Intelligent distribution transformer operation and maintenance method and system

Abstract

The invention relates to the technical field of intelligent operation and maintenance of power equipment, and discloses an intelligent distribution transformer operation and maintenance method and system, and the method comprises the steps: fusing the four-dimensional parameters of electricity, state, nameplate and environment; a dynamic attention mechanism anchored by nameplate parameters is adopted to realize adaptive fusion of key features; abnormal detection is based on a transfer learning model, and cross-device knowledge transfer is promoted; residual life prediction is completed through a dynamic method of reference life and personalized aging coefficient correction; a multi-dimensional health assessment system is constructed, quantitative health scores are generated, and a multi-stage intelligent early warning mechanism is established according to the quantitative health scores and the residual life; meanwhile, an intelligent man-machine interaction system based on a BERT model is integrated, and natural language and voice questions and answers are supported; according to the invention, omnibearing intelligent monitoring, refined state evaluation, scientific life prediction and auxiliary decision making of the distribution transformer can be realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for power equipment, and more specifically, to an intelligent operation and maintenance method and system for distribution transformers. Background Technology

[0002] As a key piece of equipment in the power system for power conversion and distribution, the reliable operation of distribution transformers directly affects power supply quality and operation and maintenance efficiency. Traditional operation and maintenance relies on periodic manual inspections and single-parameter monitoring, which generally suffers from problems such as narrow data dimensions, slow anomaly detection, and lack of health assessment and lifespan prediction. Existing online monitoring technologies mainly collect basic electrical parameters through sensors and alarm based on fixed thresholds, failing to comprehensively perceive oil quality, insulation, and environmental conditions, and also failing to effectively integrate the inherent information on the equipment nameplate with dynamic operating data, resulting in insufficient data value mining.

[0003] While some technologies incorporate machine learning or digital twins for status analysis, their anomaly detection often relies on fixed thresholds or static models, failing to dynamically adjust the judgment logic based on specific equipment nameplate parameters. This results in poor adaptability to different equipment operating conditions, particularly on new equipment lacking historical data. Furthermore, existing solutions generally lack a comprehensive, quantitative health status assessment system, making it difficult to accurately reflect the true condition of the equipment. In terms of remaining lifespan prediction, a few methods employ statistically based general models, whose aging factor weights fail to adequately consider individual differences such as equipment model and operating environment, leading to limited prediction accuracy and generalization ability.

[0004] Existing alarm mechanisms typically only trigger when limits are exceeded, without distinguishing the severity of the anomaly or linking it to the overall health status and remaining lifespan of the equipment. This makes it difficult for maintenance personnel to quickly determine the priority of handling. Furthermore, information acquisition still relies on professionals consulting documents or operating backend systems, which is inefficient and lacks intelligent decision-making support tools that support natural language interaction. Summary of the Invention

[0005] This invention provides an intelligent operation and maintenance method and system for distribution transformers, which solves the technical problems of lack of adaptability in anomaly detection, low accuracy of remaining life prediction, imperfect alarm mechanism, and low efficiency of human-computer interaction in related technologies.

[0006] This invention provides a method for the operation and maintenance of intelligent distribution transformers, comprising the following steps:

[0007] S1, based on the transformer itself and its operating environment, adopts a four-dimensional acquisition architecture to obtain a complete dataset in the cloud;

[0008] S2, based on the complete dataset in the cloud, uses a dynamic attention mechanism anchored by nameplate parameters to fuse four-dimensional features and obtain the device fused feature vector;

[0009] S3. Based on the device's fused feature vector, a transfer learning anomaly detection model anchored by the nameplate parameters is used to map the feature space to obtain personalized anomaly detection results.

[0010] S4. Based on the equipment fusion feature vector and personalized anomaly detection results, a personalized weighted adaptive life prediction model driven by nameplate parameters is adopted to obtain the remaining life value and life decay prediction curve.

[0011] S5, based on the device's fused feature vector, personalized anomaly detection results, and remaining lifespan value, uses multi-dimensional index quantification scoring and correction factor calculation to obtain a comprehensive health score and structured health report;

[0012] S6, based on the comprehensive health score and remaining life expectancy, adopts a multi-level early warning mechanism and differentiated push strategy to obtain graded alarm information and auxiliary decision-making information.

[0013] In a preferred embodiment, S1 includes:

[0014] Deploy electrical parameter units, status parameter units, nameplate parameter units, and environmental parameter units to obtain a four-dimensional time-series dataset;

[0015] A parameter calibration module is integrated inside the transformer body. The calibration algorithm unit automatically executes the calibration process periodically and uses the least squares method to perform linear fitting on the deviation data to obtain a four-dimensional time series dataset after quality calibration.

[0016] The edge gateway uses a time synchronization algorithm for data alignment, performs linear interpolation to complete the low-frequency sampled data, and obtains a time-aligned four-dimensional time-series dataset; it then performs data preprocessing to standardize the format, clean the data, and compress it.

[0017] The edge gateway is configured with dual-link redundant transmission, and the data transmission process uses encryption algorithms to encrypt data packets;

[0018] Cloud data centers implement tiered storage strategies based on data characteristics and access frequency to obtain complete datasets in the cloud.

[0019] In a preferred embodiment, S2 includes:

[0020] Extract the four-dimensional parameter data and nameplate static parameters of the target transformer from the cloud database, and use the nameplate parameters as the feature anchor points of the target transformer to establish the parameter reference coordinate system of the individual equipment.

[0021] For each operating parameter, calculate its normalized deviation from the nameplate rating;

[0022] A dynamic attention weight calculation model based on deviation sensitivity is constructed. A deviation sensitivity factor is introduced. The attention weight is equal to the absolute value of the normalized deviation multiplied by the deviation sensitivity factor, then amplified by an exponential function and normalized by softmax.

[0023] Based on dynamic attention weights, the four-dimensional parameters are weighted and fused at multiple levels, including parameter-level fusion, category-level fusion and device-level fusion, to obtain the device fused feature vector.

[0024] In a preferred embodiment, S3 includes:

[0025] Using historical operating data of transformers of the same type, the operating data of each device is standardized and mapped through its nameplate parameters. The standardized feature is equal to the device fused feature vector divided by the nameplate parameter normalized vector. The isolated forest anomaly detection model is trained in the standardized feature space to obtain a general anomaly detection basic model.

[0026] For the target transformer, perform inverse feature space mapping based on nameplate parameters, multiply the segmentation threshold of the isolation tree in the model by the corresponding component of the normalized vector of the target equipment nameplate, and obtain a personalized anomaly detection model.

[0027] For the device fusion feature vectors collected in real time, dual-space collaborative anomaly detection is performed. The comprehensive anomaly judgment adopts dual-space scoring weighted fusion to obtain personalized anomaly detection results.

[0028] In a preferred embodiment, S4 includes:

[0029] Nameplate parameter information is extracted from the equipment fusion feature vector and personalized anomaly detection results. The nameplate parameters are transformed into multi-dimensional equipment feature labels. Similar equipment groups are matched based on the equipment feature labels. The Bayesian estimation method is used to fuse the design life and actual life statistics to calculate the baseline life.

[0030] A weighted matching model for aging factors based on equipment feature labels is constructed. The random forest algorithm is used to train the weighted model for each group of equipment. A mapping table from equipment feature labels to weight vectors is established. The weights are fine-tuned according to the historical operating data characteristics of the target equipment to obtain personalized aging factor weight coefficients.

[0031] Historical operating data and real-time status parameters of the target transformer are extracted from the cloud database. The quantitative values ​​of each aging factor are calculated and normalized. The comprehensive aging coefficient and remaining life value are calculated.

[0032] Based on the current aging rate and trend extrapolation, a lifespan decline prediction curve is established.

[0033] In a preferred embodiment, S5 includes:

[0034] Establish a health assessment index system covering primary indicators and multiple secondary indicators, including electrical performance, thermal performance, insulation performance, and mechanical performance. Introduce the anomaly detection score and remaining life value from personalized anomaly detection results as correction factors for health assessment.

[0035] Each secondary indicator is quantitatively scored, the primary indicator score is calculated according to the preset weight coefficient, the preliminary comprehensive health score is calculated, and the score is corrected by introducing the anomaly detection correction factor and the remaining lifespan correction factor. The final comprehensive health score is equal to the preliminary comprehensive health score multiplied by the anomaly correction factor and the lifespan correction factor.

[0036] A structured health report is generated based on the health score and scores of each indicator.

[0037] In a preferred embodiment, S6 includes:

[0038] Based on the dual indicators of health score and remaining life expectancy, a three-level alarm threshold is set, including a first-level early warning threshold, a second-level alarm threshold, and a third-level emergency alarm threshold. The system sequentially judges whether the health score and remaining life expectancy meet the threshold conditions of each level and triggers the corresponding alarm level.

[0039] Differentiated alarm information content is generated based on the alarm level. The alarm information includes alarm level, device number, anomaly type, associated parameters, health score, remaining life estimate, and preliminary handling suggestions.

[0040] Different alarm information push strategies are implemented based on alarm levels, with different push methods and push targets for different levels;

[0041] The alarm information pushed by the Level 3 emergency alarm will be accompanied by auxiliary decision-making information, including equipment physical location information, spare parts inventory information, remaining life assessment report, and historical fault records.

[0042] In a preferred embodiment, performing dual-space collaborative anomaly detection includes:

[0043] The first spatial detection is a standardized spatial detection, which projects the device fusion feature vector onto the standardized feature space through the standardized mapping of nameplate parameters, and inputs it into the general anomaly detection model to calculate the anomaly score;

[0044] The second spatial detection is individual spatial detection, which inputs the device's fused feature vector into a personalized model to calculate anomaly scores;

[0045] The comprehensive anomaly determination adopts a weighted fusion of dual-space scoring. If the comprehensive anomaly score exceeds the preset anomaly threshold, it is determined to be an anomaly.

[0046] The anomaly type identification method involves tracing back the original parameter data of the feature vector fusion device, analyzing the deviation of each parameter and the distribution of attention weights, and identifying the anomaly type corresponding to the parameter with the largest product of deviation and attention weights as the detection result.

[0047] In a preferred embodiment, a method for operating and maintaining an intelligent distribution transformer further includes:

[0048] The system acquires user natural language or voice query requests, uses the BERT model and a structured knowledge base for semantic understanding and knowledge matching, and obtains accurate query results and visualizations; specifically including:

[0049] By integrating multi-source information such as transformer operation and maintenance manuals, equipment nameplate parameters, historical operating data, fault case databases, and life assessment standards, a structured knowledge base is constructed. Named entity recognition technology is used to extract entities from text, relation extraction technology is used to identify the relationships between entities, and knowledge graph construction technology is used to organize entities and relations into a graph structure.

[0050] The BERT pre-trained model is used for semantic understanding of user queries. The BERT model performs semantic encoding on the query text and performs query intent recognition and key entity extraction.

[0051] Based on the query intent and key entities, the system retrieves the most relevant knowledge items from the knowledge base. The retrieval method uses semantic similarity calculation. The system retrieves device parameters, health reports, and life prediction results from the cloud database in real time and populates the answer template with real-time data to generate an accurate answer.

[0052] While generating text responses, the system also displays associated visual content and supports the export of health reports, life expectancy prediction reports, and alarm records.

[0053] In a preferred embodiment, an intelligent distribution transformer operation and maintenance system is used to execute the above-described intelligent distribution transformer operation and maintenance method, including:

[0054] The data acquisition module is used to obtain a complete dataset in the cloud based on the transformer body and its operating environment using a four-dimensional acquisition architecture.

[0055] The feature fusion module is used to fuse four-dimensional features based on the complete dataset in the cloud using a dynamic attention mechanism anchored by the nameplate parameters, to obtain the device fused feature vector;

[0056] The anomaly detection module is used to obtain personalized anomaly detection results by mapping the anomaly detection model anchored by the nameplate parameters to the feature space based on the device's fused feature vector.

[0057] The life prediction module is used to obtain the remaining life value and life decay prediction curve by using a personalized weighted adaptive life prediction model driven by nameplate parameters, based on the equipment's fused feature vector and personalized anomaly detection results.

[0058] The health assessment module is used to obtain a comprehensive health score and a structured health report by using multi-dimensional index quantification scoring and correction factor calculation based on the device's fused feature vector, personalized anomaly detection results and remaining life value.

[0059] The alarm push module is used to obtain graded alarm information and auxiliary decision-making information based on the comprehensive health score and remaining life value, using a multi-level early warning mechanism and differentiated push strategy.

[0060] The beneficial effects of this invention are as follows:

[0061] By constructing a four-dimensional parameter acquisition architecture and a cloud-edge collaborative transmission mechanism, unified acquisition and fusion of electrical, status, nameplate, and environmental data have been achieved. The deep integration of static nameplate information and dynamic operation data breaks through the limitations of traditional monitoring in terms of data dimensions, providing a more comprehensive and accurate data foundation for subsequent intelligent analysis, thereby improving the completeness and accuracy of equipment status perception.

[0062] An anomaly detection model based on nameplate parameter anchoring is adopted. This model can adaptively adjust the detection threshold and judgment logic according to the inherent characteristics of different equipment, overcoming the shortcomings of traditional fixed threshold methods such as poor adaptability and high false alarm and false negative rates. At the same time, through the incremental learning mechanism, the model can quickly adapt to newly commissioned equipment, effectively improving the adaptability of anomaly detection to different working conditions and the overall reliability. Attached Figure Description

[0063] Figure 1 This is a flowchart of an intelligent distribution transformer operation and maintenance method according to the present invention;

[0064] Figure 2 This is a block diagram of an intelligent power distribution transformer operation and maintenance system according to the present invention;

[0065] Figure 3 This is a flowchart of an intelligent distribution transformer operation and maintenance method according to the present invention. Detailed Implementation

[0066] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0067] At least one embodiment of the present invention discloses an intelligent distribution transformer operation and maintenance method, such as... Figures 1 to 3 As shown, it includes the following steps:

[0068] S1, based on the transformer itself and its operating environment, adopts a four-dimensional acquisition architecture to obtain a complete dataset in the cloud;

[0069] Based on the transformer itself and its operating environment, a multi-parameter sensing module is deployed for comprehensive data acquisition. A four-dimensional acquisition architecture, encompassing electrical parameters, status parameters, nameplate parameters, and environmental parameters, is adopted to obtain a complete data profile of the equipment's operation. After processing via a self-calibration mechanism and time synchronization algorithm, the data is transmitted to the cloud through a cloud-edge collaborative architecture and managed in a hierarchical manner. Specifically, the process includes the following steps:

[0070] S11. Based on the transformer body and its operating environment, a four-dimensional acquisition architecture is used to deploy a multi-parameter sensing module to obtain a four-dimensional time series dataset.

[0071] An electrical parameter unit is deployed to collect operational electrical quantities. Three-phase current transformers and three-phase voltage transformers are installed on the high-voltage and low-voltage sides of the transformer, respectively. The current transformers are selected with a 0.2-class accuracy, covering a measurement range of 0 to 200% of the rated current. The voltage transformers are also selected with a 0.2-class accuracy, covering a measurement range of 80% to 120% of the rated voltage. A power factor control device is installed to monitor changes in the transformer's power factor in real time. For large-capacity oil-immersed transformers, a gas relay is installed on the top of the tank to monitor gas signals generated by internal transformer faults. The electrical parameter unit continuously collects operational electrical quantities such as instantaneous three-phase current values, instantaneous three-phase voltage values, real-time power factor values, and gas signal status at a preset sampling period. All collected data is appended with a unified timestamp and transmitted to the edge gateway via the Modbus-TCP protocol.

[0072] The status parameter unit is deployed to collect device status data. For oil-immersed transformers, a platinum resistance oil temperature sensor is installed in the oil temperature sensor mounting slot at the top of the tank. The measurement range is -20°C to 150°C, with an accuracy of ±0.5°C, and oil temperature data is collected at a sampling period of 1 minute. An oil quality sensor is installed on the side wall of the tank to monitor the dielectric loss factor and moisture content of the transformer oil in real time, collecting oil quality data at a sampling period of 10 minutes. An insulation aging sensor is installed at the winding lead-out end to assess the degree of insulation aging by measuring insulation resistance and dielectric loss, collecting insulation status data at a sampling period of 1 hour. A piezoelectric noise sensor is installed at a noise acquisition hole near the core on the side wall of the transformer body. The measurement range is 30 to 120 dB, with a sampling frequency of 100 Hz, continuously collecting noise signals during transformer operation and capturing aging abnormal noise characteristics. All data collected by the status parameter unit is appended with a unified timestamp and transmitted to the edge gateway via the industrial bus.

[0073] The system deploys a nameplate parameter acquisition unit to collect manufacturing parameters. An RFID reader and a high-definition barcode scanner are integrated into the nameplate parameter acquisition interface pre-installed on the transformer body sidewall. The RFID reader supports the ISO15693 protocol with a reading distance of 10 cm, while the high-definition barcode scanner achieves a recognition accuracy of 0.1 mm. Upon system startup, the system automatically executes the nameplate parameter acquisition process. The RFID reader attempts to read the electronic tag information on the nameplate. If successful, it acquires core manufacturing parameters such as rated capacity, rated voltage, rated current, temperature rise limit, short-circuit impedance, no-load loss, load loss, transformer ratio, tap changer voltage parameters, and design life. If the RFID reading fails, the high-definition barcode scanner is activated to acquire an image of the nameplate and extract the text information using optical character recognition technology. If all automatic acquisition fails, the system queries the cloud-based nameplate parameter database using the device number to retrieve historical data. If the database has no record, a manual data entry process is triggered. Maintenance personnel manually input the nameplate parameters via a mobile application. The system records the data entry time and operator information to ensure the integrity of the nameplate data. Once the nameplate parameters are collected, they are permanently stored in the local storage and then uploaded to the cloud database for backup.

[0074] An environmental parameter unit is deployed to collect environmental impact data. Temperature and humidity sensors are installed inside a protective enclosure on top of the transformer, measuring temperatures from -40°C to 85°C and humidity from 0 to 100% relative humidity. Temperature and humidity data for the transformer's operating environment are collected at 10-minute sampling intervals. A pollution level sensor is installed outside the protective enclosure to assess the environmental pollution level by measuring the leakage current on the insulator surface, collecting pollution level data at 1-hour sampling intervals. All data collected by the environmental parameter unit is appended with a unified timestamp and transmitted to the edge gateway via an industrial bus.

[0075] S12, Based on the four-dimensional time series dataset, the parameter calibration module and calibration algorithm are used to obtain the quality-calibrated four-dimensional time series dataset;

[0076] A parameter calibration module is integrated within the transformer body. This module includes a high-precision reference sensor and a calibration algorithm unit. The reference sensor uses high-precision equipment certified by national metrological standards, with a measurement accuracy one order of magnitude higher than that of the working sensor. For example, the reference temperature sensor has an accuracy of ±0.05 degrees Celsius, and the reference current sensor has an accuracy of 0.05%. The calibration algorithm unit automatically executes the calibration process quarterly: during the transformer's stable operation, the reference sensor and the working sensor simultaneously measure the same physical quantity, collecting measurement data for 30 minutes and calculating the deviation coefficient between them. The least squares method is used to linearly fit the deviation data to obtain the calibration curve equation, which describes the mapping relationship between the working sensor's measured value and the true value. The calibration algorithm unit stores the calibration curve equation parameters in local memory. During subsequent data acquisition, it performs real-time correction on the working sensor output. The correction formula is: the true value equals the measured value multiplied by the slope coefficient plus the intercept coefficient, ensuring the accuracy of the collected data. If the deviation coefficient of a working sensor exceeds a preset threshold, the system triggers a sensor anomaly alarm, prompting maintenance personnel to replace the sensor.

[0077] S13. Based on the quality-calibrated four-dimensional time-series dataset, a time synchronization algorithm is used to perform data alignment processing to obtain a time-aligned four-dimensional time-series dataset.

[0078] The edge gateway receives heterogeneous data from the electrical parameter unit, status parameter unit, and environmental parameter unit. Due to the inconsistent sampling frequencies of different sensors (100 milliseconds for electrical parameters, 1 minute for oil temperature, 10 minutes for oil quality, and 1 hour for insulation status), the data is misaligned in the time dimension. The edge gateway uses a time synchronization algorithm for data alignment: using the 100-millisecond sampling period of the electrical parameters as the base time scale, linear interpolation is performed on the low-frequency sampled data. Specifically, for oil temperature data, linear interpolation is performed between two adjacent sampling points at 100-millisecond intervals. The interpolation formula is: the current oil temperature value equals the oil temperature value at the previous sampling point plus the time difference divided by the sampling period multiplied by the oil temperature change, generating an oil temperature time series aligned with the electrical parameters. The same interpolation method is used for lower-frequency sampled data such as oil quality and insulation status. The time synchronization algorithm ensures that the four-dimensional parameter data has a corresponding value at each 100-millisecond time scale, forming a time-aligned four-dimensional time series dataset.

[0079] S14. Based on the time-aligned four-dimensional time-series dataset, a data preprocessing workflow is used to perform format standardization, data cleaning and compression to obtain the preprocessed four-dimensional time-series dataset.

[0080] The edge device employs an industrial-grade gateway equipped with a quad-core processor, supporting Modbus-TCP and DL / T645 communication protocols. After receiving the four-dimensional time-series dataset from the multi-parameter sensing module, the gateway performs a data preprocessing procedure: Format standardization is performed, converting the proprietary data formats used by different sensors into a unified JSON format. Each data record includes fields such as device number, timestamp, parameter type, parameter value, unit, and quality identifier. Next, data cleaning is performed, using a 3-standard-deviation criterion to identify outliers. The mean and standard deviation of each parameter within the sliding time window are calculated; if a data point deviates from the mean by more than 3 standard deviations, it is considered an outlier and removed. Simultaneously, a hash algorithm is used to identify and delete duplicate data records. Finally, data compression is performed using the LZ77 lossless compression algorithm to compress the JSON data, achieving a compression rate of 60% to 70%, reducing data transmission volume. The preprocessed data is temporarily stored in the gateway's local storage, awaiting uploading to the cloud.

[0081] S15. Based on the preprocessed four-dimensional time series dataset, dual-link redundant transmission and AES-256 encryption are used to obtain the encrypted four-dimensional time series dataset uploaded to the cloud.

[0082] The edge gateway is configured with a 5G SA network module and a gigabit fiber optic interface to achieve dual-link redundant transmission. The gateway prioritizes the fiber optic link for data transmission; when the fiber optic link fails or bandwidth is insufficient, it automatically switches to the 5G link to ensure uninterrupted data transmission. Data packets are encrypted using the AES-256 encryption algorithm during data transmission. The encryption key is periodically updated by the cloud key management center and distributed to the edge gateway through a secure channel. The gateway then uploads the encrypted data packets to the cloud data center via HTTPS protocol. The cloud data center is deployed on cloud servers, and the server configuration meets the data processing and storage requirements.

[0083] S16. Based on the encrypted four-dimensional time-series dataset uploaded to the cloud, a hierarchical storage strategy is adopted for data storage management to obtain a complete dataset stored in the cloud hierarchical storage.

[0084] The cloud data center uses a Hadoop distributed database for data storage management, implementing a tiered storage strategy based on data characteristics and access frequency. Real-time data storage layer: The most recent 7 days of four-dimensional time-series data are stored in a Redis high-speed cache database. Redis uses in-memory storage, with a data read latency of less than 1 millisecond, allowing maintenance personnel to quickly query the current operating status and recent parameter trends of the equipment. Historical operating data storage layer: Historical electrical parameters, status parameters, and environmental parameters from 7 days ago to 5 years ago are stored in the HDFS distributed file system. HDFS uses a three-replica redundant storage strategy to ensure data reliability. This layer of data is mainly used for AI model training, trend analysis, and fault backtracking applications. Nameplate static data storage layer: All transformer nameplate parameters are permanently stored in a MySQL relational database. The MySQL database uses a master-slave replication architecture, with the master database responsible for data writing and the slave database responsible for data reading, achieving read-write separation to improve performance. Nameplate data serves as equipment identification and analysis benchmarks and needs to be permanently stored for equipment lifecycle traceability. Environmental historical data storage layer: Historical environmental parameters from the past 3 years are stored in the HDFS distributed file system. This layer of data is mainly used for environmental impact factor analysis and lifespan prediction model training.

[0085] S17. Based on the complete dataset stored in the cloud at different levels, a role-based access control and multi-level permission management mechanism is adopted to obtain the complete dataset in the cloud.

[0086] The cloud database employs a multi-level access control mechanism, using a role-based access control model. Four user roles are defined: system administrators have the highest privileges, allowing them to perform database configuration, user management, data backup and recovery, and other operations; data analysts have data reading and analysis permissions, accessing all historical data for model training and algorithm optimization; maintenance personnel have device data query and alarm handling permissions, viewing real-time data, health reports, and alarm information for devices within their assigned area; and ordinary users have basic query permissions, only able to view basic operating parameters of authorized devices. Users must authenticate with username and password, along with two-factor authentication, upon logging into the system. After successful authentication, the system assigns appropriate permissions based on the user's role. All data access operations are recorded in the audit log, which includes information such as operation time, user, operation type, and accessed data range, for security auditing and traceability. The cloud database automatically performs full data backups periodically, with backup data stored in an off-site disaster recovery center to ensure data security and recoverability.

[0087] Based on the above sub-steps, reliable transmission and hierarchical storage management of the four-dimensional time series dataset from the edge to the cloud are achieved, resulting in a complete dataset in the cloud, which provides data support for subsequent AI intelligent analysis.

[0088] S2, based on the complete dataset in the cloud, uses a dynamic attention mechanism anchored by nameplate parameters to fuse four-dimensional features and obtain the device fused feature vector;

[0089] Based on a complete cloud-based dataset, and using nameplate parameters as feature anchors for individual devices, a dynamic attention mechanism is constructed to fuse multi-dimensional features, enabling personalized feature representation across different devices. Specifically, this includes the following steps:

[0090] S21. Based on the four-dimensional parameter data and nameplate static parameters in the complete cloud dataset, the nameplate parameters are used as feature anchors to establish a parameter reference coordinate system and obtain the normalized deviation.

[0091] The four-dimensional parameter data and nameplate static parameters of the target transformer are extracted from a cloud database. Taking a 1000 kVA 10 kV oil-immersed distribution transformer as an example, the extracted nameplate parameters include rated capacity of 1000 kVA, rated voltage of 10 kV, rated current of 144.3 Amperes, rated oil temperature of 65 degrees Celsius, temperature rise limit of 55 Kelvin, and design life of 20 years. The nameplate parameters are used as feature anchors for the equipment to establish a parameter reference coordinate system for each individual device. For each operating parameter, its normalized deviation relative to the nameplate rating is calculated. The formula is: normalized deviation equals the difference between the measured value and the rated value, divided by the rated value. This deviation reflects the degree of deviation of the current operating state of the equipment from its design state. Unlike traditional methods that use a uniform normalization interval, this method establishes a dedicated normalization benchmark for the nameplate parameters of each device, resulting in different deviations for the same measured value on different rated devices, thus achieving accurate characterization of individual device differences.

[0092] S22, Based on the normalized deviation, the attention weights of each parameter are obtained by using the deviation sensitivity factor and dynamic attention weight calculation model.

[0093] A dynamic attention weight calculation model based on deviation sensitivity is constructed. This model introduces a deviation sensitivity factor, which is dynamically adjusted according to parameter type and equipment operating stage. For electrical parameters, the default value of the deviation sensitivity factor is set to 1.2, indicating that deviations in electrical parameters have a significant impact on equipment health. For oil temperature and insulation parameters among the status parameters, the default value of the deviation sensitivity factor is set to 1.5, indicating that deviations in these parameters have a significant impact on equipment lifespan. For environmental parameters, the default value of the deviation sensitivity factor is set to 0.8, indicating that deviations in environmental parameters have a relatively small impact. The attention weight calculation formula is: attention weight equals the absolute value of the normalized deviation multiplied by the deviation sensitivity factor, then amplified by an exponential function and normalized by softmax. This calculation method allows parameters with large deviations and high sensitivity to receive higher attention weights, achieving automatic highlighting of key abnormal features. For example, when the oil temperature deviates from the rated value by 10%, its attention weight can reach 0.25, while the weight is only 0.08 when the current deviates by 5%, reflecting the difference in the impact of different parameters on equipment health.

[0094] S23. Based on the attention weights of each parameter and the four-dimensional parameter data, a multi-level weighted fusion method is used to obtain the device fusion feature vector;

[0095] A multi-level weighted fusion of four-dimensional parameters is performed based on dynamic attention weights. The first layer is parameter-level fusion, where time-series parameters are multiplied by their corresponding attention weights to obtain a weighted parameter sequence. The second layer is category-level fusion, where the weighted parameter sequence is grouped and weighted according to three categories: electrical, state, and environment. The weight coefficients are dynamically determined based on the equipment capacity level in the nameplate parameters, resulting in three category feature components. The third layer is equipment-level fusion, where the three category feature components are concatenated with the nameplate parameter encoding vector, which is formed by normalizing key nameplate parameters such as rated capacity, rated voltage, rated current, and design life. The final result is an equipment fusion feature vector containing both category feature components and nameplate encoding components. This fusion feature vector contains both the current operating status information of the equipment and embeds individual design characteristic information, achieving a deep fusion of operating data and design parameters, laying the foundation for subsequent personalized analysis based on transfer learning.

[0096] Based on the above sub-steps, the feature fusion of the dynamic attention mechanism based on nameplate parameter anchoring is completed, and the device fusion feature vector is obtained. The device fusion feature vector reflects the individual differences and operating status characteristics of the device.

[0097] S3. Based on the device's fused feature vector, a transfer learning anomaly detection model anchored by the nameplate parameters is used to map the feature space to obtain personalized anomaly detection results.

[0098] Based on device fusion feature vectors, and using nameplate parameters as anchor points for feature space mapping, an anomaly detection model for cross-device knowledge transfer is constructed to achieve accurate transfer from group knowledge to individual applications; specifically, the following steps are included:

[0099] S31. Based on the historical operating data and nameplate parameters of transformers of the same type, a general anomaly detection basic model is obtained by using nameplate parameter standardization mapping and the isolated forest algorithm.

[0100] By leveraging a large amount of historical operating data on similar transformers aggregated in a cloud data center, a universal feature space based on standardized nameplate parameters is constructed. Taking oil-immersed distribution transformers with a voltage level of 10 kV and a capacity range of 800 kVA to 1250 kVA as the same type of equipment, historical operating data for the past 5 years for over 500 such devices were extracted from the cloud database. The aim is to avoid directly using the raw operating data and instead standardize and map the operating data of each device through its nameplate parameters. The standardization mapping formula is: Standardized feature equals the device's fused feature vector divided by the nameplate parameter normalized vector, where the nameplate parameter normalized vector is composed of ratios such as rated capacity divided by the average capacity of the type, rated current divided by the average current of the type, and rated oil temperature divided by the average oil temperature of the type. This mapping projects the operating characteristics of equipment with different ratings onto a unified standardized feature space, making the operating characteristics of 800 kVA and 1250 kVA rated capacity equipment comparable. An isolated forest anomaly detection model was trained in a standardized feature space, with model parameters set to 100 isolated trees, a subsampling size of 256, and a maximum tree depth of 10 layers. After training, a general anomaly detection base model was obtained. This model learns anomaly patterns in the standardized feature space and has cross-device generalization capabilities.

[0101] S32. Based on the general anomaly detection model and the nameplate parameters of the target transformer, a personalized anomaly detection model is obtained by using inverse feature space mapping and incremental learning methods.

[0102] For the target transformer, an inverse feature space mapping based on nameplate parameters is performed to achieve knowledge transfer from the general model to the individual equipment. The nameplate parameters of the target transformer are extracted, including individual rated values ​​such as rated capacity of 1000 kVA, rated current of 144.3 A, and rated oil temperature of 65 degrees Celsius. The normalized vector of the target equipment's nameplate parameters is calculated using the formula: the normalized vector of the target equipment's nameplate parameter is equal to the ratio of the target equipment's rated capacity divided by the average capacity of the type, the target equipment's rated current divided by the average current of the type, and the target equipment's rated oil temperature divided by the average oil temperature of the type. An inverse mapping is performed on the general anomaly detection model, multiplying the segmentation threshold of all isolation trees in the model by the corresponding component of the target equipment's nameplate normalized vector, completing the mapping of the model from the standardized feature space to the individual feature space of the target equipment. This inverse mapping allows the general model to automatically adapt to the rated value characteristics of the target equipment. For example, for a 1000 kVA rated capacity device, its capacity-related anomaly judgment threshold is automatically adjusted to the ratio of the general threshold multiplied by 1000 divided by the average capacity of the type. If the target device has been running for a period of time and has accumulated historical data, extract the normal operation samples of the device in the past 3 months, input the device fusion feature vector of these samples into the model after inverse mapping for fine-tuning. The fine-tuning process adopts the incremental learning method to dynamically adjust the fine-tuning coefficient of the segmentation threshold so that the model can be further adapted to the individual operating characteristics of the device.

[0103] S33. Based on the device fusion feature vector and personalized anomaly detection model, a dual-space collaborative anomaly detection and anomaly type identification method is adopted to obtain personalized anomaly detection results.

[0104] For the real-time acquired device fusion feature vectors, dual-space collaborative anomaly detection is performed. The first spatial detection is a standardized spatial detection, where the device fusion feature vectors are projected onto a standardized feature space through a standardized mapping of nameplate parameters. This is then input into a general anomaly detection model to calculate an anomaly score, reflecting the degree of anomaly of the device relative to a group of similar devices. The second spatial detection is an individual spatial detection, where the device fusion feature vectors are directly input into a personalized model after inverse mapping to calculate an anomaly score. This score reflects the degree of anomaly of the device relative to its own historical operating state. Comprehensive anomaly determination uses a weighted fusion of dual-space scores. The formula is: comprehensive anomaly score equals standardized spatial score multiplied by 0.4 plus individual spatial score multiplied by 0.6. The weighting reflects the principle of prioritizing individual operating characteristics. If the comprehensive anomaly score exceeds a threshold, it is determined to be an anomaly. The anomaly type identification method involves backtracking the original parameter data of the device fusion feature vectors, analyzing the deviation of each parameter and the distribution of attention weights. The anomaly type corresponding to the parameter with the largest product of deviation and attention weight is the detection result. This method comprehensively considers both the degree of parameter deviation and the importance of the parameters. The output includes anomaly detection results, including anomaly type, comprehensive anomaly score, standardized spatial score, individual spatial score, correlation parameters and their deviation, and occurrence time.

[0105] Based on the above sub-steps, transfer learning anomaly detection based on nameplate parameter anchoring is completed, and personalized anomaly detection results are obtained. This method realizes the organic integration of group knowledge and individual characteristics through nameplate parameters, thereby improving the accuracy and adaptability of anomaly detection.

[0106] S4. Based on the equipment fusion feature vector and personalized anomaly detection results, a personalized weighted adaptive life prediction model driven by nameplate parameters is adopted to obtain the remaining life value and life decay prediction curve.

[0107] Based on the fused feature vectors of equipment and personalized anomaly detection results, a personalized weighted adaptive lifespan prediction model is constructed using a nameplate parameter-driven equipment feature label matching strategy. This model enables accurate identification and lifespan prediction of different equipment aging modes. Specifically, the following steps are included:

[0108] S41, Based on the equipment fusion feature vector and personalized anomaly detection results, nameplate parameter extraction and equipment feature label matching are used to obtain equipment feature labels and equipment baseline lifespan;

[0109] Nameplate parameter information was extracted from the equipment's fused feature vector and personalized anomaly detection results to construct an equipment classification system. The nameplate parameters of this 1000 kVA oil-immersed distribution transformer include a design life of 20 years, model S11-1000 kVA, fully sealed manufacturing process, Class A insulation material, and No. 25 mineral oil transformer oil. These nameplate parameters were converted into multi-dimensional equipment feature tags, including capacity level tags, voltage level tags, manufacturing process tags, insulation material tags, and cooling method tags. The capacity level tags are categorized by rated capacity as small equipment (less than 500 kVA), medium equipment (500-1500 kVA), and large equipment (greater than 1500 kVA); this equipment is tagged as medium-sized. The manufacturing process tags are categorized by structural type as open, semi-sealed, and fully sealed; this equipment is tagged as fully sealed. The insulation material tags are categorized by insulation class as Class A, Class E, Class B, Class F, and Class H; this equipment is tagged as Class A. The power equipment lifespan database was queried, and similar equipment groups were matched based on equipment feature tags. The actual lifespan distribution of the equipment in these groups was statistically analyzed, and the average actual lifespan was calculated to be 22 years with a standard deviation of 3 years. A Bayesian estimation method was used to fuse the design lifespan and actual lifespan statistics, resulting in a baseline lifespan of 21.4 years. This method achieves precise equipment grouping through nameplate parameter feature tags, making the lifespan prediction baseline more closely reflect the actual characteristics of the equipment.

[0110] S42. Based on the equipment feature labels and the equipment baseline lifespan, the random forest algorithm and weight fine-tuning method are used to obtain the personalized aging factor weight coefficients.

[0111] Six core dynamic aging factors were selected, including cumulative overload time, average operating temperature, insulation aging degree, oil quality deterioration degree, environmental pollution level, and number of abnormal failures. An aging factor weight matching model based on equipment feature tags was constructed. A large amount of full lifecycle data for similar equipment was extracted from a cloud database and grouped according to equipment feature tags. Each group contained equipment with the same or similar feature tags. For each group of equipment, a random forest algorithm was used to train the aging factor weight model, obtaining the feature weight vector for that group. A mapping table from equipment feature tags to weight vectors was established. For example, the weight vector for the group of fully sealed equipment with Class A insulation and medium capacity equipment was: average operating temperature 0.38, insulation aging degree 0.32, cumulative overload time 0.14, oil quality deterioration degree 0.09, environmental pollution level 0.05, and number of abnormal failures 0.02. This weight distribution reflects the characteristics of fully sealed equipment being sensitive to temperature and Class A insulation equipment being sensitive to insulation aging. For the target equipment, based on the equipment feature tags extracted from its nameplate parameters, the most similar equipment group was matched in the mapping table, and the feature weight vector of that group was obtained as the initial weight. Then, the weights are fine-tuned based on the historical operating data characteristics of the target equipment. The fine-tuning method involves calculating the historical trend of each aging factor of the target equipment. If an aging factor shows an accelerating deterioration trend, its weight is increased; if it shows a stable or improving trend, its weight is decreased. The fine-tuning range is limited to ±20% of the initial weight. This method achieves adaptive weight adjustment from group characteristics to individual characteristics.

[0112] S43. Based on the personalized aging factor weight coefficient and the equipment's baseline lifespan, the remaining lifespan value is obtained by extracting historical operating data, normalizing aging factors, and calculating the comprehensive aging coefficient.

[0113] Historical operating data and real-time status parameters of the target transformer were extracted from a cloud database, and the quantitative values ​​of each aging factor were calculated. The parameters included: cumulative overload time of 1200 hours, average operating temperature of 68 degrees Celsius, insulation aging degree of 0.5, oil quality deterioration degree of 0.6, environmental pollution level of 3, and an average of 4 abnormal faults per year. Each aging factor was normalized using the following methods: cumulative overload time divided by the total operating hours corresponding to the design life; average operating temperature exceeding the rated value divided by the temperature rise limit; insulation aging degree and oil quality deterioration degree directly used as calculated values; environmental pollution level divided by the highest level of 4; and abnormal fault count divided by a preset threshold of 10 times per year. The normalized aging factor values ​​were: cumulative overload time 0.068, average operating temperature 0.055, insulation aging degree 0.5, oil quality deterioration degree 0.6, environmental pollution level 0.75, and abnormal fault count 0.4. The personalized weighting coefficients obtained from matching equipment feature tags are: average operating temperature 0.40, insulation aging degree 0.30, cumulative overload time 0.13, oil quality deterioration degree 0.09, environmental pollution level 0.05, and number of abnormal failures 0.03. The formula for calculating the comprehensive aging coefficient is: comprehensive aging coefficient = 1 minus the weighted sum of the normalized values ​​of each aging factor multiplied by their corresponding weighting coefficients, resulting in a comprehensive aging coefficient of 0.78. The remaining lifespan is calculated using the formula: remaining lifespan = base lifespan minus the number of years already operated, multiplied by the comprehensive aging coefficient. Substituting the values, the remaining lifespan is calculated as: remaining lifespan = 21.4 minus 8 multiplied by 0.78, resulting in a remaining lifespan of approximately 10.5 years.

[0114] S44. Based on the remaining lifespan value and the weighting coefficient of the personalized aging factor, the exponential decay model and sensitivity analysis method are used to obtain the lifespan decay prediction curve and optimization suggestions.

[0115] Based on the current aging rate and trend extrapolation, a lifespan decline prediction curve is established. The aging rate curve is fitted using historical health scores and aging factor data, employing an exponential decay model. The aging rate constant obtained from historical data is 0.03 per year. Based on the current comprehensive aging coefficient of 0.78 and the aging rate constant of 0.03 per year, the trend of the comprehensive aging coefficient change and the corresponding remaining lifespan over the next 3 to 5 years are extrapolated, and a lifespan decline prediction curve is plotted. Sensitivity analysis is used to calculate the impact of each aging factor on the remaining lifespan. The analysis results show that average operating temperature is the main influencing factor; a 5°C reduction in average operating temperature can increase the remaining lifespan by 1.2 years. Insulation aging is a secondary influencing factor; a 0.1°C reduction in insulation aging can increase the remaining lifespan by 0.8 years. Optimization suggestions are marked on the curve: strengthening cooling system inspections and reducing the load rate by 10% can lower the average oil temperature by 5°C, slowing the annual decline rate by 15% and increasing the remaining lifespan to 11.7 years; regular transformer oil replacement and insulation maintenance can delay the insulation aging process.

[0116] Based on the above sub-steps, personalized weighted adaptive lifetime prediction driven by nameplate parameters is completed, and the remaining lifetime value and lifetime degradation prediction curve are obtained. This method realizes accurate identification of aging modes and adaptive adjustment of weights through nameplate parameter feature labels, thereby improving the accuracy of lifetime prediction.

[0117] S5, based on the device's fused feature vector, personalized anomaly detection results, and remaining lifespan value, uses multi-dimensional index quantification scoring and correction factor calculation to obtain a comprehensive health score and structured health report;

[0118] Based on device fusion feature vectors, personalized anomaly detection results, and remaining lifespan values, a comprehensive health assessment model is constructed to generate quantitative health scores and structured health reports; specifically, the following steps are included:

[0119] S51. Based on the equipment fusion feature vector, personalized anomaly detection results and remaining life value, a multi-dimensional indicator system is established to obtain the health assessment indicator system and correction factor.

[0120] A health assessment index system was established, encompassing four primary indicators and twelve secondary indicators: electrical performance, thermal performance, insulation performance, and mechanical performance. The primary electrical performance indicators include three secondary indicators: current deviation rate, voltage deviation rate, and power factor. The primary thermal performance indicators include three secondary indicators: oil temperature deviation, temperature rise trend, and hot spot temperature. The primary insulation performance indicators include three secondary indicators: insulation resistance, oil dielectric loss, and oil moisture content. The primary mechanical performance indicators include three secondary indicators: noise level (decibels), vibration intensity, and gas signal status. Data for each secondary indicator is derived from the original parameter data of the equipment's fused feature vector. Anomaly detection scores and remaining lifespan values ​​from personalized anomaly detection results are introduced as correction factors for the health assessment. The anomaly detection score reflects the degree of anomaly in the current operating state of the equipment, while the remaining lifespan value reflects the long-term health trend of the equipment.

[0121] S52. Based on the equipment fusion feature vector, anomaly detection results and remaining life prediction results, a comprehensive health score is obtained by using a multi-dimensional index quantitative scoring and correction factor calculation method.

[0122] For the target transformer, 12 secondary indicator data, anomaly detection scores, and remaining life values ​​are extracted from the equipment fusion feature vector, anomaly detection results, and remaining life prediction results. Each secondary indicator is quantitatively scored based on the deviation between the measured value and the standard value; the smaller the deviation, the higher the score, ranging from 0 to 100. Primary indicator scores are calculated using preset weighting coefficients: electrical performance 0.30, thermal performance 0.25, insulation performance 0.30, and mechanical performance 0.15. A preliminary comprehensive health score is calculated as the weighted sum of the scores for each primary indicator. Anomaly detection correction factor and remaining life correction factor are introduced for score correction. The anomaly detection correction factor is calculated as 1 minus the anomaly detection score divided by 100; a higher anomaly detection score indicates a more severe anomaly, and a smaller correction factor is applied. The remaining life correction factor is calculated as follows: the life correction factor equals the remaining life divided by the baseline life. The shorter the remaining life, the smaller the correction factor. Substituting the values, the life correction factor is approximately 0.49, calculated as 10.5 divided by 21.4. The final comprehensive health score is calculated as follows: the final comprehensive health score equals the preliminary comprehensive health score multiplied by the anomaly correction factor multiplied by the life correction factor. This formula comprehensively considers three dimensions: the current operating status of the equipment, the degree of anomaly, and the remaining life, achieving a comprehensive health assessment.

[0123] S53. Based on the comprehensive health score and the scores of each indicator, a structured health report is obtained using a structured report generation method.

[0124] Based on the health score and scores of various indicators, a structured health report is generated. The health report includes modules such as basic equipment information, comprehensive health score, primary indicator scores, secondary indicator details, anomaly details, remaining lifespan information, parameter trend curves, historical comparative analysis, and maintenance recommendations. The health report is stored in a cloud database, and maintenance personnel can view the report content through a web or mobile application. It can also be exported as a PDF document for archiving.

[0125] Based on the above sub-steps, a comprehensive health assessment is completed, resulting in a quantitative health score and a structured health report, which accurately reflects the actual operating status of the equipment and provides a basis for subsequent alarm decisions.

[0126] S6, based on the comprehensive health score and remaining life expectancy, adopts a multi-level early warning mechanism and differentiated push strategy to obtain graded alarm information and auxiliary decision-making information;

[0127] Based on comprehensive health scores and remaining life expectancy, a multi-level intelligent alarm and life expectancy-related early warning mechanism is constructed to achieve accurate early warning and efficient response; specifically, it includes the following steps:

[0128] S61, based on the dual indicators of health score and remaining life expectancy, adopts a three-level alarm threshold setting method to obtain the alarm thresholds for first-level early warning, second-level alarm, and third-level emergency alarm.

[0129] A three-tiered alarm threshold system is established based on both health score and remaining lifespan indicators. Level 1 Warning Threshold: Health score range of 80 to 90 points and remaining lifespan greater than 8 years. This level indicates that the equipment is operating well but has minor anomalies or parameter fluctuations, requiring maintenance personnel to monitor the equipment's status trends. Level 2 Alarm Threshold: Health score range of 60 to 80 points or remaining lifespan range of 5 to 8 years. This level indicates that the equipment is operating moderately or its remaining lifespan is entering the mid-stage, with obvious anomalies or signs of aging, requiring specialized inspection and targeted maintenance. Level 3 Emergency Alarm Threshold: Health score less than 60 points or remaining lifespan less than 5 years. This level indicates that the equipment is operating poorly or its remaining lifespan is about to end, posing a serious risk of anomalies or malfunctions, requiring immediate emergency response measures or equipment upgrades. The alarm level determination logic is as follows: First, determine if the health score is less than 60 points or the remaining life expectancy is less than 5 years. If either condition is met, a Level 3 emergency alarm is triggered. If the Level 3 emergency alarm condition is not met, then determine if the health score is between 60 and 80 points or the remaining life expectancy is between 5 and 8 years. If either condition is met, a Level 2 alarm is triggered. If the Level 2 alarm condition is not met, then determine if the health score is between 80 and 90 points and the remaining life expectancy is greater than 8 years. If both conditions are met, a Level 1 warning is triggered. If neither condition is met, no alarm is triggered.

[0130] S62, Based on the alarm level determination result, a structured alarm information generation method is used to obtain differentiated alarm information content containing key information;

[0131] Differentiated alarm information content is generated based on the alarm level. The alarm information is organized in a structured format and includes the following fields: alarm level, equipment number, equipment name, installation location, alarm time, anomaly type, anomaly severity, associated parameters, health score, estimated remaining lifespan, and preliminary handling suggestions. For the target transformer in this embodiment, the current health score is 76 points and the remaining lifespan is 7.9 years. A level 2 alarm is triggered based on the alarm level determination logic. The generated alarm information content is as follows: alarm level: level 2 alarm; equipment number: BYQ-2024-1001; equipment name: 1000 kVA distribution transformer; installation location: industrial park A area substation; alarm time: 14:30 on December 22, 2024; anomaly type: mild insulation aging, excessively high oil temperature; anomaly severity: moderate; associated parameters: average operating temperature 72 degrees Celsius, exceeding the rated value by 7 degrees Celsius, insulation aging degree 0.5, critical value 0.5; health score: 76 points; estimated remaining lifespan: 7.9 years; preliminary handling suggestions: strengthen cooling system inspection, reduce load rate by 10%, and arrange insulation oil testing to assess insulation status.

[0132] S63, based on the alarm level and differentiated alarm information content, adopts a multi-level push strategy and differentiated push method to obtain alarm notifications pushed to relevant personnel in a graded manner;

[0133] Differentiated alarm information push strategies are implemented based on alarm levels. Level 1 Warning Push Strategy: Alarm information is pushed to maintenance personnel in the responsible area via a mobile application. The push method is an in-app message notification; maintenance personnel can view alarm details simply by opening the application. The push priority is normal, and no sound or vibration alerts are triggered. Level 2 Alarm Push Strategy: Alarm information is pushed to both maintenance personnel in the responsible area and team leaders simultaneously. The push method is a combination of in-app message notifications and SMS notifications. The push priority is important; an sound alert is triggered, but no vibration is. The SMS content includes key information such as device number, alarm level, anomaly type, health score, and remaining lifespan. The Level 3 emergency alarm push strategy triggers the LTE-1101J audible and visual alarm device on the transformer itself. The alarm emits a high-decibel sound and flashes a red light, allowing on-site personnel to immediately detect the anomaly. Simultaneously, the alarm information is pushed in real-time to the responsible maintenance personnel, team leader, and power grid dispatch center. The push method includes in-app message notification, SMS notification, and telephone voice notification. The push priority is emergency, triggering both sound and vibration alerts. The telephone voice notification uses an automatic outbound call method; the system automatically dials the maintenance personnel and team leader, playing pre-recorded alarm audio to ensure timely delivery of information. For the Level 2 alarm triggered in this embodiment, the system pushes the alarm information to the maintenance engineer and team leader, Mr. Li, who simultaneously receive in-app message and SMS notifications.

[0134] S64, based on the three-level emergency alarm information, the auxiliary decision-making information association display method is used to obtain auxiliary decision-making information including location, inventory, life assessment and historical records;

[0135] The alarm information pushed by the Level 3 emergency alert includes associated auxiliary decision-making information to support emergency response and equipment upgrade decisions. This auxiliary decision-making information includes the following: equipment physical location information, which uses a geographic information system to mark the precise coordinates of the equipment on an electronic map, displaying the address of the substation where the equipment is located, surrounding roads, and the nearest maintenance team's location, facilitating rapid on-site arrival for maintenance personnel; spare parts inventory information, which queries the spare parts management system to check the inventory of spare parts matching the equipment model, displaying information such as transformer oil inventory, insulation material inventory, and the number of available standby transformers, assessing the material support capability for emergency response; remaining life assessment report, which retrieves a complete assessment report generated by the life prediction module, including detailed information such as remaining life value, life decay prediction curve, key influencing factors, optimization suggestions, and environmental correction coefficients, providing a scientific basis for equipment upgrade decisions; and historical fault records, which queries the equipment's historical alarm and fault handling records, displaying all abnormal events, handling measures, and handling effects that occurred in the past three years, providing a reference for the diagnosis and handling of current anomalies. When maintenance personnel and dispatch center personnel view alarm information through mobile or web applications, they can click to view associated auxiliary decision-making information to gain a comprehensive understanding of equipment status and available resources for handling the situation.

[0136] S65, based on historical alarm records and false alarm annotation data, adopts backtracking query and false alarm annotation methods to obtain an alarm information management mechanism that supports self-learning optimization;

[0137] The system supports alarm information retrospective query functionality. Maintenance personnel can query historical alarm records via mobile or web applications. Query conditions include time range, device number, alarm level, and anomaly type. Query results are displayed in list format. Each record includes fields such as alarm time, device information, alarm level, anomaly type, and handling status. Clicking on a record allows viewing detailed alarm information and the handling process. The system also supports false alarm labeling. When maintenance personnel receive an alarm and confirm it as a false alarm after on-site verification, they can label the alarm record as a false alarm in the mobile application. The labeling interface includes functions such as selecting the cause of the false alarm (sensor failure, data transmission anomaly, model misjudgment, other reasons), describing the actual device status, and uploading on-site photos. After labeling, the system stores the labeled false alarm data in a cloud database. The system automatically executes a model self-learning optimization process quarterly, statistically analyzing false alarm labeling data and the characteristics of false alarm samples. False alarm samples are then incorporated as negative samples into the training set of the anomaly detection model and health assessment model to retrain model parameters, adjust detection thresholds and scoring strategies, reduce the false alarm rate, form a self-learning closed loop, and continuously improve alarm reliability.

[0138] Based on the above sub-steps, multi-level intelligent alarms and lifespan-related early warnings based on health scores and remaining lifespan are completed, resulting in graded alarm information, differentiated push strategies, and auxiliary decision-making information, thereby achieving accurate early warning and efficient response, and improving operation and maintenance efficiency and fault handling capabilities.

[0139] In one embodiment of the present invention, in order to achieve efficient interaction between maintenance personnel and the system and improve the intelligence level of maintenance decision-making, an intelligent distribution transformer maintenance method further includes: obtaining user natural language or voice query requests, using a BERT model and a structured knowledge base for semantic understanding and knowledge matching, and obtaining accurate query results and visualization content; specifically including the following steps:

[0140] S71, based on multi-source information such as transformer operation and maintenance manuals, equipment nameplate parameters, historical operating data, fault case database, and life assessment standards, a structured knowledge base is obtained using natural language processing and knowledge graph construction technology;

[0141] A structured knowledge base is constructed by integrating multi-source information such as transformer operation and maintenance manuals, equipment nameplate parameters, historical operating data, fault case databases, and life assessment standards. The knowledge base construction process includes the following steps: collecting multi-source information, obtaining electronic documents of transformer operation and maintenance manuals from equipment manufacturers, exporting nameplate parameter data for all equipment from cloud databases, extracting typical operating conditions and parameter change patterns from historical operating data, compiling fault cases and handling solutions from fault handling records, and extracting life assessment standards and aging criteria from power industry standards; then performing structured processing using natural language processing, employing named entity recognition technology to extract entities such as equipment names, parameter names, values, units, and fault types from the text, using relation extraction technology to identify relationships between entities such as parameters belonging to equipment, faults caused by causes, and handling measures to resolve faults, and using knowledge graph construction technology to organize entities and relationships into a graph structure, with nodes representing entities and edges representing relationships, forming a transformer operation and maintenance knowledge graph; finally, establishing a knowledge entry index, organizing the information in the knowledge graph into question-answer pairs, with each knowledge entry containing fields such as question template, answer content, associated entities, and confidence level. The knowledge base covers equipment parameter queries, fault diagnosis, operation and maintenance guidance, and lifespan consultation, containing a total of more than 1,200 knowledge entries, of which more than 200 are lifespan-related special entries, including lifespan influencing factor queries, basis for remaining lifespan calculation, lifespan extension measures, the impact of different operating conditions on lifespan, and interpretation of lifespan assessment standards.

[0142] S72, based on the user's natural language or voice query request, the BERT model is used to perform semantic encoding and query intent recognition to obtain query intent classification and key entities;

[0143] A BERT pre-trained model was used for semantic understanding of user queries. The BERT model selected was the Chinese pre-trained version, BERT-Base-Chinese, with 110 million parameters. Pre-trained on a large-scale Chinese corpus, it possesses powerful semantic understanding capabilities. The model was fine-tuned specifically for the transformer operation and maintenance (O&M) field. A professional O&M corpus was collected, including O&M manuals, technical standards, fault reports, and other documents, totaling 5 million characters. Fine-tuning was performed using a masked language model task and a next-sentence prediction task, with 10 training epochs, a learning rate of 0.00002, and a batch size of 32. The fine-tuned BERT model demonstrates a better understanding of the professional terminology and expressions used in transformer O&M. Users can input queries via web, mobile, or voice. Web and mobile interfaces support text input, allowing users to enter queries such as "How many years is the current remaining lifespan?", "What is the impact of overload operation on lifespan?", and "How to extend transformer lifespan?". Voice input is supported; the system integrates the iFlytek speech recognition engine, allowing users to speak their queries through their phone microphones. The speech recognition engine converts the speech signal into text, achieving an accuracy rate of over 98%. The BERT model semantically encodes the query text, converting it into a 768-dimensional semantic vector containing the semantic information of the query. Then, query intent recognition and key entity extraction are performed. Query intents are categorized into four types: parameter query, fault diagnosis, maintenance guidance, and lifespan consultation. A classifier is used to classify the semantic vectors to identify the user's query intent. Key entity extraction uses a named entity recognition model to extract key entities such as equipment number, parameter name, and time range from the query text. For example, querying "How many years is the current remaining lifespan?" identifies the query intent as lifespan consultation, with key entities being "remaining lifespan" and "current time."

[0144] S73. Based on the query intent classification and key entities, semantic similarity calculation and knowledge base matching methods are used to obtain accurate query results and answer content;

[0145] Based on query intent and key entities, the system retrieves the most relevant knowledge entries from the knowledge base. The retrieval method employs semantic similarity calculation, performing cosine similarity calculation between the query's semantic vector and the question template semantic vectors of each knowledge entry in the knowledge base. The similarity calculation formula is: cosine similarity equals the dot product of the two vectors divided by the product of their magnitudes. The similarity value ranges from -1 to 1, with values ​​closer to 1 indicating greater semantic similarity. All knowledge entries are sorted from highest to lowest similarity, and the top 5 most similar entries are selected as candidate answers. Then, precise answer matching is performed, filtering and sorting candidate answers based on key entities. For example, if the query intent is "life expectancy consultation" and the key entity is "remaining life expectancy," knowledge entries containing content such as "remaining life expectancy calculation methods" and "factors influencing remaining life expectancy" are prioritized. The system retrieves device parameters, health reports, and life expectancy prediction results from the cloud database in real time, populating the answer template with this real-time data to generate precise answers. When asked how many years the equipment has left to live, the system retrieves the equipment's lifespan prediction results. The answer is that, according to the latest assessment, the equipment's current remaining lifespan is approximately 7.9 years. The main influencing factors are an average operating temperature of 72 degrees Celsius, exceeding the rated value by 7 degrees Celsius, and an insulation aging degree of 0.5, which is in a critical state. It is recommended to strengthen the inspection of the cooling system and reduce the load rate to extend the service life.

[0146] S74. Based on the accurate query results and answer content, the query results and visualization content are obtained by using visualization and report export methods.

[0147] While generating text responses, the system also displays associated visualizations to enhance information delivery. These visualizations include charts showing lifespan decline prediction curves, influencing factor weight distributions, and historical trend comparisons. The lifespan decline prediction curve is displayed as a line chart, with the horizontal axis representing future years and the vertical axis representing remaining lifespan. Key time points and influencing factors are marked on the curve. The influencing factor weight distribution is displayed as a pie chart, with the area of ​​each sector representing the weight percentage of each aging factor, and each sector labeled with its factor name and weight value. Historical trend comparisons are displayed using multiple line charts, comparing the changing trends of key parameters such as current health scores, oil temperature, and insulation resistance with historical values ​​for the same period. When users view query results on mobile or web applications, the text responses and visualizations are displayed simultaneously. Users can zoom and drag to view chart details using gestures. The system supports exporting data such as health reports, lifespan prediction reports, and alarm records. Users can click the export button, select the export format as PDF or Excel, and the system automatically generates a formatted document. PDF format is suitable for printing and archiving, containing complete text content, charts, tables, and other elements. Excel format is suitable for data analysis, containing original data tables and charts. The exported report documents are stored on the user's local device, meeting the needs of operation and maintenance document management and asset assessment.

[0148] Based on the above sub-steps, intelligent human-computer interaction based on natural language processing is completed, resulting in accurate query results, associated visual content, and exportable report documents, thereby improving the convenience of operation and maintenance interaction and the efficiency of information acquisition.

[0149] In one embodiment of the present invention, an application example of the above-mentioned intelligent distribution transformer operation and maintenance method is provided: A 1000 kVA 10 kV oil-immersed distribution transformer in the A area of ​​an industrial park of a provincial power grid company has been in operation for many years, with a design life of 20 years. At a certain moment, the system executes a routine monitoring and analysis process.

[0150] The multi-parameter sensing module collects four-dimensional parameters every 100 milliseconds. The data collected on that day is shown in Table 1.

[0151] Table 1, Example of four-dimensional parameter data

[0152]

[0153] After preprocessing the collected data, the edge gateway uploads it to the cloud data center via a 5G link. The data transmission latency is 18 milliseconds, and the upload success rate is 100%. The cloud-based AI intelligent core module performs analysis every 5 minutes. The analysis process at 14:30 is as follows:

[0154] The attention mechanism feature fusion module calculates the deviation of each parameter. The deviation of oil temperature is 0.127, the deviation of insulation resistance is 0.5, and the deviation of oil dielectric loss is 0.6. The attention weights of these three parameters are automatically increased to 0.18, 0.22, and 0.20, respectively, while the weights of other parameters are reduced accordingly, generating a fused feature vector.

[0155] The anomaly detection model inputs a fused feature vector and calculates an average path length of 3.2, which is less than the threshold of 4.5, thus identifying it as an anomaly. Reviewing the original parameter data, the anomaly type is identified as mild insulation aging with excessively high oil temperature. An anomaly score of 65 indicates a moderate degree of anomaly. Related parameters include an oil temperature exceeding the rated value (72 degrees Celsius), an insulation resistance below the standard value (500 megohms), and excessive oil dielectric loss and moisture content.

[0156] The lifespan prediction module calculates the remaining lifespan. The baseline lifespan is 21.4 years. After 8 years of operation, with a comprehensive aging factor of 0.78, the calculated remaining lifespan is approximately 10.5 years (21.4 minus 8). Lifespan degradation prediction curves are generated, predicting a remaining lifespan of 10.2 years after 1 year, 9.9 years after 2 years, 9.6 years after 3 years, 9.3 years after 4 years, and 9.0 years after 5 years. Key influencing factors are identified: average operating temperature is the primary degradation factor; reducing it by 5 degrees Celsius could increase the remaining lifespan to 11.7 years. Insulation aging is a secondary degradation factor; improving insulation condition could increase the remaining lifespan to 11.3 years.

[0157] The health assessment module calculates a comprehensive health score based on anomaly detection results and remaining lifespan values. It calculates scores for 12 secondary indicators: current deviation rate (95 points), voltage deviation rate (98 points), power factor (92 points), oil temperature deviation (68 points), temperature rise trend (72 points), hot spot temperature (70 points), insulation resistance (50 points), oil dielectric loss (40 points), oil moisture content (50 points), noise level (90 decibels), vibration intensity (85 points), and gas signal status (100 points). It also calculates scores for 4 primary indicators: electrical performance (95 points), thermal performance (70 points), insulation performance (47 points), and mechanical performance (92 points). The preliminary comprehensive health score is calculated as: 95 x 0.30 + 70 x 0.25 + 47 x 0.30 + 92 x 0.15 = 76 points. An anomaly detection correction factor and a remaining life correction factor are introduced to adjust the score. The anomaly correction factor is 1 minus 65 divided by 100, which equals 0.35. The remaining life correction factor is 10.5 divided by 21.4, which is approximately 0.49. The final comprehensive health score is 76 multiplied by 0.35 multiplied by 0.49, which is approximately 13 points. A structured health report is generated, which indicates that insulation performance is the primary problem, followed by thermal performance, and recommends strengthening insulation monitoring and cooling system maintenance.

[0158] The alarm module determined the health score to be 13 points, less than 60 points, triggering a Level 3 emergency alarm. Alarm information was generated: Alarm Level: Level 3 Emergency, Equipment Number, Anomaly Type: Mild Insulation Aging, High Oil Temperature, Health Score: 13 points, Remaining Life: 10.5 years, Recommended Actions: Immediately arrange a special inspection, strengthen cooling system inspections, reduce load rate by 10%, and arrange insulation oil testing to assess insulation status. The system triggered the transformer's audible and visual alarm device, emitting a high-decibel alarm and flashing red lights. Simultaneously, the alarm information was pushed in real-time to the maintenance engineer, team leader (Supervisor Li), and the power grid dispatch center via in-app message notification, SMS notification, and telephone voice notification. The system automatically dialed the maintenance personnel and team leader, playing the pre-recorded alarm voice message. The engineer and supervisor received the alarm notification at 2:31 PM.

[0159] Upon receiving the alarm, the engineer arrived at the site at 2:35 PM to investigate. He confirmed that the oil temperature was indeed high, and while the cooling fan was operating normally, excessive dust accumulation on the heat sink was affecting heat dissipation. The engineer used a mobile application to inquire via voice about how much the lifespan could be improved by reducing the load rate. The system invoked the human-computer interaction module, and the speech recognition engine converted the speech into text. The BERT model identified the query intent as a lifespan inquiry, with key entities being "reducing the load rate" and "lifespan improvement." The system retrieved relevant knowledge entries from its knowledge base, retrieved lifespan prediction results, and generated the following answer: reducing the load rate by 10% would lower the average oil temperature by 5 degrees Celsius, slow the annual degradation rate by 15%, and increase the remaining lifespan to 11.7 years, an improvement of approximately 1.2 years. The system also displayed a lifespan degradation prediction curve comparison chart, showing the lifespan change trend under the current operating mode and after reducing the load rate. Based on the system's recommendations, the engineer immediately arranged for the dust on the heat sink to be cleaned and coordinated with the power grid dispatch center to reduce the transformer's load rate from 95% to 85%. Simultaneously, an insulating oil test was scheduled for that afternoon.

[0160] After the heat sink cleaning was completed and the load rate was adjusted, the system performed monitoring and analysis again. The oil temperature dropped to 68 degrees Celsius, the anomaly score decreased to 45 points, and the health score improved to 25 points, but it remained in a level three emergency alarm state. Insulating oil testing results showed a high degree of oil deterioration, and the maintenance department decided to immediately replace the transformer oil and perform insulation maintenance. After maintenance, the system performed monitoring and analysis again. The oil dielectric loss decreased to 0.006, the oil moisture content decreased to 12 mg / L, and the insulation resistance increased to 800 megohms. The anomaly score decreased to 30 points, the health score improved to 55 points, and the alarm level decreased to level three emergency alarm. Continuous system monitoring showed that all parameters tended to stabilize, the anomaly score decreased to 20 points, the health score improved to 68 points, and the alarm level decreased to level two alarm. It is expected that through continuous optimization, the health score can be improved to over 80 points, and the remaining lifespan can be extended to over 11 years.

[0161] This embodiment achieves precise monitoring of the operating status of distribution transformers, quantitative assessment of their health status, scientific prediction of their remaining lifespan, timely early warning of abnormal situations, and intelligent support for operation and maintenance decisions through a closed-loop management system encompassing four-dimensional parameter acquisition, cloud-edge collaborative transmission, AI intelligent analysis, remaining lifespan prediction, multi-level intelligent alarms, and human-machine interactive Q&A. This improves operation and maintenance efficiency and the level of equipment lifecycle management, providing strong support for the safe and reliable operation of the power grid. In this embodiment, the system accurately identifies severe abnormal states of equipment through comprehensive health scoring, triggering a level-three emergency alarm to promptly notify relevant personnel. Maintenance personnel quickly take measures based on system recommendations, such as cleaning heat sinks, reducing load rates, and changing transformer oil, gradually improving the equipment's health status, effectively avoiding equipment failure risks, and extending equipment lifespan.

[0162] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for the operation and maintenance of an intelligent distribution transformer, characterized in that, Includes the following steps: S1, based on the transformer itself and its operating environment, adopts a four-dimensional acquisition architecture to obtain a complete dataset in the cloud; S2, based on the complete dataset in the cloud, uses a dynamic attention mechanism anchored by nameplate parameters to fuse four-dimensional features and obtain the device fused feature vector; S3. Based on the device's fused feature vector, a transfer learning anomaly detection model anchored by the nameplate parameters is used to map the feature space to obtain personalized anomaly detection results. S4. Based on the equipment fusion feature vector and personalized anomaly detection results, a personalized weighted adaptive life prediction model driven by nameplate parameters is adopted to obtain the remaining life value and life decay prediction curve. S5, based on the device's fused feature vector, personalized anomaly detection results, and remaining lifespan value, uses multi-dimensional index quantification scoring and correction factor calculation to obtain a comprehensive health score and structured health report; S6, based on the comprehensive health score and remaining life expectancy, adopts a multi-level early warning mechanism and differentiated push strategy to obtain graded alarm information and auxiliary decision-making information.

2. The intelligent distribution transformer operation and maintenance method according to claim 1, characterized in that, S1 includes: Deploy electrical parameter units, status parameter units, nameplate parameter units, and environmental parameter units to obtain a four-dimensional time-series dataset; A parameter calibration module is integrated inside the transformer body. The calibration algorithm unit automatically executes the calibration process periodically and uses the least squares method to perform linear fitting on the deviation data to obtain a four-dimensional time series dataset after quality calibration. The edge gateway uses a time synchronization algorithm for data alignment, performs linear interpolation to complete the low-frequency sampled data, and obtains a time-aligned four-dimensional time-series dataset; it then performs data preprocessing to standardize the format, clean the data, and compress it. The edge gateway is configured with dual-link redundant transmission, and the data transmission process uses encryption algorithms to encrypt data packets; Cloud data centers implement tiered storage strategies based on data characteristics and access frequency to obtain complete datasets in the cloud.

3. The intelligent distribution transformer operation and maintenance method according to claim 1, characterized in that, S2 includes: Extract the four-dimensional parameter data and nameplate static parameters of the target transformer from the cloud database, and use the nameplate parameters as the feature anchor points of the target transformer to establish the parameter reference coordinate system of the individual equipment. For each operating parameter, calculate its normalized deviation from the nameplate rating; A dynamic attention weight calculation model based on deviation sensitivity is constructed. A deviation sensitivity factor is introduced. The attention weight is equal to the absolute value of the normalized deviation multiplied by the deviation sensitivity factor, then amplified by an exponential function and normalized by softmax. Based on dynamic attention weights, the four-dimensional parameters are weighted and fused at multiple levels, including parameter-level fusion, category-level fusion and device-level fusion, to obtain the device fused feature vector.

4. The intelligent distribution transformer operation and maintenance method according to claim 1, characterized in that, S3 includes: Using historical operating data of transformers of the same type, the operating data of each device is standardized and mapped through its nameplate parameters. The standardized feature is equal to the device fused feature vector divided by the nameplate parameter normalized vector. The isolated forest anomaly detection model is trained in the standardized feature space to obtain a general anomaly detection basic model. For the target transformer, perform inverse feature space mapping based on nameplate parameters, multiply the segmentation threshold of the isolation tree in the model by the corresponding component of the normalized vector of the target equipment nameplate, and obtain a personalized anomaly detection model. For the device fusion feature vectors collected in real time, dual-space collaborative anomaly detection is performed. The comprehensive anomaly judgment adopts dual-space scoring weighted fusion to obtain personalized anomaly detection results.

5. The intelligent distribution transformer operation and maintenance method according to claim 1, characterized in that, S4 includes: Nameplate parameter information is extracted from the equipment fusion feature vector and personalized anomaly detection results. The nameplate parameters are transformed into multi-dimensional equipment feature labels. Similar equipment groups are matched based on the equipment feature labels. The Bayesian estimation method is used to fuse the design life and actual life statistics to calculate the baseline life. A weighted matching model for aging factors based on equipment feature labels is constructed. The random forest algorithm is used to train the weighted model for each group of equipment. A mapping table from equipment feature labels to weight vectors is established. The weights are fine-tuned according to the historical operating data characteristics of the target equipment to obtain personalized aging factor weight coefficients. Historical operating data and real-time status parameters of the target transformer are extracted from the cloud database. The quantitative values ​​of each aging factor are calculated and normalized. The comprehensive aging coefficient and remaining life value are calculated. Based on the current aging rate and trend extrapolation, a lifespan decline prediction curve is established.

6. The intelligent distribution transformer operation and maintenance method according to claim 1, characterized in that, S5 includes: Establish a health assessment index system covering primary indicators and multiple secondary indicators, including electrical performance, thermal performance, insulation performance, and mechanical performance. Introduce the anomaly detection score and remaining life value from personalized anomaly detection results as correction factors for health assessment. Each secondary indicator is quantitatively scored, the primary indicator score is calculated according to the preset weight coefficient, the preliminary comprehensive health score is calculated, and the score is corrected by introducing the anomaly detection correction factor and the remaining lifespan correction factor. The final comprehensive health score is equal to the preliminary comprehensive health score multiplied by the anomaly correction factor and the lifespan correction factor. A structured health report is generated based on the health score and scores of each indicator.

7. The intelligent distribution transformer operation and maintenance method according to claim 1, characterized in that, S6 includes: Based on the dual indicators of health score and remaining life expectancy, a three-level alarm threshold is set, including a first-level early warning threshold, a second-level alarm threshold, and a third-level emergency alarm threshold. The system sequentially judges whether the health score and remaining life expectancy meet the threshold conditions of each level and triggers the corresponding alarm level. Differentiated alarm information content is generated based on the alarm level. The alarm information includes alarm level, device number, anomaly type, associated parameters, health score, remaining life estimate, and preliminary handling suggestions. Different alarm information push strategies are implemented based on alarm levels, with different push methods and push targets for different levels; The alarm information pushed by the Level 3 emergency alarm will be accompanied by auxiliary decision-making information, including equipment physical location information, spare parts inventory information, remaining life assessment report, and historical fault records.

8. The intelligent distribution transformer operation and maintenance method according to claim 3, characterized in that, Performing dual-space collaborative anomaly detection includes: The first spatial detection is a standardized spatial detection, which projects the device fusion feature vector onto the standardized feature space through the standardized mapping of nameplate parameters, and inputs it into the general anomaly detection model to calculate the anomaly score; The second spatial detection is individual spatial detection, which inputs the device's fused feature vector into a personalized model to calculate anomaly scores; The comprehensive anomaly determination adopts a weighted fusion of dual-space scoring. If the comprehensive anomaly score exceeds the preset anomaly threshold, it is determined to be an anomaly. The anomaly type identification method involves tracing back the original parameter data of the feature vector fusion device, analyzing the deviation of each parameter and the distribution of attention weights, and identifying the anomaly type corresponding to the parameter with the largest product of deviation and attention weights as the detection result.

9. The intelligent distribution transformer operation and maintenance method according to claim 1, characterized in that, Also includes: The system acquires user natural language or voice query requests, uses the BERT model and a structured knowledge base for semantic understanding and knowledge matching, and obtains accurate query results and visualizations; specifically including: By integrating multi-source information such as transformer operation and maintenance manuals, equipment nameplate parameters, historical operating data, fault case databases, and life assessment standards, a structured knowledge base is constructed. Named entity recognition technology is used to extract entities from text, relation extraction technology is used to identify the relationships between entities, and knowledge graph construction technology is used to organize entities and relations into a graph structure. The BERT pre-trained model is used for semantic understanding of user queries. The BERT model performs semantic encoding on the query text and performs query intent recognition and key entity extraction. Based on the query intent and key entities, the system retrieves the most relevant knowledge items from the knowledge base. The retrieval method uses semantic similarity calculation. The system retrieves device parameters, health reports, and life prediction results from the cloud database in real time and populates the answer template with real-time data to generate an accurate answer. While generating text responses, the system also displays associated visual content and supports the export of health reports, life expectancy prediction reports, and alarm records.

10. An intelligent distribution transformer operation and maintenance system, used to execute the steps of an intelligent distribution transformer operation and maintenance method as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to obtain a complete dataset in the cloud based on the transformer body and its operating environment using a four-dimensional acquisition architecture. The feature fusion module is used to fuse four-dimensional features based on the complete dataset in the cloud using a dynamic attention mechanism anchored by the nameplate parameters, to obtain the device fused feature vector; The anomaly detection module is used to obtain personalized anomaly detection results by mapping the anomaly detection model anchored by the nameplate parameters to the feature space based on the device's fused feature vector. The life prediction module is used to obtain the remaining life value and life decay prediction curve by using a personalized weighted adaptive life prediction model driven by nameplate parameters, based on the equipment's fused feature vector and personalized anomaly detection results. The health assessment module is used to obtain a comprehensive health score and a structured health report by using multi-dimensional index quantification scoring and correction factor calculation based on the device's fused feature vector, personalized anomaly detection results and remaining life value. The alarm push module is used to obtain graded alarm information and auxiliary decision-making information based on the comprehensive health score and remaining life value, using a multi-level early warning mechanism and differentiated push strategy.

Images