A power distribution automation terminal defect handling system
By using deep learning models and optimization algorithms, combined with multi-source data and federated learning, the problem of early detection and resource optimization of latent defects in FTU batteries was solved, enabling intelligent diagnosis and efficient handling of distribution automation terminals, and forming a self-optimizing closed-loop management system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUANGGANG POWER SUPPLY COMPANY HUBEI ELECTRIC POWER
- Filing Date
- 2026-02-13
- Publication Date
- 2026-06-19
AI Technical Summary
In existing technologies, when the batteries of distribution automation terminals (FTUs) are operating in harsh outdoor environments, there is a lack of early detection and intelligent analysis of latent performance degradation, which leads to the equipment "operating with defects", affecting reliability and making it difficult to achieve dynamic optimal allocation and closed-loop management of resources.
We employ deep learning models based on LSTM and other structures to automatically diagnose latent defects by combining multi-source features. We also generate optimal handling solutions through optimization algorithms, forming a full-process digital closed loop of early warning, scheduling, execution, and verification. We utilize a federated learning architecture to perform incremental model learning to adapt to equipment aging and new defect patterns.
It enables early detection of latent defects in FTU batteries, improves diagnostic accuracy and handling efficiency, optimizes resource allocation, forms a self-optimizing closed-loop management system, adapts to equipment aging, and protects data privacy.
Smart Images
Figure CN122241462A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of automated operation and maintenance of power systems, and in particular to a defect handling system for distribution automation terminals. Background Technology
[0002] With the rapid development of smart distribution networks, the number of distribution automation terminals (FTUs), as key equipment for achieving rapid fault isolation and recovery, has surged, and their operating environments have become increasingly complex. FTU terminals (hereinafter referred to as FTUs), especially their built-in batteries, are prone to performance degradation when operating in harsh outdoor environments for extended periods. These defects (such as battery capacity decay and increased internal resistance) are latent. Traditional methods relying on regular manual inspections or passively waiting for fault reports suffer from significant delays in defect detection. Failure to detect and address defects in a timely manner leads to terminals operating with defects, severely impacting the reliability of distribution automation functions and even causing power outages.
[0003] Currently, the power industry has some asset or production management systems for equipment management, such as the equipment defect management module in a production management system (PMS) or equipment asset health management systems (EAM) provided by some manufacturers. Although these systems have achieved electronic circulation and process management of defect work orders, their defect detection mostly relies on manual reporting or simple over-limit alarms (such as voltage below a certain fixed threshold). They lack the ability to continuously and intelligently analyze and provide early warnings of equipment status, especially the slow deterioration process of battery performance. In addition, the dispatching function of such systems is usually relatively simple, mostly based on rules or experience and manual dispatching. They cannot achieve global dynamic optimal allocation of resources (human and material resources) under cross-regional and multi-constraint conditions, nor can they effectively feed the on-site handling results back to the diagnostic model to form a self-optimizing closed loop, resulting in poor practicality. Therefore, there is an urgent need for a distribution automation terminal defect handling system to improve the above problems. Summary of the Invention
[0004] To address the aforementioned technical challenges, this invention provides a power distribution automation terminal defect handling system that utilizes a deep learning model based on LSTM and other structures. This system automatically learns the degradation patterns of equipment performance from time-series data, enabling early detection of latent defects such as those in batteries. Furthermore, the model integrates multi-source features, resulting in significantly higher diagnostic accuracy and generalization ability compared to traditional methods relying on fixed thresholds or simple rules. Secondly, based on defect warnings and real-time resource status, the system automatically generates optimal handling solutions through optimization algorithms, greatly improving the efficiency of manpower and material resource allocation. It also forms a fully digital closed loop of warning-scheduling-execution-verification through mobile work orders and on-site feedback, significantly enhancing handling efficiency and management level. Finally, the system continuously and incrementally learns the diagnostic model using on-site handling results, enabling it to adapt to equipment aging and new defect patterns, becoming increasingly intelligent with use. The federated learning architecture employed allows for the aggregation of multi-source data to train a more powerful global model while ensuring data privacy.
[0005] The present invention provides a power distribution automation terminal defect handling system, comprising: Multi-source data acquisition and fusion layer: used to acquire time-series operating data and static attribute data of FTU, and perform standardization processing to construct model input samples; Intelligent Diagnosis and Early Warning Layer: Connected to the multi-source data acquisition and fusion layer, it has a built-in neural network model based on a deep learning framework, which is used to analyze input samples and output FTU health assessment results and defect early warning information. Resource scheduling and decision optimization layer: connected to the intelligent diagnosis and early warning layer, used to generate disposal plans through optimization algorithms based on the defect early warning information and real-time human resources, material inventory and geographical information; Closed-loop execution and model evolution layer: It interacts with the resource scheduling and decision optimization layer and the on-site mobile terminal respectively, and is used to push the disposal plan, track and record the disposal process, and feed back the disposal result data to the intelligent diagnosis and early warning layer for updating the neural network model.
[0006] Preferably, the multi-source data acquisition and fusion layer includes a dedicated battery monitoring module: this module is implemented through a wireless sensing node that integrates a voltage probe and an internal resistance measurement circuit, and the voltage probe and internal resistance measurement circuit are connected to the FTU battery electrodes through an aviation plug with a contact resistance of less than 2.0mΩ.
[0007] Preferably, the neural network model in the intelligent diagnosis and early warning layer is a hybrid neural network structure, including: Long Short-Term Memory Network Module: Used to extract features from the time-series running data; Multilayer perceptron module: used to extract features from the static attribute data of the device; Feature fusion and decision module: used to fuse the features output by the LSTM module and the MLP module, and output the health assessment result and defect classification probability.
[0008] Preferably, the neural network model is built using the PyTorch or TensorFlow framework, and trained using a weighted combination of the mean squared error loss function and the cross-entropy loss function as the objective function.
[0009] Preferably, the long short-term memory network module integrates an attention mechanism to evaluate the importance weights of data at different time points in the input time-series data.
[0010] Preferably, the intelligent diagnosis and early warning layer also includes a federated learning client: used to train a local neural network model based on local data, and to encrypt and upload the trained model parameters to the federated learning server for secure aggregation in order to update the global model.
[0011] Preferably, the resource scheduling and decision optimization layer is also connected to the meteorological information system, and the optimization algorithm will avoid the severe weather areas indicated by the meteorological warning information when generating the emergency response plan.
[0012] Preferably, the closed-loop execution and model evolution layer includes a model incremental learning unit: used to periodically perform incremental fine-tuning of the neural network model using the feedback processing result data and its corresponding original input samples.
[0013] Preferably, the closed-loop execution and model evolution layer further includes a mobile work order module: used to push structured work orders containing navigation information, operation steps and spare parts list to the mobile terminals of operation and maintenance personnel, and to receive feedback on on-site handling results.
[0014] Preferably, the method for handling defects in distribution automation terminals according to the distribution automation terminal defect handling system includes the following steps: S1. Data Acquisition and Preprocessing Steps: Acquire multi-source operating data and static attribute data of the FTU. The multi-source operating data includes battery timing data acquired through wireless sensor nodes deployed on-site. The wireless sensor nodes integrate voltage probes and internal resistance measurement circuits and are connected to the FTU battery electrodes through aviation plugs with a contact resistance of less than 2.0mΩ. S2. Intelligent Diagnosis and Early Warning Steps: Input samples are fed into a pre-trained deep learning neural network model. The model is a hybrid neural network built on the PyTorch or TensorFlow framework, including a Long Short-Term Memory network module for extracting features from time-series data, a multilayer perceptron module for extracting static attribute features, and a feature fusion and decision module for feature fusion and decision-making. The LSTM module integrates an attention mechanism to evaluate the importance weights of data at different time points in the input time-series data. The model outputs the health score of the FTU and the multi-label defect classification probability, and generates graded early warning information based on the health score and defect probability. S3. Optimize scheduling steps: Based on the graded early warning information, combined with the real-time acquired status of the maintenance team, spare parts inventory information, geographical information and accessed meteorological early warning information, the operation research optimization algorithm generates a response plan. The optimization algorithm avoids the severe weather areas indicated by the meteorological early warning information during the calculation process. S4. Work order execution and closed-loop steps: The handling plan is pushed to the mobile terminal of the designated maintenance personnel in the form of a structured work order through the mobile work order application. The work order includes navigation information, operation steps and spare parts list, and receives on-site handling result feedback from the mobile terminal, including the confirmed defect type and handling completion information. S5. Model evolution steps: Associate the processing results with their corresponding original input samples to form high-quality labeled training samples; Based on the training samples, the deep learning neural network model is updated using at least one of the following methods: a. Incrementally fine-tune the model locally using the aforementioned samples; b. Under the federated learning architecture, the model parameters updated based on local samples are encrypted and uploaded to the central server for secure aggregation in order to update the global model.
[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By using deep learning models based on LSTM and other structures, the system can automatically learn the degradation patterns of equipment performance from time series data, enabling early detection of latent defects such as batteries. At the same time, the model integrates multi-source features, and its diagnostic accuracy and generalization ability are significantly better than traditional methods that rely on fixed thresholds or simple rules. 2. Based on defect warnings and real-time resource status, the system automatically generates optimal handling solutions through optimization algorithms, significantly improving the efficiency of manpower and material resource allocation. Furthermore, through mobile work orders and on-site feedback, it forms a fully digital closed loop of warning-scheduling-execution-verification, significantly improving handling efficiency and management level. 3. The system uses on-site handling results to continuously and incrementally learn the diagnostic model, enabling it to adapt to equipment aging and new defect patterns, becoming increasingly intelligent with use. The federated learning architecture adopted can aggregate data from multiple sources to train a more powerful global model while ensuring data privacy. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the structure of the power distribution automation terminal defect handling system of the present invention; Figure 2 This is a schematic diagram of the neural network model in the intelligent diagnosis and early warning layer of this invention; Figure 3 This is a schematic diagram of the closed-loop execution and model evolution layer of the present invention. Detailed Implementation
[0017] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. The present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.
[0018] Example: A power distribution automation terminal defect handling system, comprising: Multi-source data acquisition and fusion layer: used to acquire time-series operating data and static attribute data of FTU, and perform standardization processing to construct model input samples; Intelligent Diagnosis and Early Warning Layer: Connected to the multi-source data acquisition and fusion layer, it has a built-in neural network model based on a deep learning framework, which is used to analyze input samples and output FTU health assessment results and defect early warning information. Resource scheduling and decision optimization layer: connected to the intelligent diagnosis and early warning layer, used to generate disposal plans through optimization algorithms based on the defect early warning information and real-time human resources, material inventory and geographical information; Closed-loop execution and model evolution layer: It interacts with the resource scheduling and decision optimization layer and the on-site mobile terminal respectively, and is used to push the disposal plan, track and record the disposal process, and feed back the disposal result data to the intelligent diagnosis and early warning layer to update the neural network model; The multi-source data acquisition and fusion layer includes a dedicated battery monitoring module: this module is implemented through a wireless sensing node that integrates a voltage probe and an internal resistance measurement circuit, and the voltage probe and internal resistance measurement circuit are connected to the FTU battery electrodes through an aviation plug with a contact resistance of less than 2.0mΩ. The neural network model in the intelligent diagnosis and early warning layer is a hybrid neural network structure, including: Long Short-Term Memory Network Module: Used to extract features from the time-series running data; Multilayer perceptron module: used to extract features from the static attribute data of the device; Feature fusion and decision module: used to fuse the features output by the LSTM module and the MLP module, and output the health assessment result and defect classification probability; The neural network model is built using the PyTorch or TensorFlow framework and trained using a weighted combination of the mean squared error loss function and the cross-entropy loss function as the objective function. The long short-term memory network module integrates an attention mechanism to evaluate the importance weights of data at different time points in the input time-series data; The intelligent diagnosis and early warning layer also includes a federated learning client: used to train a local neural network model based on local data, and to encrypt and upload the trained model parameters to the federated learning server for secure aggregation in order to update the global model; The resource scheduling and decision optimization layer is also connected to the meteorological information system. When generating the emergency response plan, the optimization algorithm will avoid the severe weather areas indicated by the meteorological warning information. The closed-loop execution and model evolution layer includes a model incremental learning unit: used to periodically perform incremental fine-tuning of the neural network model using the feedback processing result data and its corresponding original input samples; The closed-loop execution and model evolution layer also includes a mobile work order module: used to push structured work orders containing navigation information, operation steps and spare parts list to the mobile terminals of operation and maintenance personnel, and to receive feedback on on-site handling results; The method for handling defects in distribution automation terminals according to the distribution automation terminal defect handling system includes the following steps: S1. Data Acquisition and Preprocessing Steps: Acquire multi-source operating data and static attribute data of the FTU. The multi-source operating data includes battery timing data acquired through wireless sensor nodes deployed on-site. The wireless sensor nodes integrate voltage probes and internal resistance measurement circuits and are connected to the FTU battery electrodes through aviation plugs with a contact resistance of less than 2.0mΩ. S2. Intelligent Diagnosis and Early Warning Steps: Input samples are fed into a pre-trained deep learning neural network model. The model is a hybrid neural network built on the PyTorch or TensorFlow framework, including a Long Short-Term Memory network module for extracting features from time-series data, a multilayer perceptron module for extracting static attribute features, and a feature fusion and decision module for feature fusion and decision-making. The LSTM module integrates an attention mechanism to evaluate the importance weights of data at different time points in the input time-series data. The model outputs the health score of the FTU and the multi-label defect classification probability, and generates graded early warning information based on the health score and defect probability. S3. Optimize scheduling steps: Based on the graded early warning information, combined with the real-time acquired status of the maintenance team, spare parts inventory information, geographical information and accessed meteorological early warning information, the operation research optimization algorithm generates a response plan. The optimization algorithm avoids the severe weather areas indicated by the meteorological early warning information during the calculation process. S4. Work order execution and closed-loop steps: The handling plan is pushed to the mobile terminal of the designated maintenance personnel in the form of a structured work order through the mobile work order application. The work order includes navigation information, operation steps and spare parts list, and receives on-site handling result feedback from the mobile terminal, including the confirmed defect type and handling completion information. S5. Model evolution steps: Associate the processing results with their corresponding original input samples to form high-quality labeled training samples; Based on the training samples, the deep learning neural network model is updated using at least one of the following methods: a. Incrementally fine-tune the model locally using the aforementioned samples; b. Under the federated learning architecture, the model parameters updated based on local samples are encrypted and uploaded to the central server for secure aggregation in order to update the global model.
[0019] It mainly includes four logical levels: Multi-source data acquisition and fusion layer: Battery data is periodically acquired by intelligent sensing units (integrating high-precision voltage and internal resistance measurement circuits and using aviation plug interfaces with contact resistance <2.0mΩ) deployed at the FTU site and uploaded via LoRa communication. At the same time, FTU telemetry and teleindication data are obtained from the power distribution master station. This layer aligns and normalizes all data according to a unified time reference to form a standard sample containing time sequence and static tags. Intelligent Diagnosis and Early Warning Layer: Deployed on a cloud server, its core model is built using PyTorch. The model input consists of a window of time-series data from the past 30 days and a static vector. The LSTM time-series network is a two-layer stacked structure that captures performance degradation trends. The MLP static network has two layers that process device attributes. The output features of both are concatenated and then fed into a feature fusion and decision layer with Dropout. The final output is a health score (0-100) and the probability of six types of defects, such as battery failure and communication interruption. When "battery failure probability > 0.8 and score < 60", a serious warning is generated. Resource scheduling and decision optimization layer: Real-time acquisition of personnel location, skills, work queues, and warehouse spare parts inventory for each team. Upon receiving an alert, with the goal of minimizing the total handling time, an improved genetic algorithm is used to solve the VRPTW model, assigning the most suitable team and planning routes for each defect, and generating an electronic part number. Closed-loop execution and model evolution layer: Operation and maintenance personnel receive work orders through a mobile APP, follow the navigation to the location, scan the code to confirm the equipment, and handle it according to the standard operating procedure. After the handling is completed, they upload the result photos and the confirmed defect type. The system automatically stores the case (from raw data to final confirmation result) in a high-quality sample library. Every week, the system uses new samples to perform incremental fine-tuning on the diagnostic model to continuously improve the model's ability to identify new defect patterns. The entire process forms a data-driven intelligent closed loop. In addition, the initial training and regular global updates of the model adopt federated learning. Each provincial company uses local historical data to train local models, and uploads the model parameters to the State Grid Cloud in encryption every month for secure aggregation. After generating a stronger global model, it is distributed to each node.
[0020] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A power distribution automation terminal defect handling system, characterized in that, include: Multi-source data acquisition and fusion layer: used to acquire time-series operating data and static attribute data of FTU, and perform standardization processing to construct model input samples; Intelligent Diagnosis and Early Warning Layer: Connected to the multi-source data acquisition and fusion layer, it has a built-in neural network model based on a deep learning framework, which is used to analyze input samples and output FTU health assessment results and defect early warning information. Resource scheduling and decision optimization layer: connected to the intelligent diagnosis and early warning layer, used to generate disposal plans through optimization algorithms based on the defect early warning information and real-time human resources, material inventory and geographical information; Closed-loop execution and model evolution layer: It interacts with the resource scheduling and decision optimization layer and the on-site mobile terminal respectively, and is used to push the disposal plan, track and record the disposal process, and feed back the disposal result data to the intelligent diagnosis and early warning layer for updating the neural network model.
2. The power distribution automation terminal defect handling system as described in claim 1, characterized in that, The multi-source data acquisition and fusion layer includes a dedicated battery monitoring module: this module is implemented through a wireless sensing node that integrates a voltage probe and an internal resistance measurement circuit, and the voltage probe and internal resistance measurement circuit are connected to the FTU battery electrodes through an aviation plug with a contact resistance of less than 2.0mΩ.
3. The power distribution automation terminal defect handling system as described in claim 1, characterized in that, The neural network model in the intelligent diagnosis and early warning layer is a hybrid neural network structure, including: Long Short-Term Memory Network Module: Used to extract features from the time-series running data; Multilayer perceptron module: used to extract features from the static attribute data of the device; Feature fusion and decision module: used to fuse the features output by the LSTM module and the MLP module, and output the health assessment result and defect classification probability.
4. The power distribution automation terminal defect handling system as described in claim 3, characterized in that, The neural network model is built using the PyTorch or TensorFlow framework and trained using a weighted combination of the mean squared error loss function and the cross-entropy loss function as the objective function.
5. A power distribution automation terminal defect handling system as described in claim 3, characterized in that, The long short-term memory network module integrates an attention mechanism to evaluate the importance weights of data at different time points in the input time-series data.
6. The power distribution automation terminal defect handling system as described in claim 1, characterized in that, The intelligent diagnosis and early warning layer also includes a federated learning client: used to train a local neural network model based on local data, and to encrypt and upload the trained model parameters to the federated learning server for secure aggregation in order to update the global model.
7. The power distribution automation terminal defect handling system as described in claim 1, characterized in that, The resource scheduling and decision optimization layer is also connected to the meteorological information system. When generating emergency response plans, the optimization algorithm will avoid the severe weather areas indicated by the meteorological warning information.
8. The power distribution automation terminal defect handling system as described in claim 1, characterized in that, The closed-loop execution and model evolution layer includes a model incremental learning unit: used to periodically perform incremental fine-tuning of the neural network model using the feedback processing result data and its corresponding original input samples.
9. A power distribution automation terminal defect handling system as described in claim 8, characterized in that, The closed-loop execution and model evolution layer also includes a mobile work order module: used to push structured work orders containing navigation information, operation steps and spare parts list to the mobile terminals of operation and maintenance personnel, and to receive feedback on on-site handling results.
10. A power distribution automation terminal defect handling system as described in claim 9, characterized in that, The method for handling defects in distribution automation terminals based on the distribution automation terminal defect handling system includes the following steps: S1. Data Acquisition and Preprocessing Steps: Acquire multi-source operating data and static attribute data of the FTU. The multi-source operating data includes battery timing data acquired through wireless sensor nodes deployed on-site. The wireless sensor nodes integrate voltage probes and internal resistance measurement circuits and are connected to the FTU battery electrodes through aviation plugs with a contact resistance of less than 2.0mΩ. S2. Intelligent Diagnosis and Early Warning Steps: Input samples are fed into a pre-trained deep learning neural network model. The model is a hybrid neural network built on the PyTorch or TensorFlow framework, including a Long Short-Term Memory network module for extracting features from time-series data, a multilayer perceptron module for extracting static attribute features, and a feature fusion and decision module for feature fusion and decision-making. The LSTM module integrates an attention mechanism to evaluate the importance weights of data at different time points in the input time-series data. The model outputs the health score of the FTU and the multi-label defect classification probability, and generates graded early warning information based on the health score and defect probability. S3. Optimize scheduling steps: Based on the graded early warning information, combined with the real-time acquired status of the maintenance team, spare parts inventory information, geographical information and accessed meteorological early warning information, the operation research optimization algorithm generates a response plan. The optimization algorithm avoids the severe weather areas indicated by the meteorological early warning information during the calculation process. S4. Work order execution and closed-loop steps: The handling plan is pushed to the mobile terminal of the designated maintenance personnel in the form of a structured work order through the mobile work order application. The work order includes navigation information, operation steps and spare parts list, and receives on-site handling result feedback from the mobile terminal, including the confirmed defect type and handling completion information. S5. Model evolution steps: Associate the processing results with their corresponding original input samples to form high-quality labeled training samples; Based on the training samples, the deep learning neural network model is updated using at least one of the following methods: a. Incrementally fine-tune the model locally using the aforementioned samples; b. Under the federated learning architecture, the model parameters updated based on local samples are encrypted and uploaded to the central server for secure aggregation in order to update the global model.