Switch cabinet partial discharge and temperature abnormity real-time monitoring method and system

Abstract

The invention discloses a switch cabinet partial discharge and temperature anomaly real-time monitoring method and system, and the method comprises the steps: S1, cooperating with a plurality of power enterprise nodes through a federated learning server, carrying out the combined training of a double-branch fault recognition model under the condition of protecting data privacy, and generating an initial strategy mapping library of related fault information and linkage strategies; s2, deploying the model in an edge computing device, performing cross-model fault identification on real-time monitoring data, and matching a corresponding linkage instruction from a strategy mapping library when the confidence coefficient meets a condition; s3, issuing the linkage instruction to the secondary loop equipment for execution, and collecting execution effect data as linkage feedback data; and S4, the linkage feedback data is processed and transmitted back for triggering iterative training of the fault identification model and strategy parameter updating of the strategy mapping library. According to the invention, closed-loop intelligent monitoring from fault identification and accurate linkage to strategy self-optimization is realized.

Description

Technical Field ,

[0006]

[0001] The present invention relates to the technical field of power system operation monitoring and fault prevention and control, and particularly relates to a method and system for real-time monitoring of partial discharge and temperature abnormality in a switch cabinet. Background Art

[0002] The switch cabinet is a key device in the power transmission and distribution link of the power system, and the stability of its operation state directly relates to the safe and reliable operation of the power grid. During long-term operation, affected by various factors such as equipment aging, environmental changes, and operation condition fluctuations, partial discharge phenomena and temperature abnormalities often become important indicators before the insulation performance deteriorates and potential faults occur. Therefore, continuously monitoring, accurately identifying, and timely disposing of relevant abnormal states during the operation of the switch cabinet is one of the core requirements for reducing fault risks and improving the operation safety of the power system.

[0003] With the development of sensing and information processing technologies, the monitoring means for the operation state of the switch cabinet have gradually evolved from single-parameter monitoring to multi-dimensional comprehensive perception. By collecting different types of operation information, the analysis and early warning of abnormal states are realized. At the same time, the secondary circuit system supporting the switch cabinet, as a protection, measurement, and execution unit, plays an important role in adjusting the operation state and suppressing the expansion of risks after an abnormal state appears. The timeliness of its response and the rationality of the strategy have a direct impact on the fault evolution result.

[0004] However, in actual engineering applications, due to significant differences in equipment models, operation environments, management specifications, and historical operation data among different power operation entities, relevant monitoring and disposal systems are usually built and operated in an independent manner, and the overall intelligence and coordination are still limited. Especially in aspects such as the recognition accuracy of abnormal states, the disposal response efficiency, and the long-term adaptability of the system, existing solutions still face many restrictions.

[0005] On the one hand, the operation data and fault information of the switch cabinet have high industry sensitivity, and it is difficult for different operation entities to directly share data, resulting in abnormal recognition and analysis often relying on limited data samples and being difficult to comprehensively cover complex and changeable operation conditions. When the equipment model or operation conditions change, the original recognition ability is likely to decline, and the system adaptability is insufficient. On the other hand, the monitoring and subsequent disposal processes of abnormal states are usually relatively fragmented, and the recognition results mostly stay at the warning level. Manual participation in decision-making and execution is still required, and it is difficult to achieve a rapid linkage response after an abnormal state appears, resulting in certain lag and uncertainty.

[0006] Furthermore, different switchgear models differ in structural characteristics, parameter configurations, and secondary circuit control logic. Existing handling strategies are often designed for specific equipment or scenarios, limiting their versatility and scalability. Large-scale deployment of multiple equipment models typically requires repeated configuration and manual debugging, resulting in high implementation costs. Simultaneously, related monitoring and handling strategies often employ static settings, lacking continuous feedback and dynamic optimization mechanisms for actual operational effectiveness. As equipment status changes and the operating environment evolves, their identification accuracy and handling effectiveness tend to gradually decline, making it difficult to maintain good operational performance in the long term.

[0007] Therefore, how to improve the accuracy and adaptability of switchgear abnormality identification while ensuring data security and operational independence, achieve efficient linkage between abnormality identification and secondary circuit handling, and possess continuous optimization capabilities based on operational feedback has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for real-time monitoring of partial discharge and temperature anomalies in switchgear. This method improves the accuracy and adaptability of switchgear anomaly identification while protecting data privacy, achieves efficient linkage between anomaly identification and secondary circuit handling, and has continuous optimization capabilities based on operational feedback.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for real-time monitoring of partial discharge and temperature anomalies in switchgear, comprising: Step S1, Federated Collaborative Model Training: The federated learning server collaborates with multiple power enterprise nodes to jointly train a dual-branch fault identification model based on a privacy protection mechanism. Each power enterprise node uses its local de-identified monitoring data to train its local model, generating encrypted model parameters which are then uploaded to the federated learning server. The federated learning server aggregates the encrypted model parameters uploaded by all nodes to update the global model parameters and distributes them. The training process of the dual-branch fault identification model simultaneously incorporates the historical secondary circuit data of each node to generate an initial policy mapping library of associated fault information and linkage strategies. Step S2, Edge Identification and Command Matching: The trained dual-branch fault identification model is deployed on an edge computing device; the edge computing device collects real-time monitoring data of the target switchgear and calls the model corresponding to the target switchgear model for fault identification; when the confidence level of the identification result is higher than the set threshold, the corresponding secondary circuit linkage command is matched from the strategy mapping library according to the identified fault information. Step S3, Command Execution and Data Feedback: The matched secondary circuit linkage command is sent to the secondary circuit equipment of the target switchgear for execution via a secure communication link; the equipment status change data after command execution is collected as linkage feedback data; Step S4, closed-loop iterative optimization: the linkage feedback data obtained by each node is desensitized to form feedback training samples; the feedback training samples are sent back to the federated learning server to trigger iterative training of the dual-branch fault identification model and updating of linkage policy parameters of the policy mapping library.

[0010] Furthermore, in step S1, the training of the federated collaborative model is performed under a horizontal federated learning architecture, which includes a federated learning center node and at least two local nodes of the power companies. Each of the power company's local nodes uses de-identified local multimodal monitoring data to train a local model copy, and only uploads the encrypted model parameters generated during training to the federated learning center node through a secure communication channel. The federated learning center node aggregates the received encrypted model parameters to update the global model parameters and distribute them to each local node. The joint training process synchronously inputs the secondary circuit equipment parameters and historical linkage data of each node, so that the trained dual-branch fault identification model has the ability to identify faults across models and generates the initial strategy mapping library.

[0011] Furthermore, the dual-branch fault identification model specifically includes a common feature extraction branch, a unique feature extraction branch, and a fusion output layer based on an attention mechanism; The common feature extraction branch is collaboratively optimized by the federated learning center nodes based on model parameters uploaded by multiple nodes, and is used to learn the common features of faults in different types of switchgear. The common features include predetermined signal frequency band features and temperature change features. The personalized feature extraction branch is optimized locally by each power company's local node based on its own local data of specific model switchgear, in order to learn the personalized fault characteristics of that specific model of equipment. The fusion output layer is configured to dynamically fuse the general features output by the common feature extraction branch and the personalized features output by the individual feature extraction branch, and output an identification result containing core fault information and related secondary loop linkage strategy suggestions. The core fault information includes at least the fault type, location, and severity.

[0012] Furthermore, the step S2 of calling a model matching the target switchgear model for fault identification specifically includes: A model identification unit is deployed in the edge computing device, which obtains the model information of the target switchgear by reading the equipment file of the target switchgear. Based on the acquired model information, the corresponding individual feature extraction branch is dynamically configured and activated from the deployed dual-branch fault identification model to form a combined identification network of common feature extraction branch adapted to the specific model and activated individual feature extraction branch, and to identify faults in real-time monitoring data.

[0013] Furthermore, the process of matching and generating secondary loop linkage instructions in step S2 also includes a confidence-based hierarchical decision-making mechanism, which includes: The edge computing device determines the confidence level of the recognition result output by the combined recognition network; If the confidence level is higher than the first confidence level threshold, it is determined to be a valid identification and triggers the subsequent linkage instruction matching process; If the confidence level is lower than the first confidence level threshold but higher than the second confidence level threshold, it is determined to be in a pending state, and the enhanced monitoring mode is activated. The enhanced monitoring mode includes increasing the collection frequency of the real-time monitoring data. If the confidence level is lower than the second confidence level threshold, it is determined to be an invalid identification, and only data is recorded without triggering subsequent linkage processes.

[0014] Furthermore, the process of issuing the secondary circuit linkage command through the secure communication link in step S3 specifically includes: The matched secondary circuit linkage command is encapsulated into a data frame conforming to the standard communication protocol of the power system. The data frames are sent to the secondary circuit equipment corresponding to the target switchgear via a data transmission network based on optical fiber.

[0015] Furthermore, the process of collecting the linkage feedback data in step S3 specifically includes: The feedback unit located in the secondary circuit equipment or switch cabinet body collects equipment status change data after the execution of the secondary circuit linkage command. The device status change data is associated with the fault identification result that triggered the linkage and the issued linkage command information to generate a structured feedback data record, which serves as the linkage feedback data.

[0016] Furthermore, in step S4, the linkage feedback data is constructed as an associated recognition-linkage-effect triplet training sample; The triggering of iterative training of the dual-branch fault identification model is specifically a differentiated collaborative optimization process, which includes: The federated learning server performs centralized iterative training on the common feature extraction branch of the model based on the triplet training samples from multiple nodes to optimize its general fault representation capability. Each power enterprise node performs localized fine-tuning of the model's individual feature extraction branch based on the triplet training sample subset related to its own equipment model, in order to optimize its adaptation accuracy for specific models.

[0017] Furthermore, the step S4 of updating the linkage strategy parameters of the strategy mapping library specifically includes a closed-loop strategy optimization mechanism based on execution effect evaluation: Based on the performance data in the triplet training samples, evaluate whether the execution effect of the corresponding linkage instruction meets the preset effectiveness standard. If the validity criteria are met, the recognition-linkage relationship in the triplet training sample is used as a reinforcement sample to improve the confidence of the corresponding policy entry in the policy mapping library. If the validity criteria are not met, it is determined to be an abnormal linkage case, triggering the policy correction process: generating corrected linkage instruction parameters based on the optimization algorithm to update the policy mapping library, and using the abnormal linkage case and the corrected instruction parameters as new training samples, feeding them back to the federated learning framework to trigger targeted fine-tuning of the relevant models.

[0018] A real-time monitoring method for partial discharge and temperature anomalies in switchgear, used in the aforementioned real-time monitoring method for partial discharge and temperature anomalies in switchgear, includes: The model training collaboration module is used to collaborate with the federated learning server and multiple power enterprise nodes to jointly train a dual-branch fault identification model based on a privacy protection mechanism. Each power enterprise node uses its local de-identified monitoring data to train its local model, generating encrypted model parameters which are then uploaded to the federated learning server. The federated learning server aggregates the encrypted model parameters uploaded by all nodes to update the global model parameters and distributes them. The training process of the dual-branch fault identification model synchronously incorporates historical secondary circuit data from each node to generate an initial policy mapping library of associated fault information and linkage strategies. An edge recognition and matching module, connected to the model training collaboration module, is used to deploy the trained dual-branch fault recognition model on an edge computing device; the edge computing device collects real-time monitoring data of the target switchgear and calls the model corresponding to the target switchgear model for fault recognition; when the confidence level of the recognition result is higher than a set threshold, the corresponding secondary loop linkage instruction is matched from the strategy mapping library based on the identified fault information. The instruction execution and feedback interface module is connected to the edge recognition and matching module. It is used to send the matched secondary circuit linkage instruction to the secondary circuit equipment of the target switch cabinet for execution through a secure communication link; and to collect the equipment status change data after instruction execution as linkage feedback data. The closed-loop optimization processing module connects the instruction execution and feedback interface module and the model training collaboration module. It is used to desensitize the linkage feedback data obtained by each node to form feedback training samples. The feedback training samples are then sent back to the federated learning server to trigger iterative training of the dual-branch fault identification model and updating of the linkage policy parameters of the policy mapping library.

[0019] The beneficial effects of this invention are: 1. This invention trains a model by having a federated learning server collaborate with multiple power enterprise nodes, enabling each node to participate in model building without sharing the original monitoring data. This avoids the security risks caused by the direct flow of operational data and integrates the fault characteristics and operating condition information of different operating entities, thereby improving the accuracy and adaptability of the fault identification model in various operating environments.

[0020] 2. This invention generates a strategy mapping library that associates fault information with linkage strategies by synchronously combining historical data of secondary circuits during the model training phase. This enables the fault identification results to be directly associated with executable secondary circuit linkage instructions, avoiding the identification results from merely remaining at the alarm level and effectively improving the application value of anomaly identification results in actual operation control.

[0021] 3. This invention deploys the trained fault identification model on an edge computing device, which can complete fault identification and command matching on-site after collecting real-time monitoring data of the target switchgear. This reduces the delay caused by data backhaul and centralized processing, enables rapid judgment of abnormal states and timely issuance of linkage commands, and helps to suppress the further development of faults.

[0022] 4. By setting a confidence threshold for the fault identification results, the matching and issuance of secondary circuit linkage commands are only triggered when the identification results meet the reliability requirements, thereby avoiding malfunctions caused by misidentification and improving the safety of secondary circuit linkage execution and the stability of system operation.

[0023] 5. By collecting equipment status change data after the execution of secondary loop linkage commands and sending it back to the federated learning server as feedback training samples, the system can iteratively update the fault identification model and strategy mapping library based on the actual linkage effect, thereby maintaining high identification accuracy and linkage effectiveness under changes in equipment status and evolution of operating conditions.

[0024] 6. By forming a collaborative optimization mechanism based on feedback data among multiple power enterprise nodes, the fault identification model and linkage strategy can continuously evolve over time, avoiding the performance degradation problem caused by long-term static use of models and strategies, and improving the overall reliability and long-term operation effect of the switchgear partial discharge and temperature anomaly monitoring system. Attached Figure Description

[0025] Figure 1 This is a flowchart of the steps in the real-time monitoring method for partial discharge and temperature anomalies in switchgear of the present invention; Figure 2 This is a schematic diagram of the structure of the real-time monitoring system for partial discharge and temperature anomalies in switchgear of the present invention.

[0026] Attached reference numerals: 1. Model training collaboration module; 2. Edge recognition and matching module; 3. Instruction execution and feedback interface module; 4. Closed-loop optimization processing module. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom surface," "top surface," "inner," and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0028] Example 1, refer to Figure 1 This is the first embodiment of the present invention, which provides a method for real-time monitoring of partial discharge and temperature anomalies in switchgear. The method of this embodiment is operated in the following implementation environment: The power companies participating in the joint training include Company A, Company B, and Company C. The three companies operate three different models of medium-voltage switchgear: KYN28-12, XGN66-12, and GG-1A(F). Each company has deployed no fewer than 10 switchgear devices for long-term operation data collection and model training.

[0029] The hardware environment includes: The federated learning server, serving as a node in the federated learning center, is configured with an Intel Xeon Gold 6330 processor and 32GB of memory. Each power company sets up a local node, configured with an Intel Core i7-12700 processor and 16GB of memory; The edge computing device deployed at the substation site uses an ARM Cortex-A76 processor and has 8GB of memory; The multimodal sensing nodes include an ultra-high frequency partial discharge sensor with an operating frequency band of 500MHz-1.5GHz and a fiber optic thermometer with a temperature measurement accuracy of ±0.5℃. The secondary circuit equipment includes protection devices and measurement and control devices, which are used to perform protection parameter adjustments and operating status regulation, respectively.

[0030] The software environment includes: A federated learning framework built on TensorFlow Federated; Deep learning models implemented using PyTorch; The edge computing device and the secondary loop equipment use the IEC61850MMS communication protocol. The Federated Learning Center node and each enterprise's local node establish a secure communication link using the SSL / TLS 1.3 protocol, and the model parameters are encrypted using AES-256.

[0031] Working principle of Example 1: The method in this embodiment includes four stages in chronological order: federated collaborative model training, edge detection and instruction matching, instruction execution and data feedback, and closed-loop iterative optimization. Its overall working principle is as follows: By training collaboratively with multiple power companies, a dual-branch fault identification model with strong generalization ability is formed. This model enables real-time identification of partial discharge and temperature anomalies in switchgear at the edge, and directly links the identification results to secondary circuit linkage commands. At the same time, the linkage execution effect is used to continuously optimize the model and strategy, thereby realizing the integrated closed-loop operation of anomaly monitoring and handling.

[0032] (I) System Setup Phase (T0–T1, T1 is 7 days): First, the federated learning server is deployed with local nodes of each power company, and an encrypted communication channel is established between the nodes. The time synchronization unit is used to uniformly synchronize the federated learning server, local nodes, edge computing devices and sensor nodes to ensure that the time synchronization accuracy is no higher than 1ms, so as to ensure the consistency of multi-source data in the time dimension.

[0033] Subsequently, multimodal sensing nodes were installed on key parts such as busbar joints, insulating bushings, and contacts of the switchgear in each enterprise to collect partial discharge signals and temperature data, which were then transmitted to the local nodes of the enterprise via shielded cables. At the same time, the edge computing device was connected to the corresponding secondary circuit equipment through the IEC61850 interface.

[0034] During the model initialization phase, a dual-branch fault identification model is constructed, including a common feature branch for learning common characteristics of different switchgear models, and a unique feature branch for characterizing model-specific differences. Initially, the weight of the common feature branch is set to 0.7, and the weight of the unique feature branch is set to 0.3 to ensure that the model prioritizes learning universally applicable abnormal features in the early stages of training. Simultaneously, a "fault-model-linkage strategy" mapping library is initialized to store the correspondence between different fault types and secondary circuit linkage commands.

[0035] (II) Step S1, Federated Collaborative Model Training Phase (T1–T2, T2 is 30 days): The federated learning server collaborates with multiple power enterprise nodes to jointly train a dual-branch fault identification model based on a privacy protection mechanism; each power enterprise node uses local de-identified monitoring data to train its local model, generates encrypted model parameters, and uploads them to the federated learning server; the federated learning server aggregates the encrypted model parameters uploaded by all nodes to update the global model parameters and distributes them; the training process of the dual-branch fault identification model is combined with the historical secondary circuit data of each node to generate an initial policy mapping library of associated fault information and linkage strategies.

[0036] Specifically, in this phase, each power company's local nodes train their local models based on multimodal monitoring data collected by the company. The monitoring data includes normal operation data as well as partial discharge and temperature anomaly data obtained through experiments or historical data reproduction, covering various fault modes such as corona discharge, surface discharge, and poor contact. All data is anonymized before entering the training process to remove information related to the company's identity.

[0037] Each enterprise's local node only updates its local model parameters and uploads the encrypted parameter gradients to the federated learning server. The federated learning server merges model parameters from different enterprises through a parameter aggregation mechanism, updates the parameters of the common feature branches in the dual-branch model, and distributes the updated parameters to each enterprise's local node. Simultaneously, each enterprise fine-tunes its individual feature branches based on its limited local data to enhance the model's ability to identify the characteristics of its own switchgear models.

[0038] The above-mentioned federated training process is executed cyclically, with no fewer than 100 training rounds, until the model's fault identification accuracy on the validation set reaches or exceeds 92%. During model training, the correspondence between faults and linkage instructions in the policy mapping library is simultaneously optimized and updated based on historical secondary loop action records and operational results, thereby forming an initially usable linkage policy system.

[0039] The technical effect of this stage is that, without sharing the original monitoring data, it makes full use of the operating experience of multiple enterprises and significantly improves the model's ability to identify abnormal states of multiple switchgear models.

[0040] (III) Step S2, Edge Identification and Command Matching Stage (T2–T3, Long-term Operation): Deploy the trained dual-branch fault identification model on the edge computing device; The edge computing device collects real-time monitoring data of the target switchgear and calls the model corresponding to the target switchgear model for fault identification; When the confidence level of the identification result is higher than the set threshold, the corresponding secondary circuit linkage command is matched from the strategy mapping library according to the identified fault information.

[0041] Specifically, the trained dual-branch fault identification model is deployed to the edge computing devices of various enterprises. The edge computing devices have built-in model recognition units to read the equipment file information of the target switchgear and match the corresponding unique feature branches accordingly.

[0042] During actual operation, multimodal sensing nodes collect partial discharge signals and temperature data of the target switchgear in real time and transmit them to the edge computing device. The edge computing device calls the dual-branch fault identification model corresponding to the target switchgear model to perform fault identification and outputs the identification results, including the fault type and its corresponding confidence level.

[0043] The confidence threshold is set to 0.85. When the confidence level of the identification result is higher than this threshold, the identification result is considered to have sufficient reliability. At this time, the edge computing device matches the secondary loop linkage instruction corresponding to the fault type and device model from the policy mapping library based on the identified fault information.

[0044] The technical benefits of this stage are: to achieve a direct link between anomaly identification and secondary loop control, reduce manual intervention, and improve the real-time performance of anomaly handling.

[0045] (iv) Step S3, instruction execution and data feedback stage: The matched secondary circuit linkage instruction is sent to the secondary circuit equipment of the target switch cabinet for execution through the secure communication link; the equipment status change data after instruction execution is collected as linkage feedback data.

[0046] Specifically, the edge computing device sends the matched secondary circuit linkage command to the secondary circuit equipment of the target switchgear via a secure communication link, executing control actions including load adjustment and protection threshold adjustment. After completing the command execution, the secondary circuit equipment collects the corresponding equipment status change data and feeds it back to the edge computing device.

[0047] The edge computing device de-identifies the feedback data, forms linked feedback data, and periodically transmits it back to the federated learning server for subsequent model and strategy optimization.

[0048] (v) Step S4, closed-loop iterative optimization (continued after T3): desensitize the linkage feedback data obtained by each node to form feedback training samples; send the feedback training samples back to the federated learning server to trigger iterative training of the dual-branch fault identification model and update of linkage strategy parameters of the strategy mapping library.

[0049] Specifically, the federated learning server iteratively optimizes the common feature branches in the dual-branch model every 7 days based on feedback data from various enterprises; each enterprise's local node fine-tunes its individual feature branches every 15 days. When the system detects that the linkage execution effect is not ideal, it triggers an anomaly correction process, adjusts the corresponding linkage instruction parameters, and updates the strategy mapping library synchronously.

[0050] When a new model of switchgear is connected to the system, only about 500 sets of local monitoring data need to be collected, and the individual characteristic branches need to be fine-tuned in a short period of time (no more than 2 days) to achieve accurate identification and linkage control of the new model of equipment.

[0051] Verification of the effect of Example 1: Through actual deployment and verification in companies A, B, and C, this embodiment achieved the following results: The anomaly identification accuracies for different switchgear models reached 94%, 93%, and 92%, respectively, with an average identification accuracy of 93%. The average latency from the completion of anomaly identification to the completion of the secondary loop linkage instruction execution is 85ms, which is significantly shorter than that of manual handling. The containment rate of early minor faults reaches 95%, effectively preventing the fault from escalating; The system deployment cost in multi-model scenarios is reduced by about 65% compared to a single-model customized solution, and the debugging time is shortened by about 72%.

[0052] Example 2 is the second embodiment of the present invention. Based on Example 1, Example 2 further describes in detail the training mechanism of the federated collaborative model, the internal structure of the dual-branch fault identification model and its mathematical expression.

[0053] The core working principle of this embodiment is as follows: Under the horizontal federated learning architecture, a dual-branch fault identification model is collaboratively trained by local nodes of multiple power companies without sharing the original monitoring data. This enables the model to simultaneously learn the common fault characteristics of different switchgear models and the personalized characteristics of specific models. Furthermore, secondary circuit equipment parameters and historical linkage data are embedded during the model training process, so that the model output not only contains the core fault information but can also directly generate linkage strategy suggestions, thereby providing a reliable basis for real-time identification and secondary circuit linkage at the edge.

[0054] Training of federated collaborative models under a horizontal federated learning architecture: (a) Horizontal Federated Learning Architecture; This embodiment adopts a horizontal federated learning architecture, which includes a federated learning center node and at least two local nodes of power companies. The data feature spaces of each local node of a power company are consistent, but the sample sources are different.

[0055] Each power company's local nodes deploy local training units and data encryption units, while the federated learning center nodes deploy parameter aggregation units and model management units.

[0056] During model training, each power company's local nodes only use the anonymized local multimodal monitoring data to train the local model copy, without uploading any original monitoring data; the model parameters generated during training are encrypted and uploaded to the federated learning center node through a secure communication channel.

[0057] The federated learning center node aggregates the received encrypted model parameters to form global model parameters and distributes them to each local node, thereby completing one federated training iteration.

[0058] This architecture achieves "data not leaving the domain and models being collaborative" in its working mechanism, effectively solving the problem that power companies cannot share operational data due to data privacy restrictions, while significantly expanding the effective sample size that the model can utilize.

[0059] The joint training process synchronously inputs the secondary circuit equipment parameters and historical linkage data of each node, so that the trained dual-branch fault identification model has the ability to identify faults across models and generates an initial strategy mapping library.

[0060] Structure and function of the dual-branch fault identification model: The dual-branch fault identification model includes an input layer, a common feature extraction branch, a unique feature extraction branch, an attention-based fusion output layer, and an output layer.

[0061] Common Feature Extraction Branch: This branch is jointly optimized by the federated learning center nodes based on model parameters uploaded by multiple enterprise local nodes to learn the common features of partial discharge and temperature anomalies in different types of switchgear.

[0062] The general characteristics include at least: the spectral distribution characteristics of the partial discharge signal in the 500MHz–1.5GHz frequency band; and a temperature change rate greater than or equal to... Abnormal trend characteristics.

[0063] The parameters obtained through training this branch are shared among enterprises, enabling the model to have basic recognition capabilities across models and enterprises.

[0064] Personalized Feature Extraction Branch: This branch is used by local nodes of each power company to perform local optimization based on local data of their specific switchgear models, in order to learn personalized fault features related to equipment structure, materials and layout.

[0065] Each switchgear model only requires no less than 500 local samples to complete the effective fine-tuning of its individual feature branches, thereby significantly reducing the adaptation cost when new models of equipment are connected to the system.

[0066] The attention-based fusion output layer is configured to dynamically fuse the general features output from the common feature extraction branch and the personalized features output from the individual feature extraction branch, and output the identification result containing the core fault information and the secondary loop linkage strategy suggestions directly related to it. The core fault information includes at least the fault type, location and severity.

[0067] The dynamic allocation range of feature weights is set as follows: common feature weights are 0.6–0.8; individual feature weights are 0.2–0.4.

[0068] This dynamic fusion method ensures the model's generalization ability while maintaining the recognition accuracy for specific device models.

[0069] Mathematical expression and mechanism of action of the federated collaborative model: In this embodiment, the combined output of the federated collaborative model after one global training iteration can be represented by the following integration equation: ; in, The output value of the model fusion is used to characterize the comprehensive risk level of the current abnormal state of the target switchgear. Its value range is from 0 to 1, where close to 0 indicates normal operation and close to 1 indicates significant abnormal risk. The shape parameters are obtained based on the statistical analysis of the partial discharge energy spectrum and are derived from multimodal sensing data. The complexity parameter for the frequency domain distribution of multi-enterprise samples is used in the Riemann Zeta function to characterize the spectral attenuation characteristics across samples. Here, is the normalized parameter for the rate of temperature change; its physical dimension is dimensionless. d is the oscillation intensity parameter of the partial discharge signal within the characteristic frequency band, used for Bessel functions to characterize periodic characteristics. e is the time decay coefficient, used for exponential functions to describe the influence weight of abnormal characteristics over time. t is the time variable, in seconds. This is a parameter for normalizing the load fluctuation amplitude, used in the logarithmic function to suppress the effects of extreme operating conditions. For feature mapping parameters related to equipment structure, The modulus of the elliptic function is used to characterize the structural differences between different models. For local sample indexing, This represents the total number of local samples. This is an information filtering function built based on historical linkage effects, and its input is... For the first The intensity of the linkage feedback of each sample For the sample noise intensity parameter, This is the global model stability adjustment coefficient. These are the error compensation coefficients, used in the complementary error function to characterize the uncertainty in anomaly detection. For cross-enterprise parameter aggregation index, The number of companies participating in federal training. For the first Model complexity adjustment parameters for individual enterprises For the first Normalization factor for enterprise equipment size.

[0070] The working principle of this formula is as follows: by continuously integrating and mapping the multi-source monitoring signals in the time domain and feature domain with complex nonlinear functions, a comprehensive modeling of partial discharge intensity, temperature change trend and operating condition disturbance is achieved; the fractional structure is used to normalize and suppress abnormal signals, so that the model output is comparable among different enterprises and different models of equipment, thereby improving the stability and robustness of the federated model in the horizontal collaborative training scenario.

[0071] Through the formula of the above mathematical model, the federated learning center node can uniformly model the abnormal characteristics of switchgear from different enterprises and different models without accessing the original data. The output of this model serves as an important input to the dual-branch fault identification model fusion output layer, which is used to further generate core fault information and linkage strategy suggestions, thereby effectively solving the problems of insufficient model generalization ability and poor consistency of cross-model anomaly identification under cross-enterprise data isolation conditions.

[0072] Overall technical effects of Example 2: The method in this second embodiment enables collaborative training of anomaly identification models for multiple enterprises and multiple switchgear models while protecting the data privacy of power companies. By combining a dual-branch structure with a complex federated collaborative model, the accuracy and stability of anomaly identification in cross-model scenarios are significantly improved. The model output is directly associated with secondary loop linkage strategy suggestions, enabling the anomaly identification results to serve actual operation control in real time; The system has good scalability. When new models of devices are connected, there is no need to reconstruct the global model. Only a few local fine-tunings are required before it can be put into use.

[0073] Example 3 is the third embodiment of the present invention. This embodiment focuses on the specific implementation of steps S2 and S3 in Example 1 at the enterprise edge.

[0074] Its core working principle is: The dual-branch fault identification model, trained by federation, is deployed on the enterprise edge computing device. The model identification unit dynamically activates the individual feature extraction branch that matches the target switch cabinet model, thus constructing a combined identification network of "common feature extraction branch + activated individual feature extraction branch". After fault identification of real-time multimodal monitoring data, a confidence-based hierarchical decision-making mechanism is introduced to distinguish between valid identification, undetermined state and invalid identification, and to decide whether to trigger secondary loop linkage accordingly. When the linkage conditions are met, the linkage command is accurately matched through the cross-model linkage strategy adaptation unit and sent to the secondary circuit equipment for execution via a secure communication link that conforms to power system standards. At the same time, the execution effect is collected in a structured feedback manner, thereby realizing the closed-loop operation of anomaly identification, linkage handling and operation feedback.

[0075] Edge-side implementation of cross-model adaptive fault identification; (a) Dynamic configuration of model identification unit and model; Deploy model identification units in the edge computing devices of various power companies.

[0076] The model identification unit obtains the model information of the target switchgear by reading its equipment file. The equipment file includes at least the switchgear model, manufacturer, and corresponding secondary circuit parameters.

[0077] After obtaining the model information of the target switchgear, the edge computing device dynamically configures and activates the individual feature extraction branch corresponding to the model from the deployed dual-branch fault identification model based on the model information, while keeping the common feature extraction branch unchanged, thereby forming a combined identification network adapted to the specific model.

[0078] Using the above method, different models of switchgear do not need to be reloaded or fully trained on the edge side. Model adaptation can be completed simply by activating the corresponding individual feature extraction branch, which significantly reduces the model deployment and operation costs.

[0079] (ii) Fault identification of real-time monitoring data; Multimodal sensing nodes collect partial discharge signals, temperature signals, and operating condition data of the target switchgear in real time, and transmit the real-time monitoring data to the edge computing device.

[0080] Edge computing devices utilize a combined recognition network to perform inference calculations on real-time monitoring data and output recognition results that include fault type, fault location, fault severity, and corresponding confidence level.

[0081] The confidence level is used to measure the reliability of the current identification result, and its value ranges from 0 to 100%.

[0082] A confidence-based hierarchical decision-making mechanism; In this embodiment, the edge computing device classifies the confidence level of the recognition results output by the combined recognition network, specifically including the following rules: Valid identification determination: When the confidence level of the identification result is greater than 85%, it is determined to be a valid identification, and the current abnormal state identification result is considered to have high reliability. The edge computing device triggers the subsequent secondary loop linkage instruction matching process.

[0083] Undetermined state determination: When the confidence level of the recognition result is greater than or equal to 60% and less than or equal to 85%, it is determined to be in an undetermined state.

[0084] In this state, the edge computing device does not immediately trigger the linkage command, but instead starts the enhanced monitoring mode, increasing the acquisition frequency of real-time monitoring data from 10kHz to 50kHz to obtain higher resolution monitoring data, and continuously performs fault identification calculations to improve the certainty of the identification results.

[0085] Invalid identification criteria: When the confidence level of the identification result is lower than 60%, it is considered an invalid identification.

[0086] In this case, the edge computing device only records and stores the relevant monitoring data without triggering any secondary loop linkage operations, thereby avoiding the risk of malfunction caused by misidentification.

[0087] Through the aforementioned hierarchical decision-making mechanism, while ensuring timely handling of anomalies, the probability of false triggering is effectively reduced, thereby improving the security and stability of system operation.

[0088] Cross-model linkage strategy with precise matching mechanism; Once the identification result is determined to be valid, the edge computing device calls the cross-model linkage strategy adaptation unit to match the linkage command.

[0089] The cross-model linkage strategy adaptation unit has a built-in "fault-model-linkage strategy" mapping library generated and continuously optimized during the training phase of the federated collaborative model. The mapping library covers at least 10 common switchgear models and 5 typical fault types.

[0090] The adaptation unit uses the K-nearest neighbor algorithm for strategy matching based on the fault type, equipment model and fault severity in the effective identification results, where K is 3.

[0091] By calculating the feature similarity between the current recognition result and the historical strategy samples in the mapping library, the three strategy records with the highest similarity are selected, and the optimal linkage instruction is determined by combining their execution effects.

[0092] For example: When the identification result is a poor contact fault, the severity level is 3, and the equipment model is KYN28-12, the matching linkage instructions include reducing the load current by 10%–15% as priority 1 operation, and lowering the overcurrent action threshold of the protection device by 5% as priority 2 operation. When the identification result is a surface discharge fault, the severity level is 2, and the equipment model is XGN66-12, the matching linkage instructions include increasing the sampling frequency of the insulation monitoring device to 100Hz as a priority 1 operation, and starting backup power supply switching preparation as a priority 2 operation. When the identification result is a corona discharge fault, the severity is level 1, and the equipment model is GG-1A(F), only the linkage command to start local ventilation and cooling is matched, and the load or protection parameter adjustment is not executed.

[0093] Through the above methods, differentiated and refined linkage handling of different models of switchgear under abnormal conditions is achieved.

[0094] Command issuance and execution under secure communication; The edge computing device encapsulates the matched secondary circuit linkage commands into data frames that conform to the standard communication protocol of the power system.

[0095] The data frames are sent to the secondary circuit equipment corresponding to the target switchgear via a fiber optic-based data transmission network.

[0096] In this embodiment, the communication process uses an encrypted communication interface that conforms to the IEC61850 standard and supports the MMS protocol to ensure the real-time performance and security of the linkage commands during transmission.

[0097] After receiving the linkage command, the secondary circuit equipment executes the corresponding control action according to the command content.

[0098] Collection and structured processing of linkage feedback data; After the secondary circuit equipment completes the execution of the linkage command, the feedback unit set in the secondary circuit equipment or the switch cabinet body is used to collect the equipment status change data in real time. The data at least includes the current value after load adjustment, the temperature change of the fault location, and the insulation monitoring result.

[0099] The edge computing device associates the equipment status change data with the fault recognition result that triggers this linkage and the issued linkage command information, and generates a structured feedback data record as the linkage feedback data.

[0100] The linkage feedback data is then used in the closed-loop iterative optimization process in step S4.

[0101] Integrated technical effects of Embodiment 3: 1. Achieved cross-model adaptive deployment of the dual-branch fault recognition model on the edge side, significantly improving the applicability of the model in scenarios of multi-model switch cabinets; 2. The confidence-based hierarchical decision-making mechanism effectively balances the timeliness of abnormal handling and the system operation safety, reducing the risk of mis-linkage; 3. The cross-model linkage strategy adaptation unit enables the secondary circuit linkage command to accurately match different equipment characteristics and fault severity, improving the pertinence of abnormal prevention and control; 4. The secure communication and structured feedback mechanism provides a reliable data basis for the continuous optimization of subsequent models and strategies.

[0102] Embodiment 4, this embodiment corresponds to step S4 in Embodiment 1, and focuses on continuously iteratively optimizing the dual-branch fault recognition model and the linkage strategy by using the linkage execution effect during the actual operation of the system.

[0103] Its core working principle is as follows: Associate the linkage execution effect data collected on the edge side with the corresponding fault recognition result and the linkage command, and construct a "recognition-linkage-effect" triple training sample; under the premise of ensuring data privacy, feedback the triple training sample to the federated learning server through the federated learning framework; the federated learning server and each power enterprise node respectively perform differential collaborative optimization on the common feature extraction branch and the individual feature extraction branch of the dual-branch fault recognition model, and at the same time perform closed-loop update on the linkage strategy parameters in the strategy mapping library based on the execution effect evaluation result, so that the system has the ability to continuously adaptively evolve with the change of operating conditions.

[0104] Construction and privacy processing of linkage feedback data; After the secondary circuit equipment completes the execution of the linkage command, the edge terminals of each power company uniformly process the linkage execution data transmitted back.

[0105] The data for coordinated execution should include at least the following: The fault identification results that trigger the linkage (including fault type, fault location, fault severity and confidence level); The actual secondary circuit linkage commands and their parameters that are issued and executed; Data on equipment status changes after linkage execution, including temperature changes at fault locations, load current changes, and insulation monitoring results.

[0106] The edge device first performs secondary desensitization processing on the above data, removing all information that can be used to identify specific enterprises, specific equipment or operating locations, and retaining only the technical parameters related to fault characteristics, linkage methods and execution effects.

[0107] After anonymization, the data is constructed into training samples of associated "recognition-linkage-effect" triples, where: "Identification" is used to characterize the core fault information output by the model; "Linkage" is used to characterize the secondary loop control action performed in response to the identification result; "Effect" is used to characterize the degree to which abnormal states are suppressed or improved after the linkage is executed.

[0108] By employing the above processing methods, the effectiveness of the feedback data in subsequent training and optimization is ensured, while avoiding the privacy risks associated with directly exposing the runtime data.

[0109] Differentiated collaborative iterative optimization of the dual-branch fault identification model; In this embodiment, the desensitized triplet training samples are fed back to the federated learning server through a federated learning encrypted channel to trigger iterative training of the dual-branch fault identification model.

[0110] The iterative training process is a differentiated collaborative optimization process, which specifically includes the following two aspects: Centralized iterative training of common feature extraction branch: The federated learning server performs centralized iterative training on the common feature extraction branch in the dual-branch fault identification model based on triple training samples from multiple power enterprise nodes.

[0111] In this process, the server focuses on utilizing the common response characteristics of different enterprises and different models of equipment after coordinated execution to optimize and update the general features in the model used to characterize partial discharge and temperature anomalies, so as to improve the overall generalization ability of the model in cross-enterprise and cross-model scenarios.

[0112] Local fine-tuning of the personality feature extraction branch: Each power enterprise node performs local fine-tuning on the personality feature extraction branch in the model based on the subset of triple training samples related to its own equipment model.

[0113] This fine-tuning process mainly optimizes the response differences of specific model switch cabinets after linkage execution, enabling the model to further improve the adaptation accuracy to the equipment characteristics of the enterprise while maintaining global consistency.

[0114] Through the above differential collaborative optimization mechanism, the synchronous evolution of the model at the global and local levels is achieved, avoiding the limitations brought by single centralized training or single local training.

[0115] Closed-loop optimization of the linkage strategy based on execution effect evaluation; In this embodiment, the update of the policy mapping library is completed based on the closed-loop policy optimization mechanism of execution effect evaluation.

[0116] The system pre-sets a linkage effect evaluation threshold to determine whether the execution effect of the linkage instruction reaches the expected prevention and control target. For example, the following validity criteria are set: Within 10 minutes after the execution of the linkage instruction, the temperature drop amplitude at the fault location is greater than or equal to , then it is determined that the execution of the linkage instruction is effective; If the temperature does not drop or shows an upward trend within the above time window, it is determined that the execution of the linkage instruction is ineffective.

[0117] When the effect data in the triple training sample meets the validity criteria, the system uses the recognition-linkage relationship in this triple sample as a reinforcement sample to increase the confidence of the corresponding linkage policy entry in the policy mapping library, thereby increasing the priority of this policy being selected in subsequent matching processes.

[0118] When the effect data does not meet the validity criteria, the system determines it as an abnormal linkage case and automatically triggers the policy correction process.

[0119] In the policy correction process, the system adjusts the linkage instruction parameters based on an optimization algorithm. For example, it searches and corrects parameters such as the load adjustment amplitude and the protection threshold adjustment range through the particle swarm optimization algorithm, optimizes the original load adjustment amplitude from 12% to 15%, and reissues the corrected linkage instruction until the expected prevention and control effect is achieved.

[0120] The corrected linkage instruction parameters and the corresponding abnormal linkage cases are incorporated into the policy mapping library and simultaneously fed back to the federated learning framework as new training samples to trigger targeted fine-tuning of relevant model parameters, thereby continuously improving the system's handling ability under complex working conditions.

[0121] Supporting units and their synergistic effects; To ensure the reliability and security of the closed-loop optimization process in Example 4, the system is also equipped with the following supporting units: Clock synchronization unit: It adopts GPS and Beidou dual-mode time synchronization to perform unified time calibration on federated learning server, local nodes of various power companies, edge computing devices and secondary loop equipment to ensure that the time synchronization accuracy is no more than 1ms.

[0122] By using high-precision clock synchronization, data association errors and misjudgments due to linkage execution delays caused by time deviations are avoided, thus improving the accuracy of triplet training sample construction.

[0123] Security Protection Unit: Firewalls and intrusion detection systems are deployed in the federated learning parameter transmission channel and the secondary loop communication channel to monitor abnormal access and malicious attacks in real time; Meanwhile, digital signature technology is used for model parameters and linkage strategy parameters to prevent them from being tampered with during transmission and storage, thereby ensuring the security and reliability of the model iterative optimization process.

[0124] Overall technical effects of Example 4: 1. It enables the continuous feedback of the linkage execution effect to the fault identification model and linkage strategy, giving the system long-term self-optimization capability; 2. The differentiated collaborative optimization mechanism improves the model's generalization ability while maintaining high adaptability to specific equipment models; 3. The closed-loop strategy optimization based on execution effect evaluation significantly improved the effectiveness of secondary loop linkage prevention and control, and reduced the probability of ineffective linkage and repeated debugging; 4. The supporting clock synchronization and security protection mechanisms provide reliable engineering guarantees for closed-loop iterative optimization.

[0125] A real-time monitoring method for partial discharge and temperature anomalies in switchgear, used for the aforementioned real-time monitoring of partial discharge and temperature anomalies in switchgear, refers to... Figure 2 ,include: Model training collaboration module 1 is used to collaborate with the federated learning server and multiple power enterprise nodes to jointly train a dual-branch fault identification model based on a privacy protection mechanism. Each power enterprise node uses its local de-identified monitoring data to train its local model, generates encrypted model parameters, and uploads them to the federated learning server. The federated learning server aggregates the encrypted model parameters uploaded by all nodes to update the global model parameters and distributes them. The training process of the dual-branch fault identification model is simultaneously combined with the historical secondary circuit data of each node to generate an initial policy mapping library of associated fault information and linkage strategies. Edge recognition and matching module 2, connected to model training collaboration module 1, is used to deploy the trained dual-branch fault recognition model on the edge computing device; the edge computing device collects real-time monitoring data of the target switchgear and calls the model corresponding to the target switchgear model for fault recognition; when the confidence level of the recognition result is higher than the set threshold, the corresponding secondary circuit linkage instruction is matched from the strategy mapping library according to the identified fault information. The instruction execution and feedback interface module 3 is connected to the edge recognition and matching module 2. It is used to send the matched secondary circuit linkage instructions to the secondary circuit equipment of the target switch cabinet for execution through a secure communication link; and to collect the equipment status change data after instruction execution as linkage feedback data. The closed-loop optimization processing module 4 is connected to the instruction execution and feedback interface module 3. It is used to desensitize the linkage feedback data obtained by each node and form feedback training samples. The feedback training samples are sent back to the federated learning server to trigger iterative training of the dual-branch fault identification model and update of linkage strategy parameters of the strategy mapping library.

[0126] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for real-time monitoring of partial discharge and temperature anomalies in switchgear, characterized in that, include: Step S1, Federated Collaborative Model Training: The federated learning server collaborates with multiple power enterprise nodes to jointly train a dual-branch fault identification model based on a privacy protection mechanism. Each power enterprise node uses its local de-identified monitoring data to train its local model, generating encrypted model parameters which are then uploaded to the federated learning server. The federated learning server aggregates the encrypted model parameters uploaded by all nodes to update the global model parameters and distributes them. The training process of the dual-branch fault identification model simultaneously incorporates the historical secondary circuit data of each node to generate an initial policy mapping library of associated fault information and linkage strategies. Step S2, Edge Identification and Command Matching: The trained dual-branch fault identification model is deployed on an edge computing device; the edge computing device collects real-time monitoring data of the target switchgear and calls the model corresponding to the target switchgear model for fault identification; when the confidence level of the identification result is higher than the set threshold, the corresponding secondary circuit linkage command is matched from the strategy mapping library according to the identified fault information. Step S3, Command Execution and Data Feedback: The matched secondary circuit linkage command is sent to the secondary circuit equipment of the target switchgear for execution via a secure communication link; the equipment status change data after command execution is collected as linkage feedback data; Step S4, closed-loop iterative optimization: the linkage feedback data obtained by each node is desensitized to form feedback training samples; the feedback training samples are sent back to the federated learning server to trigger iterative training of the dual-branch fault identification model and updating of linkage policy parameters of the policy mapping library.

2. The method for real-time monitoring of partial discharge and temperature anomalies in switchgear according to claim 1, characterized in that: In step S1, the training of the federated collaborative model is performed under a horizontal federated learning architecture, which includes a federated learning center node and at least two local nodes of power companies. Each of the power company's local nodes uses de-identified local multimodal monitoring data to train a local model copy, and only uploads the encrypted model parameters generated during training to the federated learning center node through a secure communication channel. The federated learning center node aggregates the received encrypted model parameters to update the global model parameters and distribute them to each local node. The joint training process synchronously inputs the secondary circuit equipment parameters and historical linkage data of each node, so that the trained dual-branch fault identification model has the ability to identify faults across models and generates the initial strategy mapping library.

3. The method for real-time monitoring of partial discharge and temperature anomalies in switchgear according to claim 2, characterized in that: The dual-branch fault identification model specifically includes a common feature extraction branch, a unique feature extraction branch, and a fusion output layer based on an attention mechanism; The common feature extraction branch is collaboratively optimized by the federated learning center nodes based on model parameters uploaded by multiple nodes, and is used to learn the common features of faults in different types of switchgear. The common features include predetermined signal frequency band features and temperature change features. The personalized feature extraction branch is optimized locally by each power company's local node based on its own local data of specific model switchgear, in order to learn the personalized fault characteristics of that specific model of equipment. The fusion output layer is configured to dynamically fuse the general features output by the common feature extraction branch and the personalized features output by the individual feature extraction branch, and output an identification result containing core fault information and related secondary loop linkage strategy suggestions. The core fault information includes at least the fault type, location, and severity.

4. The method for real-time monitoring of partial discharge and temperature anomalies in switchgear according to claim 1, characterized in that: The step S2, which involves calling a model that matches the target switchgear model for fault identification, specifically includes: A model identification unit is deployed in the edge computing device, which obtains the model information of the target switchgear by reading the equipment file of the target switchgear. Based on the acquired model information, the corresponding individual feature extraction branch is dynamically configured and activated from the deployed dual-branch fault identification model to form a combined identification network of common feature extraction branch adapted to the specific model and activated individual feature extraction branch, and to identify faults in real-time monitoring data.

5. The method for real-time monitoring of partial discharge and temperature anomalies in switchgear according to claim 4, characterized in that: The process of matching and generating secondary loop linkage instructions in step S2 further includes a confidence-based hierarchical decision-making mechanism, which includes: The edge computing device determines the confidence level of the recognition result output by the combined recognition network; If the confidence level is higher than the first confidence level threshold, it is determined to be a valid identification and triggers the subsequent linkage instruction matching process; If the confidence level is lower than the first confidence level threshold but higher than the second confidence level threshold, it is determined to be in a pending state, and the enhanced monitoring mode is activated. The enhanced monitoring mode includes increasing the collection frequency of the real-time monitoring data. If the confidence level is lower than the second confidence level threshold, it is determined to be an invalid identification, and only data is recorded without triggering subsequent linkage processes.

6. The method for real-time monitoring of partial discharge and temperature anomalies in switchgear according to claim 1, characterized in that: The process of issuing the secondary circuit linkage command through the secure communication link in step S3 specifically includes: The matched secondary circuit linkage command is encapsulated into a data frame conforming to the standard communication protocol of the power system. The data frames are sent to the secondary circuit equipment corresponding to the target switchgear via a data transmission network based on optical fiber.

7. The method for real-time monitoring of partial discharge and temperature anomalies in switchgear according to claim 6, characterized in that: The process of collecting the linkage feedback data in step S3 specifically includes: The feedback unit located in the secondary circuit equipment or switch cabinet body collects equipment status change data after the execution of the secondary circuit linkage command. The device status change data is associated with the fault identification result that triggered the linkage and the issued linkage command information to generate a structured feedback data record, which serves as the linkage feedback data.

8. The method for real-time monitoring of partial discharge and temperature anomalies in switchgear according to claim 1, characterized in that: In step S4, the linkage feedback data is constructed as an associated recognition-linkage-effect triplet training sample. The triggering of iterative training of the dual-branch fault identification model is specifically a differentiated collaborative optimization process, which includes: The federated learning server performs centralized iterative training on the common feature extraction branch of the model based on the triplet training samples from multiple nodes to optimize its general fault representation capability. Each power enterprise node performs localized fine-tuning of the model's individual feature extraction branch based on the triplet training sample subset related to its own equipment model, in order to optimize its adaptation accuracy for specific models.

9. The method for real-time monitoring of partial discharge and temperature anomalies in switchgear according to claim 8, characterized in that: The step S4, which involves updating the linkage strategy parameters of the strategy mapping library, specifically includes a closed-loop strategy optimization mechanism based on execution performance evaluation. Based on the performance data in the triplet training samples, evaluate whether the execution effect of the corresponding linkage instruction meets the preset effectiveness standard. If the validity criteria are met, the recognition-linkage relationship in the triplet training sample is used as a reinforcement sample to improve the confidence of the corresponding policy entry in the policy mapping library. If the validity criteria are not met, it is determined to be an abnormal linkage case, triggering the policy correction process: generating corrected linkage instruction parameters based on the optimization algorithm to update the policy mapping library, and using the abnormal linkage case and the corrected instruction parameters as new training samples, feeding them back to the federated learning framework to trigger targeted fine-tuning of the relevant models.

10. A real-time monitoring method for partial discharge and temperature anomalies in switchgear, used to implement the real-time monitoring method for partial discharge and temperature anomalies in switchgear as described in any one of claims 1-9, characterized in that, include: The model training collaboration module is used to collaborate with the federated learning server and multiple power enterprise nodes to jointly train a dual-branch fault identification model based on a privacy protection mechanism. Each power enterprise node uses its local de-identified monitoring data to train its local model, generating encrypted model parameters which are then uploaded to the federated learning server. The federated learning server aggregates the encrypted model parameters uploaded by all nodes to update the global model parameters and distributes them. The training process of the dual-branch fault identification model synchronously incorporates historical secondary circuit data from each node to generate an initial policy mapping library of associated fault information and linkage strategies. An edge recognition and matching module, connected to the model training collaboration module, is used to deploy the trained dual-branch fault recognition model on an edge computing device; the edge computing device collects real-time monitoring data of the target switchgear and calls the model corresponding to the target switchgear model for fault recognition; when the confidence level of the recognition result is higher than a set threshold, the corresponding secondary loop linkage instruction is matched from the strategy mapping library based on the identified fault information. The instruction execution and feedback interface module is connected to the edge recognition and matching module. It is used to send the matched secondary circuit linkage instruction to the secondary circuit equipment of the target switch cabinet for execution through a secure communication link; and to collect the equipment status change data after instruction execution as linkage feedback data. The closed-loop optimization processing module connects the instruction execution and feedback interface module and the model training collaboration module. It is used to desensitize the linkage feedback data obtained by each node to form feedback training samples. The feedback training samples are then sent back to the federated learning server to trigger iterative training of the dual-branch fault identification model and updating of the linkage policy parameters of the policy mapping library.

Images