Familial defect prediction system and method based on relay protection device

Abstract

The invention discloses a familial defect prediction system and method based on a relay protection device, and belongs to the technical field of electrical equipment monitoring. Comprising an acquisition processing module, a container management module, a distribution processing module, a cooperative training module, a defect prediction module, a prediction optimization module, a diagnosis alarm module, a decision support module, a record auditing module and a monitoring interaction module. According to the method, the security problem caused by data leakage or transmission risk can be avoided, the limitation of single site data is broken through, the generalization ability of the model to the familial defects is enhanced, the system operation cost is reduced, the deployment flexibility is improved, and the adaptability to the familial defects of the whole network is improved; misjudgment caused by only depending on static analysis is avoided, prediction errors are reduced, the response capability of the system to a complex defect mode is enhanced, a clear risk level evaluation basis is provided for equipment operation and maintenance, the anti-interference capability of the system is enhanced, and the practicability in a complex power grid environment is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical equipment monitoring, and in particular to a family defect prediction system and method based on a relay protection device. Background Art

[0002] In modern power grids, relay protection devices are important equipment to ensure the safe and stable operation of power systems. Their main tasks are to monitor the operating status of the system, detect faults in a timely manner, and take protective measures. However, due to the design complexity of relay protection devices and the diversity of working environments, such devices may form family defects due to similar design defects or common operating environments. These defects are often difficult to detect in the early stages, but after accumulating to a certain extent, they will cause serious power grid accidents. Traditional equipment status monitoring and defect prediction methods usually rely on data analysis of a single device and cannot effectively capture the family association characteristics between devices. In addition, with the increasing demand for data privacy protection, data sharing between different substations or regions faces huge challenges, which further limits the joint analysis and prediction of family defects.

[0003] The existing family defect prediction system and method based on relay protection devices have many security issues due to data leakage or transmission risks. The model has poor generalization ability for family defects, which increases the system operation cost and reduces deployment flexibility. In addition, the existing family defect prediction system and method based on relay protection devices have misjudgments caused by relying only on static analysis, and the system has a slow response capability to complex defect modes, and cannot provide a clear risk level assessment basis for equipment operation and maintenance. For this reason, we propose a family defect prediction system and method based on relay protection devices. Summary of the invention

[0004] The purpose of the present invention is to solve the defects in the prior art and to propose a family defect prediction system and method based on a relay protection device.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The family defect prediction system based on relay protection device includes acquisition processing module, container management module, distribution processing module, collaborative training module, defect prediction module, prediction optimization module, diagnosis alarm module, decision support module, record audit module and monitoring interaction module;

[0007] The acquisition and processing module is used to collect the operation data of the relay protection device in real time by using the Internet of Things technology and the sensor network, and to pre-process the collected data;

[0008] The container management module is used to isolate the operating environments of different modules and provide an independent operating environment for each analysis module;

[0009] The distributed processing module is used for distributed storage and processing of each group of historical operation data;

[0010] The collaborative training module is used to integrate data of relay protection devices in multiple substations or regions and conduct collaborative model training at the same time;

[0011] The defect prediction module is used to predict the defects of the relay protection device;

[0012] The prediction optimization module is used to optimize the parameters of the defect prediction module;

[0013] The diagnostic alarm module is used to predict the occurrence of future defects in the equipment, monitor the operating status of the relay protection device in real time, trigger an alarm and generate a diagnostic report when a potential defect is found;

[0014] The decision support module is used to provide a variety of optimization decision solutions based on the prediction results, including equipment maintenance plans, upgrade suggestions or operation adjustment plans;

[0015] The recording and auditing module is used to record system operation logs and user access behaviors;

[0016] The monitoring interaction module is used to display the device operation status, prediction results and decision suggestions in the form of interactive charts.

[0017] As a further solution of the present invention, the specific steps of isolating the operating environments of different modules by the container management module are as follows:

[0018] S1.1: For each module, create a dedicated Docker image file and define the software, libraries, dependencies, and run configuration required by the module in the Dockerfile. Then use the formula Manage images in a hierarchical manner, where I represents the total size of the image and L represents the total size of the image. i represents the size of the i-th layer of the image, and n represents the total number of layers of the image;

[0019] S1.2: Define the container operating environment of each module in the Docker Compose file, including the network, mounted volume path, port mapping, and environment variables. At the same time, estimate the demand for system network resources for container deployment and isolate each group of modules through virtual networks.

[0020] S1.3: In the startup script of each module, use environment variables or configuration files to define the service addresses of other modules, and use the built-in network function of Docker to communicate between modules. Then allocate CPU and memory resources to each container, and use the Docker cpus and memory parameters to limit the CPU and memory resources occupied by each container.

[0021] S1.4: Use cgroups technology to monitor the resource usage of each container, and use Docker Swarm or Kubernetes group orchestration tools to automatically deploy, expand and manage each group of containers, set the number of copies and load balancing strategy for each container, monitor the running status of the container in real time, and use Docker's logging mechanism to centrally store the running logs of each module.

[0022] As a further solution of the present invention, the specific steps of the collaborative model training of the collaborative training module are as follows:

[0023] S2.1: The central server generates an initial global model w0 based on a hybrid deep learning model of convolutional neural network and recurrent neural network, and distributes it to all participating nodes, i.e., substations. Each node then loads the initial model locally, where w0 = RandomInit(θ), w0 represents the parameter vector of the initial model, RandomInit() represents the random initialization function, and θ represents the parameter set used for random initialization;

[0024] S2.2: Each node receives the initial model distributed by the central server, loads it locally for training, and initializes the model parameters. After that, the node loads the collected device operation data and real label data from the local storage, and normalizes, cleans, and batches the data. Each node sets a local loss function based on its own task objectives.

[0025] S2.3: The node uses the processed running data and label data to train the initial model. The initial model obtains the predicted value of the running data through the forward propagation algorithm, and calculates the loss value between the predicted value and the real label data through the local loss function, and calculates the partial derivative of the loss value with respect to the model parameters to obtain the gradient of each parameter.

[0026] S2.4: The node uses the gradient descent method to iteratively update the model parameters. After each update, the model parameters are adjusted in the direction of reducing the loss function. Through multiple iterative training, until the model loss value converges to the preset range, the model parameters after multiple rounds of training and update are saved as w i , and encrypt it and upload it to the central server;

[0027] S2.5: The central server collects the encrypted model parameters uploaded by all nodes, decrypts them, aggregates the model parameters uploaded by each node through the weighted average method, generates new global model parameters, and selects the test data set to perform performance evaluation on the final aggregated global model;

[0028] S2.6: If the performance evaluation result does not meet the preset indicators, the central server will send the global model after the parameter update in this round to all participating nodes as the initial model for the next round of local training. Each node repeats the above local training and uploading process until the model performance reaches the preset standard.

[0029] As a further solution of the present invention, the specific steps of the prediction optimization module optimizing the parameters of the defect prediction module are as follows:

[0030] S3.1: Based on all the hyperparameters to be optimized of the hybrid deep learning model, define the search space Θ of the optimization problem, initialize multiple groups of search entities, each of which randomly selects a solution θ∈Θ as the starting point, and then initialize the pheromone concentration of each node to indicate the quality of the hyperparameter combination;

[0031] S3.2: Calculate the heuristic information of each node transfer based on the potential for improving model performance, and then use the pheromone concentration and heuristic information to Calculate the selection probability of each group of selectable paths for the current node of the search body, and each group of search bodies selects the next hyperparameter combination according to the current solution path, where P ij represents the probability of moving from the current node i to node j, τ ij represents the pheromone concentration moving from node i to node j, η ij represents the heuristic information for moving from node i to node j, τ ik represents the pheromone concentration moving from node i to node k, η ik represents the heuristic information of moving from node i to node k, α and β represent the importance weights of regulating pheromone and heuristic information respectively;

[0032] S3.3: Each search entity configures the model according to the selected hyperparameter combination and trains the hybrid deep learning model, evaluates the model performance using the validation set, calculates the evaluation index, and updates the pheromone concentration of each group of nodes based on the calculated evaluation index;

[0033] S3.4: Repeat the process of path selection, model training and evaluation, and pheromone update multiple times until the set maximum number of iterations is reached or the evaluation index is improved to converge to the preset value. After the iteration, the path with the highest pheromone concentration is output as the optimal hyperparameter combination of the hybrid deep learning model and replaces the original hybrid learning model hyperparameters.

[0034] Based on the family defect prediction method of relay protection device, the specific steps of the prediction method are as follows:

[0035] Ⅰ. Use sensors and data loggers installed in relay protection devices to collect operation data in real time and integrate historical fault records of different substations or regions;

[0036] II. Preprocess the running data through edge computing devices and use deep learning autoencoders to extract key features of the data;

[0037] III. Use a data distributed processing framework to process large-scale data in parallel and build a hybrid deep learning model to predict family defects of relay protection devices;

[0038] IV. Optimize the hyperparameters of the hybrid deep learning model. The central server aggregates the parameters uploaded by the local models of each substation and updates the federated learning model.

[0039] V. Predict the collected data in real time, identify whether the equipment has the risk of family defects, analyze the dynamic evolution of the prediction results, and generate future defect patterns and risk probabilities;

[0040] VI. When a high-risk defect is detected, the alarm mechanism is immediately triggered, the maintenance personnel are notified, and a diagnostic report is automatically generated, while providing a variety of optimization decision suggestions.

[0041] As a further solution of the present invention, the specific steps of identifying whether the device has a risk of familial defects in step V are as follows:

[0042] S4.1: Collect real-time operation data from the relay protection device, including current, voltage, frequency, and fault information data, clean and normalize the real-time data, and extract feature information reflecting the performance and status of the equipment from the processed real-time data, including time domain features, frequency domain features, and embedded extraction features;

[0043] S4.2: Load the trained hybrid learning model and input the extracted features into the hybrid learning model. The model outputs the predicted probability of familial defect risk of each group of devices through forward propagation calculation;

[0044] S4.3: Based on the set threshold, the model prediction results are classified into risk levels, and the corresponding diagnosis report is generated. At the same time, the alarm mechanism is triggered and maintenance personnel are also alerted.

[0045] As a further solution of the present invention, the specific steps of analyzing the dynamic evolution of the prediction results in step V and generating future defect modes and risk probabilities are as follows:

[0046] S5.1: Collect all states of the relay protection device and construct the state space of the system, where each state represents the operation status and defect level of the equipment. Then, based on historical data and expert experience, construct a state transition probability matrix to describe the transition probability between states, and the transition probability matrix satisfies that the sum of each row is 1;

[0047] S5.2: Obtain the current state distribution through the prediction results of the hybrid learning model. According to the current state distribution and the transition probability matrix, calculate the state distribution at the next moment through matrix operations. Check whether all elements in the generated state distribution are non-negative and the sum is 1. If not, check the correctness of the transition probability matrix or the current state distribution.

[0048] S5.3: Repeat the prediction iteration to predict the state distribution of multiple time steps in the future, and judge the defect mode evolution trend based on the predicted future state distribution, and select the state with serious defects for key attention, and then calculate the cumulative defect risk probability in the preset future time interval as a comprehensive evaluation indicator of the equipment health status;

[0049] S5.4: Classify the prediction results according to the future state distribution and risk probability, and formulate corresponding maintenance or treatment measures to ensure the safe operation of the equipment.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. The central server of the present invention generates an initial global model based on a hybrid deep learning model of a convolutional neural network and a recurrent neural network, and distributes it to all participating nodes, i.e., substations. After that, each node loads the initial model locally. Each node receives the initial model distributed by the central server, loads it locally for training, and initializes the model parameters. After that, the node loads the collected equipment operation data and real label data from the local storage. Each node sets a local loss function based on its own task objectives. The node uses the processed operation data and label data to train the initial model, and uses the gradient descent method to iteratively update the model parameters. The updated model parameters after multiple rounds of training are encrypted and uploaded to the central server. The central server collects all nodes. The encrypted model parameters uploaded by each node are collected and decrypted, and the model parameters uploaded by each node are aggregated through the weighted average method to generate new global model parameters. The central server selects a test data set and performs performance evaluation on the final aggregated global model. If the performance evaluation result does not meet the preset indicators, the central server will send the global model after the parameter update in this round to all participating nodes as the initial model for the next round of local training. Each node repeats the above local training and uploading process until the model performance reaches the preset standard. This can avoid security issues caused by data leakage or transmission risks, break through the data limitations of a single site, enhance the model's generalization ability for family defects, reduce system operating costs, improve deployment flexibility, and improve the adaptability to family defects in the entire network.

[0052] 2. The present invention collects all states of the relay protection device and constructs the state space of the system, wherein each state represents the operation status and defect level of the equipment. Then, based on historical data and expert experience, a state transition probability matrix is ​​constructed, and the current state distribution is obtained through the prediction result of the hybrid learning model. According to the current state distribution and the transition probability matrix, the state distribution at the next moment is calculated through matrix operation, and it is detected whether all elements in the generated state distribution are non-negative and the sum is 1. If not, the correctness of the transition probability matrix or the current state distribution is checked, and the prediction iteration is repeated to predict the state distribution of multiple time steps in the future. According to the predicted future state distribution, the evolution trend of the defect mode is judged, and the state with serious defects is selected for key attention. Then, the cumulative defect risk probability in the preset future time interval is calculated as a comprehensive evaluation index of the health status of the equipment. According to the future state distribution and risk probability, the prediction results are classified, and corresponding maintenance or treatment measures are formulated to ensure the safe operation of the equipment, avoid misjudgment caused by relying solely on static analysis, reduce prediction errors, enhance the system's responsiveness to complex defect modes, provide a clear risk level assessment basis for equipment operation and maintenance, enhance the system's anti-interference ability, and improve practicality in complex power grid environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0054] Figure 1 This is a system block diagram of the family defect prediction system based on relay protection device proposed by the present invention;

[0055] Figure 2 This is a flowchart of the family defect prediction method based on relay protection device proposed in the present invention. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0057] Example 1

[0058] Reference Figure 1 , a family defect prediction system based on relay protection devices, includes an acquisition and processing module, a container management module, a distributed processing module, a collaborative training module, a defect prediction module, a prediction optimization module, a diagnosis and alarm module, a decision support module, a record audit module and a monitoring interaction module.

[0059] The acquisition and processing module is used to use the Internet of Things technology and sensor networks to collect the operating data of the relay protection device in real time, and pre-process the collected data; the container management module is used to isolate the operating environment of different modules and provide an independent operating environment for each analysis module.

[0060] Specifically, for each module, create a dedicated Docker image file, and define the software, libraries, dependencies, and run configuration required by the module in the Dockerfile, and then use the formula Manage images in a hierarchical manner, where I represents the total size of the image and L represents the total size of the image. i Represents the size of the i-th layer of the image, and n represents the total number of layers of the image. The container operating environment of each module is defined in the Docker Compose file, including the network, mounted volume path, port mapping, and environment variables. At the same time, the demand for system network resources for container deployment is estimated, and each group of modules is isolated through a virtual network. In the startup script of each module, the service address of other modules is defined using environment variables or configuration files, and the modules communicate with each other through the built-in network function of Docker. Then, CPU and memory resources are allocated to each container, and the cpus and memory parameters of Docker are used to limit the CPU and memory resources occupied by each container. The cgroups technology is used to monitor the resource usage of each container, and the orchestration tools of Docker Swarm or Kubernetes are used to automatically deploy, expand, and manage each group of containers, and the number of copies and load balancing strategies of each container are set. The running status of the container is monitored in real time, and the running logs of each module are stored centrally using the logging mechanism of Docker.

[0061] The distributed processing module is used for distributed storage and processing of each group of historical operation data; the collaborative training module is used to integrate the data of relay protection devices in multiple substations or regions, and perform collaborative model training at the same time.

[0062] Specifically, the central server generates an initial global model w0 based on a hybrid deep learning model of convolutional neural networks and recurrent neural networks, and distributes it to all participating nodes, i.e., substations. Each node then loads the initial model locally, where w0 = RandomInit(θ), w0 represents the parameter vector of the initial model, RandomInit() represents the random initialization function, and θ represents the parameter set for random initialization. Each node receives the initial model distributed by the central server, loads it locally for training, and initializes the model parameters. After that, the node loads the collected equipment operation data and real label data from the local storage, and performs training on the data. Normalization, cleaning, batch processing, each node sets a local loss function based on its own task objectives, and the node uses the processed running data and label data to train the initial model. The initial model obtains the predicted value of the running data through the forward propagation algorithm, and calculates the loss value between the predicted value and the real label data through the local loss function. The partial derivative of the loss value with respect to the model parameter is calculated to obtain the gradient of each parameter. The node uses the gradient descent method to iteratively update the model parameters. After each update, the model parameters are adjusted in the direction of reducing the loss function. Through multiple iterative training, until the model loss value converges to the preset range, the model parameters after multiple rounds of training are saved as w i , and encrypt it and upload it to the central server. The central server collects the encrypted model parameters uploaded by all nodes, decrypts them, aggregates the model parameters uploaded by each node through the weighted average method, and generates new global model parameters. The central server selects a test data set and performs performance evaluation on the final aggregated global model. If the performance evaluation result does not meet the preset indicators, the central server will send the global model after the parameter update in this round to all participating nodes as the initial model for the next round of local training. Each node repeats the above local training and uploading process until the model performance reaches the preset standard.

[0063] The defect prediction module is used to predict the defects of the relay protection device; the prediction optimization module is used to optimize the parameters of the defect prediction module.

[0064] Specifically, according to all the hyperparameters to be optimized of the hybrid deep learning model, the search space Θ of the optimization problem is defined, and multiple groups of search bodies are initialized. Each search body randomly selects a solution θ∈Θ as the starting point, and then the pheromone concentration of each node is initialized to indicate the pros and cons of the hyperparameter combination. The heuristic information of each node transfer is calculated according to the potential for improving the model performance. Then, based on the pheromone concentration and heuristic information, Calculate the selection probability of each group of selectable paths for the current node of the search body, and each group of search bodies selects the next hyperparameter combination according to the current solution path, where P ij represents the probability of moving from the current node i to node j, τ ijrepresents the pheromone concentration moving from node i to node j, η ij represents the heuristic information for moving from node i to node j, τ ik represents the pheromone concentration moving from node i to node k, η ik represents the heuristic information of moving from node i to node k, α and β represent the importance weights of regulating pheromone and heuristic information respectively. Each search entity configures the model according to the selected hyperparameter combination and trains the hybrid deep learning model. The model performance is evaluated using the validation set, the evaluation index is calculated, and the pheromone concentration of each group of nodes is updated according to the calculated evaluation index. The process of path selection, model training and evaluation, and pheromone update is repeated many times until the set maximum number of iterations is reached or the evaluation index is improved to converge to the preset value. After the iteration, the path with the highest pheromone concentration is output as the optimal hyperparameter combination of the hybrid deep learning model, and the original hybrid learning model hyperparameters are replaced.

[0065] The diagnostic alarm module is used to predict future defects in the equipment, monitor the operating status of the relay protection device in real time, trigger an alarm and generate a diagnostic report when potential defects are found; the decision support module is used to provide a variety of optimization decision-making plans based on the prediction results, including equipment maintenance plans, upgrade suggestions or operation adjustment plans; the record audit module is used to record system operation logs and user access behaviors; the monitoring interaction module is used to display the device operating status, prediction results and decision suggestions in the form of interactive charts.

[0066] Example 2

[0067] Reference Figure 2 , based on the family defect prediction method of relay protection device, the specific steps of the prediction method are as follows:

[0068] Utilize sensors and data loggers installed in relay protection devices to collect operation data in real time and integrate historical fault records of different substations or regions.

[0069] The running data is preprocessed through edge computing devices, and the key features of the data are extracted using deep learning autoencoders.

[0070] A data distributed processing framework is used to process large-scale data in parallel, and a hybrid deep learning model is constructed to predict familial defects of relay protection devices.

[0071] The hyperparameters of the hybrid deep learning model are optimized. The central server aggregates the parameters uploaded by the local models of each substation and updates the federated learning model.

[0072] Make predictions on the collected data in real time, identify whether the equipment is at risk of familial defects, analyze the dynamic evolution of the prediction results, and generate future defect patterns and risk probabilities.

[0073] Specifically, real-time operating data is collected from the relay protection device, including current, voltage, frequency, and fault information data. The real-time data is then cleaned and normalized, and characteristic information reflecting the performance and status of the equipment is extracted from the processed real-time data, including time domain features, frequency domain features, and embedded extracted features. The trained hybrid learning model is loaded, and the extracted features are input into the hybrid learning model. The model outputs the predicted probability of the family defect risk of each group of equipment through forward propagation calculation. Based on the set threshold, the model prediction results are divided into risk levels, and the corresponding diagnostic report is generated. At the same time, the alarm mechanism is triggered and maintenance personnel are alerted.

[0074] Specifically, all states of the relay protection device are collected, and the state space of the system is constructed, in which each state represents the operation status and defect level of the equipment. Then, based on historical data and expert experience, a state transition probability matrix is ​​constructed to describe the transition probability between states, and the transition probability matrix satisfies that the sum of each row is 1. The current state distribution is obtained through the prediction results of the hybrid learning model. According to the current state distribution and the transition probability matrix, the state distribution at the next moment is calculated through matrix operations. It is detected whether all elements in the generated state distribution are non-negative and the sum is 1. If not, the correctness of the transition probability matrix or the current state distribution is checked, and the prediction iteration is repeated to predict the state distribution of multiple time steps in the future. According to the predicted future state distribution, the defect mode evolution trend is judged, and the state with serious defects is selected for key attention. Then, the accumulated defect risk probability in the preset future time interval is calculated as a comprehensive evaluation indicator of the health status of the equipment. According to the future state distribution and risk probability, the prediction results are classified, and corresponding maintenance or treatment measures are formulated to ensure the safe operation of the equipment.

[0075] When a high-risk defect is detected, the alarm mechanism is immediately triggered, the maintenance personnel are notified, and a diagnostic report is automatically generated, while providing a variety of optimization decision suggestions.

Claims

1. A family defect prediction system based on relay protection device, characterized in that: It includes acquisition and processing module, container management module, distributed processing module, collaborative training module, defect prediction module, prediction optimization module, diagnosis and alarm module, decision support module, record audit module and monitoring interaction module; The acquisition and processing module is used to collect the operation data of the relay protection device in real time by using the Internet of Things technology and the sensor network, and to pre-process the collected data; The container management module is used to isolate the operating environments of different modules and provide an independent operating environment for each analysis module; The distributed processing module is used for distributed storage and processing of each group of historical operation data; The collaborative training module is used to integrate data of relay protection devices in multiple substations or regions and conduct collaborative model training at the same time; The defect prediction module is used to predict the defects of the relay protection device; The prediction optimization module is used to optimize the parameters of the defect prediction module; The diagnostic alarm module is used to predict the occurrence of future defects in the equipment, monitor the operating status of the relay protection device in real time, trigger an alarm and generate a diagnostic report when a potential defect is found; The decision support module is used to provide a variety of optimization decision solutions based on the prediction results, including equipment maintenance plans, upgrade suggestions or operation adjustment plans; The recording and auditing module is used to record system operation logs and user access behaviors; The monitoring interaction module is used to display the device operation status, prediction results and decision suggestions in the form of interactive charts.

2. The family defect prediction system based on relay protection device according to claim 1 is characterized in that: The specific steps of the container management module isolating the operating environment of different modules are as follows: S1.1: For each module, create a dedicated Docker image file and define the software, libraries, dependencies, and run configuration required by the module in the Dockerfile. Then use the formula Manage images in a hierarchical manner, where I represents the total size of the image and L represents the total size of the image. i represents the size of the i-th layer of the image, and n represents the total number of layers of the image; S1.2: Define the container operating environment of each module in the Docker Compose file, including the network, mounted volume path, port mapping, and environment variables. At the same time, estimate the demand for system network resources for container deployment and isolate each group of modules through virtual networks. S1.3: In the startup script of each module, use environment variables or configuration files to define the service addresses of other modules, and use the built-in network function of Docker to communicate between modules. Then allocate CPU and memory resources to each container, and use the Docker cpus and memory parameters to limit the CPU and memory resources occupied by each container. S1.4: Use cgroups technology to monitor the resource usage of each container, and use Docker Swarm or Kubernetes group orchestration tools to automatically deploy, expand and manage each group of containers, set the number of copies and load balancing strategy for each container, monitor the running status of the container in real time, and use Docker's logging mechanism to centrally store the running logs of each module.

3. The family defect prediction system based on relay protection device according to claim 2 is characterized in that: The specific steps of the collaborative model training of the collaborative training module are as follows: S2.1: The central server generates an initial global model w0 based on a hybrid deep learning model of convolutional neural network and recurrent neural network, and distributes it to all participating nodes, i.e., substations. Each node then loads the initial model locally, where w0 = RandomInit(θ), w0 represents the parameter vector of the initial model, RandomInit() represents the random initialization function, and θ represents the parameter set used for random initialization; S2.2: Each node receives the initial model distributed by the central server, loads it locally for training, and initializes the model parameters. After that, the node loads the collected device operation data and real label data from the local storage, and normalizes, cleans, and batches the data. Each node sets a local loss function based on its own task objectives. S2.3: The node uses the processed running data and label data to train the initial model. The initial model obtains the predicted value of the running data through the forward propagation algorithm, and calculates the loss value between the predicted value and the real label data through the local loss function, and calculates the partial derivative of the loss value with respect to the model parameters to obtain the gradient of each parameter. S2.4: The node uses the gradient descent method to iteratively update the model parameters. After each update, the model parameters are adjusted in the direction of reducing the loss function. Through multiple iterative training, until the model loss value converges to the preset range, the model parameters after multiple rounds of training and update are saved as w i , and encrypt it and upload it to the central server; S2.5: The central server collects the encrypted model parameters uploaded by all nodes, decrypts them, aggregates the model parameters uploaded by each node through the weighted average method, generates new global model parameters, and selects the test data set to perform performance evaluation on the final aggregated global model; S2.6: If the performance evaluation result does not meet the preset indicators, the central server will send the global model after the parameter update in this round to all participating nodes as the initial model for the next round of local training. Each node repeats the above local training and uploading process until the model performance reaches the preset standard.

4. The family defect prediction system based on relay protection device according to claim 3 is characterized in that: The specific steps of the prediction optimization module to optimize the parameters of the defect prediction module are as follows: S3.1: Based on all the hyperparameters to be optimized of the hybrid deep learning model, define the search space Θ of the optimization problem, initialize multiple groups of search entities, each of which randomly selects a solution θ∈Θ as the starting point, and then initialize the pheromone concentration of each node to indicate the quality of the hyperparameter combination; S3.2: Calculate the heuristic information of each node transfer based on the potential for improving model performance, and then use the pheromone concentration and heuristic information to Calculate the selection probability of each group of selectable paths for the current node of the search body, and each group of search bodies selects the next hyperparameter combination according to the current solution path, where P ij represents the probability of moving from the current node i to node j, τ ij represents the pheromone concentration moving from node i to node j, η ij represents the heuristic information for moving from node i to node j, τ ik represents the pheromone concentration moving from node i to node k, η ik represents the heuristic information of moving from node i to node k, α and β represent the importance weights of regulating pheromone and heuristic information respectively; S3.3: Each search entity configures the model according to the selected hyperparameter combination and trains the hybrid deep learning model, evaluates the model performance using the validation set, calculates the evaluation index, and updates the pheromone concentration of each group of nodes based on the calculated evaluation index; S3.4: Repeat the process of path selection, model training and evaluation, and pheromone update multiple times until the set maximum number of iterations is reached or the evaluation index is improved to converge to the preset value. After the iteration, the path with the highest pheromone concentration is output as the optimal hyperparameter combination of the hybrid deep learning model and replaces the original hybrid learning model hyperparameters.

5. A family defect prediction method based on a relay protection device, used to implement the family defect prediction system function based on a relay protection device as described in any one of claims 1 to 4, characterized in that: The specific steps of the prediction method are as follows: Ⅰ. Use sensors and data loggers installed in relay protection devices to collect operation data in real time and integrate historical fault records of different substations or regions; II. Preprocess the running data through edge computing devices and use deep learning autoencoders to extract key features of the data; III. Use a data distributed processing framework to process large-scale data in parallel and build a hybrid deep learning model to predict family defects of relay protection devices; IV. Optimize the hyperparameters of the hybrid deep learning model. The central server aggregates the parameters uploaded by the local models of each substation and updates the federated learning model. V. Predict the collected data in real time, identify whether the equipment has the risk of family defects, analyze the dynamic evolution of the prediction results, and generate future defect patterns and risk probabilities; VI. When a high-risk defect is detected, the alarm mechanism is immediately triggered, the maintenance personnel are notified, and a diagnostic report is automatically generated, while providing a variety of optimization decision suggestions.

6. The method for predicting family defects based on relay protection devices according to claim 5, characterized in that: The specific steps for identifying whether the device has a risk of familial defects as described in step V are as follows: S4.1: Collect real-time operation data from the relay protection device, including current, voltage, frequency, and fault information data, clean and normalize the real-time data, and extract feature information reflecting the performance and status of the equipment from the processed real-time data, including time domain features, frequency domain features, and embedded extraction features; S4.2: Load the trained hybrid learning model and input the extracted features into the hybrid learning model. The model outputs the predicted probability of familial defect risk of each group of devices through forward propagation calculation; S4.3: Based on the set threshold, the model prediction results are classified into risk levels, and the corresponding diagnosis report is generated. At the same time, the alarm mechanism is triggered and maintenance personnel are also alerted.

7. The method for predicting family defects based on relay protection devices according to claim 6, characterized in that: The specific steps for analyzing the dynamic evolution of the prediction results described in step V and generating future defect modes and risk probabilities are as follows: S5.1: Collect all states of the relay protection device and construct the state space of the system, where each state represents the operation status and defect level of the equipment. Then, based on historical data and expert experience, construct a state transition probability matrix to describe the transition probability between states, and the transition probability matrix satisfies that the sum of each row is 1; S5.2: Obtain the current state distribution through the prediction results of the hybrid learning model. According to the current state distribution and the transition probability matrix, calculate the state distribution at the next moment through matrix operations. Check whether all elements in the generated state distribution are non-negative and the sum is 1. If not, check the correctness of the transition probability matrix or the current state distribution. S5.3: Repeat the prediction iteration to predict the state distribution of multiple time steps in the future, and judge the defect mode evolution trend based on the predicted future state distribution, and select the state with serious defects for key attention, and then calculate the cumulative defect risk probability in the preset future time interval as a comprehensive evaluation indicator of the equipment health status; S5.4: Classify the prediction results according to the future state distribution and risk probability, and formulate corresponding maintenance or treatment measures to ensure the safe operation of the equipment.

Images