risk prediction report: Based on the historical operation data, maintenance records and current status assessment results of the equipment, the risk prediction module can use the multimodal reasoning framework to call the local machine learning algorithm to predict the future failure risk of the equipment. Specifically, the system uses a time series-based prediction model combined with the current operation trend of the equipment to predict the possible time and type of failure.

[0100] The risk prediction report includes: (1) Risk assessment results: The system generates a fault risk score (0-100) for each device, which is based on multiple factors such as the current operating status of the device and historical fault records. (2) Potential fault types: Based on the risk assessment results, the system lists the possible fault types, such as motor overheating and component wear. (3) Prediction time window: Based on the model's prediction, the system gives a time window in which the fault may occur, helping operation and maintenance personnel to formulate maintenance plans in advance and avoid faults.

[0101] Maintenance optimization recommendation report: In order to help operation and maintenance personnel better manage equipment maintenance, the system will automatically generate maintenance optimization recommendations based on equipment status assessment, abnormality analysis and risk prediction results. These recommendations include not only short-term repairs and maintenance, but also long-term equipment management and maintenance strategy optimization.

[0102] Maintenance optimization suggestions include: (1) Maintenance cycle adjustment: Based on the actual operating status of the equipment, the system will recommend adjusting the maintenance cycle. For example, for equipment in poor operating conditions, the system will recommend shortening the maintenance cycle; while for equipment in good condition, the system may recommend appropriately extending the maintenance cycle to save maintenance costs. (2) Spare parts recommendations: Based on the risk prediction results of the equipment, the system will recommend that the company prepare wearing parts or parts that may need to be replaced in advance to avoid production being affected by lack of spare parts when equipment fails. (3) Resource allocation optimization: Based on the health status and risk assessment results of different equipment, the system will recommend that more maintenance resources (such as technicians, funds, etc.) be allocated to equipment with higher risks first, thereby improving resource utilization efficiency.

[0103] In addition, substation inspection reports can also be output in multiple formats supported by the module (such as PDF, HTML, DOCX), making it easier for users to read and share.

[0104] Among them, the multimodal reasoning framework can use the existing LangChain framework or other frameworks. The multimodal reasoning framework is used to analyze and process a large amount of unstructured data such as equipment inspection data, equipment status data, operation and maintenance data. Through this framework, the system can introduce various large models as intelligent analysis brains according to the actual situation on site. For example, the domestic open source model ChatGLM-6B, Qwen2-7B and other large models, the system can identify key information in the text, such as equipment model, fault type, maintenance time, abnormal status, etc. In addition, the multimodal reasoning framework can call the natural language query interface through the Function method, allowing users to query in natural language. For example, users can directly ask "What is the most recent fault of a certain device?" or "What is the current health status of the device?" The system will extract information from multiple data sources based on these questions and generate corresponding answers. This interactive method greatly improves the ease of use and intelligence level of the system.

[0105] The present invention can achieve the following technical effects and advantages:

[0106] Data fusion and real-time analysis: The data acquisition module obtains the substation's patrol data, equipment status data, and operation and maintenance data in real time, and performs fusion analysis on multimodal data based on the multimodal reasoning framework model to ensure the comprehensiveness and real-time nature of data processing. The system can not only process data from a single source, but also simultaneously analyze multi-source data from multiple systems, thereby improving the accuracy and timeliness of anomaly detection.

[0107] Accurate anomaly detection: Using the anomaly detection module, the system can monitor the equipment status data of the substation in real time, and evaluate the data through a multi-algorithm model based on ensemble learning to promptly detect potential equipment failures and abnormal conditions; through the real-time feedback mechanism, the system can dynamically adjust the detection parameters according to the actual situation to reduce the possibility of false alarms and missed alarms.

[0108] Intelligent analysis and fault location: The analysis module uses the thinking chain prompt technology of the multimodal reasoning framework, combined with the self-learning operation and inspection knowledge base, to form a dynamic knowledge system through the learning of patrol data, equipment status data, operation and maintenance data and the integration of new data to support real-time decision-making and analysis. Further in-depth analysis of the abnormal detection results helps users understand the root cause of the abnormality. The system combines the knowledge base with historical data to provide a more accurate explanation for abnormal data and put forward targeted maintenance suggestions, thereby optimizing the operation and maintenance management process. The knowledge base not only records equipment failure cases and treatment plans, but also analyzes the occurrence patterns of similar failures to form a self-optimizing learning mechanism.

[0109] Automatic report generation: The report generation module can automatically generate inspection reports that meet industry standards based on user needs. The report content covers the complete anomaly detection and analysis process, and provides an effective decision-making basis for the maintenance and upgrade of the power system. The system also supports multiple report output formats to facilitate the sharing and archiving of inspection reports.

[0110] Technical advantages: Strong data fusion and analysis capabilities: By integrating patrol data, equipment status data and operation and maintenance data, the system can achieve all-round data fusion and analysis, greatly improving the accuracy of equipment status assessment and anomaly detection.

[0111] Strong natural language processing capabilities: The introduction of the multimodal reasoning framework enables the system to process large amounts of unstructured data and support natural language interaction, greatly improving user experience and operational efficiency.

[0112] Intelligent maintenance recommendations: The system can not only detect equipment anomalies in real time, but also provide forward-looking maintenance recommendations to help companies optimize maintenance strategies and reduce the risk of equipment failure.

[0113] The report generation module can automatically generate inspection reports based on user-defined templates and inspection specifications, including detected abnormal conditions, analysis results and maintenance recommendations, to facilitate subsequent decision-making and archiving.

[0114] The intelligent analysis system of the present invention can not only improve the efficiency of patrol report generation, but also improve the overall quality of patrol work through a self-learning mechanism, thereby ensuring the safe and stable operation of the power system.

[0115] The present invention combines a self-learning operation and maintenance knowledge base in the intelligent analysis module, forms a dynamic knowledge system through the learning of historical maintenance cases and the integration of new data, and supports real-time updating of the maintenance knowledge base; the knowledge base not only records equipment failure cases and treatment plans, but also can analyze the occurrence patterns of similar failures to form a self-optimizing learning mechanism. The report generation module of the present invention will combine the knowledge in the self-learning operation and maintenance knowledge base to improve the intelligence of the report, making the maintenance suggestions more targeted and forward-looking. The present invention integrates patrol data, equipment status data and operation and maintenance data to achieve all-round data fusion and analysis, greatly improving the accuracy of equipment status assessment and anomaly detection. The present invention can efficiently solve the problems existing in the prior art, such as incomplete data analysis, low data analysis efficiency, untimely response, inability to fully utilize the existing knowledge base, and insufficient depth of reports.

[0116] The present invention is described in detail below in conjunction with examples:

[0117] Data collection and preprocessing: In the intelligent patrol system of the substation, through real-time monitoring: the temperature sensor of the substation collects the signal that the transformer temperature is too high, the current and voltage data of the switchgear show abnormal fluctuations, and the equipment status log records the time of the last maintenance and other information.

[0118] Anomaly detection and preliminary diagnosis: All collected real-time data will be input into the anomaly detection module. This module uses machine learning and deep learning models (such as neural networks, time series prediction algorithms, etc.) to analyze data to detect whether there is equipment behavior that exceeds the normal range. For example, based on historical data and real-time sensor data, the anomaly detection model finds that the transformer temperature continues to rise above the threshold, and there may be an overheating problem. At the same time, the fluctuation pattern of current and voltage is consistent with the historical fault pattern, and the diagnosis of electrical fault is more likely. This step uses time series analysis to capture long-term trends and emergencies in the data, and obtains indications of potential faults through the output of the model, for example: the transformer temperature will exceed the threshold in 30 minutes.

[0119] Self-learning operation and maintenance knowledge base: In traditional intelligent operation and maintenance systems, fault detection and analysis usually rely on manual rules and static expert knowledge bases. However, the self-learning operation and maintenance knowledge base overcomes this limitation through dynamic learning and adjustment. Based on machine learning methods, the system can automatically update and optimize the knowledge base based on historical equipment data and operation and maintenance experience. For example, the system will continuously update its understanding of problems such as transformer overheating and electrical failures based on real-time feedback and new fault cases. If a certain device in the substation often fails under specific environmental conditions (such as temperature, humidity, load, etc.), the system will automatically add these new patterns to the knowledge base. The self-learning mechanism enables the knowledge base to not only support static rules, but also to continuously optimize over time to improve the accuracy of fault diagnosis and prediction.

[0120] Intelligent report generation and decision support: Once an anomaly is detected, the multimodal reasoning framework obtains input through large model integration prompt templates, calling system function tools, indexing knowledge base experience, etc., and automatically generates intelligent reports for output and provides them to operation and maintenance personnel. These reports not only contain fault diagnosis information, but also recommend the next step based on the analysis of the intelligent analysis system (such as whether to stop for inspection, replace equipment, adjust parameters, etc.). For example, if the temperature of the transformer is too high, the intelligent report will point out the specific reasons for the high temperature (such as insufficient heat dissipation, excessive current load, etc.), and recommend appropriate maintenance plans based on historical data and failure modes (such as cleaning the radiator, replacing equipment with larger loads, etc.).

[0121] Real-time feedback and decision adjustment: The generation of intelligent reports and decision suggestions are only preliminary diagnoses. The system will also adjust decision strategies in real time based on the feedback from operators. For example, if operators modify the diagnostic results in the report or choose a different operation plan, the system will record these feedbacks and further adjust its knowledge base. In this way, the system can not only generate reports, but also continuously optimize its detection and decision-making process through the operations of operators. For example, after the operator selects "equipment is fault-free" in the report and performs equipment operation debugging, the system will update its knowledge base based on this decision and improve detection and report generation in similar situations in the future.

[0122] The multimodal reasoning framework adopted in the present invention supports different types of data input (such as text, images, voiceprint data, etc.), realizes information fusion and intelligent decision-making through a cross-modal reasoning process, and can analyze and process information provided by different data sources. It is suitable for application in multi-source data environments such as substations to improve the efficiency and accuracy of substation inspection report generation, and solve the problems existing in the prior art such as low data analysis efficiency, inability to fully utilize the existing knowledge base, and insufficient depth of reports.

[0123] The present invention can effectively realize deep learning and understanding of massive data by combining large-scale model intelligent technology, thereby improving the accuracy and efficiency of analysis. The existing multimodal reasoning framework can give full play to its multi-module fusion capabilities, allowing the system to flexibly integrate different types of data and models, and enhance the effect of intelligent analysis. In addition, the introduction of the self-learning operation and inspection knowledge base can enable the system to continuously optimize its own knowledge structure when processing new cases, improve adaptability and real-time performance, and ensure that effective maintenance cases can be provided for large models in different devices and environments, making the reports output by large models more intelligent and accurate.

[0124] A method for generating a substation inspection report includes the following steps:

[0125] Based on the big model of the multimodal reasoning framework, the inspection data, equipment status data and operation and maintenance data of the substation are analyzed in real time through integrated image recognition technology and integrated time series prediction algorithm. The abnormal data obtained by the real-time analysis is deeply analyzed through the thinking chain prompt technology of the big model, and the historical operation and maintenance data, the knowledge base called by the big model and at least one of the predefined standards are combined to obtain the in-depth analysis results. The substation inspection report is generated by integrating the real-time analysis results and the in-depth analysis results in combination with the inspection template; the inspection data includes at least one of the images, videos and voiceprint information of the substation equipment.

[0126] Specifically, during in-depth analysis, relevant contextual information used to provide a basis for the interpretation of abnormal data is obtained from the knowledge base through retrieval enhancement generation technology based on vector retrieval.

[0127] Specifically, the knowledge base can use an existing expert database or a self-learning operation and inspection knowledge base. The self-learning operation and inspection knowledge base is constructed in the following way:

[0128] Utilizing the natural language processing capabilities of the big model, historical maintenance cases and equipment failure records are analyzed to extract fault feature data. The trained machine learning model is combined with operation and inspection data to classify and predict situations where there is existing equipment status data but the abnormal causes, diagnostic methods, and repair methods are unknown, thus forming a self-learning operation and inspection knowledge base. The machine learning model takes fault feature data as input and the classification results of the fault feature data as output.

[0129] Specifically, the self-learning operation and maintenance knowledge base is also updated based on the real-time feedback of the operation and maintenance personnel on the substation inspection reports.

[0130] Specifically, the machine learning model may select a BP neural network classification model or other existing classification models.

[0131] Specifically, the substation inspection report may include at least one or more of status assessment, abnormality analysis, risk prediction and maintenance optimization suggestions for substation equipment. Of course, other contents may also be included according to actual needs.

[0132] Specifically, the large model includes ChatGLM-6B and Qwen2-7B.

[0133] Specifically, before real-time analysis, the patrol data, equipment status data, and operation and maintenance data are preliminarily processed through edge computing technology, and the results of the preliminary processing are used as the patrol data, equipment status data, and operation and maintenance data for the final real-time analysis.

[0134] Specifically, the integrated image recognition technology includes convolutional neural network and YOLO algorithm.

[0135] Specifically, the integrated time series prediction algorithms include long short-term memory networks, recurrent neural networks, and Prophet algorithms.

[0136] Specifically, the inspection template is obtained according to the inspection standard template and / or user requirements, and the inspection standard template selects an industry standard template.

[0137] Specifically, the LangChain framework is used as the multimodal reasoning framework.

[0138] An embodiment of a method for generating a substation inspection report:

[0139] A method for generating a substation inspection report has been described in detail in an embodiment of a substation inspection report generating system, and will not be repeated here.

Claims

1. A method for generating a substation inspection report, characterized in that: The steps include: A big model based on a multimodal reasoning framework performs real-time analysis on the inspection data, equipment status data and operation and maintenance data of the substation through integrated image recognition technology and integrated time series prediction algorithm. The abnormal data obtained from the real-time analysis is deeply analyzed through the thinking chain prompt technology of the big model, and the in-depth analysis results are obtained by combining the historical operation and maintenance data, the knowledge base called by the big model and at least one of the predefined standards. The substation inspection report is generated by integrating the real-time analysis results and the in-depth analysis results in combination with the inspection template; the inspection data includes at least one of the images, videos and voiceprint information of the substation equipment.

2. The method for generating a substation inspection report according to claim 1, characterized in that: During in-depth analysis, relevant context information for providing a basis for interpreting abnormal data is obtained from the knowledge base through a retrieval enhancement generation technology based on vector retrieval.

3. The method for generating a substation inspection report according to claim 1 or 2, characterized in that: The knowledge base is a self-learning operation and inspection knowledge base, which is constructed in the following way: Utilize the natural language processing capabilities of the big model to analyze historical maintenance cases and equipment failure records, extract fault feature data, use the trained machine learning model combined with operation and inspection data, classify and predict situations where there is existing equipment status data but the abnormal cause, diagnosis method, and repair method are unknown, and form the self-learning operation and inspection knowledge base; The machine learning model takes fault feature data as input and takes classification results of the fault feature data as output.

4. The method for generating a substation inspection report according to claim 3, characterized in that: The self-learning operation and inspection knowledge base is also updated according to the real-time feedback of the operation and maintenance personnel on the substation inspection report.

5. The method for generating a substation inspection report according to claim 1, characterized in that: The substation inspection report includes status assessment, abnormality analysis, risk prediction and maintenance optimization suggestions for substation equipment.

6. The method for generating a substation inspection report according to claim 1, characterized in that: Before real-time analysis, the patrol data, equipment status data, and operation and maintenance data are preliminarily processed through edge computing technology, and the results of the preliminary processing are used as the patrol data, equipment status data, and operation and maintenance data for the final real-time analysis.

7. The method for generating a substation inspection report according to claim 1, characterized in that: The integrated image recognition technology includes convolutional neural network and YOLO algorithm; the integrated time series prediction algorithm includes long short-term memory network, recurrent neural network and Prophet algorithm.

8. The method for generating a substation inspection report according to claim 1, characterized in that: The inspection template is obtained according to the inspection standard template and / or user requirements, and the inspection standard template selects an industry standard template.

9. The method for generating a substation inspection report according to claim 1, characterized in that: The multimodal reasoning framework uses the LangChain framework.

10. A substation inspection report generation system, comprising a processor, characterized in that: The processor is used to execute a computer program to implement the steps of the substation inspection report generating method according to any one of claims 1 to 9.

Images