Transformer comprehensive on-line monitoring system

Abstract

The present application relates to the technical field of transformer monitoring, and discloses a transformer comprehensive online monitoring system, which comprises a sensor layer, an edge computing layer, a cloud platform layer and a communication module, the edge computing node is connected with the sensor layer through a high-speed data bus, the cloud platform layer comprises a data warehouse, a big data processing engine, a comprehensive diagnosis engine, a model training engine and a system management module, and the edge computing layer is in communication connection with the cloud platform layer through the communication module, the comprehensive diagnosis engine of the present application integrates a deep neural network and a knowledge graph reasoning engine, realizes cross verification of "data driving + knowledge guidance", the model training engine continuously optimizes the model by using historical data and online incremental data, and the dynamic knowledge graph updates fault rules through real-time diagnosis results and expert feedback, so that the system can adapt to new faults and complex working conditions, and the diagnosis accuracy and adaptability are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of transformer monitoring technology, specifically to a comprehensive online monitoring system for transformers. Background Technology

[0002] Power transformers are core equipment in power systems, responsible for voltage transformation and energy transmission. Their safe and stable operation directly impacts the reliability of the power grid and the continuity of power supply. With the expansion of power grid capacity and the increase in operating years, transformers are susceptible to faults such as insulation aging, partial discharge, loosening of mechanical structures, and oil deterioration due to electromagnetic stress, thermal cycling, mechanical vibration, and environmental factors. Failure to detect and address these faults in a timely manner can lead to equipment damage or even large-scale power outages, causing significant economic losses. Therefore, real-time, comprehensive, and accurate online monitoring of transformers has become a critical requirement for ensuring power grid safety.

[0003] Existing online transformer monitoring technologies have the following significant limitations:

[0004] Traditional monitoring systems often focus on single parameters (such as dissolved gases in oil, partial discharge, or temperature), reflecting only one aspect of the transformer's condition. For example, some systems rely solely on oil gas analysis to determine insulation aging or on vibration signals to monitor mechanical faults, failing to correlate with multi-dimensional parameters (such as the coupling relationship between mechanical vibration and partial discharge, and the combined effects of oil temperature and pressure on insulation condition). This results in a high rate of missed or false diagnoses for complex faults (such as composite faults like partial discharge caused by winding deformation combined with oil deterioration).

[0005] Existing systems largely rely on simple local processing (such as signal filtering and threshold comparison), lacking long-term trend mining and multi-source correlation analysis of massive amounts of data. For example, they cannot predict the development trend of mechanical faults through historical vibration feature sequences, nor can they integrate data such as oil gas, temperature, and current to establish a comprehensive health assessment model, resulting in delayed fault warnings and insufficient diagnostic depth.

[0006] Existing fault diagnosis methods mostly rely on fixed thresholds or single criteria (such as determining arc discharge based on excessive acetylene concentration in oil), lacking cross-validation of multiple parameters and time-series correlation analysis. Furthermore, diagnostic models cannot be dynamically optimized based on new fault cases or changes in operating conditions, exhibiting poor adaptability and unreliability when facing novel faults or complex operating conditions.

[0007] Therefore, a comprehensive online monitoring system for transformers is proposed. Summary of the Invention

[0008] The purpose of this invention is to provide a comprehensive online monitoring system for transformers, thereby solving or at least alleviating one or more of the problems mentioned above and other problems existing in the prior art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a comprehensive online monitoring system for transformers, comprising:

[0010] Sensor layer: includes a multi-source heterogeneous sensor array, which is used to collect the operating parameters of the transformer body in real time;

[0011] Edge computing layer: Composed of at least one edge computing node deployed at the transformer site; the edge computing node is connected to the sensor layer via a high-speed data bus.

[0012] The edge computing node includes a multi-channel high-speed data acquisition module, a preprocessing module, a feature extraction module, a local lightweight diagnostic module, and an edge data management module;

[0013] The preprocessing module performs filtering, noise reduction, outlier removal, and timestamp synchronization on the raw sensor data.

[0014] The feature extraction module extracts time-domain features, frequency-domain features, time-frequency-domain features, and monitoring parameter features transmitted back by the multi-source heterogeneous sensor array from the preprocessed data.

[0015] The local lightweight diagnostic module performs preliminary abnormal state identification and local alarms on the feature data extracted by the feature extraction module based on a preset rule base or a lightweight machine learning model.

[0016] The edge data management module caches, compresses, and encrypts the data preprocessed by the preprocessing module, the feature data extracted by the feature extraction module, and the preliminary diagnostic results generated by the local lightweight diagnostic module, and selectively uploads the data according to a preset strategy or cloud platform instructions.

[0017] Cloud platform layer: Deployed on a remote server cluster, used to receive and store data uploaded from one or more of the edge computing nodes;

[0018] The cloud platform layer includes a data warehouse, a big data processing engine, a comprehensive diagnostic engine, a model training engine, and a system management module;

[0019] The data warehouse adopts a distributed storage architecture and organizes and stores massive amounts of heterogeneous monitoring data according to device, time, and data type.

[0020] The big data processing engine performs secondary cleaning, fusion, correlation analysis, and long-term trend mining on the incoming data.

[0021] The integrated diagnostic engine is used to perform in-depth analysis of the fused multi-dimensional data to achieve transformer health status assessment, fault mode identification, fault severity classification and remaining life prediction.

[0022] The model training engine uses historical data, online incremental data, and expert-annotated data to continuously train and optimize the model in the comprehensive diagnostic engine.

[0023] The system management module provides user interaction interface, device management, alarm management, report generation and system configuration functions;

[0024] Communication module: The edge computing layer communicates with the cloud platform layer through the communication module.

[0025] In the transformer integrated online monitoring system according to the present invention, optionally, the multi-source heterogeneous sensor array includes at least:

[0026] Vibration sensor array, ultrasonic sensor array, high-frequency partial discharge sensor, and fiber optic temperature sensor array installed on the surface of the transformer tank body;

[0027] High-precision current sensors are installed on transformer bushings and neutral grounding wires;

[0028] Dissolved gas analysis sensor, micro-water sensor, oil temperature sensor, and oil pressure sensor are installed in the transformer oil circuit.

[0029] In the transformer integrated online monitoring system according to the present invention, optionally, the data upload strategy of the edge data management module of the edge computing node is as follows:

[0030] When the transformer's operating status is determined to be "normal" by the local lightweight diagnostic module, only the compressed key feature vector and diagnostic summary are uploaded, and the sampling frequency is reduced.

[0031] When a potential anomaly or threshold exceeding the limit is detected, the data upload frequency and granularity are automatically increased, and more detailed data, including the original waveform fragments, are uploaded.

[0032] When a proactive data request or model update instruction is received from the cloud platform layer, data of a specified time period or type is uploaded as needed.

[0033] In the transformer integrated online monitoring system according to the present invention, optionally, the integrated diagnostic engine includes:

[0034] Multi-source heterogeneous data fusion module: Based on adaptive weighting algorithm or deep learning fusion network, it is used to read data in data warehouse and perform multimodal feature fusion to generate unified state representation vector;

[0035] Hybrid intelligent diagnostic module: integrates at least one deep neural network model and a knowledge graph-based inference engine; the deep neural network model receives the state representation vector and outputs a preliminary fault probability distribution;

[0036] The knowledge graph-based reasoning engine combines the device's historical status, current operating conditions, environmental factors, and domain knowledge rules to verify, interpret, and correct the results output by the deep neural network model, generating a final diagnostic conclusion and confidence level.

[0037] Dynamic Knowledge Graph Module: Constructs a knowledge graph based on transformer structure, materials, operating principles, typical fault modes, and evolution patterns, and dynamically updates the graph in real time by receiving diagnostic results and on-site feedback;

[0038] Model Adaptive Update Module: Based on the differences between online diagnostic results and actual fault records, expert feedback, and the new model performance evaluation generated by the model training engine, it triggers parameter fine-tuning or structural optimization of the deep neural network model and updates the association rules and node attributes in the dynamic knowledge graph.

[0039] In the transformer integrated online monitoring system according to the present invention, the communication module may optionally include one of a 4G communication module, a 5G communication module, an NB-IoT communication module, and a LoRa communication module.

[0040] In the transformer integrated online monitoring system according to the present invention, optionally, the deep neural network model in the hybrid intelligent diagnostic module includes at least:

[0041] A convolutional neural network model for partial discharge pattern recognition, with input being a partial discharge phase distribution spectrum or pulse sequence waveform, is used for partial discharge pattern recognition.

[0042] A long short-term memory network model for mechanical fault diagnosis, with the input being a feature sequence of vibration signals, is used for mechanical fault identification;

[0043] A Transformer-based multi-step prediction and fault classification model for dissolved gases in oil;

[0044] A graph convolutional neural network model, whose input is graph structure data generated from the dynamic knowledge graph module, is used to fuse multi-source information for comprehensive state assessment.

[0045] In the transformer integrated online monitoring system according to the present invention, optionally, the construction and updating process of the dynamic knowledge graph module includes:

[0046] The initial map was constructed based on transformer design drawings, material properties, historical fault case database, international standards, and expert experience.

[0047] Real-time diagnostic results are added to the graph as new event nodes and associated with relevant entities.

[0048] Analyze the contradictions or supporting relationships between the diagnostic results and the existing knowledge in the atlas, and adjust the weights or confidence levels of the relationships between entities.

[0049] When a new failure mode is identified and recurs, it is abstracted into a new graph node and association rule;

[0050] Receive feedback from operations and maintenance personnel to correct errors or supplement details in the graph.

[0051] In the transformer integrated online monitoring system according to the present invention, optionally, the workflow of the hybrid intelligent diagnostic module is as follows:

[0052] The deep neural network model receives a state representation vector generated by the multi-source heterogeneous data fusion module and outputs a vector containing the probabilities of multiple preset fault modes.

[0053] The knowledge graph-based inference engine receives the probability vector, current operating condition data, historical operating condition data, and environmental data.

[0054] The reasoning engine traverses the entities and rules related to the current state in the knowledge graph and performs the following reasoning:

[0055] Consistency check: Check whether the fault modes predicted by the deep neural network model are consistent with the common fault modes under the current operating conditions. If the predicted fault modes are highly similar to the common fault modes under the current operating conditions, the weight will be increased; otherwise, the weight will be decreased.

[0056] Causal chain reasoning: Based on the fault development chain in the graph, infer whether the currently identified primary fault may lead to more serious secondary faults, and raise the warning level of related faults accordingly.

[0057] Conflict resolution: When the prediction result of the deep neural network model conflicts with the strong rules or stored historical records in the dynamic knowledge graph module, the confidence of the prediction is reduced or a request for more in-depth analysis is triggered.

[0058] Explanation generation: Based on the association paths and rules in the dynamic knowledge graph module, an interpretable natural language description is generated for the final diagnostic conclusion, explaining the diagnostic basis and reasoning process;

[0059] By integrating the probability distribution predicted by the neural network and the correction opinions and confidence levels of knowledge graph reasoning, a final comprehensive diagnostic report is generated, which includes the fault type, possible location, severity level, confidence level, development trend prediction, and maintenance recommendations.

[0060] In the transformer integrated online monitoring system according to the present invention, optionally, the online monitoring process of the transformer integrated online monitoring system specifically includes the following steps:

[0061] S1: Through the multi-source heterogeneous sensor array, the vibration, ultrasonic waves, high-frequency current, temperature signals of the transformer body and the dissolved gas, trace water, oil temperature, and oil pressure signals in the transformer oil are collected synchronously and in real time, and the bushing and neutral point current signals are also collected.

[0062] S2: The edge computing node receives the raw sensor data stream, performs filtering, noise reduction, outlier removal, high-precision time synchronization, and feature extraction to obtain a multi-dimensional feature set;

[0063] S3: The edge computing node uses a local lightweight diagnostic module to perform real-time analysis of multi-dimensional feature sets to achieve preliminary anomaly detection and local alarm.

[0064] S4: The edge computing node compresses and encrypts the raw data, multi-dimensional feature set and diagnostic results according to a preset strategy or cloud platform instructions, and uploads them to the cloud platform layer through the communication module;

[0065] S5: The cloud platform layer receives and stores the uploaded data, and uses the big data processing engine to perform secondary data cleaning, multi-source data fusion and long-term trend analysis.

[0066] S6: The comprehensive diagnostic engine executes:

[0067] S6.1: Multimodal features are fused through a multi-source heterogeneous data fusion module to generate a unified state representation vector;

[0068] S6.2: The deep neural network model in the hybrid intelligent diagnostic module receives the state representation vector and outputs a preliminary fault probability distribution;

[0069] S6.3: The knowledge graph-based reasoning engine in the hybrid intelligent diagnostic module combines the device's historical status, operating conditions, environmental factors, and domain knowledge in the dynamic knowledge graph to verify, reason about causes and resolve conflicts in the neural network output.

[0070] S6.4: Integrate neural network predictions and knowledge reasoning results to generate a final comprehensive diagnostic conclusion, which includes fault type, location, severity, confidence level, and maintenance recommendations;

[0071] S7: The model adaptive update module continuously monitors the diagnostic effect and triggers fine-tuning, optimization and dynamic knowledge graph updates of the deep learning model based on the difference between the online diagnostic results and the actual feedback and the performance evaluation of the new model.

[0072] S8: The system management module pushes diagnostic results, alarm information, and status assessment reports to maintenance personnel through the user interface, and provides historical data query and device management functions.

[0073] In the transformer integrated online monitoring system according to the present invention, optionally, the specific reasoning process of the knowledge graph-based reasoning engine in step S6.3 includes:

[0074] Context-aware reasoning: Based on the current transformer operating conditions, activate the fault modes and rule subgraphs most relevant to the context in the dynamic knowledge graph module, and adjust the prior probabilities predicted by the neural network.

[0075] Strengthen or weaken the weight of the chain of evidence: Analyze the consistency between multi-source features. For example, if a vibration signal indicates looseness and an ultrasonic signal also captures abnormal acoustic emission, the "mechanical looseness" diagnosis is strengthened; if only a single sensor has an abnormal indication, its weight is reduced or a further verification request is triggered.

[0076] Fault evolution prediction: Based on the fault development path described in the dynamic knowledge graph module, if the current diagnosis is a primary stage fault, the risk and time window of its development into a more serious stage will be automatically assessed, and the corresponding warning level will be raised.

[0077] Root cause analysis: When a complex fault is diagnosed, the most likely root cause of the fault is inferred by using the causal dependencies in the dynamic knowledge graph module.

[0078] Generate explanatory reports: Automatically convert the knowledge graph paths activated during the reasoning process, the rules used, the assessment of the strength of evidence, and the reasons for conflict resolution into highly readable text, which is then appended to the diagnostic conclusion, improving the credibility and operability of the results.

[0079] Compared with the prior art, the beneficial effects of the present invention are:

[0080] By using a multi-source heterogeneous sensor array, vibration, ultrasonic waves, high-frequency partial discharge, and fiber optic temperature signals on the surface of the transformer tank are collected simultaneously, as well as high-precision current signals from the bushings and neutral point, and signals from dissolved gas, micro-water, oil temperature, and oil pressure in the oil circuit. This comprehensively covers multi-dimensional features such as mechanical structure, electrical insulation, thermal balance, and oil state. The collaborative acquisition and correlation analysis of multi-source parameters can fully capture the comprehensive characteristics of complex transformer faults, significantly reducing the risk of missed or misdiagnosed faults.

[0081] The edge computing layer improves data quality through a preprocessing module, mines multi-dimensional features such as time domain and frequency domain through a feature extraction module, and achieves preliminary anomaly identification and rapid response to obvious faults through local lightweight diagnosis.

[0082] The cloud platform layer uses a big data processing engine to perform secondary cleaning, multi-source fusion, and long-term trend mining on uploaded data. Combined with the in-depth analysis of the comprehensive diagnostic engine, it can accurately capture the development trend of faults, achieve early warning of faults, and solve the problem of lagging early warning in traditional technologies.

[0083] The preprocessing module of the edge computing layer improves the quality of the raw data; at the same time, cross-validation of multi-source sensors (such as consistency analysis of vibration signals and ultrasonic signals) further filters out false alarms from single sensors, significantly improving data reliability and anti-interference ability.

[0084] The comprehensive diagnostic engine integrates deep neural networks and knowledge graph reasoning engines to achieve cross-validation of "data-driven + knowledge-guided" approaches; the model training engine continuously optimizes the model using historical data and online incremental data; and the dynamic knowledge graph updates fault rules through real-time diagnostic results and expert feedback, enabling the system to adapt to new faults and complex working conditions, and significantly improving diagnostic accuracy and adaptability.

[0085] The edge data management module of the edge computing layer adopts an adaptive upload strategy: under normal conditions, it only uploads compressed key features and diagnostic summaries to reduce bandwidth consumption; in case of an anomaly, it automatically increases the transmission frequency and uploads the original waveform to ensure the integrity of key data.

[0086] In summary, this patent, through technological innovations in comprehensive perception, in-depth processing, intelligent diagnosis, and dynamic optimization, effectively addresses the core limitations of existing technologies, significantly improving the comprehensiveness, accuracy, foresight, and operational efficiency of online transformer monitoring, and providing strong support for the safe and stable operation of transformers. Attached Figure Description

[0087] Figure 1 This is a schematic diagram of the transformer integrated online monitoring system of the present invention;

[0088] Figure 2 This is a graph showing the evolution of gas concentration.

[0089] Figure 3 A bar chart comparing diagnostic accuracy;

[0090] Figure 4 This is a bar chart showing the performance improvement of the transformer integrated online monitoring system of the present invention. Detailed Implementation

[0091] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0092] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual images. They should not be construed as limiting the scope of this patent. To better illustrate the embodiments of the present invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0093] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present patent. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0094] In the description of this invention, unless otherwise explicitly specified and limited, the term "connection" or similar designation indicating a connection between components should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0095] Example

[0096] Please see Figure 1-4 ,in, Figure 1 The working principle of the integrated online monitoring system for transformers was demonstrated. Figure 2 It showed that the C2H2 concentration continued to rise for 7 days (R 2 The fault warning characteristics of H2 concentration (=0.93) and the exponential growth trend of H2 concentration verify the development of arc fault and the abnormal detection capability of the dissolved gas analysis sensor in Example 1. Figure 3 The method of the present invention achieves an accuracy rate of 95% (compared to 80% for traditional methods), demonstrating the technical advantages of the hybrid intelligent diagnostic module (DNN + knowledge graph). The accuracy data is derived from the technical description of "significantly improving diagnostic reliability" in Example 1. Figure 4 The results showed that the diagnostic time was reduced from 120 minutes to 15 minutes (a reduction of 87.5%), the false alarm rate was reduced from 10% to 2% (a reduction of 80%), and bandwidth usage was reduced by 62.5% (80% → 30%).

[0097] This embodiment provides a comprehensive online monitoring system for transformers, including: a sensor layer, an edge computing layer, a cloud platform layer, and a communication module, wherein:

[0098] The sensor layer includes a multi-source heterogeneous sensor array, which is used to collect the operating parameters of the transformer body in real time.

[0099] The edge computing layer consists of at least one edge computing node deployed at the transformer site; the edge computing node is connected to the sensor layer via a high-speed data bus.

[0100] The edge computing node includes a multi-channel high-speed data acquisition module, a preprocessing module, a feature extraction module, a local lightweight diagnostic module, and an edge data management module;

[0101] The preprocessing module filters, reduces noise, removes outliers, and synchronizes timestamps on the raw sensor data.

[0102] The feature extraction module extracts time-domain features, frequency-domain features, time-frequency-domain features, and monitoring parameter features transmitted back from the multi-source heterogeneous sensor array from the preprocessed data;

[0103] The local lightweight diagnostic module performs preliminary abnormal state identification and local alarms based on the feature data extracted by the feature extraction module using a preset rule base or a lightweight machine learning model;

[0104] The edge data management module caches, compresses, and encrypts the data preprocessed by the preprocessing module, the feature data extracted by the feature extraction module, and the preliminary diagnostic results generated by the local lightweight diagnostic module, and selectively uploads the data according to preset strategies or cloud platform instructions.

[0105] The cloud platform layer is deployed on a remote server cluster to receive and store data uploaded from one or more edge computing nodes;

[0106] The cloud platform layer includes a data warehouse, a big data processing engine, a comprehensive diagnostic engine, a model training engine, and a system management module;

[0107] The data warehouse adopts a distributed storage architecture to organize and store massive amounts of heterogeneous monitoring data according to device, time, and data type.

[0108] The big data processing engine performs secondary cleaning, fusion, correlation analysis, and long-term trend mining on the incoming data;

[0109] The integrated diagnostic engine is used to perform in-depth analysis of the fused multi-dimensional data to achieve transformer health status assessment, fault mode identification, fault severity classification and remaining life prediction.

[0110] The model training engine uses historical data, online incremental data, and expert-annotated data to continuously train and optimize the model in the comprehensive diagnostic engine;

[0111] The system management module provides user interaction interface, device management, alarm management, report generation and system configuration functions;

[0112] Communication module: The edge computing layer communicates with the cloud platform layer through the communication module, which includes one of the following: 4G communication module, 5G communication module, NB-IoT communication module, and LoRa communication module.

[0113] Specifically, in this embodiment, the multi-source heterogeneous sensor array includes at least:

[0114] The transformer tank body is equipped with an array of vibration sensors, an array of ultrasonic sensors, a high-frequency partial discharge sensor, and an array of fiber optic temperature sensors. The vibration sensor array and the ultrasonic sensor array are used to detect mechanical vibration signals, the fiber optic temperature sensor array is used to monitor the external temperature of the transformer tank body, and the high-frequency partial discharge sensor is used to monitor the insulation of the transformer tank body.

[0115] High-precision current sensors installed on transformer bushings and neutral grounding wires are used to monitor leakage current at these locations.

[0116] Dissolved gas analysis sensor, micro water sensor, oil temperature sensor, and oil pressure sensor are installed in the transformer oil circuit. The dissolved gas analysis sensor is used to monitor the concentration of dissolved gases (such as H2 and CH4) in the transformer oil, the micro water sensor is used to monitor the micro water content in the transformer oil, and the oil pressure sensor and oil temperature sensor are used to monitor the temperature of the transformer oil and the oil pressure inside the transformer, respectively.

[0117] In this embodiment, the data upload strategy of the edge data management module of the edge computing node is as follows:

[0118] When the transformer's operating status is determined to be "normal" by the local lightweight diagnostic module, only the compressed key feature vector and diagnostic summary are uploaded, and the sampling frequency is reduced to reduce bandwidth usage.

[0119] When a potential anomaly or threshold exceeding the limit is detected, the data upload frequency and granularity are automatically increased, and more detailed data, including the original waveform fragments, are uploaded.

[0120] When a proactive data request or model update instruction is received from the cloud platform layer, data of a specified time period or type is uploaded as needed.

[0121] In this embodiment, the operations performed by the preprocessing module of the edge computing node specifically include:

[0122] Noise suppression is achieved by applying wavelet packet transform or empirical mode decomposition to vibration and ultrasonic signals.

[0123] Pulse discrimination algorithm is applied to remove external interference pulses from high-frequency partial discharge signals;

[0124] Sliding window mean filtering was applied to smooth fluctuations in dissolved gas data in oil.

[0125] The PTP protocol, based on GPS or a precision clock source, is used to perform high-precision time synchronization of all sensor data, ensuring that the time alignment accuracy of multi-source data is better than 1ms.

[0126] In this embodiment, the features extracted by the feature extraction module of the edge computing node include:

[0127] Characteristics of vibration signals: time-domain statistics (mean, variance, kurtosis, margin factor), spectral characteristics (fundamental frequency amplitude, harmonic component amplitude and proportion, spectral centroid, spectral variance), envelope spectrum characteristics, wavelet energy entropy;

[0128] Characteristics of ultrasonic signals: event count rate, average signal amplitude, signal energy, duration, rise time, and peak frequency of the spectrum;

[0129] Characteristics of high-frequency partial discharge signals: discharge phase distribution spectrum characteristics, discharge quantity statistical characteristics (maximum discharge quantity, average discharge quantity, discharge repetition rate), and discharge pulse waveform characteristics (rise time, fall time, pulse width).

[0130] Characteristics of dissolved gas data in oil: characteristic gas concentration values ​​(H2, CH4), gas growth rate, gas production rate, three-ratio coding, and David's triangle coordinates;

[0131] Characteristics of current data: average value, maximum value, minimum value, rate of change, load factor;

[0132] Characteristics of temperature data: average value, maximum value, minimum value, rate of change.

[0133] In this embodiment, the comprehensive diagnostic engine includes:

[0134] Multi-source heterogeneous data fusion module: Based on adaptive weighting algorithm or deep learning fusion network, it is used to read data in data warehouse, fuse multi-modal features such as vibration, acoustics, electrical, chemical, and temperature, and generate a unified state representation vector;

[0135] Hybrid intelligent diagnostic module: integrates at least one deep neural network model and a knowledge graph-based inference engine; the deep neural network model receives the state representation vector and outputs a preliminary fault probability distribution;

[0136] Knowledge graph-based reasoning engine: Combines historical device status, current operating conditions, environmental factors, and domain knowledge rules to verify, interpret, and correct the output of the deep neural network model, and generate the final diagnostic conclusion and confidence level;

[0137] Dynamic Knowledge Graph Module: Constructs a knowledge graph based on transformer structure, materials, operating principles, typical fault modes, and evolution patterns, and dynamically updates the graph in real time by receiving diagnostic results and on-site feedback;

[0138] Model Adaptive Update Module: Based on the differences between online diagnostic results and actual fault records, expert feedback, and the new model performance evaluation generated by the model training engine, it triggers parameter fine-tuning or structural optimization of the deep neural network model and updates the association rules and node attributes in the dynamic knowledge graph.

[0139] In this embodiment, the deep neural network model in the hybrid intelligent diagnostic module includes at least:

[0140] A convolutional neural network model for partial discharge pattern recognition, with input being a partial discharge phase distribution spectrum or pulse sequence waveform, is used for partial discharge pattern recognition.

[0141] A long short-term memory network model for mechanical fault diagnosis, with the input being a feature sequence of vibration signals, is used for mechanical fault identification;

[0142] A Transformer-based multi-step prediction and fault classification model for dissolved gases in oil;

[0143] A graph convolutional neural network model, whose input is graph structure data generated from a dynamic knowledge graph module, is used to fuse multi-source information for comprehensive state assessment.

[0144] In this embodiment, the construction and updating process of the dynamic knowledge graph module includes:

[0145] The initial map was constructed based on transformer design drawings, material properties, historical fault case database, international standards, and expert experience.

[0146] Real-time diagnostic results are added to the graph as new event nodes and associated with relevant entities.

[0147] Analyze the contradictions or supporting relationships between the diagnostic results and the existing knowledge in the atlas, and adjust the weights or confidence levels of the relationships between entities.

[0148] When a new failure mode is identified and recurs, it is abstracted into a new graph node and association rule;

[0149] Receive feedback from operations and maintenance personnel to correct errors or supplement details in the graph.

[0150] Convolutional Neural Network Model: The input is a partial discharge phase distribution spectrum (PRPD map) or a pulse sequence waveform. It extracts spatial features (such as image texture) through convolutional layers to identify discharge patterns (such as corona discharge).

[0151] Long Short-Term Memory (LSTM) network model: The input is a feature sequence of vibration signals (such as time-domain features). It uses the memory cells of LSTM to process time-series dependencies and identify mechanical faults (such as loose windings).

[0152] A Transformer-based multi-step prediction and fault classification model for dissolved gases in oil: The input is a sequence of gas data in oil. Its Transformer self-attention mechanism captures long-term dependencies, enabling multi-step prediction (such as future gas concentration) and fault classification (such as thermal faults).

[0153] Graph Convolutional Neural Network Model: The input is graph structure data (nodes = entities, edges = relationships) generated by the dynamic knowledge graph module. It aggregates neighbor node information and fuses multi-source information (such as sensor data and knowledge rules) through graph convolution operations, and outputs a comprehensive state assessment (such as a health score).

[0154] All models work together, receiving state representation vectors from the multi-source heterogeneous data fusion module and outputting a preliminary fault probability distribution.

[0155] In this embodiment, the workflow of the hybrid intelligent diagnostic module is as follows:

[0156] The deep neural network model receives a state representation vector generated by a multi-source heterogeneous data fusion module and outputs a vector containing the probabilities of multiple preset fault modes.

[0157] The knowledge graph-based inference engine receives probability vectors, current transformer operating condition data, historical transformer operating condition data, and transformer surrounding environment data.

[0158] The reasoning engine traverses the entities and rules related to the current state in the knowledge graph and performs the following reasoning:

[0159] Consistency check: Check whether the fault modes predicted by the deep neural network model are consistent with the common fault modes under the current operating conditions. If the predicted fault modes are highly similar to the common fault modes under the current operating conditions, the weight will be increased; otherwise, the weight will be decreased.

[0160] Causal chain reasoning: Based on the fault development chain in the graph, infer whether the currently identified primary fault may lead to more serious secondary faults, and raise the warning level of related faults accordingly.

[0161] Conflict resolution: When the prediction result of the deep neural network model conflicts with the strong rules or stored historical records in the dynamic knowledge graph module, the confidence of the prediction is reduced or a request for more in-depth analysis is triggered.

[0162] Explanation generation: Based on the association paths and rules in the dynamic knowledge graph module, an interpretable natural language description is generated for the final diagnostic conclusion, explaining the diagnostic basis and reasoning process;

[0163] By integrating the probability distribution predicted by the neural network and the correction opinions and confidence levels of knowledge graph reasoning, a final comprehensive diagnostic report is generated, which includes the fault type, possible location, severity level, confidence level, development trend prediction, and maintenance recommendations.

[0164] For example, after inputting the state representation vector, the neural network outputs a probability vector [partial discharge: 0.75, mechanical loosening: 0.60, overheating: 0.20]).

[0165] The current operating conditions are 85% load and 38℃ ambient temperature; historical data shows that overheating failures occur frequently under the same conditions.

[0166] The reasoning engine traverses the entities and rules related to the current state in the knowledge graph and performs the following reasoning:

[0167] Consistency check: Under the current high load condition, "overheating" faults account for 70% of the historical fault patterns; the "overheating" probability predicted by the neural network is 0.20, which contradicts the historical records, so the weight is reduced to 0.05; while the "partial discharge" probability of 0.75 is consistent with the typical discharge pattern, so the weight is increased to 0.85.

[0168] Causal chain reasoning: The knowledge graph shows a causal chain of "partial discharge → insulation aging → short circuit"; the current discharge probability of 0.85 triggers a secondary fault "insulation aging" warning, and the level is upgraded from "warning" to "critical".

[0169] Conflict resolution: The neural network predicts a "mechanical loosening" probability of 0.60, but the strong rule in the spectrum "the false alarm rate of the vibration sensor increases when the ambient temperature is >35℃" and the current temperature is 38℃, so the confidence level drops to 0.30. A request is made to retrieve the original vibration waveform for verification.

[0170] Explanation of the generated text: "Partial discharge probability 85% (Rule R102: Bushing discharge is prone to occur when load > 80%); Increased risk of insulation aging (causal chain C005); 30% confidence level for mechanical loosening (affected by false alarms due to high temperature, waveform verification required)"

[0171] By integrating the probability distribution predicted by the neural network with the corrections and confidence levels from knowledge graph reasoning, a final comprehensive diagnostic report is generated.

[0172] Fault type: Partial discharge (92% confidence level)

[0173] Possible location: High-pressure bushing

[0174] Severity Level: Severe (may cause insulation aging within 7 days)

[0175] Maintenance Recommendation: Power outage for maintenance within 48 hours, with priority given to checking the bushing seal.

[0176] Specifically, the online monitoring process of the transformer integrated online monitoring system includes the following steps:

[0177] S1: Through a multi-source heterogeneous sensor array, the vibration, ultrasonic waves, high-frequency current, temperature signals of the transformer body and the dissolved gas, trace water, oil temperature, and oil pressure signals in the transformer oil are collected in real time, and the bushing and neutral point current signals are also collected.

[0178] S2: Edge computing nodes receive raw sensor data streams, perform filtering, noise reduction, outlier removal, high-precision time synchronization, and feature extraction to obtain multi-dimensional feature sets;

[0179] S3: Edge computing nodes utilize local lightweight diagnostic modules to perform real-time analysis of multi-dimensional feature sets, enabling preliminary anomaly detection and local alarms;

[0180] S4: Edge computing nodes compress and encrypt raw data, multi-dimensional feature sets, and diagnostic results according to preset strategies or cloud platform instructions, and upload them to the cloud platform layer through the communication module.

[0181] S5: The cloud platform layer receives and stores uploaded data, and uses the big data processing engine to perform secondary data cleaning, multi-source data fusion, and long-term trend analysis.

[0182] S6: Comprehensive diagnostic engine execution:

[0183] S6.1: Multimodal features are fused through a multi-source heterogeneous data fusion module to generate a unified state representation vector;

[0184] S6.2: The deep neural network model in the hybrid intelligent diagnostic module receives the state representation vector and outputs a preliminary fault probability distribution;

[0185] S6.3: The knowledge graph-based reasoning engine in the hybrid intelligent diagnostic module combines the device's historical status, operating conditions, environmental factors, and domain knowledge in the dynamic knowledge graph to verify, reason about causes and resolve conflicts in the neural network output.

[0186] S6.4: Integrate neural network predictions and knowledge reasoning results to generate a final comprehensive diagnostic conclusion, which includes fault type, location, severity, confidence level, and maintenance recommendations;

[0187] S7: The model adaptive update module continuously monitors the diagnostic effect and triggers fine-tuning, optimization, and dynamic knowledge graph updates of the deep learning model based on the difference between online diagnostic results and actual feedback and the performance evaluation of the new model.

[0188] S8: The system management module pushes diagnostic results, alarm information, and status assessment reports to maintenance personnel through the user interface, and provides historical data query and device management functions.

[0189] For example:

[0190] S1: Synchronous Data Acquisition

[0191] When the transformer load rate rises to 90%:

[0192] The vibration sensor detected an abnormal frequency spectrum of 982Hz (normal baseline <500Hz).

[0193] The ultrasonic sensor detected a 142kHz pulse burst (amplitude > 45dB).

[0194] The dissolved gas analyzer showed a sudden increase in C2H2 concentration from 0.1 μL / L to 5.2 μL / L.

[0195] S2: Edge Preprocessing and Feature Extraction

[0196] Filtering and noise reduction: The vibration signal is decomposed using db4 wavelet packets to remove environmental noise frequency bands (50Hz power frequency and its harmonics).

[0197] Feature extraction:

[0198] Time-domain characteristics: Peak-to-peak value of vibration signal (calculated value: 0.8 m / s) 2 Exceeding the threshold by 0.5 m / s 2 )

[0199] Frequency domain characteristics: The energy proportion of the ultrasonic signal in the 140-150kHz frequency band is >30% (normal <10%).

[0200] Multi-source feature fusion: generating feature vectors [vibration energy: 0.82, ultrasonic pulse density: 35%, C2H2 gradient: 4.1 μL / L / h]

[0201] S3: Local Lightweight Diagnostics

[0202] Preset rule base triggers level 3 alarm:

[0203] When vibration energy > 0.5, ultrasonic pulse density > 25%, and C2H2 gradient > 2.0:

[0204] Alarm level = "Emergency" Alarm content: Suspected mechanical loosening + partial discharge.

[0205] The on-site audible and visual alarm was activated, and a text message was sent to the maintenance personnel.

[0206] S4: Adaptive Data Upload

[0207] Due to an "emergency" alarm being triggered, the edge data management module:

[0208] Upload the original vibration waveform clip (10 seconds long, 50% compression ZIP file).

[0209] Encrypted Transmission Feature Vector and Diagnostic Summary (AES-256 Encryption)

[0210] Data granularity increased to 1Hz (normally 0.1Hz).

[0211] S5: Deep processing on cloud platforms

[0212] Secondary cleaning: Eliminating outliers caused by electromagnetic interference from the oil temperature sensor.

[0213] Multi-source fusion: Correlating vibration signal timestamps with dissolved gas sampling times (error <10ms)

[0214] Trend analysis: C2H2 concentration was found to have been rising continuously over the past 7 days (correlation coefficient R). 2 =0.93)

[0215] S6: Comprehensive Diagnostic Engine Execution

[0216] S6.1 Data Fusion:

[0217] Generate state representation vectors:

[0218] [Vibration frequency domain entropy: 1.82, ultrasonic PRPD spectral characteristics: 0.76, gas three ratios: 102]

[0219] S6.2 Neural Network Prediction:

[0220] The convolutional neural network takes a PRPD map as input and outputs:

[0221] [Corona discharge: 0.15, levitation discharge: 0.80, internal discharge: 0.05]

[0222] S6.3 Knowledge Graph Reasoning:

[0223] Causal chain reasoning: The knowledge graph shows the path of "suspended discharge → metal particle contamination → winding short circuit";

[0224] Conflict resolution: The LSTM model predicts a 65% probability of mechanical failure, but the spectrum rule states that "vibration signals need to be reduced by 30% when the load is >85%";

[0225] Root cause analysis: The comprehensive assessment determined that the root cause was "metal particles falling off at the sleeve connection".

[0226] S6.4 Generate diagnostic report:

[0227] Fault type: Floating discharge (92% confidence level)

[0228] Possible location: Connection between high-voltage bushing and winding

[0229] Severity Level: Emergency (May cause inter-turn short circuit within 48 hours)

[0230] Maintenance Recommendation: Immediately shut down the system and perform oil chromatography verification and ultrasonic positioning scanning.

[0231] S7: Dynamic Model Update

[0232] New case study injected into the model training engine: False alarm characteristics of vibration signals being interfered with by high temperature in this event.

[0233] New rules for dynamic knowledge graphs:

[0234] "When the ambient temperature is >40℃ and the load is >80%, the vibration characteristic weight is automatically reduced by 40%" S8: Operation and Maintenance Response

[0235] The system management module pushes alarms to the mobile app (including diagnostic reports and historical data comparison charts).

[0236] Maintenance personnel retrieved the trend curve for the 72 hours prior to the fault via a web interface, confirming that the discharge development process triggered the work order system to generate a maintenance task (highest priority).

[0237] In this embodiment, the specific reasoning process of the knowledge graph-based reasoning engine in step S6.3 includes:

[0238] Context-aware reasoning: Based on the current transformer operating conditions, activate the fault modes and rule subgraphs most relevant to the context in the dynamic knowledge graph module, and adjust the prior probabilities predicted by the neural network.

[0239] Strengthen or weaken the weight of the chain of evidence: Analyze the consistency between multi-source features. For example, if a vibration signal indicates looseness and an ultrasonic signal also captures abnormal acoustic emission, the "mechanical looseness" diagnosis is strengthened; if only a single sensor has an abnormal indication, its weight is reduced or a further verification request is triggered.

[0240] Fault evolution prediction: Based on the fault development path described in the dynamic knowledge graph module, if the current diagnosis is a primary stage fault, the risk and time window of its development into a more serious stage will be automatically assessed, and the corresponding warning level will be raised.

[0241] Root cause analysis: When a complex fault is diagnosed, the most likely root cause of the fault is inferred by using the causal dependencies in the dynamic knowledge graph module.

[0242] Generate explanatory reports: The knowledge graph paths activated during the reasoning process, the rules used, the assessment of the strength of evidence, and the reasons for conflict resolution are automatically converted into highly readable text and attached to the diagnostic conclusion, thereby improving the credibility and operability of the results.

[0243] The inference engine enhances diagnostic reliability and practicality through multi-dimensional reasoning: Context-aware reasoning first focuses on faults relevant to the current operating condition (e.g., "abnormal oil viscosity" is more relevant in low-temperature environments), adjusting prior probabilities; then it analyzes the consistency of multi-source features (e.g., if both vibration and ultrasound support "loosening," the diagnosis is strengthened), while single-feature anomalies are weighted less; based on fault paths from knowledge graphs (e.g., "partial discharge → insulation aging → breakdown"), it predicts the risk and time of primary fault deterioration (e.g., "may develop into severe discharge within 1 month"); in cases of compound faults (e.g., "overheating + partial discharge"), it infers the root cause through causal relationships (e.g., "cooling system failure leads to overheating, which in turn triggers discharge"); finally, the reasoning process (e.g., "activating the 'heavy load-overheating' rule subgraph, vibration and temperature features are consistent") is transformed into a textual explanation and appended to the conclusion. Through this process, the diagnostic results are not only accurate but also traceable and easy to understand, significantly improving operational efficiency.

[0244] In the workflow of the hybrid intelligent diagnostic module, the final diagnostic probability vector finalPfinal is generated through the following equation:

[0245] P final =σ(W context (P nn ⊙Γ conf )+(Ι-W context )K kg )

[0246] The parameters are defined as follows:

[0247] P nn The original fault probability vector output by the deep neural network model (dimension n×1, where n is the number of fault types);

[0248] K kg : The corrected probability vector (dimension n×1) generated by the knowledge graph reasoning engine;

[0249] Γ conf The confidence calibration vector (dimension n×1, value [0,1]) provided by the knowledge graph is calculated from the historical rule matching degree and the entity association strength;

[0250] W context Context-aware weight matrix (n×n diagonal matrix), its diagonal elements w ii It is dynamically determined by the correlation between the current operating conditions and the failure modes;

[0251] ⊙: Hadamard product (element-by-element multiplication);

[0252] σ(·): Softmax normalization function, ensuring that the sum of the output probabilities is 1;

[0253] I: Identity matrix.

[0254] Additional condition for conflict resolution: When ||P nn -K kg When ||2>δ (δ is a preset threshold), the original data review process is triggered, and the diagnostic conclusion is frozen until the verification is completed.

[0255] Example: Scenario description:

[0256] The neural network detects partial discharge with the probability Pnn = [0.75, 0.60, 0.20]. T (Fault type: partial discharge, mechanical loosening, overheating), but the knowledge graph, based on the current high load condition (85% load rate) and historical data, determines that the probability of overheating is underestimated, and outputs a correction vector K. kg =[0.85,0.30,0.05] T Knowledge graph confidence level Γ conf =[0.9,0.6,0.3] T Context weight matrix W context =diag(0.7,0.5,0.2) (Current operating conditions are strongly correlated with discharge).

[0257] Calculation process:

[0258] 1. Confidence calibration:

[0259] P nn ⊙Γ conf = [0.75×0.9, 0.60×0.6, 0.20×0.3] T =[0.675,0.36,0.06] T

[0260] 2. Weighted fusion:

[0261] W context (P nn ⊙Γ conf = [0.7 × 0.675, 0.5 × 0.36, 0.2 × 0.06] T =[0.4725,0.18,0.012] T (IW context )K kg = [0.3×0.85, 0.5×0.30, 0.8×0.05] T =[0.255,0.15,0.04] T

[0262] 3. Normalized output:

[0263] P final =σ([0.4725+0.255,0.18+0.15,0.012+0.04] T )

[0264] =σ([0.7275,0.33,0.052]) T )

[0265] =[0.66,0.30,0.04] T

[0266] Diagnostic conclusion: Partial discharge probability 66% (confidence increased), mechanical loosening probability 30% (confidence decreased), overheating probability 4% (inhibited).

[0267] Triggering conflict detection (due to ||P) nn -K kg (||2=0.54>δ=0.5), the system automatically requests a verification of the original vibration waveform.

[0268] Technical effect

[0269] 1. Dynamic weight allocation:

[0270] W context The weighting of data-driven and knowledge-guided approaches is adjusted based on real-time operating conditions (such as load rate and temperature). For example, under high load, more reliance is placed on knowledge graphs (because historical failure patterns are clear), while under new operating conditions, neural networks are preferred (to adapt to unknown patterns).

[0271] This addresses the shortcomings of traditional static fusion methods, which cannot adapt to changes in operating conditions.

[0272] 2. Confidence calibration:

[0273] Γ conf γ is dynamically generated based on the historical accuracy of rules in the knowledge graph (e.g., if a fault rule has been correctly validated 9 out of the past 10 times, then γ...). i =0.9).

[0274] Suppressing false alarms in low-confidence neural networks (such as when high temperature causes false alarms in vibration sensors) i ↓).

[0275] 3. Conflict arbitration mechanism:

[0276] When the differences between the neural network and the knowledge graph are too large (||Δ||2>δ), data verification is automatically triggered to avoid the propagation of errors from a single model.

[0277] Improve the diagnostic reliability of complex faults (such as compound faults).

[0278] 4. Enhanced interpretability:

[0279] Each parameter (W) in the equation context ,Γ conf All of these are derived from the traceable rules of the knowledge graph, supporting the logic for generating the final conclusion.

[0280] All parts not described in this invention are the same as or can be implemented using existing technology. Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A comprehensive online monitoring system for transformers, characterized in that, include: Sensor layer: includes a multi-source heterogeneous sensor array, which is used to collect the operating parameters of the transformer body in real time; Edge computing layer: Composed of at least one edge computing node deployed at the transformer site; the edge computing node is connected to the sensor layer via a high-speed data bus. The edge computing node includes a multi-channel high-speed data acquisition module, a preprocessing module, a feature extraction module, a local lightweight diagnostic module, and an edge data management module; The preprocessing module performs filtering, noise reduction, outlier removal, and timestamp synchronization on the raw sensor data. The feature extraction module extracts time-domain features, frequency-domain features, time-frequency-domain features, and monitoring parameter features transmitted back by the multi-source heterogeneous sensor array from the preprocessed data. The local lightweight diagnostic module performs preliminary abnormal state identification and local alarms on the feature data extracted by the feature extraction module based on a preset rule base or a lightweight machine learning model. The edge data management module caches, compresses, and encrypts the data preprocessed by the preprocessing module, the feature data extracted by the feature extraction module, and the preliminary diagnostic results generated by the local lightweight diagnostic module, and selectively uploads the data according to a preset strategy or cloud platform instructions. Cloud platform layer: Deployed on a remote server cluster, used to receive and store data uploaded from one or more of the edge computing nodes; The cloud platform layer includes a data warehouse, a big data processing engine, a comprehensive diagnostic engine, a model training engine, and a system management module; The data warehouse adopts a distributed storage architecture and organizes and stores massive amounts of heterogeneous monitoring data according to device, time, and data type. The big data processing engine performs secondary cleaning, fusion, correlation analysis, and long-term trend mining on the incoming data. The integrated diagnostic engine is used to perform in-depth analysis of the fused multi-dimensional data to achieve transformer health status assessment, fault mode identification, fault severity classification and remaining life prediction. The model training engine uses historical data, online incremental data, and expert-annotated data to continuously train and optimize the model in the comprehensive diagnostic engine. The system management module provides user interaction interface, device management, alarm management, report generation and system configuration functions; Communication module: The edge computing layer communicates with the cloud platform layer through the communication module; The comprehensive diagnostic engine includes: Multi-source heterogeneous data fusion module: Based on adaptive weighting algorithm or deep learning fusion network, it is used to read data in data warehouse and perform multimodal feature fusion to generate unified state representation vector; Hybrid intelligent diagnostic module: integrates at least one deep neural network model and a knowledge graph-based inference engine; the deep neural network model receives the state representation vector and outputs a preliminary fault probability distribution; The knowledge graph-based reasoning engine combines the device's historical status, current operating conditions, environmental factors, and domain knowledge rules to verify, interpret, and correct the results output by the deep neural network model, generating a final diagnostic conclusion and confidence level. Dynamic Knowledge Graph Module: Constructs a knowledge graph based on transformer structure, materials, operating principles, typical fault modes, and evolution patterns, and dynamically updates the graph in real time by receiving diagnostic results and on-site feedback; Model Adaptive Update Module: Based on the differences between online diagnostic results and actual fault records, expert feedback, and the new model performance evaluation generated by the model training engine, it triggers parameter fine-tuning or structural optimization of the deep neural network model and updates the association rules and node attributes in the dynamic knowledge graph.

2. The transformer integrated online monitoring system according to claim 1, characterized in that, The multi-source heterogeneous sensor array includes at least: Vibration sensor array, ultrasonic sensor array, high-frequency partial discharge sensor, and fiber optic temperature sensor array installed on the surface of the transformer tank body; High-precision current sensors are installed on transformer bushings and neutral grounding wires; Dissolved gas analysis sensor, micro-water sensor, oil temperature sensor, and oil pressure sensor are installed in the transformer oil circuit.

3. The transformer integrated online monitoring system according to claim 1, characterized in that, The data upload strategy of the edge data management module of the edge computing node is as follows: When the transformer's operating status is determined to be "normal" by the local lightweight diagnostic module, only the compressed key feature vector and diagnostic summary are uploaded, and the sampling frequency is reduced. When a potential anomaly or threshold exceeding the limit is detected, the data upload frequency and granularity are automatically increased, and more detailed data, including the original waveform fragments, are uploaded. When a proactive data request or model update instruction is received from the cloud platform layer, data of a specified time period or type is uploaded as needed.

4. The transformer integrated online monitoring system according to claim 1, characterized in that, The communication module includes one of the following: a 4G communication module, a 5G communication module, an NB-IoT communication module, and a LoRa communication module.

5. The transformer integrated online monitoring system according to claim 1, characterized in that, The deep neural network model in the hybrid intelligent diagnostic module includes at least: A convolutional neural network model for partial discharge pattern recognition, with input being a partial discharge phase distribution spectrum or pulse sequence waveform, is used for partial discharge pattern recognition. A long short-term memory network model for mechanical fault diagnosis, with the input being a feature sequence of vibration signals, is used for mechanical fault identification; A Transformer-based multi-step prediction and fault classification model for dissolved gases in oil; A graph convolutional neural network model, whose input is graph structure data generated from the dynamic knowledge graph module, is used to fuse multi-source information for comprehensive state assessment.

6. The transformer integrated online monitoring system according to claim 1, characterized in that, The construction and updating process of the dynamic knowledge graph module includes: The initial map was constructed based on transformer design drawings, material properties, historical fault case database, international standards, and expert experience. Real-time diagnostic results are added to the graph as new event nodes and associated with relevant entities. Analyze the contradictions or supporting relationships between the diagnostic results and the existing knowledge in the atlas, and adjust the weights or confidence levels of the relationships between entities. When a new failure mode is identified and recurs, it is abstracted into a new graph node and association rule; Receive feedback from operations and maintenance personnel to correct errors or supplement details in the graph.

7. The transformer integrated online monitoring system according to claim 1, characterized in that, The workflow of the hybrid intelligent diagnostic module is as follows: The deep neural network model receives a state representation vector generated by the multi-source heterogeneous data fusion module and outputs a vector containing the probabilities of multiple preset fault modes. The knowledge graph-based inference engine receives the probability vector, current operating condition data, historical operating condition data, and environmental data. The reasoning engine traverses the entities and rules related to the current state in the knowledge graph and performs the following reasoning: Consistency check: Check whether the fault modes predicted by the deep neural network model are consistent with the common fault modes under the current operating conditions. If the predicted fault modes are highly similar to the common fault modes under the current operating conditions, the weight will be increased; otherwise, the weight will be decreased. Causal chain reasoning: Based on the fault development chain in the graph, infer whether the currently identified primary fault may lead to more serious secondary faults, and raise the warning level of related faults accordingly. Conflict resolution: When the prediction result of the deep neural network model conflicts with the strong rules or stored historical records in the dynamic knowledge graph module, the confidence of the prediction is reduced or a request for more in-depth analysis is triggered. Explanation generation: Based on the association paths and rules in the dynamic knowledge graph module, an interpretable natural language description is generated for the final diagnostic conclusion, explaining the diagnostic basis and reasoning process; By integrating the probability distribution predicted by the neural network and the correction opinions and confidence levels of knowledge graph reasoning, a final comprehensive diagnostic report is generated, which includes the fault type, possible location, severity level, confidence level, development trend prediction, and maintenance recommendations.

8. The transformer integrated online monitoring system according to any one of claims 1-7, characterized in that, The online monitoring process of the transformer integrated online monitoring system specifically includes the following steps: S1: Through the multi-source heterogeneous sensor array, the vibration, ultrasonic waves, high-frequency current, temperature signals of the transformer body and the dissolved gas, trace water, oil temperature, and oil pressure signals in the transformer oil are collected synchronously and in real time, and the bushing and neutral point current signals are also collected. S2: The edge computing node receives the raw sensor data stream, performs filtering, noise reduction, outlier removal, high-precision time synchronization, and feature extraction to obtain a multi-dimensional feature set; S3: The edge computing node uses a local lightweight diagnostic module to perform real-time analysis of multi-dimensional feature sets to achieve preliminary anomaly detection and local alarm. S4: The edge computing node compresses and encrypts the raw data, multi-dimensional feature set and diagnostic results according to a preset strategy or cloud platform instructions, and uploads them to the cloud platform layer through the communication module; S5: The cloud platform layer receives and stores the uploaded data, and uses the big data processing engine to perform secondary data cleaning, multi-source data fusion and long-term trend analysis. S6: The comprehensive diagnostic engine executes: S6.1: Multimodal features are fused through a multi-source heterogeneous data fusion module to generate a unified state representation vector; S6.2: The deep neural network model in the hybrid intelligent diagnostic module receives the state representation vector and outputs a preliminary fault probability distribution; S6.3: The knowledge graph-based reasoning engine in the hybrid intelligent diagnostic module combines the device's historical status, operating conditions, environmental factors, and domain knowledge in the dynamic knowledge graph to verify, reason about causes and resolve conflicts in the neural network output. S6.4: Integrate neural network predictions and knowledge reasoning results to generate a final comprehensive diagnostic conclusion, which includes fault type, location, severity, confidence level, and maintenance recommendations; S7: The model adaptive update module continuously monitors the diagnostic effect and triggers fine-tuning, optimization and dynamic knowledge graph updates of the deep learning model based on the difference between the online diagnostic results and the actual feedback and the performance evaluation of the new model. S8: The system management module pushes diagnostic results, alarm information, and status assessment reports to maintenance personnel through the user interface, and provides historical data query and device management functions.

9. The transformer integrated online monitoring system according to claim 8, characterized in that, The specific reasoning process of the knowledge graph-based reasoning engine described in step S6.3 includes: Context-aware reasoning: Based on the current transformer operating conditions, activate the fault modes and rule subgraphs most relevant to the context in the dynamic knowledge graph module, and adjust the prior probabilities predicted by the neural network. Strengthen or weaken the weight of the chain of evidence: Analyze the consistency between multi-source features. For example, if a vibration signal indicates looseness and an ultrasonic signal also captures abnormal acoustic emission, the "mechanical looseness" diagnosis is strengthened; if only a single sensor has an abnormal indication, its weight is reduced or a further verification request is triggered. Fault evolution prediction: Based on the fault development path described in the dynamic knowledge graph module, if the current diagnosis is a primary stage fault, the risk and time window of its development into a more serious stage will be automatically assessed, and the corresponding warning level will be raised. Root cause analysis: When a complex fault is diagnosed, the most likely root cause of the fault is inferred by using the causal dependencies in the dynamic knowledge graph module. Generate explanatory reports: The knowledge graph paths activated during the reasoning process, the rules used, the assessment of the strength of evidence, and the reasons for conflict resolution are automatically converted into highly readable text and attached to the diagnostic conclusion, thereby improving the credibility and operability of the results.

Images