A predictive maintenance method and system for high-voltage electrical equipment

Abstract

The present application belongs to the traction power supply system, and particularly relates to a kind of high-voltage electrical equipment predictive maintenance method and system;The present application adopts the predictive maintenance method based on multilayer gated recurrent unit (GRU), according to its network characteristics, attention mechanism is added to improve the prediction accuracy of model, the running state and parameter of equipment are monitored in real time by sensor, and the fog platform close to the equipment end and having certain computing power and data storage capacity is used to realize data preprocessing and other operations, then the preprocessed data is sent to cloud for real-time analysis and processing, finally through data analysis and algorithm model, the predictive maintenance of equipment is realized, the diagnostic accuracy is higher, can accurately judge the health status of high-voltage electrical equipment under running state, has important significance for reducing equipment operation and maintenance cost and intelligent transformation, ensuring the safe and economic operation of railway system.

Description

Technical Field

[0001] This invention pertains to traction power supply systems, and more specifically, relates to a predictive maintenance method and system for high-voltage electrical equipment. Background Technology

[0002] Electrical transient processes are frequent in traction power supply systems, such as inrush current, automatic phase-break overvoltage, harmonic resonant voltage, and circuit breaker overvoltage, which seriously affect the operational reliability and lifespan of high-voltage electrical equipment in railway electric locomotives. Accidents involving the explosion of high-voltage equipment on high-speed trains severely impact their safe operation and may even cause traction power supply station tripping or overhead contact line outages.

[0003] The invention patent CN100401080C, entitled "Online Monitoring Method and System for High-Voltage Electrical Equipment in Substations," discloses an online monitoring method for capacitive equipment, metal oxide surge arresters, and circuit breakers, including system structure, online monitoring terminal, and system communication. However, it does not involve data preprocessing using a platform with sufficient computing power and data storage capabilities. Overall, the monitoring accuracy is not high, and it cannot fully and accurately determine the health status of high-voltage electrical equipment under operating conditions.

[0004] The invention patent with publication number CN113267711B, "Online Monitoring System and Monitoring Method for Insulation Status of High-Voltage Electrical Equipment in Substations," includes system structure and system communication, but does not involve using recurrent neural networks or other methods to improve prediction accuracy. Similarly, it cannot fully and accurately determine the health status of high-voltage electrical equipment during operation. Summary of the Invention

[0005] The purpose of this invention is to address the problem that high-voltage electrical equipment is in a monitoring blind spot during the operation of railway electric locomotives and EMUs, and to provide a predictive maintenance method and system for high-voltage electrical equipment based on recurrent neural networks and cloud-fog collaboration.

[0006] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0007] A predictive maintenance method for high-voltage electrical equipment, the method comprising the following steps:

[0008] (1) Select variable selection method based on information entropy theory as sensing parameters to evaluate the amount of information and the trend of change of different sensing parameters;

[0009] (2) Preprocess the acquired data, the preprocessing being smoothing and data normalization;

[0010] (3) Perform sliding window processing on the preprocessed data;

[0011] (4) Input the data that has been processed by the sliding window into the multi-layer GRU model for training;

[0012] In step (4), the forward inference of the multi-layer GRU model is as follows:

[0013]

[0014] Wherein, the update gate of the nth layer GRU network at time t The hidden state from the previous time step and the (n-1)th layer at this moment A linear transformation is performed using the weight matrix, and the result is input into the activation function to determine the degree of information retention; reset gate The degree of information forgetting is determined using the same method; Hidden state at this moment The process of selecting and memorizing input;

[0015] W Z W r and W h These are the three fully connected matrices used in the update gate, reset gate, and hidden state update process, respectively; W dense b is the weight of the dense layer; b is the bias; if the state of the last GRU network at the last time step t is... The final output y is determined by It is calculated in a fully connected layer with ReLU activation function.

[0016] The hidden state H(h1,h2,…h) of the last layer of the multi-layer GRU model t As input to the attention layer, the formula for the correlation between the hidden state at different times and the current time is:

[0017] h t,t' =tanh(W t h t +W t' h t' +b t );

[0018] Corr(h t ,h t' )=σ(W a h t,t' +b a );

[0019] Where tanh is the activation function, W t Let be the weight matrix, t'∈(1,2,…,t), representing different times from t'=1 to t'=t, Corr(h t ,ht' ) represents the hidden state h at time t'. t' With the hidden state h at time t t The degree of similarity;

[0020] Calculate all h t' The weight parameter a t,t' The calculation formula is:

[0021]

[0022] The hidden state c of the final output is obtained by weighted summation. t ;

[0023] By c t The model output RUL is obtained after passing through a fully connected layer with an activation function. pre ;

[0024] (5) The evaluation criteria for predictive maintenance is a maintenance evaluation criterion based on a custom target loss function, wherein the custom target loss function is a target function that integrates the root mean square error evaluation criterion and the scoring function.

[0025] Let h i The bias of the model prediction for the i-th sample data is expressed as:

[0026] h i =RUL pre,i -RUL real,i ;

[0027] The root mean square error method in the root mean square error evaluation standard calculates the same penalty weight for prediction deviations greater than 0 and less than 0, that is, predicting RUL by the same amount of time in advance or lag will result in the same cumulative penalty.

[0028] The calculation formula for the root mean square error method is as follows:

[0029]

[0030] The expression for the scoring function is:

[0031]

[0032] For h i <0 means that a relatively small penalty will be imposed for the predicted failure time being earlier than expected, for h i A value greater than 0 indicates that a relatively large penalty will be imposed for the predicted failure time lag.

[0033] During model training, a custom objective function is designed to minimize the custom objective function to complete the network parameter update and training. Let Θ be the weight parameter, f(x) i;Θ) represents the model with respect to parameter x i Given the predicted value, J(Θ) is the regularization term. Therefore, the objective function calculation formula based on the root mean square error method is:

[0034]

[0035] The formula for calculating the objective function based on the evaluation function is:

[0036]

[0037] The objective loss function based on two different evaluation criteria is expressed as follows:

[0038] L cus (Θ)=αL RMSE (Θ)+(1-α)L SCORE +J(Θ).

[0039] Furthermore, in step (1), the kernel probability density estimation method is used to estimate the distribution function of the continuous random variable, assuming that the true probability density function is f(x). Let f(x) be an estimate. The calculation formula is:

[0040]

[0041] Where N is the sample space size, h is the bandwidth, and K is the kernel function;

[0042] Represented using the Gaussian kernel function:

[0043]

[0044] In this case, the optimal bandwidth for probability density estimation methods of one-dimensional random variables is:

[0045]

[0046] Furthermore, in step (2), the data smoothing process employs a moving average method. Let the smoothed data of the one-dimensional sensor sequence be X = (x1, x2, ..., x...). t ), where x t For the smoothed data at time t, when the smoothing window size is 1, the formula for calculating the moving average is:

[0047]

[0048] Furthermore, in step (2), the data normalization process employs the minimum-maximum normalization method, where x... iLet x be a member of the sensor attribute x, where max(x) and min(x) represent the maximum and minimum values ​​of the attribute, respectively. Then the normalization formula is:

[0049]

[0050] Further, in step (3), let the entire operating cycle of the device be T, the time step of the multi-layer GRU model be time_steps, and the feature dimension be feature_nums. Then, the first training data to be constructed is all sensor data from time 1 to time_steps. Then, the time window is slid to the right by one unit to construct the second training data with the dimension of time_steps*feature_nums from time 2 to time_steps+1, until the last training data at time T-time_steps.

[0051] A predictive maintenance system for high-voltage electrical equipment that applies the above-described maintenance method, the system comprising a sensing layer, a transmission layer, and an application layer;

[0052] The sensing layer is configured to enable sensor monitoring, and the sensor monitoring includes real-time monitoring of the voltage, current and partial discharge parameters of the high-voltage electrical equipment.

[0053] The transmission layer includes a fog-end processing module and a cloud-end processing module. The fog-end processing module is configured to preprocess the data collected by the perception layer, perform visual management of all data, real-time fault diagnosis, and complete data transmission. The cloud-end processing module is configured to train a fault diagnosis model, schedule the resources of the fog-end processing module, and provide visual management functions.

[0054] In the fault diagnosis task, the data interaction between the cloud processing module and the fog terminal processing module includes the transmission and management of real-time data streams and the migration of diagnostic model parameters;

[0055] The perception layer monitors the health status of the high-voltage electrical equipment in real time and obtains health monitoring data. The health monitoring data is intercepted and diagnosed at the edge of the cloud processing module to obtain tagged diagnostic data. The tagged diagnostic data is then used to train the diagnostic model in the fog terminal processing module and the cloud processing module, and stored in the databases of the cloud processing module and the fog terminal processing module.

[0056] The migration of diagnostic model parameters includes training the fault diagnosis model in the cloud processing module, transferring the trained parameters to the corresponding fog terminal processing module to determine whether personalized training is needed in the fog terminal processing module, and finally transferring the trained model parameters to the corresponding edge terminal controller in the fog terminal processing module to complete the fault diagnosis of the high-voltage electrical equipment.

[0057] Furthermore, the cloud processing module continuously updates and trains the fault diagnosis model based on training sample data and computing resources, and obtains a universal diagnostic model, which serves as an intermediate result for different diagnostic scenarios.

[0058] Furthermore, the migration of diagnostic model parameters from the cloud processing module to the fog processing module is driven by both the accumulated amount of training data and the interval time.

[0059] Furthermore, the cloud processing module includes a user management unit and a device management unit;

[0060] The user management unit is configured to divide users into three categories based on their permissions: regular users, trial users, and super administrators.

[0061] The equipment management unit is configured to record equipment information and statistically analyze equipment operating time and operating status.

[0062] The fog terminal processing module includes a data management unit and a fault diagnosis unit;

[0063] The data management unit is configured to provide users with data query methods;

[0064] The fault diagnosis unit is configured to perform data analysis and processing, monitor the operating status of the equipment in real time, issue early warnings when equipment malfunctions, complete equipment fault diagnosis, and propose corresponding maintenance suggestions based on the diagnosis results.

[0065] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0066] This invention employs a predictive maintenance method based on multi-layer gated cyclic units (GRUs). Based on the network characteristics, an attention mechanism is incorporated to improve the model's predictive accuracy. Sensors monitor the equipment's operating status and parameters in real time. A fog platform, located close to the equipment and possessing sufficient computing power and data storage capabilities, is used for data preprocessing. The preprocessed data is then sent to the cloud for real-time analysis and processing. Finally, through data analysis and algorithmic models, predictive maintenance of the equipment is achieved, resulting in higher diagnostic accuracy. This method can accurately determine the health status of high-voltage electrical equipment during operation, which is of great significance for reducing equipment maintenance costs, promoting intelligent transformation, and ensuring the safe and economical operation of railway systems. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of the framework of the predictive maintenance method described in this invention;

[0068] Figure 2 This is a schematic diagram of the sliding window processing process in the predictive maintenance method described in this invention;

[0069] Figure 3 This is a schematic diagram of the cloud-fog collaborative framework in the predictive maintenance system described in this invention;

[0070] Figure 4 This is a schematic diagram of the system module structure of the predictive maintenance system described in this invention;

[0071] Figure 5 This is a schematic diagram illustrating the implementation process of the predictive maintenance system described in this invention;

[0072] Figure 6 This is a schematic diagram of the hardware structure of the predictive maintenance system described in this invention. Detailed Implementation

[0073] The present invention will be further described below with reference to specific embodiments.

[0074] The predictive maintenance method for high-voltage electrical equipment described in this invention includes surge arresters, high-voltage voltage transformers, and high-voltage cable termination equipment. The core predictive maintenance method is based on a multi-layer gated cyclic unit (GRU) predictive maintenance method, which incorporates an attention mechanism to improve model prediction accuracy based on the network characteristics.

[0075] Specifically, such as Figure 1 As shown, this method includes the following steps:

[0076] (1) Select the variable selection method based on information entropy theory as the sensing parameter to evaluate the information content and change trend of different sensing parameters, so as to select the sensor parameters that are more suitable for predictive maintenance.

[0077] Furthermore, we employ kernel probability density estimation to estimate the distribution function of continuous random variables. Compared to the histogram method, this method is not limited by subjective interval divisions and achieves better estimation results for the true probability density function of continuous variables. Assuming the true probability density function is f(x), Let f(x) be an estimate. The calculation formula is:

[0078]

[0079] Where N is the sample space size, h is the bandwidth, and K is the kernel function;

[0080] Represented using the Gaussian kernel function:

[0081]

[0082] In this case, the optimal bandwidth for probability density estimation methods of one-dimensional random variables is:

[0083]

[0084] (2) The acquired data is preprocessed, which includes smoothing and data normalization. The smoothing process uses a moving average method. Let the smoothed data of the one-dimensional sensor sequence be X = (x1, x2, ..., x...). t ), where x t For the smoothed data at time t, when the smoothing window size is 1, the formula for calculating the moving average is:

[0085]

[0086] The data normalization process employs the minimum-maximum normalization method, which is more convenient and simpler than the z-score normalization method, but it is more sensitive to outliers. After smoothing the sensor data, the influence of outliers has been eliminated. Let x... i Let x be a member of the sensor attribute x, where max(x) and min(x) represent the maximum and minimum values ​​of the attribute, respectively. Then the normalization formula is:

[0087]

[0088] (3) Figure 2As shown, a sliding window process is applied to the preprocessed data. Let the total operating cycle of the device be T, the time step of the multi-layer GRU model be time_steps, and the feature dimension be feature_nums. The first training data is all sensor data from time 1 to time_steps. Then, the time window is slid one unit to the right to construct the second training data with the dimension of time_steps*feature_nums from time 2 to time_steps+1, until the last training data at time T-time_steps.

[0089] (4) The data processed by the sliding window is input into the multi-layer GRU model for training. The combination of the attention mechanism and the multi-layer GRU model can effectively solve the problem of insufficient memory capacity of recurrent neural networks.

[0090] In step (4), the forward inference of the multi-layer GRU model is as follows:

[0091]

[0092] Wherein, the update gate of the GRU network in layer n at time t The hidden state from the previous time step and the (n-1)th layer at this moment A linear transformation is performed using the weight matrix, and the result is input into the activation function to determine the degree of information retention; reset gate The degree of information forgetting is determined using the same method; For this moment, it is in a hidden state. The process of selecting and memorizing input;

[0093] W Z W r and W h These are the three fully connected matrices used in the update gate, reset gate, and hidden state update process, respectively; W dense b is the weight of the dense layer; b is the bias; if the state of the last time step t of the last GRU network is... The final output y is determined by It is calculated in a fully connected layer with ReLU activation function.

[0094] As a preferred approach, a Dropout layer can be added during training. As a commonly used regularization method, the Dropout layer can interrupt certain neural connections in the network with a certain probability during the training of the deep learning model, thereby reducing the network's over-dependence on certain neurons and features, thus improving the model's generalization ability and preventing overfitting.

[0095] The multi-layer GRU model that incorporates an attention mechanism calculates the hidden state h at each time step of each layer based on the forward propagation model of the multi-layer GRU. t The hidden states H(h1,h2,…h) of the last layer of the multilayer GRU model. t As input to the attention layer, the formula for the correlation between the hidden state at different times and the current time is:

[0096] h t,t' =tanh(W t h t +W t' h t' +b t );

[0097] Corr(h t ,h t' )=σ(W a h t,t' +b a );

[0098] Where tanh is the activation function, W t Let be the weight matrix, t'∈(1,2,…,t), representing different times from t'=1 to t'=t, Corr(h t ,h i' ) represents the hidden state h at time t'. t' With the hidden state h at time t t The degree of similarity;

[0099] Calculate all h t' The weight parameter a t,t' The calculation formula is:

[0100]

[0101] The hidden state c of the final output is obtained by weighted summation. t ;

[0102] By c t The model output RUL is obtained after passing through a fully connected layer with an activation function. pre (Distance of the device or system from the event in which the failure occurred);

[0103] (5) The evaluation criteria for predictive maintenance is a maintenance evaluation criterion based on a custom target loss function, wherein the custom target loss function is a target function that integrates the root mean square error evaluation criterion and the scoring function.

[0104] The structure of the multi-layer GRU model ultimately outputs a predicted label RULpre through a fully connected layer with an activation function, while the actual label is RULreal. Let h... iThe bias of the model prediction for the i-th sample data is expressed as:

[0105] h i =RUL pre,i -RUL real,i ;

[0106] The root mean square error (RMSE) method in the root mean square error evaluation standard calculates the same penalty weight for prediction deviations greater than 0 and less than 0. That is, predicting RUL by the same amount of time in advance or lag will result in the same cumulative penalty.

[0107] A higher RMSE value indicates lower prediction accuracy. The calculation formula for the root mean square error method is as follows:

[0108]

[0109] The expression for the scoring function is:

[0110]

[0111] For h i <0 means that a relatively small penalty will be imposed for the predicted failure time being earlier than expected, for h i A value greater than 0 means that a predicted failure time lag will incur a relatively large penalty. Here, "relatively small" and "relatively large" refer to their relative relationship.

[0112] The numbers 13 and 10 in the above formula are based on empirical formulas.

[0113] During model training, a custom objective function is designed to minimize the custom objective function to complete the network parameter update and training. Let Θ be the weight parameter, f(x) i ;Θ) represents the model with respect to parameter x i Given the predicted value, J(Θ) is the regularization term. Therefore, the objective function calculation formula based on the root mean square error method is:

[0114]

[0115] The formula for calculating the objective function based on the evaluation function is:

[0116]

[0117] The objective loss function based on two different evaluation criteria is expressed as follows:

[0118] L cus (Θ)=αL RMSE (Θ)+(1-α)L SCORE +J(Θ).

[0119] like Figure 3 and Figure 4 As shown, the high-voltage electrical equipment predictive maintenance system of the present invention, which applies the above-mentioned maintenance method, relies on the Internet of Things and is based on the cloud-fog collaborative framework. The system is designed with a three-layer architecture, namely the perception layer, the transmission layer and the application layer.

[0120] First, the sensing layer is configured to implement sensor monitoring, which includes real-time monitoring of the voltage, current, and partial discharge parameters of the high-voltage electrical equipment. It consists of a large number of sensors, and the collected data is connected to the intelligent terminal acquisition module via an RS485 module. The Modbus protocol, commonly used between electronic devices, is selected to transmit the data to the intelligent terminal device for caching and preprocessing.

[0121] The transport layer is responsible for transmitting the processed data over long distances and displaying relevant parameters on the smart terminal's screen. The smart terminal consists of a data acquisition unit and a computing unit (fog nodes), encapsulating data in JSON format. The acquisition unit connects to the perception layer, transmitting the acquired data to the local computing unit, which then processes the data and uploads it to the cloud server. The communication network can be a 5G network to meet real-time requirements, using the MQTT protocol as the transport protocol to wirelessly transmit data to the server platform via the 5G network.

[0122] The application layer consists of a database, a web platform, a data analysis unit, a terminal configuration unit, and an MQTT service unit. The database is used to store information, the web platform is used to realize human-computer interaction and meet the needs of remote monitoring, and the data analysis unit can perform information mining on the data and use relevant algorithms to make predictions.

[0123] Specifically, the process implementation of this system can be referred to Figure 5 Sensors monitor parameters such as voltage, current, and partial discharge of high-voltage electrical equipment in real time and transmit the data to the edge computing node. The edge computing node processes, analyzes, and stores the sensor data received at the node, generating data models and features locally. Cloud collaboration transmits the processed data to the cloud server for collaborative processing with cloud data, enabling deeper data analysis and decision-making. Intelligent decision-making utilizes cloud-based artificial intelligence algorithms to predict equipment status and diagnose faults, generating intelligent maintenance plans. Preventive maintenance, through a web platform or mobile application, enables remote monitoring and management of equipment status, timely response to anomalies, and the development of preventive maintenance plans based on equipment status predictions and fault diagnosis results. This allows for proactive equipment maintenance, reducing failure rates and improving equipment reliability and safety.

[0124] from Figure 4As can be seen, this system is an online monitoring system for high-voltage electrical equipment based on cloud-fog collaboration, which fully utilizes the computing and storage capabilities of the cloud and fog terminals. Considering the computing and storage resources involved, we define it as a cloud processing module and a fog terminal processing module.

[0125] Specifically, the fog-end processing module is configured to preprocess the data collected by the perception layer, perform visual management of all data, real-time fault diagnosis, and complete data transmission; the cloud processing module is configured to train the fault diagnosis model, schedule the resources of the fog-end processing module, and perform visual management functions.

[0126] In the initial stage of cluster construction, the training sample data mainly consisted of publicly available fault diagnosis datasets. As fog terminal data accumulated in the cluster, it was replaced with fault datasets collected by each fog terminal. Based on the abundant training sample resources and computing resources of the cloud processing module, the fault diagnosis model was continuously trained and updated, resulting in a corresponding universal diagnostic model, which serves as an intermediate result for different diagnostic scenarios.

[0127] When the fog terminal processing module updates the diagnostic model in the cloud, it can actively pull matching data from the cloud processing module's database based on the type of fog terminal high-voltage electrical equipment it is responsible for. It then uses this data to complete the personalized training of different diagnostic models and finally transmits the trained fog terminal diagnostic model parameters to the corresponding fog terminal processing module.

[0128] In fault diagnosis tasks, the data interaction between the cloud processing module and the fog processing module includes the transmission and management of real-time data streams and the migration of diagnostic model parameters.

[0129] The transmission and management of real-time data streams refers to the system's need to constantly monitor the health status of high-voltage electrical equipment to ensure its normal operation, thus generating a large amount of health monitoring data. This data needs to be intercepted and diagnosed at the edge, and then the labeled diagnostic data is used to train diagnostic models at the edge and in the cloud, and stored in databases at the edge and in the cloud.

[0130] Real-time data streaming is an important data interaction method between the cloud and the terminal during the fault diagnosis process.

[0131] The transfer of diagnostic model parameters refers to the training of a fault diagnosis model in the cloud and then transferring the trained model parameters to the corresponding edge device to determine whether personalized training is needed on the edge. Finally, the trained model parameters are transferred to a specific edge controller to complete the fault diagnosis of high-voltage electrical equipment.

[0132] The parameter transfer from the cloud to the fog is driven by both the amount of accumulated training data and the time interval.

[0133] The design of cloud-fog collaborative data flow collaboration methods includes the design of data acquisition system solutions, real-time data management system solutions, and storage system solutions.

[0134] The data acquisition system uses Flume to monitor all database tables on the fog terminal. Flume is a highly available, highly reliable, distributed system for collecting, aggregating, and transmitting massive logs based on a streaming architecture. It is flexible, simple, and easy to deploy. When new data is added to the database, the newly added data is transmitted completely through two channels. One channel connects to the subsequent cloud storage system, and the other connects to the subsequent real-time data management system.

[0135] The real-time data management system adopts a publish / subscribe model based Kafka, which can meet the needs of this design. After data is generated, it is published to the corresponding groups. When the data reaches a certain quantity, it can be pushed to the consumer for consumption. The Kafka system can also play a role in peak shaving.

[0136] In the real-time data management solution, when the Flume data acquisition system sends messages to the Kafka group, the Producer interface provided by Kafka is selected, which uses an asynchronous sending method.

[0137] The storage system uses the HDFS system architecture for data storage to achieve requirements such as high fault tolerance, high reliability, high scalability, high availability, and high throughput.

[0138] The storage system primarily uses the HDFS 2.0 framework, including Datanode, Namenode, Journalnode, and Zookeeper.

[0139] Datanodes are primarily nodes used for data storage. They store the actual data blocks, perform read / write operations on these blocks, and report their dynamic block location information to both Namenodes. The main function of the Namenode is to synchronize and manage metadata, such as managing the mapping information of all data blocks across the Datanodes it manages. In addition, it configures the replication strategy for the entire HDFS and handles client read / write requests. In the HDFS 2.0 architecture, there are typically two Namenodes: one active and one standby. The active Namenode receives client read / write requests in real time and simultaneously writes its log information to all Journalnodes via log files. If more than half of the log files in the Journalnodes are consistent, the data is considered valid. The standby Namenode synchronizes its logs through the Journalnodes.

[0140] Journalnodes are log server nodes, primarily used to keep Standby Namenodes and Active Namenodes synchronized. They typically use a network filesystem for log file synchronization. ZooKeeper is a distributed coordination system used to manage the overall health of HDFS, consistent with the Flume management system. ZooKeeper monitors the health status of Namenodes through a health check mechanism. When an active Namenode encounters an error, a failover mechanism activates the standby Namenode to take over management duties.

[0141] In this embodiment, the cloud processing module includes a user management unit and a device management unit.

[0142] The user management unit can be divided into three categories based on user type and permissions: regular users, trial users, and super administrators. Regular users can change their usernames and passwords in this module; trial users only have query permissions and cannot make any modifications, such as viewing existing information or querying current information; super administrators can perform more operations, such as changing passwords, setting user permissions, resetting regular users' passwords, and registering new users.

[0143] The equipment management unit is mainly responsible for recording basic information such as equipment model, manufacturing date, manufacturer, and user unit, and for compiling statistics on equipment operating time and status.

[0144] The fog terminal processing module includes a data management unit and a fault diagnosis unit.

[0145] The data management unit helps users query data over a specific period, which can be set by the user. It also analyzes the data and presents the results in tables and graphs for intuitive understanding. Furthermore, this unit supports importing and exporting different data formats to meet the needs of diverse users.

[0146] The fault diagnosis unit is the core component of the system. Deploying it at the cloud endpoint enables rapid data analysis and processing, reducing the amount of information transmitted to the cloud and facilitating quick response to emergencies. This unit primarily monitors the real-time operating status of equipment, provides early warnings of equipment faults, diagnoses equipment faults, and proposes corresponding maintenance suggestions based on the diagnostic results.

[0147] The hardware structure of this system can be referenced. Figure 6 It collects electrical characteristic quantities through a current sensor, a 300 / 5V voltage transformer, and an ultrasonic partial discharge sensor.

[0148] In addition, the corresponding intelligent terminal software design is mainly divided into data acquisition unit software design and computing unit software design.

[0149] After the acquisition unit is powered on, it reads and initializes the hardware system and configuration information, displays the information on the serial port screen, connects to the computing unit, mounts the SD card file system, and starts running the main program after the preparation work is completed. When the timer overflows, it reads the parameters collected by each sensor, calculates and filters them, packages the filtered data, sends it to the computing unit, displays the collected data on the serial port screen, and caches the collected data on the local SD card. The main program checks whether any personnel have modified the parameters of the touch screen, such as the IP address and the acquisition instrument parameters. If no modifications are made, it returns to the main program and continues to collect sensor data and other operations. If modifications are made, the configuration is rewritten, the information is re-initialized, and then the data acquisition, packaging, and transmission work continues.

[0150] After the computing unit initializes its program, it first receives data packets from the acquisition unit, parses and preprocesses the data, uploads normal data to the cloud server, and generates alarm information locally for abnormal data, which is simultaneously transmitted to the cloud server. The anomaly detection model in the computing unit adopts an unsupervised learning density peak clustering algorithm. The algorithm first manually determines the cluster centers based on the decision graph, and then reasonably allocates the remaining points to obtain the clustering results.

[0151] Equipment operating status assessment includes surge arrester status diagnosis, voltage transformer status diagnosis, and cable termination status diagnosis. Surge arrester status diagnosis includes characteristic resistance diagnosis and characteristic current phase diagnosis; voltage transformer status diagnosis includes excitation current diagnosis and partial discharge diagnosis; the health status of cable terminations is mainly determined by the degree of partial discharge.

Claims

1. A method for predictive maintenance of high-voltage electrical equipment, characterized in that, The method comprises the following steps: (1) obtaining initial data through a sensor, selecting a variable selection method based on information entropy theory as a sensor parameter to evaluate the information amount and variation trend contained in different sensor parameters; (2) pre-processing the obtained data, wherein the pre-processing is smoothing processing and data normalization processing; (3) performing sliding window processing on the data after pre-processing; (4) inputting the data after sliding window processing into a multi-layer GRU model for training; In step (4), the forward inference of the multi-layer GRU model is: Wherein the update gate of the GRU network of the nth layer at time t The hidden state at the last time And the hidden state of the (n-1)th layer at this time The reset gate is determined by linear transformation through a weight matrix and input into an activation function to determine the degree of information retention The forget gate is determined by the same method to determine the degree of information forgetting The hidden state at this time The selective memory process on the input W Z , W r , and W h are three fully connected matrices in update gate, reset gate and hidden state update process, respectively; W dense is the dense layer weight; b is the bias; if the state of the last time t of the last layer GRU network is , then the final output y is calculated by inputting into the fully connected layer with relu activation function. The hidden state H (h1, h2, … h t ) of the last layer of the multi-layer GRU model is input to the attention layer, and the correlation degree of the hidden state at different times with the current time is calculated according to the following formula: h t，t′ = tanh(W t h t + W t′ h t′ + b t ); Corr(h t , h t′ ) = σ(W a h t，t′ + b a ); wherein tanh is an activation function, W t is a weight matrix, t' ∈ (1, 2, …, t) represents different time points from t' = 1 to t' = t, Corr(h t , h t′ ) represents the similarity between the hidden state h t′ at t' time and the hidden state h t at t time; The weight parameter a of all h is calculated t′ t，t′ The calculation formula is:​ The final output hidden state c is obtained by a weighted sum t ; By c t Through the fully connected layer with activation function to get the model output RUL pre ; (5) the evaluation criterion of the predictive maintenance is a maintenance evaluation criterion based on a self-defined target loss function, and the self-defined target loss function is a target function based on the root mean square error evaluation criterion and the scoring function; Let h i The bias for the prediction of the i-th sample data model is expressed as: h i = RUL pre，i - RUL real，i ; The root mean square error method in the root mean square error evaluation criterion calculates the same penalty weight for the prediction deviation greater than 0 and less than 0, that is, the same penalty is accumulated when the prediction RUL is advanced or delayed by the same time; The calculation formula of the root mean square error method is: The expression of the scoring function is: For h i <0, i.e. the predicted failure time is ahead, a relatively small penalty is given to h i >0, i.e. the predicted failure time is behind, a relatively large penalty is given to h A custom objective function is designed during model training to minimize the custom objective function to complete the update and training of network parameters, where Θ is the weight parameter, f(x i ; Θ) is the predicted value of the model for parameter x i , and J(Θ) is the regularization term. The objective function calculation formula based on the root mean square error method is as follows: The target function calculation formula based on the evaluation function is: The target loss function expression formula based on two different evaluation criteria is: L cus (Θ) = aL RMSE (Θ) + (1 - a)L SCORE + J(Θ).

2. The predictive maintenance method of high voltage electrical equipment according to claim 1, characterized in that, In step (1), the distribution function of continuous random variable is estimated by using the method of kernel probability density estimation, assuming that the real probability density function is f(x), The estimated value of f(x) is The calculation formula is: Wherein, N is the sample space size, h is the bandwidth, and K is the kernel function; The Gaussian kernel function is expressed as: In this case, the optimal bandwidth of the probability density estimation method of one-dimensional random variables is:

3. The predictive maintenance method of high voltage electrical equipment according to claim 1, characterized in that, In step (2), the data smoothing processing adopts a sliding average manner, assuming that the smoothed data of a one-dimensional sensor sequence is X = (x1, x2, …, x t ), wherein x t is the data at time t after smoothing, and when the smoothing window size is 1, the calculation formula of the sliding average is:

4. The predictive maintenance method of high voltage electrical equipment according to claim 1, characterized in that, In step (2), the data normalization process employs the min-max normalization method, let x i is a member of the sensor attribute x, where max(x) and min(x) represent the maximum and minimum values in the attribute, respectively, then the normalization formula is:

5. The predictive maintenance method of high voltage electrical equipment according to claim 1, characterized in that, In step (3), the entire running cycle of the equipment is T, the time step of the multi-layer GRU model is time_steps, and the feature dimension is feature_nums. The first training data is constructed from all sensor data from time 1 to time_steps. Then, the time window is slid to the right by one unit, and the second training data from time 2 to time_steps+1 is constructed. time_steps*feature_nums dimension training data, until the last training data at the T-time_steps moment.

6. A predictive maintenance system for high voltage electrical equipment applying the maintenance method according to any one of claims 1 to 5, characterized in that, The system comprises a perception layer, a transmission layer and an application layer; The perception layer is configured to realize sensor monitoring, and the content of the sensor monitoring includes real-time monitoring of voltage, current and partial discharge parameters of the high-voltage electrical equipment; The transmission layer comprises a fog end processing module and a cloud end processing module, the fog end processing module is configured to pre-process the data collected by the perception layer, and to visually manage, real-time fault diagnose and complete data transmission of all data; the cloud end processing module is configured to train fault diagnosis, that is, the multi-layer GRU model, to schedule resources of the fog end processing module and to visually manage functions; In the fault diagnosis task, the data interaction between the cloud end processing module and the fog end processing module includes transmission and management of real-time data flow, migration of diagnosis model parameters; The perception layer monitors the health status of the high-voltage electrical equipment in real time and obtains health monitoring data, the health monitoring data is intercepted and diagnosed at the edge of the cloud processing module to obtain labeled diagnosis data; then the labeled diagnosis data is used for training of a diagnosis model in the fog processing module and the cloud processing module, and is stored in the database of the cloud processing module and the fog processing module; The migration of the diagnosis model parameters includes that when the cloud processing module needs to train a fault diagnosis model and needs to migrate the parameters obtained after training to the corresponding fog processing module to determine whether individual training needs to be performed in the fog processing module, and finally the parameters of the trained model are migrated to the controller of the corresponding edge in the fog processing module to complete fault diagnosis of the high-voltage electrical equipment.

7. The predictive maintenance system for high voltage electrical apparatus according to claim 6, characterized in that, The cloud processing module continuously updates and trains the fault diagnosis model based on training sample data and computing resources, and obtains a universal diagnosis model, which serves as an intermediate result for different diagnosis scenarios.

8. The predictive maintenance system for high voltage electrical apparatus according to claim 6, characterized in that, The migration of the diagnosis model parameters from the cloud processing module to the fog processing module is jointly driven by the accumulation amount of training data and the interval time.

9. The predictive maintenance system for high voltage electrical apparatus according to claim 6, characterized in that, The cloud processing module includes a user management unit and a device management unit. The user management unit is configured to divide users into three categories of ordinary users, experience users and super administrators according to permissions. The device management unit is configured to record device information and count device running time and running status. The fog processing module includes a data management unit and a fault diagnosis unit. The data management unit is configured to provide data query methods for users. The fault diagnosis unit is configured to complete data analysis and processing, monitor the running status of the device in real time, and give an early warning when the device fails, complete device fault diagnosis, and propose corresponding maintenance suggestions according to the diagnosis results.

Images