Fault diagnosis method and system for 500kV bus protection device

Abstract

The invention provides a fault diagnosis method and system for a 500kV bus protection device, and the method comprises the following steps: obtaining the operation data of the 500kV bus protection device, and carrying out the preprocessing of the obtained operation data; according to the method, fault feature extraction is carried out by adopting the convolutional neural network in combination with an attention mechanism, fault information can be accurately captured, fault diagnosis and classification are carried out in combination with a support vector machine and a fuzzy logic theory, and the diagnosis accuracy of complex hidden faults is effectively improved; the fault location can be quickly and accurately determined through a fault location algorithm based on the graph theory, a fault isolation strategy is automatically generated, equipment action is controlled, fault expansion is avoided, and the reliability and stability of a power system are improved; by establishing the prediction model, the future operation state of the protection device can be predicted, early warning and prevention of faults are realized, a scientific and reasonable basis is provided for maintenance and repair of equipment, and the maintenance cost and power failure loss of the equipment are reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system protection, in particular to a fault diagnosis method and system for a 500kV bus protection device. BACKGROUND

[0002] In the power system, the 500kV bus is a key hub for power transmission, and its safe and stable operation is crucial. The bus protection device is an important equipment to ensure the safety of the bus. However, due to the complex operating environment of the bus, the protection device may have various faults, such as hardware failure, software failure, communication failure, etc. Timely and accurate diagnosis of the fault of the bus protection device is of great significance to improve the reliability and stability of the power system. However, there are still some problems: 1. The existing fault diagnosis method for 500k bus protection device mainly relies on traditional fault detection and diagnosis technology, such as threshold comparison-based method, signal analysis-based method, etc. These methods can detect some obvious faults to some extent, but the diagnosis effect is not ideal for some complex and hidden faults, which reduces the reliability and stability of the power system; 2. The existing fault location method usually relies on manual analysis of protection action signals and wiring diagrams for troubleshooting, which is not only time-consuming, but also has low positioning accuracy, resulting in low processing efficiency and expanded fault impact; 3. The existing diagnosis technology only responds passively to the current fault and cannot provide proactive warning before hardware failure occurs, resulting in a lack of preventive maintenance and high maintenance cost, long power recovery time, etc. SUMMARY

[0003] The purpose of the present application is to provide a fault diagnosis method and system for a 500kV bus protection device to solve one of the problems raised in the background art.

[0004] In order to achieve the above technical features, the purpose of the present application is achieved as follows: a fault diagnosis method for a 500kV bus protection device, comprising the following steps: Obtain the operating data of the 500kV bus protection device and pre-process the obtained operating data; Based on a convolutional neural network combined with an attention mechanism, the pre-processed operating data is subjected to feature recognition and extraction of a fault feature vector; the convolutional neural network includes at least three residual connection layers; Based on principal component analysis, the extracted fault feature vector is subjected to dimensionality reduction processing, and the principal components with a cumulative variance contribution rate of ≥95% are retained; The fault feature vector after dimensionality reduction processing is input into a support vector machine model for fault type recognition, and the fault severity is evaluated by a fuzzy membership function based on fuzzy logic theory; Based on graph theory, a graph structure model of a 500kV busbar protection device is constructed. By analyzing and identifying the fault types and severity, and combining the action information of the 500kV busbar protection device, the fault path is searched in the graph structure, the fault location is located, and the fault location result is generated. Based on the fault location results, a fault isolation strategy is generated to control the operation of circuit breakers and disconnect switches; at the same time, the future operating status of the 500kV bus protection device is predicted through time series analysis algorithms and neural network prediction models.

[0005] Preferably, the method for extracting fault feature vectors includes the following steps: Two-dimensional feature extraction is performed on the preprocessed running data through the convolutional layers in the convolutional neural network to generate an initial feature map; The attention weights of each region in the initial feature map are calculated using an attention mechanism. The attention weights are multiplied element-wise with the initial feature map to obtain a weighted feature map; The weighted feature map is nonlinearly transformed by at least three residual connection layers in the convolutional neural network to output a fault feature vector; the residual connection layers alleviate the gradient vanishing problem through a skip connection structure.

[0006] Preferably, the support vector machine model uses a radial basis function as the kernel function, and optimizes the penalty parameter C and the kernel function parameter γ using a k-fold cross-validation method; the fuzzy membership function is generated based on historical fault data, inputs the dimensionality-reduced fault feature vector, and outputs the fault severity.

[0007] Preferably, the graph structure model is constructed with the components of the protection device and the busbar segments as nodes, and the electrical connection relationships and signal transmission logic between nodes as edges; the fault location is determined by analyzing the fault type, fault severity, and the action information of the 500kV busbar protection device, and by traversing and searching for the abnormal signal propagation path in the graph structure based on a graph theory algorithm; the graph theory algorithm is either breadth-first search or depth-first search.

[0008] Preferably, the operating data includes the input and output signals, internal electrical parameters, hardware status parameters of the 500kV bus protection device, and voltage and current data of the 500kV bus; the preprocessing includes denoising and normalization, wherein the denoising is performed using the least mean square adaptive filtering algorithm, and the normalization is performed using the Z-score normalization method to map the operating data to a standard normal distribution.

[0009] Preferably, the neural network prediction model is a long short-term memory network model. The long short-term memory network model takes historical time-series data of the 500kV bus protection device as input, learns the long-term dependencies in the historical time-series data through training, and outputs the predicted value of the change trend of key parameters in the next 72 hours. If the predicted value exceeds the preset threshold, a fault warning is triggered. The historical time-series data includes temperature, voltage and number of actions.

[0010] Preferably, the dimensionality reduction process calculates the covariance matrix of the fault feature vectors, performs eigenvalue decomposition on the covariance matrix, and selects the fault feature vectors corresponding to the principal components with a cumulative variance contribution rate ≥ 95% to form the projection matrix.

[0011] Another aspect of the present invention provides a fault diagnosis system for a 500kV busbar protection device, comprising: a data acquisition and processing module, a feature extraction module, a feature dimensionality reduction module, a fault diagnosis module, a fault location module, and a fault processing and prediction module; The data acquisition and processing module is configured to acquire the operating data of the 500kV busbar protection device and preprocess the acquired operating data. The feature extraction module is configured to perform feature recognition on the preprocessed running data and extract fault feature vectors based on an improved convolutional neural network combined with an attention mechanism; the improved convolutional neural network includes at least three residual connection layers. The feature dimensionality reduction module is configured to perform dimensionality reduction processing on the extracted fault feature vector based on principal component analysis, retaining principal components with a cumulative variance contribution rate ≥ 95%. The fault diagnosis module is configured to input the dimensionality-reduced fault feature vector into the support vector machine model for fault type identification, and combine fuzzy logic theory to evaluate the severity of the fault through fuzzy membership function. The fault location module is configured to construct a graph structure model of the 500kV busbar protection device based on graph theory. By analyzing and identifying the fault type and severity, and combining the action information of the 500kV busbar protection device, it searches for the fault path in the graph structure, locates the fault location, and generates the fault location result. The fault handling and prediction module is configured to generate a fault isolation strategy based on the fault location result and control the operation of the circuit breaker and disconnector; at the same time, it predicts the future operating status of the 500kV bus protection device through time series analysis algorithm and neural network prediction model.

[0012] Preferably, the attention mechanism in the feature extraction module is a dynamic adaptive attention mechanism.

[0013] Preferably, the fault handling and prediction module includes: a strategy generation unit, a prediction error correction unit, and an edge computing node; The strategy generation unit is used to construct a fault isolation strategy optimization model based on a deep Q-network; The prediction error correction unit is used to compare the predicted value of the long short-term memory network model with the actual operating data of the device in real time through a recurrent neural network. The edge computing node is used to deploy reinforcement learning models and time series analysis algorithms on the local edge computing unit of the protection device.

[0014] The present invention has the following beneficial effects: 1. This invention uses convolutional neural networks combined with attention mechanisms for fault feature extraction, which can more accurately capture key information related to faults. Combined with support vector machines and fuzzy logic theory for fault diagnosis and classification, it effectively improves the diagnostic accuracy of complex hidden faults. 2. This invention uses a graph theory-based fault location algorithm to quickly and accurately determine the fault location, automatically generate fault isolation strategies and control equipment actions, thereby shortening fault handling time, preventing fault escalation, and improving the reliability and stability of the power system. 3. By establishing a predictive model, this invention can predict the future operating status of the protection device, realize early warning and prevention of faults, provide a scientific and reasonable basis for equipment maintenance and repair, and reduce equipment maintenance costs and power outage losses. Attached Figure Description

[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0016] Figure 1 This is a flowchart illustrating a fault diagnosis method for a 500kV busbar protection device according to the present invention.

[0017] Figure 2 This is a flowchart illustrating the method for extracting fault feature vectors according to the present invention.

[0018] Figure 3 This is a schematic diagram of the functional modules of a fault diagnosis system for a 500kV busbar protection device according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example Figure 1This is a flowchart illustrating a fault diagnosis method for a 500kV busbar protection device according to an embodiment of the present invention. It should be noted that if substantially the same result is obtained, the method of this application is not necessarily identical. Figure 1 The illustrated process sequence is limited. For example... Figures 1-2 The following is a method for fault diagnosis of a 500kV busbar protection device, comprising the following steps: Step 1: Obtain the operating data of the 500kV busbar protection device and preprocess the obtained operating data; Specifically, data is collected from both the protection device itself and the bus electrical quantities: input and output signals are obtained through the protection device's DI / DO module and RS-485 serial communication interface, internal electrical parameters are retrieved using a microprocessor, and hardware status parameters are collected using built-in temperature, voltage, and other sensors; voltage and current signals are obtained through the PT and CT of the 500kV bus, and after conversion by the secondary circuit, they are sampled by the protection device's analog input module at a sampling frequency of 1000Hz, while relevant data is obtained by accessing the substation system through communication protocols such as IEC 61850. Then, the least mean square adaptive filtering algorithm is used to denoise the acquired operating data to suppress transient interference during bus faults. Next, the Z-score normalization method is used to normalize continuous data such as voltage and current to eliminate the influence of dimensional differences on subsequent models. Finally, the data is verified and completed. If there are missing sampling points, for data with a missing time interval ≤ 5ms, linear interpolation is used to fill in the missing data based on the data before and after the missing time. For key parameters such as voltage and current, outliers are identified by the 3σ principle. If the current value at a certain moment exceeds 3 times the standard deviation of the mean, it is replaced with the median of the adjacent valid values ​​to ensure the reliability of the input data.

[0021] Step 2: Based on the convolutional neural network combined with the attention mechanism, perform feature recognition on the preprocessed running data and extract fault feature vectors; the convolutional neural network includes at least three residual connection layers; Specifically, firstly, the preprocessed operational data (such as time-series signals like voltage and current waveforms) is converted into a two-dimensional structure suitable for convolutional neural network processing. For example, multi-dimensional parameters at different time points (such as three-phase voltage and zero-sequence current) are arranged into a matrix form similar to an image, which serves as the network input. The first layer of the convolutional neural network scans the input data through multiple convolutional kernels (similar to "feature detectors"). Each convolutional kernel is responsible for capturing specific local features (such as voltage spikes and current harmonic components) and generating corresponding initial feature maps. These feature maps are equivalent to "feature snapshots" of data extracted from different angles. For example, some focus on the trend of voltage changes, while others focus on abnormal fluctuations in current. Then, an attention mechanism is introduced to optimize the initial feature map. The attention mechanism is similar to the "focusing" ability of human vision, which can automatically identify the regions and channels in the feature map that are more important for fault diagnosis. The specific process is to first compress the global information of each feature map to obtain a vector that reflects the overall features, and then calculate the attention weight of each position and channel through a multi-layer neural network. These weights represent the "importance" of each region in the feature map. For example, in a bus short-circuit fault, the time point when the current suddenly increases and the corresponding current channel will be given higher weights, while the weight of the stable region during normal operation will be lower. Next, the calculated attention weights are multiplied element-wise with the initial feature map to obtain a weighted feature map. This step is equivalent to "filtering" and "enhancing" the original features based on the attention weights: the signals of key fault feature regions are preserved or amplified, while irrelevant noise or secondary information is suppressed. For example, when identifying bus grounding faults, the weighted feature map will significantly highlight the abnormal characteristics of zero-sequence current, while weakening the fluctuations of other normal parameters, making the fault characteristics more obvious. Finally, the weighted feature map is processed through at least three residual connection layers. The residual connection layers adopt a "skip connection" structure, which allows the network to directly pass features from the previous layer to the next layer, avoiding the "vanishing gradient" problem common in deep network training. Each residual connection layer performs non-linear transformations on the feature map through operations such as convolution and activation functions, gradually extracting multi-level features from simple to complex (such as from waveform distortion features to fault type features). After three layers of processing, the feature map is input into a global average pooling layer and compressed into a fixed-length fault feature vector. This vector condenses the key fault-related information in the original data and can be used for subsequent fault type identification and severity assessment. By combining the automatic feature extraction capability of CNN with the focusing characteristics of the attention mechanism, the fault characteristics in the operation data of 500kV bus protection devices were effectively captured. In particular, the feature extraction capability of early latent faults (such as insulation aging and poor contact) was significantly improved, providing a key basis for subsequent fault diagnosis.

[0022] Step 3: Perform dimensionality reduction on the extracted fault feature vectors based on principal component analysis, and retain principal components with a cumulative variance contribution rate ≥ 95%; Specifically, firstly, the extracted fault feature vectors are standardized and preprocessed; the high-dimensional feature vectors (e.g., 128-dimensional) obtained in step two are normalized to zero mean and processed with unit variance by column to eliminate the influence of differences in dimensions between different feature dimensions; for example, for voltage-related features and current-related features, after standardization, their mean is 0 and their standard deviation is 1 to ensure the fairness of subsequent PCA calculations. Then, the covariance matrix of the standardized eigenvectors is calculated; the covariance matrix reflects the correlation between each feature dimension, and its elements... Indicates the first The first feature and the second The covariance of each feature; by calculating the covariance matrix, we can analyze which feature dimensions have strong linear relationships, providing a basis for subsequent principal component extraction; Next, eigenvalue decomposition is performed on the covariance matrix; the eigenvalues ​​and eigenvectors of the covariance matrix are solved, where the eigenvalues ​​represent the variance of the corresponding principal component (i.e., the amount of information carried by the principal component), and the eigenvectors represent the direction of the principal component; the eigenvalues ​​are then sorted in descending order to obtain the eigenvalue sequence. and its corresponding eigenvectors ; Subsequently, the number of principal components to be retained is determined; the variance contribution rate of each principal component (i.e., the proportion of a single eigenvalue to the sum of all eigenvalues) and the cumulative variance contribution rate are calculated; Based on the principle that the cumulative variance contribution rate is ≥95%, the top [percentage] is selected. Each principal component makes ; For example, if the cumulative variance contribution rate of the first 10 principal components reaches 96%, then these 10 principal components are retained, and the original 128-dimensional feature vector is reduced to 10 dimensions. Finally, the dimensionality-reduced feature vectors are generated through projection transformation; the standardized original feature vectors are then projected onto a selected... In each principal component direction, the dimensionality-reduced feature vector is obtained. This vector retains more than 95% of the information in the original data while significantly reducing the feature dimension, effectively reducing the computational complexity and overfitting risk of subsequent classification models. PCA dimensionality reduction not only removes redundant information from the original features (such as highly correlated voltage and current harmonic components), but also extracts the most representative comprehensive features (such as fault energy features and phase change features), providing more efficient input for subsequent fault type identification. For example, in the fault diagnosis of bus protection devices, the dimensionality-reduced feature vectors can more clearly distinguish different types of faults (such as short circuit, grounding, and open circuit), improving the accuracy and real-time performance of the diagnostic model.

[0023] Step 4: Input the dimensionality-reduced fault feature vector into the support vector machine model for fault type identification, and combine fuzzy logic theory to evaluate the severity of the fault through the fuzzy membership function. Specifically, firstly, the dimensionality-reduced fault feature vector is input into a support vector machine (SVM) model for fault type identification. The SVM model has been trained using historical fault data (such as typical fault samples like short circuits, grounding, and overloads) to construct the optimal classification hyperplane among different fault types. For the input dimensionality-reduced feature vector, the SVM model calculates its distance to each hyperplane based on the classification boundary obtained during training, and assigns it to the fault category represented by the hyperplane with the closest distance, thereby achieving automatic identification of the current fault type of the bus protection device. Then, the severity of the fault is assessed using fuzzy logic theory. For each fault type, multiple fuzzy subsets (such as "minor," "moderate," and "severe") are defined, and a corresponding fuzzy membership function is designed for each subset. For example, for a short-circuit fault, the membership function can be defined based on features such as the fault current amplitude and duration. Minor short circuit: When the fault current is less than twice the rated current, the membership degree gradually increases from 0 to 1; Medium short circuit: When the fault current is between 2 and 5 times the rated current, the membership degree increases from 0 to 1 and then decreases to 0. Severe short circuit: When the fault current is greater than 5 times the rated current, the membership degree gradually increases from 0 to 1; Next, the reduced fault feature vector is substituted into each membership function to calculate its membership value for different severity levels. For example, the feature vector of a short-circuit fault has membership values ​​of 0.1, 0.7, and 0.2 for "minor", "moderate", and "severe" respectively, indicating that the fault is more likely to be at the "moderately severe" level. Subsequently, fuzzy inference rules are applied to comprehensively assess the severity of the fault; a fuzzy rule base is developed based on expert experience, for example: Rule 1: If the fault type is "grounding" and the zero-sequence current membership is "high", then the fault severity is "severe". Rule 2: If the fault type is "overload" and the load current membership degree is "medium", then the fault severity is "medium". The final fuzzy evaluation result is obtained by aggregating the conclusions of all applicable rules (e.g., by weighted averaging). Finally, defuzzification is performed; the fuzzy evaluation results are converted into specific severity values ​​(such as a 0-100 score). For example, the centroid method is used to calculate the centroid position under the membership function curve to obtain a quantitative fault severity score. This score can be used to guide operation and maintenance decisions, such as determining maintenance priorities and early warning levels. By combining SVM with fuzzy logic, this method can not only accurately identify fault types, but also quantitatively assess the severity of faults, making it particularly suitable for early fault warning and preventive maintenance. For example, for minor discharge faults in the early stages of insulation aging, the system can issue an early warning through fuzzy evaluation to prevent the fault from developing into a serious short circuit accident.

[0024] Step 5: Construct a graph structure model of the 500kV busbar protection device based on graph theory. By analyzing and identifying the fault types and severity, and combining the action information of the 500kV busbar protection device, search for fault paths in the graph structure, locate the fault location, and generate fault location results. Specifically, firstly, a graph structure model of the 500kV busbar protection device is constructed based on graph theory; electrical equipment such as busbars, circuit breakers, and transformers are abstracted as nodes in the graph, and the electrical connections between equipment are abstracted as edges, forming a topology graph; attributes are assigned to each node and edge, such as equipment type and operating status for nodes, and impedance and capacity for edges; at the same time, the fault type and severity identified in step four are used as node attributes, combined with the action information of the protection device (such as circuit breaker trip signals and protection relay action time sequences), to add dynamic state information to the graph structure; Then, the fault path is searched in the graph structure; based on the fault type and the action logic of the protection device, possible fault propagation paths are determined; for example, if a bus short-circuit fault is identified and a branch circuit breaker trips, the search starts from the bus and proceeds outward along the tripped branch, while considering the action sequence of the protection device (e.g., the main protection takes precedence over the backup protection); using graph search algorithms such as breadth-first search (BFS) or Dijkstra's algorithm, combined with the equipment status and fault severity, the fault probability or weight of each path is calculated, and the most likely fault path is selected; Next, locate the fault location; on the fault path, combine equipment status and changes in electrical parameters to further refine the location; for example, by comparing the voltage distribution curves before and after the fault, find the node or edge with the most obvious voltage drop in the graph structure; or analyze the flow direction of the fault current to determine the smallest electrical isolation area where the fault point is located; for complex faults (such as multiple faults), use a hierarchical graph model or fault tree analysis method to gradually narrow down the fault range until the specific equipment or line segment is accurately located. Finally, the fault location results are generated and presented in a visual manner, such as highlighting the faulty equipment and path on a single-line diagram, and labeling the fault type, severity, and location reliability. At the same time, a text report containing information such as fault location, possible causes, and scope of impact is generated to provide decision support for maintenance personnel. For example, if the location results show a fault in a branch line of a busbar, the report will suggest checking vulnerable components such as insulators and cable joints of that line and assessing the impact on adjacent equipment. By combining graph theory models with fault characteristics, this method makes full use of the topology of the power grid and the coordinated action information of protection devices. It is especially suitable for complex 500kV bus systems with multiple power sources and multiple branches, and can quickly and accurately locate faults, reduce fault recovery time, and improve power grid reliability.

[0025] Step 6: Generate a fault isolation strategy based on the fault location results to control the operation of circuit breakers and disconnect switches; at the same time, predict the future operating status of the 500kV bus protection device through time series analysis algorithms and neural network prediction models. Specifically, firstly, a fault isolation strategy is generated based on the fault location results. Based on the fault paths and equipment states identified in the graph structure model, combined with the real-time power grid topology and operational constraints (such as power supply reliability requirements and power flow distribution limitations), the optimal isolation strategy is automatically generated through a deep Q-network (DQN) reinforcement learning model. This model aims to minimize the power outage range and minimize the isolation time. It takes the fault location, type, severity, and power grid topology parameters as input and outputs the action sequence of circuit breakers and disconnect switches. For example, if a fault is located in a bus section switch, the model prioritizes disconnecting the section switch and the disconnect switches on both sides, while maintaining normal power supply to the non-faulty bus. Then, the isolation strategy is executed and the action effect is verified; opening and closing commands are sent to the circuit breaker and disconnector via GOOSE messages or hard contact signals from the substation automation system (SAS); after the action is completed, the switch position signals and bus voltage / current data are collected in real time to verify whether the fault has been isolated (e.g., no current at the fault point, voltage in the non-fault area returns to normal); if the verification fails (e.g., there is over-tripping or incomplete isolation), the backup strategy is triggered to re-execute to ensure that the fault is quickly eliminated. Next, time series analysis and future state prediction are initiated. A Long Short-Term Memory (LSTM) network model is used to train the protection device on historical time-series data (such as temperature, voltage, and number of actions over the past 7 days) to learn the long-term dependencies of parameter changes. The model input is time-series data within a sliding time window (such as 1440 sampling points from the last 24 hours), and the output is predicted values ​​for key parameters for the next 72 hours (such as hourly predictions). For example, if it is predicted that the CPU temperature of a protection device will exceed 85°C (a preset threshold) after 48 hours, a "hardware overheating warning" is triggered, prompting maintenance personnel to check the cooling system in advance. Meanwhile, an error correction mechanism is introduced to improve prediction accuracy. The recurrent neural network (RNN) is used to compare the LSTM predictions with the actual operating data of the device in real time. If the error exceeds 10% (e.g., the deviation between the predicted voltage and the measured value is >5%), the model parameters are automatically updated and online training is performed using the latest data. The update cycle is 10 minutes. In addition, the seasonal decomposition (STL) algorithm is combined to remove periodic components (e.g., daily load fluctuations) from the time series data, thereby improving the sensitivity to abnormal trends, such as identifying non-periodic hardware aging characteristics. Finally, a prediction report is generated and linked to the operation and maintenance system. The future state prediction results and fault warning information are integrated into a visual report, which is pushed through the on-site monitoring interface and the remote operation and maintenance platform. The report content includes: key parameter trend curves, warning levels (such as green "normal", yellow "warning", red "emergency"), and maintenance suggestions (such as "replace aging capacitors within 3 days"). For emergency warnings, the system automatically links to the spare parts management system to allocate resources in advance and shorten maintenance preparation time. Through the above process, the solution achieves automated and intelligent fault isolation, and at the same time, it transforms the operation and maintenance mode from "post-event repair" to "pre-event prevention" through forward prediction, which significantly improves the reliability and operation and maintenance efficiency of the 500kV bus system.

[0026] In this embodiment, specifically, the method for extracting the fault feature vector in step two includes the following steps: Step 201: Extract two-dimensional features from the preprocessed running data using the convolutional layers in the convolutional neural network to generate an initial feature map; Specifically, firstly, the preprocessed running data is organized into a two-dimensional tensor structure suitable for convolution operations; for time-series data (such as voltage and current waveforms), a sliding window technique is used to convert them into a two-dimensional matrix, where rows represent time series points and columns represent different monitoring parameters (such as three-phase voltage and zero-sequence current); for example, a sliding window with a length of 200ms is selected (corresponding to 100 sampling points and a sampling frequency of 1000Hz), and the window step size is set to 20ms. The data in each window is reshaped into a 10×10 two-dimensional matrix, and the matrix elements are normalized electrical quantity values; Then, the convolutional layer parameters are designed for feature extraction; the first convolutional layer is configured with 16 3×3 convolutional kernels, a stride of 1, and padding set to "same" to maintain the output size consistent with the input; each convolutional kernel performs a convolution operation on the input matrix through a sliding window, calculating a linear combination of local regions to generate the corresponding feature map; for example, the first convolutional layer... Output feature map of each convolutional kernel It can be represented as: ; in, For convolution kernel weights, For the input matrix, The bias term is ReLU, and the activation function is ReLU. Next, batch normalization is applied to optimize the feature maps; the output of each convolutional layer is normalized to adjust the feature value distribution to a standard normal distribution with a mean of 0 and a variance of 1, as shown in the formula: ; in, and These are the mean and variance of the batch data, respectively. To prevent small constants with a denominator of zero, and These are learnable scaling and translation parameters; this operation accelerates network convergence and enhances model stability. Finally, nonlinearity is introduced through the ReLU activation function; the ReLU function is applied element-wise to the batch-normalized feature maps. By setting negative eigenvalues ​​to 0 and retaining positive eigenvalues, the network can learn complex nonlinear relationships. For example, for bus short-circuit faults, the ReLU activation function can highlight the characteristics of sudden current changes and suppress fluctuation noise during normal operation.

[0027] Step 202: Calculate the attention weights of each region in the initial feature map using the attention mechanism; Specifically, firstly, global information aggregation is performed on the initial feature maps; then, Global Average Pooling (GAP) is used to aggregate the information from each feature map. Compressing it into a scalar value yields the channel descriptor: ; Here, H and W represent the height and width of the feature map, respectively; this operation compresses spatial dimensional information while preserving the global statistical properties of the channel dimension, forming a feature vector. (C represents the number of channels); Then, an attention weight generation network is constructed; a two-layer fully connected network (FCN) is designed to perform nonlinear transformations on the channel descriptors: The first layer of the FCN maps z to a lower-dimensional space, reducing the dimension to... (r is the compression ratio, set to 16), the activation function uses ReLU: ; The second-layer FCN restores the features to the original channel dimension C and normalizes the output values ​​to the [0,1] interval using the Sigmoid function to generate channel attention weights. : ; Next, spatial attention weights are calculated; for the initial feature map Aggregate along the channel dimension and calculate the mean and maximum values ​​respectively: , ; Concatenate these two feature maps along the channel dimension to obtain Then, the number of channels is compressed to 1 using a 7×7 convolutional layer, and the spatial attention weights are generated using the Sigmoid function. : ; Finally, the channel and spatial attention weights are merged; the channel attention weights are... Expanded to the same spatial dimension as the feature map, with spatial attention weights. Element-wise multiplication yields the final attention weight matrix. : ; This weight matrix represents the importance of each position and channel in the initial feature map, and can be used for feature enhancement later.

[0028] Step 203: Multiply the attention weights element-wise with the initial feature map to obtain a weighted feature map; Specifically, first, confirm that the attention weight matrix has the same dimensions as the initial feature map; assuming the initial feature map is... (H is height, W is width, C is number of channels), the attention weight matrix calculated in step 202 is as follows: , where each element Represents the position in the feature map Attention weight values ​​for channel k (range [0,1]); Then, perform element-wise multiplication; for each spatial location in the feature map. And channel k, the initial eigenvalues With the corresponding attention weights Multiply to obtain the weighted eigenvalues. The calculation formula is: ; This operation achieves adaptive enhancement and suppression of the initial feature map: For fault-related key feature regions (such as current abrupt changes and voltage phase transition regions), if attention weights If it is close to 1, then The original feature values ​​are preserved, and even further enhanced due to the weights being slightly greater than 1; For background noise or irrelevant feature regions, if attention weights If it is close to 0, then It was significantly suppressed, approaching 0; Taking a single-phase ground fault on the busbar as an example, the spatial location of the zero-sequence current path (k=zero-sequence current path) in the initial characteristic diagram at the time of the fault is... attention weights The value can reach above 0.9, and the zero-sequence current characteristic value in this region is significantly enhanced after multiplication; while in the voltage fluctuation region during normal operation, the attention weight may be as low as below 0.1, and the corresponding characteristic value is suppressed to less than 10% of the original value. Finally, a weighted feature map is generated. The weighted feature map is then output to the subsequent residual connection layer. This weighted feature map not only preserves the spatial distribution and channel information of the original features, but also achieves the effect of "highlighting key features and weakening noise background" through the attention mechanism. It provides a more discriminative input for the nonlinear transformation of the subsequent residual connection layer, and in particular improves the feature representation ability of weak latent faults (such as the slow increase of contact resistance of relay contacts).

[0029] Step 204: Perform a nonlinear transformation on the weighted feature map through at least three residual connection layers in the convolutional neural network to output the fault feature vector; the residual connection layers alleviate the gradient vanishing problem through a skip connection structure; Specifically, first, construct the residual connection layer architecture; design at least three residual blocks, each containing two convolutional operations and one skip connection; taking the first residual block as an example, its structure is as follows: 1. Batch Normalization layer, which processes the input feature map. Standardize the process; 2. The ReLU activation function introduces nonlinear characteristics; 3. A 3×3 convolutional layer with 64 kernels and "same" padding to keep the feature map size constant; 4. Batch standardization layer; 5. ReLU activation function; 6. 3×3 convolutional layers, with the same number of 64 kernels; 7. Shortcut connection: This connects the original input... Adding it to the convolution output yields the residual block output. : ; Then, multiple layers of residual blocks are stacked; the second layer of residual blocks receives the output of the first layer. The structure is similar to the first layer, but the number of convolutional kernels increases to 128; the number of convolutional kernels in the third residual block further increases to 256; each residual block directly adds the input to the output through skip connections, forming... and The recursive structure enables the network to learn the residual mapping between input and output, effectively mitigating the gradient vanishing problem in deep networks. Next, global average pooling is performed; the last residual block is then output. Global average pooling is performed along the spatial dimensions (height and width) to compress the two-dimensional feature map of each channel into a scalar value, resulting in a feature vector of length 256. : , ; Finally, output the fault feature vector; and perform global average pooling on the feature vector. L2 normalization is performed to eliminate the influence of vector magnitude differences, resulting in the final fault feature vector. : ; Each dimension of this vector corresponds to a different type of fault feature. For example, the first 32 dimensions may represent short-circuit fault features, the middle 64 dimensions correspond to grounding fault features, and the last 16 dimensions are related to hardware anomalies of the protection device. Through the deep nonlinear transformation of the residual connection layer, this feature vector can capture the complex patterns of the bus protection device's operating status, especially showing good characterization ability for early latent faults (such as partial discharge features in the early stage of insulation aging).

[0030] In practical applications, this fault feature vector will be input into the subsequent principal component analysis (PCA) module for dimensionality reduction, further extracting the most representative fault features, providing a basis for the final fault type identification and severity assessment.

[0031] In this embodiment, specifically: in step four, the support vector machine model uses the radial basis function as the kernel function, and optimizes the penalty parameter C and the kernel function parameter γ through the k-fold cross-validation method; Specifically, when dealing with non-linearly separable data, Support Vector Machines (SVMs) need to map the original features to a high-dimensional space using kernel functions to find the linear classification boundary. Radial Basis Function (RBF) is a commonly used kernel function, and its form is: ; Where γ is the kernel function parameter; In fault diagnosis scenarios, the RBF kernel function can capture the complex nonlinear relationships between fault feature vectors. For example, bus short-circuit faults and ground faults may be linearly separable regions in high-dimensional space. By adjusting the γ value, the sensitivity of the model to local features can be controlled: a smaller γ makes the model consider a wider neighborhood and enhances the generalization ability; a larger γ makes the model focus on local details, which may lead to overfitting. The performance of an SVM model is highly dependent on the penalty parameter C and the kernel function parameter γ. The penalty parameter C controls the model's tolerance for classification errors: a larger C forces the model to classify strictly, potentially leading to overfitting; a smaller C allows for more misclassifications, improving generalization ability. The kernel function parameter γ determines the "width" of the RBF kernel function, affecting the granularity of the model's feature space partitioning. k-fold cross-validation is a method for evaluating the model's generalization ability; the specific steps are as follows: The dataset is divided into k subsets. The model is trained using k-1 subsets in turn, and the model performance is verified using the remaining subset. Finally, the average value of the k verification results is taken as the model evaluation metric. By iterating through different combinations of C and γ (e.g., C∈[0.1,100], γ∈[0.001,10]), the parameter combination that minimizes the cross-validation error can be selected to ensure that the model performs optimally on unknown data. In the fault diagnosis of 500kV bus protection devices, the optimized parameters can significantly improve the classification accuracy of SVM for different fault types (such as short circuit, overload, and equipment abnormality).

[0032] The fuzzy membership function is generated based on historical fault data. The input is the fault feature vector after dimensionality reduction, and the output is the severity of the fault. Specifically, fuzzy logic theory allows things to exist in a way that "partially belongs" to a certain category, and this uncertainty is quantified by membership functions. In fault severity assessment, multiple severity levels (such as "minor", "moderate", and "severe") are defined for each fault type (such as short circuit and grounding), and each level corresponds to a membership function. These functions map the fault feature vector to the interval [0,1], representing the degree to which the feature vector belongs to a specific severity level. For example, for the "severe short circuit" level, the membership function may output a value close to 1 when the fault current amplitude exceeds 5 times the rated current, and a value close to 0 when the current amplitude is small. The parameters of the fuzzy membership function are obtained through training on historical fault data. First, a large number of fault cases are collected, each containing a dimensionality-reduced fault feature vector (e.g., a 10-dimensional vector after PCA dimensionality reduction) and manually labeled severity levels. Then, optimization algorithms (e.g., least squares method, genetic algorithm) are used to adjust the parameters of the membership function (e.g., the mean and standard deviation of the Gaussian function) to maximize the matching degree between the function output and the historical labels. For example, for a Gaussian membership function for the "moderate short circuit" level, the training process will search for the optimal mean and standard deviation so that when a feature vector historically labeled as "moderate short circuit" is input, the function output value is close to 1, while the output value is close to 0 when feature vectors of other levels are input. After training, for a new fault feature vector, the most likely severity level can be determined by calculating its membership value to each severity level (e.g., taking the level with the highest membership degree), or a quantified severity score (e.g., 0-100 points) can be obtained through weighted averaging, providing an accurate basis for operation and maintenance decisions.

[0033] In this embodiment, specifically: in step five, the graph structure model is constructed with the components and busbar segments of the protection device as nodes, and the electrical connection relationships and signal transmission logic between nodes as edges; Specifically, firstly, the physical components of the 500kV busbar protection device (such as circuit breakers, relays, CT / PT sensors) and busbar segments are abstracted as nodes in the diagram; each node contains attribute information, such as equipment type, real-time status (normal / abnormal), and fault severity score (from the fuzzy logic evaluation results in step four); for example, the busbar segment node contains electrical parameters such as voltage and current, and the protection device node contains operating status (such as tripped / not tripped), operating timestamp, etc. Then, the electrical connection edge (such as the connection line between the busbar and the circuit breaker) carries parameters such as impedance and capacity; the signal transmission edge (such as the control signal between the protection device and the circuit breaker) contains logical relationships (such as "protection action → triggering circuit breaker tripping"); the weight of the edge can be dynamically adjusted according to the equipment reliability, fault propagation probability, etc. For example, during a short circuit fault, the edge weight on the fault current path increases significantly. In the dual busbar connection method, each busbar segment is treated as an independent node and connected to the circuit breaker node; the protection device node is connected to the corresponding circuit breaker node through the signal edge, forming a directed graph structure that fully reflects the operation logic of the protection device and the fault propagation path.

[0034] The fault location is determined by analyzing the fault type, fault severity, and the operation information of the 500kV bus protection device, and by using graph theory algorithms to traverse and search for the abnormal signal propagation path in the graph structure. Specifically, the fault type (such as "bus short circuit") and fault severity score (such as 85 points) identified in step four, along with the real-time action information of the protection device (such as "circuit breaker C1 trips at time t" and "protection relay R2 issues an overcurrent signal"), are used as inputs. Based on a graph structure model, starting from the node with the most significant fault characteristics (such as the bus node with the largest voltage drop), and combining the operating timing of the protection device, the propagation path of the abnormal signal is traced in reverse; for example: If a short circuit is detected at a bus node (such as voltage approaching zero and current surge), and the circuit breaker C1 connected to it trips, then C1 is taken as the starting point. Tracing back along the signal transmission edge to the protection relay R1 that triggered C1 to trip, check the matching degree of R1's action record and fault type; Continue searching for adjacent devices along the electrical connection edge to determine whether the fault was caused by a fault in an adjacent line or device. By analyzing the fault characteristic intensity (such as the amplitude of current surge and voltage deviation) of each node on the abnormal signal propagation path and the rationality of the protection device operation (such as the main protection taking priority and whether the backup protection is malfunctioning), the minimum isolation area where the fault occurs can be determined. For example, if the main protection and circuit breaker of a certain line both operate and the voltage at the end of the line is abnormal, the fault point is located inside the line.

[0035] Graph theory algorithms include breadth-first search or depth-first search; Among them, Breadth-First Search (BFS) starts from the starting node (such as the bus node with the strongest fault characteristics) and traverses adjacent nodes layer by layer, prioritizing the access to the node closest to the starting node; it is suitable for quickly locating the scope of a fault. For example, in the event of a bus short circuit, BFS can quickly determine which circuit breakers should operate to isolate the fault; the hierarchical traversal characteristic of BFS ensures that directly connected critical equipment is checked first, avoiding redundant searches. Depth-first search (DFS) starts from the starting node and visits nodes as deep as possible along a path until it cannot continue or the fault point is found, and then backtracks. It is suitable for verifying the rationality of the fault propagation path, such as checking whether the action of the protection device conforms to the logical order of "main protection → near backup protection → far backup protection". DFS is more efficient when searching complex networks and can quickly eliminate impossible paths. In practical applications, the advantages of both algorithms can be combined: first, use BFS to quickly locate the fault area, and then use DFS to accurately verify the fault path; for example, in a 500kV bus protection system, when multiple sets of protection devices operate, BFS is used to determine the suspected fault area, and then DFS is used to backtrack and verify the correctness of the operation of each protection device, and finally locate the fault point.

[0036] In this embodiment, specifically: in step one, the operating data includes the input and output signals of the 500kV bus protection device, internal electrical parameters, hardware status parameters, and voltage and current data of the 500kV bus; The input signal refers to the switching and analog signals that are externally connected to the protection device, such as: Switching signals: circuit breaker position signal (open / closed), disconnector status, and linkage signals of other protection devices (such as adjacent bay protection action contacts); Analog signals: Secondary voltage / current signals (such as three-phase bus voltage and zero-sequence current) from voltage transformers (PT) and current transformers (CT), used for fault discrimination logic of protection devices; Output signals refer to the control signals and status signals sent by the protection device to the outside, such as: Control signals: circuit breaker trip command, reclosing command; Status signals: protection action signals (such as differential protection action, overcurrent protection action), device self-test abnormal signals (such as communication interruption, hardware fault alarm); Internal electrical parameters refer to the key electrical calculation parameters processed internally by the protection device, such as: Sampled value (SV): The original voltage / current sampled value after analog-to-digital conversion, used for real-time calculation of the protection algorithm; Differential current: The vector sum of the currents in each branch calculated in the bus differential protection, used to determine whether the fault is inside or outside the bus zone; Harmonic components: Voltage / current harmonic content (such as the 2nd harmonic and the 5th harmonic) extracted by Fourier transform, used to identify non-fault anomalies such as inrush current; Phase difference: The phase relationship between the currents in each branch, used to determine the type of fault (such as phase-to-phase short circuit, single-phase grounding). Hardware status parameters are parameters that reflect the health status of the protection device's own hardware, and are monitored in real time through built-in sensors, for example: Temperature: CPU chip temperature, power module temperature, relay contact temperature, used to warn of hardware overheating and aging; Voltage / Current: Internal power supply voltage of the device (e.g., 5V, 12V), operating current of each module, used to detect power supply abnormalities or component short circuits; Mechanical condition: plug contact resistance (measured by spring probe or Hall sensor), fan speed (to determine if the cooling system is working properly); Storage status: remaining Flash memory capacity, EEPROM write count, used to assess the lifespan of the storage medium; The voltage and current data of the 500kV busbar are primary electrical quantities that directly reflect the operating status of the busbar. They are collected by PT / CT and connected to the protection device via the secondary circuit. For example: Voltage data: bus three-phase voltage amplitude, zero-sequence voltage, and voltage waveform distortion rate, used to determine whether there is overvoltage, undervoltage, or grounding fault on the bus. Current data: Three-phase current amplitude of each branch of the busbar.

[0037] Preprocessing includes denoising and normalization. The denoising process uses the least mean square adaptive filtering algorithm, and the normalization process uses the Z-score normalization method to map the running data to a standard normal distribution. Specifically, the noise reduction method using the Least Mean Square (LMS) adaptive filtering algorithm is as follows: Set the filter order N (usually 16-64, adjust according to noise characteristics); By using the recursive formula: ; The filter weights are updated in real time, where: This is the step size factor (ranging from 0.001 to 0.1, which needs to balance convergence speed and steady-state error). This is the error signal at the current moment (the difference between the original data and the filtered output); Input the data vector for the current time step; By filtering noisy signals using LMS filtering, electromagnetic interference (EMI), white noise and other high-frequency noise can be suppressed, and the signal-to-noise ratio (SNR) can be improved by 10-15dB, effectively suppressing transient interference during bus faults.

[0038] The specific method for mapping the running data to a standard normal distribution using the Z-score normalization method is as follows: The formula for calculating the normalization method is: ; in, The mean of the data. The standard deviation of the data; By calculating the mean and standard deviation of continuous data such as voltage and current dimension by dimension, dimensionless processing is achieved. After processing, the mean of the data is 0 and the standard deviation is 1, ensuring the stability of the CNN input layer.

[0039] In this embodiment, specifically: In step six, the neural network prediction model is a long short-term memory network model. The long short-term memory network model is input with the historical time series data of the 500kV bus protection device. Through training, it learns the long-term dependencies in the historical time series data and outputs the predicted value of the change trend of key parameters in the next 72 hours. If the predicted value exceeds the preset threshold, a fault warning is triggered. The historical time series data includes temperature, voltage, and number of actions. Specifically, a Long Short-Term Memory (LSTM) network model is used to predict the future operating status of the 500kV bus protection device. This model takes the device's historical time-series data (including temperature, voltage, and number of operations) as input and trains by mining long-term dependencies in the data (such as the temperature change trend over time, the periodicity of voltage fluctuations, and the correlation between the number of operations and the device status). It can output predicted values ​​of the change trends of key parameters in the next 72 hours. When the predicted value exceeds the preset safety threshold (such as excessively high temperature, voltage exceeding limits, abnormally frequent operations, etc.), the system will automatically trigger a fault warning, thereby helping maintenance personnel to discover potential equipment anomalies in advance, providing a basis for preventive maintenance, and effectively improving the reliability of the bus protection device and the safety and stability of the power grid.

[0040] In this embodiment, specifically: in step three, the dimensionality reduction process calculates the covariance matrix of the fault feature vectors, performs eigenvalue decomposition on the covariance matrix, and selects the fault feature vectors corresponding to the principal components with a cumulative variance contribution rate ≥ 95% to form the projection matrix. Specifically, first, the covariance matrix of the fault feature vectors is calculated, which reflects the correlation between the features. Next, the covariance matrix is ​​decomposed into eigenvalues ​​to obtain eigenvalues ​​and their corresponding eigenvectors, where the magnitude of the eigenvalue represents the variance contribution rate (i.e., the proportion of information carried) of the corresponding principal component. Then, the eigenvalues ​​are sorted from largest to smallest, and the variance contribution rates are accumulated sequentially until the accumulated variance contribution rate reaches or exceeds 95%. The top k eigenvectors at this point are selected to form a projection matrix. Finally, the original fault feature vectors are mapped to a k-dimensional space through this projection matrix, achieving data dimensionality reduction. This significantly reduces the feature dimension while retaining more than 95% of the key information, improving the efficiency of subsequent fault diagnosis and the model's generalization ability.

[0041] In summary, the fault diagnosis method for a 500kV busbar protection device provided by this invention acquires and preprocesses operational data, extracts fault feature vectors using a convolutional neural network combined with an attention mechanism, reduces dimensionality through principal component analysis, identifies fault types and assesses severity using support vector machines and fuzzy logic, constructs a model based on graph theory to locate the fault location, and finally generates an isolation strategy and predicts future operating status using a long short-term memory network. This method accurately captures fault information, improves diagnostic accuracy, quickly locates faults, and enables early warning, providing a scientific basis for equipment maintenance, improving the reliability and stability of the power system, and reducing maintenance costs and power outage losses.

[0042] Figure 3 This is a functional module diagram of a fault diagnosis system for a 500kV busbar protection device according to an embodiment of this application, as shown below. Figure 3 As shown, a fault diagnosis system for a 500kV busbar protection device includes: a data acquisition and processing module, a feature extraction module, a feature dimensionality reduction module, a fault diagnosis module, a fault location module, and a fault processing and prediction module. The data acquisition and processing module is configured to acquire the operating data of the 500kV busbar protection device and preprocess the acquired operating data. The feature extraction module is configured to perform feature recognition on the preprocessed running data and extract fault feature vectors based on an improved convolutional neural network combined with an attention mechanism; the improved convolutional neural network includes at least three residual connection layers. The feature dimensionality reduction module is configured to perform dimensionality reduction on the extracted fault feature vectors based on principal component analysis, retaining principal components with a cumulative variance contribution rate of ≥95%. The fault diagnosis module is configured to input the dimensionality-reduced fault feature vector into the support vector machine model for fault type identification, and combine fuzzy logic theory to evaluate the severity of the fault through fuzzy membership function. The fault location module is configured to construct a graph structure model of the 500kV busbar protection device based on graph theory. By analyzing and identifying the fault type and severity, and combining the action information of the 500kV busbar protection device, it searches for the fault path in the graph structure, locates the fault location, and generates the fault location result. The fault handling and prediction module is configured to generate fault isolation strategies based on fault location results and control the operation of circuit breakers and disconnect switches; at the same time, it predicts the future operating status of the 500kV bus protection device through time series analysis algorithms and neural network prediction models.

[0043] In this embodiment, specifically: the attention mechanism in the feature extraction module is a dynamic adaptive attention mechanism; The specific method for optimizing fault feature weight allocation using a dynamic adaptive attention mechanism is as follows: Construct a multi-branch attention subnetwork and calculate channel-level attention weights and spatial-level attention weights respectively; Introducing a gating mechanism, through formulas The weights of the two branches are combined, where, For channel feature vectors, For spatial characteristic matrix, Use the Sigmoid activation function; The attention weight distribution is dynamically adjusted based on real-time fault data to enhance the ability to focus on transient fault characteristics.

[0044] In this embodiment, the fault handling and prediction module specifically includes: a strategy generation unit, a prediction error correction unit, and an edge computing node; The strategy generation unit is used to build a fault isolation strategy optimization model based on a deep Q-network. With the goal of minimizing fault isolation time and power outage range, it takes the fault type, fault severity and power grid topology as input and outputs the optimal action sequence of circuit breakers and disconnectors. The prediction error correction unit is used to compare the predicted values ​​of the long short-term memory network model with the actual operating data of the device in real time through the recurrent neural network; when the prediction error exceeds the preset threshold, it automatically triggers the model parameter update, with an update cycle of 10 minutes. Edge computing nodes are used to deploy reinforcement learning models and time series analysis algorithms on the local edge computing unit of the protection device, enabling millisecond-level generation and execution of fault isolation strategies and reducing reliance on cloud servers.

[0045] In summary, the fault diagnosis system for a 500kV busbar protection device provided by this invention acquires and preprocesses operational data through a data acquisition and processing module; a feature extraction module extracts fault feature vectors based on an improved convolutional neural network combined with a dynamic adaptive attention mechanism; a feature dimensionality reduction module uses principal component analysis for dimensionality reduction; a fault diagnosis module performs diagnostic evaluation through support vector machines and fuzzy logic; a fault location module constructs a model based on graph theory to locate faults; and a fault handling and prediction module generates isolation strategies and deploys models using edge computing nodes to predict future states. The various modules of this system work collaboratively to automate and intelligently manage fault diagnosis, location, handling, and prediction, improving fault handling efficiency and equipment maintenance levels, and ensuring the safe and stable operation of the power grid.

[0046] For further details regarding the implementation of the technical solutions for each module in the fault diagnosis system of a 500kV busbar protection device in the above embodiments, please refer to the description in the fault diagnosis method of a 500kV busbar protection device in the above embodiments, which will not be repeated here.

[0047] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For system-type embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0048] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fault diagnosis method for a 500kV busbar protection device, characterized in that, Includes the following steps: Acquire the operating data of the 500kV busbar protection device and preprocess the acquired operating data; Based on a convolutional neural network combined with an attention mechanism, feature recognition is performed on the preprocessed running data, and fault feature vectors are extracted; the convolutional neural network includes at least three residual connection layers. Principal component analysis was used to reduce the dimensionality of the extracted fault feature vectors, retaining principal components with a cumulative variance contribution rate of ≥95%. The fault feature vector after dimensionality reduction is input into the support vector machine model for fault type identification, and the severity of the fault is evaluated by fuzzy membership function in combination with fuzzy logic theory. Based on graph theory, a graph structure model of a 500kV busbar protection device is constructed. By analyzing and identifying the fault types and severity, and combining the action information of the 500kV busbar protection device, the fault path is searched in the graph structure, the fault location is located, and the fault location result is generated. Based on the fault location results, a fault isolation strategy is generated to control the operation of circuit breakers and disconnect switches; at the same time, the future operating status of the 500kV bus protection device is predicted through time series analysis algorithms and neural network prediction models.

2. The fault diagnosis method for a 500kV busbar protection device according to claim 1, characterized in that, The method for extracting fault feature vectors includes the following steps: Two-dimensional feature extraction is performed on the preprocessed running data through the convolutional layers in the convolutional neural network to generate an initial feature map; The attention weights of each region in the initial feature map are calculated using an attention mechanism. The attention weights are multiplied element-wise with the initial feature map to obtain a weighted feature map; The weighted feature map is nonlinearly transformed by at least three residual connection layers in the convolutional neural network to output a fault feature vector; the residual connection layers alleviate the gradient vanishing problem through a skip connection structure.

3. The fault diagnosis method for a 500kV busbar protection device according to claim 1, characterized in that, The support vector machine model uses radial basis functions as kernel functions and optimizes the penalty parameter C and kernel function parameter γ through k-fold cross-validation. The fuzzy membership function is generated based on historical fault data, inputs the dimensionality-reduced fault feature vector, and outputs the fault severity.

4. The fault diagnosis method for a 500kV busbar protection device according to claim 1, characterized in that, The graph structure model is constructed with the components of the protection device and the busbar segments as nodes, and the electrical connection relationships and signal transmission logic between nodes as edges. The fault location is determined by analyzing the fault type, fault severity, and the action information of the 500kV busbar protection device, and by traversing and searching for the abnormal signal propagation path in the graph structure based on graph theory algorithms. The graph theory algorithm is either breadth-first search or depth-first search.

5. The fault diagnosis method for a 500kV busbar protection device according to claim 1, characterized in that, The operating data includes the input and output signals, internal electrical parameters, hardware status parameters of the 500kV bus protection device, and voltage and current data of the 500kV bus. The preprocessing includes denoising and normalization. The denoising is performed using the least mean square adaptive filtering algorithm, and the normalization is performed using the Z-score normalization method to map the operating data to a standard normal distribution.

6. The fault diagnosis method for a 500kV busbar protection device according to claim 1, characterized in that, The neural network prediction model is a long short-term memory network model. The long short-term memory network model takes historical time series data of the 500kV bus protection device as input, learns the long-term dependencies in the historical time series data through training, and outputs the predicted value of the change trend of key parameters in the next 72 hours. If the predicted value exceeds the preset threshold, a fault warning is triggered. The historical time-series data includes temperature, voltage, and number of actions.

7. The fault diagnosis method for a 500kV busbar protection device according to claim 1, characterized in that, The dimensionality reduction process calculates the covariance matrix of the fault feature vectors, performs eigenvalue decomposition on the covariance matrix, and selects the fault feature vectors corresponding to the principal components with a cumulative variance contribution rate ≥ 95% to form the projection matrix.

8. A fault diagnosis system for a 500kV busbar protection device, applied to the fault diagnosis method for a 500kV busbar protection device according to any one of claims 1-7, characterized in that, include: The system includes a data acquisition and processing module, a feature extraction module, a feature dimensionality reduction module, a fault diagnosis module, a fault location module, and a fault handling and prediction module. The data acquisition and processing module is configured to acquire the operating data of the 500kV busbar protection device and preprocess the acquired operating data. The feature extraction module is configured to perform feature recognition on the preprocessed running data and extract fault feature vectors based on an improved convolutional neural network combined with an attention mechanism; the improved convolutional neural network includes at least three residual connection layers. The feature dimensionality reduction module is configured to perform dimensionality reduction processing on the extracted fault feature vector based on principal component analysis, retaining principal components with a cumulative variance contribution rate ≥ 95%. The fault diagnosis module is configured to input the dimensionality-reduced fault feature vector into the support vector machine model for fault type identification, and combine fuzzy logic theory to evaluate the severity of the fault through fuzzy membership function. The fault location module is configured to construct a graph structure model of the 500kV busbar protection device based on graph theory. By analyzing and identifying the fault type and severity, and combining the action information of the 500kV busbar protection device, it searches for the fault path in the graph structure, locates the fault location, and generates the fault location result. The fault handling and prediction module is configured to generate a fault isolation strategy based on the fault location result and control the operation of the circuit breaker and disconnector; at the same time, it predicts the future operating status of the 500kV bus protection device through time series analysis algorithm and neural network prediction model.

9. A fault diagnosis system for a 500kV busbar protection device according to claim 8, characterized in that, The attention mechanism in the feature extraction module is a dynamic adaptive attention mechanism.

10. A fault diagnosis system for a 500kV busbar protection device according to claim 8, characterized in that, The fault handling and prediction module includes: a strategy generation unit, a prediction error correction unit, and an edge computing node; The strategy generation unit is used to construct a fault isolation strategy optimization model based on a deep Q-network; The prediction error correction unit is used to compare the predicted value of the long short-term memory network model with the actual operating data of the device in real time through a recurrent neural network. The edge computing node is used to deploy reinforcement learning models and time series analysis algorithms on the local edge computing unit of the protection device.

Images