Comprehensive online monitoring system for cable branch box

Abstract

The invention discloses a cable branch box integrated on-line monitoring system comprising a data acquisition module used for acquiring operation monitoring data and historical operation data of a cable branch box; the data transmission module is used for transmitting the acquired data to the central monitoring platform; the data preprocessing module is used for preprocessing the data; the analysis module is used for analyzing the data by applying an improved transformation convolutional neural network and identifying the fault and abnormal state of the cable branch box; the prediction module is used for combining historical operation data and current monitoring data, and performing fault prediction by applying an improved graph neural network and combining an attention mechanism; the central monitoring platform is used for generating fault early warning information; and the maintenance optimization module is used for providing corresponding prevention measures and maintenance suggestions. According to the method, the improved convolutional neural network and the graph neural network are combined, the data of the cable branch box are intelligently analyzed, fault prediction and early warning are realized, and the method has the advantages of high precision, high real-time performance and optimal maintenance.

Description

Technical Field

[0001] The present invention relates to the field of intelligent monitoring of power equipment, and particularly to an integrated on-line monitoring system for cable branch boxes. Background Art

[0002] As an important part of the power system, cable branch boxes are responsible for distributing electric energy from the main cable to each electrical equipment, so they play a crucial role in power transmission and distribution. With the increasing demand for electricity, the scope of use of cable branch boxes has gradually expanded, and their status in the power transmission and distribution network has become more and more important. However, most traditional monitoring methods for cable branch boxes rely on manual inspections or static detection methods, which have problems such as poor real-time performance, low accuracy, and high maintenance costs. These traditional monitoring methods not only make it difficult to detect faults or abnormal states in cable branch boxes in a timely manner, but also cannot effectively predict faults and give early warnings, resulting in the instability of the power system operation and the risk of unexpected shutdown of equipment.

[0003] At present, the monitoring of cable branch boxes mainly relies on manual inspections and regular checks. The method of manual inspection not only has low efficiency, but also is prone to missed detection of faults due to the interference of human factors. Regular checks rely on set time periods, but this periodic inspection method cannot understand the operating status of the equipment in real time. Especially when the equipment has been running for a long time, the probability of failure is relatively high. In addition, traditional monitoring means often rely on the monitoring of single-sensor data and do not fully consider multi-dimensional operating parameters of the equipment such as current, voltage, temperature, and humidity, thus reducing the predictability and response speed of fault occurrence.

[0004] In order to improve the monitoring efficiency of cable branch boxes and the ability to prevent faults, more and more research has begun to apply intelligent monitoring methods, especially using modern deep learning technologies for fault detection and prediction. For example, deep convolutional neural networks are applied to identify images or signal data of power equipment faults, but they are mainly applied to the pattern recognition of signal data and have limitations in the analysis of time series data. At the same time, graph neural networks have gradually been introduced into the monitoring and fault prediction of power equipment, especially in the modeling of the relationship between nodes in the power network. Although graph neural networks have strong structured data processing capabilities, they still face many challenges in practical applications, especially when dealing with large-scale data, it is easy to have problems of excessive computational complexity.

[0005] Some current intelligent monitoring systems usually transmit the data collected by sensors to the monitoring platform and conduct fault detection and analysis through data-based models. However, these methods still have some deficiencies. For example, most existing fault prediction models rely on static training data sets and fixed model parameters and cannot adaptively adjust their prediction capabilities according to the actual operating status and real-time data of the equipment. In addition, existing systems often ignore the temporal correlation between data and the dynamic changes in equipment status during the analysis process, resulting in the inability to update the fault prediction results in a timely manner and reducing the accuracy of fault early warning.

[0006] Upon further analysis, the common deep learning-based monitoring methods in the prior art rely on traditional convolutional neural networks or simple graph neural network models. The limitations of these methods are mainly reflected in the processing of complex time-series data. Although traditional convolutional neural networks perform excellently in the fields of image and signal processing, their ability to model time-series data is weak, and it is difficult to accurately capture the time-varying characteristics in cable branch boxes. Although graph neural networks can process graph-structured data and establish relationships between nodes, most traditional graph neural network models fail to effectively consider the dynamic changes between nodes and the weight differences of different nodes, resulting in limited accuracy of prediction results. Therefore, most of the fault prediction models in the prior art cannot take into account both the temporal characteristics of equipment status and the complexity of relationships between equipment, leaving a large room for optimization.

[0007] Although some intelligent technologies, such as neural network-based prediction models, have been introduced in the existing cable branch box monitoring methods, several key defects are still faced. First, traditional monitoring methods often cannot conduct real-time fault early warning, and the operation status information of the equipment is updated untimely, resulting in the inability to predict and prevent potential faults in advance. Second, existing methods mostly rely on the data of a single sensor and fail to effectively integrate multi-dimensional sensing data, which limits the accuracy and reliability of fault detection. Finally, the models of existing fault prediction systems are usually relatively fixed and lack adaptability, and cannot optimize the prediction strategy in real time according to the different operating environments and status changes of cable branch boxes.

[0008] Therefore, how to combine deep learning technologies, especially convolutional neural networks and graph neural networks, to establish a more accurate, efficient and adaptable cable branch box monitoring system has become the research focus of current technologies. Summary of the Invention

[0009] An object of the present invention is to provide a comprehensive on-line monitoring system for cable branch boxes. By combining an improved transform convolutional neural network and an improved graph neural network, the present invention uses an attention mechanism to deeply analyze data and predict faults. This method collects multi-dimensional monitoring data of cable branch boxes in real time, and through intelligent analysis and prediction, realizes early warning of faults and optimization of maintenance strategies, improving the safety, reliability and economy of the power system.

[0010] A comprehensive on-line monitoring system for cable branch boxes according to an embodiment of the present invention includes:

[0011] A data acquisition module for collecting operation monitoring data and historical operation data of the cable branch box;

[0012] A data transmission module for transmitting the collected operation monitoring data and historical operation data to the central monitoring platform through Internet of Things technology;

[0013] A data preprocessing module for preprocessing the real-time transmitted operation monitoring data;

[0014] An analysis module for analyzing the preprocessed data by applying an improved transform convolutional neural network to identify faults and abnormal states of the cable branch box;

[0015] A prediction module for combining historical operation data and current monitoring data and using an improved graph neural network combined with an attention mechanism for fault prediction;

[0016] A central monitoring platform for receiving the fault prediction results output by the prediction module, generating fault warning information, and displaying the operation status of the cable branch box through a visualization interface;

[0017] A maintenance optimization module for optimizing the maintenance strategy according to the fault prediction results, combining the equipment maintenance plan, and proposing corresponding preventive measures and maintenance suggestions.

[0018] Optionally, the modules are implemented by the following method:

[0019] S1. Collect operation monitoring data and historical operation data of the cable branch box through sensors;

[0020] S2. Transmit the collected operation monitoring data and historical operation data to the central monitoring platform through Internet of Things technology;

[0021] S3. Preprocess the real-time transmitted operation monitoring data to generate preprocessed data;

[0022] S4. Construct an improved transform convolutional neural network, apply the improved transform convolutional neural network to analyze the preprocessed data, use transform convolutional operations to extract features from the preprocessed data, and introduce residual connections into the improved transform convolutional neural network. Identify faults and abnormal states of the cable branch box by the fully connected layer of the improved transform convolutional neural network;

[0023] S5. Based on historical operation data and current operation monitoring data, combine the attention mechanism and the improved graph neural network. Update the node features of the improved graph neural network, adjust the information transmission between nodes, and perform fault prediction using the pooling layer and fully connected layer of the improved graph neural network by graph convolutional operations and calculating attention weights to generate prediction results;

[0024] S6. Generate fault warning information according to the analysis and prediction results, and display the operation status of the cable branch box through the monitoring platform;

[0025] S7. Combine the prediction results with the equipment maintenance plan, optimize the maintenance strategy, and propose corresponding preventive measures and maintenance suggestions.

[0026] Optionally, the operation monitoring data includes current, voltage, temperature, and humidity.

[0027] Optionally, S3 includes the following specific steps:

[0028] S31. Conduct a preliminary check on the received operation monitoring data to ensure the integrity and validity of the data. For values that do not meet the format requirements or show abnormalities, mark them through preset thresholds, and perform processing or elimination of missing values;

[0029] S32. Perform filtering and denoising processing on the operation monitoring data using a Kalman filter to remove existing random noise:

[0030]

[0031] Among them, represents the current estimated value, represents the estimated value at the previous moment, y k represents the current observed value, H k represents the state transition matrix, H k represents the Kalman gain;

[0032] S33. Perform normalization processing on the denoised data to eliminate the differences in data under different dimensions, so that the range of operation monitoring data is within a unified standard:

[0033]

[0034] Among them, x' represents the normalized data, x represents the original data value, min( x ) and max ( x ) respectively represent the minimum and maximum values in the original data values;

[0035] S34. Use the sliding window method to smooth the normalized data. Set a window size N and calculate the average value of the N data points before and after each data point:

[0036]

[0037] where y i represents the smoothed data point, x j represents the original data point, and P represents the window size;

[0038] S35. Perform time series processing on the smoothed data, sort the data in chronological order and construct time stamps;

[0039] S36. Perform difference processing on the time series data through the first-order difference formula:

[0040] Δx t =x t -x t-1 ;

[0041] where Δx t represents the data after difference, x t represents the data at the current moment, and x t-1 represents the data at the previous moment;

[0042] S37. Perform outlier detection on the processed data, use the Z-score method to measure the deviation of each data point, and if the Z-score is greater than the preset threshold, the data point is considered an outlier:

[0043]

[0044] where Z c represents the value of the Z-score, x represents the data point, μ represents the mean of the data set, and σ represents the standard deviation of the data set.

[0045] Optionally, S4 includes the following specific steps:

[0046] S41. Construct an improved transform convolutional neural network, and the improved transform convolutional neural network includes:

[0047] Input layer, input the preprocessed data and enter the network for processing;

[0048] Transform convolutional layer, extract local features and dynamically adjust the convolutional kernel;

[0049] The temporal convolutional layer extracts time-dependent features and captures the change patterns of time series;

[0050] The pooling layer downsamples the features;

[0051] The multi-scale convolutional layer captures features at different scales and processes global and local features;

[0052] The fully connected output layer integrates the extracted features and uses the ReLU activation function to output the prediction results;

[0053] S42. Process the preprocessed data using transform convolutional operations, in which the convolutional kernel is dynamically adjusted:

[0054]

[0055] Among them, represents the feature map extracted by the convolutional layer, M represents the height of the convolutional kernel, N represents the width of the convolutional kernel, represents the preprocessed input data, represents the convolutional kernel, which is used to perform convolutional operations with the input data to extract local features, b (1) represents the bias term, which is added to the convolutional result to adjust the output, m and n represent the position indices of the elements in the convolutional kernel, and i and j represent the position indices of the elements in the output feature map;

[0056] S43. Through the max pooling method, select the maximum value from each pooling window to generate the pooled feature map:

[0057]

[0058] Among them, represents the output feature map after the pooling operation, which reduces the spatial dimension, represents the output feature map of the convolutional operation, pool window represents the pooling window, max represents the maximum value, and ∈ represents the set operation "belongs to";

[0059] S44. Use the convolutional kernel to process the pooled feature map to capture the time-dependent relationship through temporal convolutional operations:

[0060]

[0061] Among them, represents the output of the temporal convolution, which is the captured time-dependent feature, T represents the length of the temporal convolutional kernel, represents the weight of the temporal convolutional kernel, represents the pooled feature map, b(3) Represents the bias term of the temporal convolution;

[0062] S45. Extract features simultaneously at different scales through multiple convolutional kernels:

[0063]

[0064] Among them, represents the output of the multi-scale convolution, containing feature information of multiple scales, represents the output of the temporal convolution operation and serves as the input to the multi-scale convolution, T m represents the length of the m-th scale convolutional kernel, b (4) represents the bias term of the multi-scale convolution, represents the weight of the m-th scale convolutional kernel, where m represents the number of scale convolutional kernels;

[0065] S46. Introduce a residual connection, and the features after the residual connection are passed to the fully connected layer for processing:

[0066]

[0067] Among them, Y (5) represents the output of the fully connected layer, the feature representation after the residual connection processing, ReLU represents the non-linear activation function, W (5) represents the weight matrix of the fully connected layer, Y (4) represents the output of the multi-scale convolution, b (5 represents the bias term of the fully connected layer, represents the input of the residual connection, which is the input feature from the multi-scale convolutional layer;

[0068] S47. According to the output of the fully connected layer, set a threshold. When the failure probability of a certain category exceeds the set threshold, it is considered that the cable branch box has this type of failure.

[0069] Optionally, the S5 includes the following specific steps:

[0070] S51. Generate an input data matrix by splicing historical operation data and current operation monitoring data along the time axis. The input data matrix represents the historical state of the cable branch box and the monitoring data at the current moment. Each row represents the data of a time step, and each column corresponds to a monitoring data:

[0071]

[0072] Among them, X represents the input data matrix, X history represents the historical operation data, X current represents the current operation monitoring data;

[0073] S52. Input the input data matrix X into the improved graph convolutional network and map it to the node feature space of the graph structure through a mapping function. The mapping function linearly maps the input data matrix X through a weight matrix and a bias term, and strengthens the relationship between nodes by combining an adaptively adjusted adjacency matrix:

[0074] H = f(X) = XW node + b node + αA adj ·X;

[0075] Among them, H represents the node feature matrix of the graph neural network, which is the state information of the nodes. W node represents the mapping weight matrix, which is responsible for mapping the input data to the node space. b node represents the bias term, which adjusts the node features. α represents the adaptive coefficient, which determines the influence degree of the adjacency matrix on the node features. A adj represents the adjacency matrix of the graph, which describes the connection relationship between nodes in the graph. f represents the mapping function. αA adj represents the interaction between the adjacency matrix and the input feature matrix, which is used to dynamically adjust the features of nodes according to the neighbor relationship of nodes;

[0076] S53. Introduce an attention mechanism so that each node can adaptively focus on the most relevant neighbor node features during update. By calculating the attention weights between neighbor nodes, adjust the information transmission between nodes and update the node features:

[0077]

[0078] Among them, H k+1 represents the feature matrix of node k + 1. represents the normalized adjacency matrix. N ( k ) represents the set of neighbor nodes of node k. ReLU represents the non-linear activation function. W k represents the weight matrix of node k. ε represents the learning parameter of the attention mechanism, which is used to adjust the attention weights and describe the influence of node i on node k. H k and H i respectively represent the feature matrices of node k and node i. Attention ( H k , H i) represents the attention score between node k and neighbor node i:

[0079]

[0080] Among them, exp represents the natural exponential function, LeakyReLU represents the non-linear activation function, a represents the parameter for calculating the attention weight, T represents the transpose operation, || represents the operation of concatenating two vectors, and W represents the weight matrix;

[0081] S54. After several layers of graph convolution and attention mechanism, the node features are updated, and the scale of the graph is reduced through the graph pooling operation, and the node information is aggregated:

[0082] H pooled = Pooling ( ∑ i∈V α i ·H k ) ;

[0083] Among them, H pooled represents the node feature matrix after graph pooling, which is the graph information compressed after pooling. Pooling represents the graph pooling operation, V represents the set of nodes in the graph, α i represents the weight applied to node i during the pooling process, and H k represents the feature matrix of node k;

[0084] S55. The pooled feature matrix is used for the final fault prediction through a fully connected layer:

[0085] Y output = Softmax ( W out H pooled + b out ) ;

[0086] Among them, Y output represents the predicted probability of the fault category, which is used to predict the fault type of the cable branch box. W out represents the weight matrix of the output layer, which maps the pooled node features to the fault category space. Softmax represents normalization, and b out represents the bias term of the output layer.

[0087] The beneficial effects of the present invention are:

[0088] First, the intelligent monitoring method of the present invention effectively solves the deficiencies in the processing of multi-dimensional data of cable distribution boxes in traditional methods by improving the transformed convolutional neural network and the improved graph neural network. The improved transformed convolutional neural network can dynamically adjust the convolutional kernel to adapt to the changes of different types of data, thereby extracting more accurate features. At the same time, the improved graph neural network combined with the attention mechanism can adaptively consider the complex relationships between nodes and the importance of different nodes, optimizing the accuracy of fault prediction. Through the combination of the improved transformed convolutional neural network and the improved graph neural network, the faults and abnormal states of cable distribution boxes can be identified more comprehensively and accurately, avoiding the accuracy and real-time problems brought by traditional monitoring methods that only rely on a single data source or a static model.

[0089] Secondly, the present invention combines historical operation data and current monitoring data for fault prediction, enabling fault early warning not only based on the current operating state but also on the historical operation performance of the equipment and environmental changes. This comprehensive data analysis method greatly improves the accuracy of fault prediction, enabling the power system to give early warnings before equipment failures, thereby achieving early intervention and reducing sudden shutdown events.

[0090] Finally, by using the improved graph neural network and the attention mechanism, fault prediction can not only identify potential problems but also adaptively optimize according to the operating mode of cable distribution box equipment. The system can dynamically adjust the prediction strategy according to real-time data to adapt to the changes of different equipment and environments, improving the robustness and flexibility of the prediction system. In addition, combined with the maintenance optimization module, it can optimize the maintenance strategy of the equipment according to the prediction results, reasonably arrange the maintenance cycle and maintenance plan, thereby improving the operation and maintenance efficiency and resource utilization rate and reducing the maintenance cost of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0092] Figure 1 is a flowchart of a method for a comprehensive on-line monitoring system of a cable distribution box proposed by the present invention;

[0093] Figure 2 is a flowchart of fault prediction for a comprehensive on-line monitoring system of a cable distribution box proposed by the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0094] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0095] ReferenceFigure 1 and Figure 2 , a comprehensive on-line monitoring system for cable branch boxes, comprising:

[0096] A data acquisition module for acquiring the operation monitoring data and historical operation data of the cable branch box;

[0097] A data transmission module for transmitting the acquired operation monitoring data and historical operation data to the central monitoring platform through Internet of Things technology;

[0098] A data preprocessing module for preprocessing the real-time transmitted operation monitoring data;

[0099] An analysis module for analyzing the preprocessed data by applying an improved transform convolutional neural network to identify the faults and abnormal states of the cable branch box;

[0100] A prediction module for combining the historical operation data and the current monitoring data and performing fault prediction by using an improved graph neural network combined with an attention mechanism;

[0101] A central monitoring platform for receiving the fault prediction results output by the prediction module, generating fault warning information, and displaying the operation status of the cable branch box through a visualization interface;

[0102] A maintenance optimization module for optimizing the maintenance strategy according to the fault prediction results, combining the equipment maintenance plan, and proposing corresponding preventive measures and maintenance suggestions.

[0103] In this embodiment, the modules are implemented by the following methods:

[0104] S1. Acquire the operation monitoring data and historical operation data of the cable branch box through sensors;

[0105] S2. Transmit the acquired operation monitoring data and historical operation data to the central monitoring platform through Internet of Things technology;

[0106] S3. Preprocess the real-time transmitted operation monitoring data to generate preprocessed data;

[0107] S4. Construct an improved transform convolutional neural network, apply the improved transform convolutional neural network to analyze the preprocessed data, use transform convolutional operations to extract features from the preprocessed data, and introduce residual connections into the improved transform convolutional neural network, and identify the faults and abnormal states of the cable branch box by the fully connected layer of the improved transform convolutional neural network;

[0108] S5. Based on historical operation data and current operation monitoring data, combined with the attention mechanism and improved graph neural network, update the node features of the improved graph neural network by graph convolution operation and calculating attention weights, adjust the information transmission between nodes, and use the pooling layer and fully connected layer of the improved graph neural network for fault prediction to generate prediction results;

[0109] S6. Generate fault warning information according to the analysis and prediction results, and display the operation status of the cable distribution box through the monitoring platform;

[0110] S7. Combine the prediction results with the equipment maintenance plan, optimize the maintenance strategy, and put forward corresponding preventive measures and maintenance suggestions.

[0111] In this embodiment, the operation monitoring data includes current, voltage, temperature, and humidity.

[0112] In this embodiment, S3 includes the following specific steps:

[0113] S31. Conduct a preliminary check on the received operation monitoring data to ensure the integrity and validity of the data. For values that do not meet the format requirements or show abnormalities, mark them through a preset threshold, and perform processing or elimination of missing values;

[0114] S32. Perform filtering and denoising processing on the operation monitoring data using a Kalman filter to remove existing random noise:

[0115]

[0116] Among them, represents the current estimated value, represents the estimated value at the previous moment, y k represents the current observed value, H k represents the state transition matrix, H k represents the Kalman gain;

[0117] S33. Perform normalization processing on the denoised data to eliminate the differences in data under different dimensions, so that the range of the operation monitoring data is within a unified standard:

[0118]

[0119] Among them, x' represents the normalized data, x represents the original data value, min ( x ) and max ( x ) represent the minimum and maximum values in the original data value respectively;

[0120] S34. Use the sliding window method to smooth the normalized data. Set a window size N and calculate the average of N data points before and after each data point:

[0121]

[0122] Among them, y i represents the smoothed data point, x j represents the original data point, and P represents the window size;

[0123] S35. Perform time series processing on the smoothed data, sort the data in chronological order and construct timestamps;

[0124] S36. Perform difference processing on the time series data through the first-order difference formula:

[0125] Δx t = x t - x t-1 ;

[0126] Among them, Δx t represents the data after difference, x t represents the data at the current moment, and x t-1 represents the data at the previous moment;

[0127] S37. Detect outliers in the processed data, use the Z-score method to measure the deviation of each data point, and if the Z-score is greater than the preset threshold, the data point is considered an outlier:

[0128]

[0129] Among them, Z c represents the value of the Z-score, x represents the data point, μ represents the mean of the data set, and σ represents the standard deviation of the data set.

[0130] In this embodiment, the S4 includes the following specific steps:

[0131] S41. Construct an improved transform convolutional neural network, and the improved transform convolutional neural network includes:

[0132] Input layer, input the preprocessed data and enter the network for processing;

[0133] Transform convolutional layer, extract local features and dynamically adjust the convolutional kernel;

[0134] Temporal convolutional layer, extract time-dependent features and capture the change patterns of time series;

[0135] Pooling layer, perform downsampling on the features;

[0136] Multi-scale convolutional layer, capturing features at different scales and processing global and local features;

[0137] Fully connected output layer, integrating the extracted features and using the ReLU activation function to output the prediction result;

[0138] S42. Process the preprocessed data using transform convolutional operations, in which the convolutional kernel is dynamically adjusted:

[0139]

[0140] Among them, represents the feature map extracted by the convolutional layer, M represents the height of the convolutional kernel, N represents the width of the convolutional kernel, represents the preprocessed input data, represents the convolutional kernel, used to perform convolutional operations with the input data to extract local features, b (1) represents the bias term, added to the convolutional result to adjust the output, m and n represent the position indices of the elements in the convolutional kernel, and i and j represent the position indices of the elements in the output feature map;

[0141] S43. Through the max pooling method, select the maximum value from each pooling window to generate the pooled feature map:

[0142]

[0143] Among them, represents the output feature map after the pooling operation, reducing the spatial dimension, represents the output feature map of the convolutional operation, pool window represents the pooling window, max represents the maximum value, and ∈ represents the set operation "belongs to";

[0144] S44. Use the convolutional kernel to process the pooled feature map to capture temporal dependencies through temporal convolutional operations:

[0145]

[0146] Among them, represents the output of the temporal convolution, which are the captured temporal dependency features, T represents the length of the temporal convolutional kernel, represents the weights of the temporal convolutional kernel, represents the pooled feature map, b (3) represents the bias term of the temporal convolution;

[0147] S45. Simultaneously extract features at different scales through multiple convolutional kernels:

[0148]

[0149] Among them, represents the output of the multi-scale convolution, which contains feature information of multiple scales. represents the output of the temporal convolution operation and serves as the input to the multi-scale convolution, T m represents the length of the m-th scale convolution kernel, b (4) represents the bias term of the multi-scale convolution. represents the weight of the m-th scale convolution kernel, where m represents the number of scale convolution kernels;

[0150] S46. Introduce a residual connection, and the features after the residual connection processing are passed to the fully connected layer for processing:

[0151]

[0152] Among them, Y (5) represents the output of the fully connected layer, the feature representation after the residual connection processing, ReLU represents the non-linear activation function, W (5) represents the weight matrix of the fully connected layer, Y (4) represents the output of the multi-scale convolution, b (5 represents the bias term of the fully connected layer. represents the input of the residual connection, which is the input feature from the multi-scale convolution layer;

[0153] S47. According to the output of the fully connected layer, set a threshold. When the failure probability of a certain category exceeds the set threshold, it is considered that the cable branch box has this type of failure.

[0154] In this embodiment, the S5 includes the following specific steps:

[0155] S51. Generate an input data matrix by splicing historical operation data and current operation monitoring data along the time axis. The input data matrix represents the historical state of the cable branch box and the monitoring data at the current moment. Each row represents the data of one time step, and each column corresponds to a monitoring data:

[0156]

[0157] Among them, X represents the input data matrix, X history represents the historical operation data, X current represents the current operation monitoring data;

[0158] S52. Input the input data matrix X into the improved graph convolutional network and map it to the node feature space of the graph structure through a mapping function. The mapping function linearly maps the input data matrix X through a weight matrix and a bias term, and strengthens the relationship between nodes by combining an adaptively adjusted adjacency matrix:

[0159] H = f ( X ) = XW node + b node + αA adj · X;

[0160] Among them, H represents the node feature matrix of the graph neural network, which is the state information of the node, W node represents the mapping weight matrix, which is responsible for mapping the input data into the node space, b node represents the bias term, which adjusts the node features, α represents the adaptive coefficient, which determines the influence degree of the adjacency matrix on the node features, A adj represents the adjacency matrix of the graph, which describes the connection relationship between the nodes in the graph, f represents the mapping function, αA adj represents the interaction between the adjacency matrix and the input feature matrix, which is used to dynamically adjust the features of the nodes according to the neighbor relationship of the nodes;

[0161] S53. Introduce the attention mechanism so that each node can adaptively focus on the most relevant neighbor node features during update. By calculating the attention weights between neighbor nodes, adjust the information transmission between nodes, and update the node features:

[0162]

[0163] Among them, H k+1 represents the feature matrix of node k + 1, represents the normalized adjacency matrix, N ( k ) represents the set of neighbor nodes of node k, ReLU represents the non-linear activation function, W k represents the weight matrix of node k, ε represents the learning parameter of the attention mechanism, which is used to adjust the attention weights and describe the influence of node i on node k, H k and H i respectively represent the feature matrices of node k and node i, Attention ( H k , H i) represents the attention score between node k and neighbor node i:

[0164]

[0165] Among them, exp represents the natural exponential function, LeakyReLU represents the non-linear activation function, a represents the parameter for calculating the attention weights, T represents the transpose operation, || represents the operation of concatenating two vectors together, and W represents the weight matrix;

[0166] S54. After several layers of graph convolution and attention mechanism, the node features are updated. The scale of the graph is reduced through graph pooling operation, and the node information is aggregated:

[0167] H pooled = Pooling ( ∑ i∈V α i ·H k ) ;

[0168] where, H pooled represents the node feature matrix after graph pooling, which is the graph information compressed after pooling. Pooling represents the graph pooling operation, V represents the set of nodes in the graph, and α i represents the weight applied to node i during the pooling process, and H k represents the feature matrix of node k;

[0169] S55. The pooled feature matrix is used for the final fault prediction through a fully connected layer:

[0170] Y output = Softmax ( W out H pooled + b out ) ;

[0171] where, Y output represents the predicted probability of the fault category, which is used to predict the fault type of the cable branch box. W out represents the weight matrix of the output layer, which maps the pooled node features to the fault category space. Softmax represents normalization, and b out represents the bias term of the output layer.

[0172] Example 1:

[0173] To verify the feasibility of the present invention in implementation, the present invention is applied to the cable branch box monitoring system of a certain power company in the core area of the city. In order to ensure the stable operation of the power grid, the power company has deployed multiple cable branch boxes, which are responsible for providing power distribution services to residents and commercial users in different areas. Since the cable branch boxes are exposed outdoors for a long time and the operating conditions are complex, the power company needs an intelligent system that can monitor the device status in real time, identify faults in time and make predictions. The traditional manual inspection and regular maintenance methods cannot detect potential faults in time, and there are cases of misdetection and missed detection, and cannot effectively reduce the sudden faults of the equipment.

[0174] In this context, power companies have introduced an online monitoring system for cable branch boxes based on the Internet of Things (IoT) and deep learning. By installing sensors on each cable branch box, the system can collect multiple parameters such as current, voltage, temperature, and humidity in real time, and use IoT technology to transmit the data to a central monitoring platform. On this platform, combined with an improved transform convolutional neural network and a graph neural network, the system can analyze the data and predict faults.

[0175] In the initial stage of system deployment, 10 cable branch boxes were selected as monitoring objects, and data was continuously monitored for two months. By comparing the monitoring data with the actual fault occurrence of the equipment, the accuracy of the fault prediction model was verified. Table 1 shows the monitoring data of the cable branch boxes.

[0176] Table 1 Monitoring Data of Cable Branch Boxes

[0177]

[0178]

[0179] Table 2 Comparison of Fault Prediction and Actual Status of Cable Branch Boxes

[0180] Equipment Number Actual Failure Status Fault Prediction Status Fault Prediction Probability Whether the Prediction is Accurate 1 Normal Normal 5 Yes 2 Normal Normal 4 Yes 3 Normal Normal 6 Yes 4 Abnormal Abnormal 85 Yes 5 Normal Normal 7 Yes 6 Normal Normal 6 Yes 7 Abnormal Abnormal 80 Yes 8 Normal Normal 5 Yes 9 Abnormal Abnormal 82 Yes 10 Normal Normal 6 Yes

[0181] Table 2 shows the content in five dimensions: equipment number, actual fault status, fault prediction status, fault prediction probability, and whether the prediction is accurate. By comparing the actual fault status and the fault prediction status, it can be clearly seen that the prediction results of all equipment are consistent with the actual fault status, indicating that the system has high accuracy in fault identification and prediction. For example, the actual fault status of equipment 4 is "abnormal", and the system prediction result is also "abnormal", with a prediction probability of 85%. This high-probability prediction indicates that the system can effectively identify the potential fault risks of the equipment and issue warnings in a timely manner, which helps with the early maintenance and fault prevention of the equipment.

[0182] The "fault prediction probability" column in the table shows the predicted probability of each equipment having a fault. The data in this column indicates that the prediction results are consistent with the actual fault status. In cases where the fault prediction probability is relatively high (such as the prediction probabilities of equipment 4 and equipment 7 are 85% and 80% respectively), the system accurately identified the fault status of these equipment and issued successful early warnings. For other equipment, the fault prediction probability is relatively low (such as equipment 1, 2, and 3, etc.), and the prediction results are normal, which also conforms to the actual operating status of the equipment, indicating that these equipment did not have faults during the monitoring period.

[0183] The column of "Whether the prediction is accurate" confirms the accuracy of the system prediction. According to the table data, the prediction results of all devices are "Yes", that is, the fault prediction is completely consistent with the actual fault status. It can be seen that the on-line monitoring system of cable branch boxes has demonstrated efficient fault identification and prediction capabilities in this test.

[0184] In summary, through the comparative analysis of the experimental data in this test, the accuracy and reliability of the cable branch box fault prediction system proposed by the present invention are verified. This system can not only monitor the operation status of equipment in real time, but also accurately predict equipment faults, improve the operation and maintenance efficiency of the power system, reduce the occurrence of sudden equipment failures, and ensure the stability and safety of power supply.

[0185] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A comprehensive online monitoring system for cable branch boxes, characterized in that: include: Data acquisition module, used to collect operation monitoring data and historical operation data of the cable branch box; Data transmission module, used to transmit the collected operation monitoring data and historical operation data to the central monitoring platform through the Internet of Things technology; A data preprocessing module is used to preprocess the real-time transmitted operation monitoring data; An analysis module, used to apply an improved transform convolutional neural network to analyze the preprocessed data and identify faults and abnormal conditions of the cable branch box; The prediction module is used to combine historical operation data and current monitoring data, and use the improved graph neural network combined with the attention mechanism to predict faults; The central monitoring platform is used to receive the fault prediction results output by the prediction module, generate fault warning information, and display the operating status of the cable branch box through a visual interface; The maintenance optimization module is used to optimize the maintenance strategy and propose corresponding preventive measures and maintenance suggestions based on the fault prediction results and equipment maintenance plan.

2. A cable branch box comprehensive online monitoring system according to claim 1, characterized in that: The modules are implemented in the following ways: S1. Collect the operation monitoring data and historical operation data of the cable branch box through sensors; S2, transmitting the collected operation monitoring data and historical operation data to the central monitoring platform through the Internet of Things technology; S3, preprocessing the real-time transmitted operation monitoring data to generate preprocessed data; S4. Construct an improved transform convolutional neural network, apply the improved transform convolutional neural network to analyze the preprocessed data, use the transform convolution operation to extract features from the preprocessed data, and introduce residual connection into the improved transform convolutional neural network, so that the fault and abnormal state of the cable branch box can be identified by the fully connected layer of the improved transform convolutional neural network; S5. Based on historical operation data and current operation monitoring data, combined with the attention mechanism and improved graph neural network, through graph convolution operation and calculation of attention weights, the node features of the improved graph neural network are updated, the information transmission between nodes is adjusted, and the pooling layer and fully connected layer of the improved graph neural network are used to perform fault prediction and generate prediction results; S6. Generate fault warning information based on the analysis and prediction results, and display the operating status of the cable branch box through the monitoring platform; S7. Combine the prediction results with the equipment maintenance plan, optimize the maintenance strategy, and put forward corresponding preventive measures and maintenance suggestions.

3. A cable branch box comprehensive online monitoring system according to claim 2, characterized in that: The operation monitoring data includes current, voltage, temperature and humidity.

4. A cable branch box comprehensive online monitoring system according to claim 2, characterized in that: The S3 includes the following specific steps: S31. Perform a preliminary check on the received operation monitoring data to ensure the integrity and validity of the data. For values ​​that do not meet the format requirements or are abnormal, mark them using a preset threshold and process or remove missing values. S32, use Kalman filter to filter and denoise the operation monitoring data to remove the existing random noise: in, represents the current estimated value, represents the estimated value at the previous moment, y k represents the current observation value, H k represents the state transfer matrix, H k represents the Kalman gain; S33. Normalize the denoised data to eliminate the differences in data under different dimensions and make the range of operation monitoring data within a unified standard: Where x' represents the normalized data, x represents the original data value, min(x) and max(x) represent the minimum and maximum values ​​in the original data value, respectively; S34, use the sliding window method to smooth the normalized data, set a window size N, and calculate the average value of the N data points before and after each data point: Among them, y i represents the smoothed data point, x j represents the original data point, and P represents the window size; S35, performing time series processing on the smoothed data, sorting the data in chronological order and constructing a timestamp; S36. Perform differential processing on the time series data, using the first difference formula: Δx t =x t -x t-1 ; Where Δx t Represents the data after difference, x t Represents the data at the current moment, x t-1 Represents the data of the previous moment; S37. Perform outlier detection on the processed data and use the Z score method to measure the deviation of each data point. If the Z score is greater than the preset threshold, the data point is considered to be an outlier: Among them, Z c represents the value of the Z score, x represents the data point, μ represents the mean of the data set, and σ represents the standard deviation of the data set.

5. A cable branch box comprehensive online monitoring system according to claim 2, characterized in that: The S4 comprises the following specific steps: S41, constructing an improved transformation convolutional neural network, wherein the improved transformation convolutional neural network comprises: Input layer, input preprocessed data into the network for processing; Transform the convolution layer, extract local features, and dynamically adjust the convolution kernel; The temporal convolution layer extracts time-dependent features and captures the changing patterns of the time series; Pooling layer, downsampling the features; Multi-scale convolutional layer, which captures features of different scales and processes global and local features; The fully connected output layer integrates the extracted features and uses the ReLU activation function to output the prediction results; S42, using a transform convolution operation to process the preprocessed data, in which the convolution kernel is dynamically adjusted: in, Represents the feature map extracted by the convolution layer, M represents the height of the convolution kernel, and N represents the width of the convolution kernel. Represents the preprocessed input data, Represents the convolution kernel, which is used to perform convolution operations with the input data to extract local features, b( 1 ) represents the bias term, which is added to the convolution result to adjust the output, m and n represent the position index of the element in the convolution kernel, and i and j represent the position index of the element in the output feature map; S43. Using the maximum pooling method, the maximum value is selected from each pooling window to generate a pooled feature map: in, Represents the output feature map after the pooling operation, reducing the spatial dimension. Represents the output feature map of the convolution operation, poolwindow represents the pooling window, max represents the maximum value, and ∈ represents the set operation "belongs to"; S44. Use convolution kernel Processing the feature map after pooling Capturing temporal dependencies through sequential convolution operations: in, represents the output of the temporal convolution, which is the captured time-dependent features, T represents the length of the temporal convolution kernel, represents the weight of the temporal convolution kernel, represents the feature map after pooling, b( 3 ) represents the bias term of temporal convolution; S45, extract features simultaneously at different scales through multiple convolution kernels: in, Represents the output of multi-scale convolution, which contains feature information of multiple scales. Represents the output of the temporal convolution operation, which is used as the input of the multi-scale convolution. m represents the length of the m-th scale convolution kernel, b( 4 ) represents the bias term of multi-scale convolution, Represents the weight of the convolution kernel of the mth scale, and m represents the number of scale convolution kernels; S46, introduce residual connection, the features processed by residual connection are passed to the fully connected layer for processing: Among them, Y( 5 ) represents the output of the fully connected layer, the feature representation after residual connection processing, ReLU represents the nonlinear activation function, W( 5 ) represents the weight matrix of the fully connected layer, Y( 4 ) represents the output of multi-scale convolution, b( 5 represents the bias term of the fully connected layer, Represents the input of the residual connection, which comes from the input features of the multi-scale convolutional layer; S47. According to the output of the fully connected layer, a threshold is set. When the probability of a certain type of fault exceeds the set threshold, it is considered that the cable branch box has this type of fault.

6. A cable branch box comprehensive online monitoring system according to claim 2, characterized in that: The S5 comprises the following specific steps: S51. Generate an input data matrix by splicing the historical operation data and the current operation monitoring data together along the time axis. The input data matrix represents the historical state of the cable branch box and the monitoring data at the current moment. Each row represents the data of one time step, and each column corresponds to one monitoring data: Where X represents the input data matrix, X history Represents historical operation data, X current Indicates the current operation monitoring data; S52, inputting the input data matrix X into the improved graph convolutional network, and mapping it to the node feature space of the graph structure through a mapping function, wherein the mapping function linearly maps the input data matrix X through a weight matrix and a bias term, and strengthens the relationship between nodes in combination with the adaptively adjusted adjacency matrix: H=f(X)=XW node +b node +αA adj ·X; Among them, H represents the node feature matrix of the graph neural network, is the state information of the node, and W node represents the mapping weight matrix, which is responsible for mapping the input data to the node space, b node represents the bias term, which adjusts the node features, α represents the adaptive coefficient, which determines the influence of the adjacency matrix on the node features, and A adj Represents the adjacency matrix of the graph, describing the connection relationship between nodes in the graph, f represents the mapping function, αA adj Represents the interaction between the adjacency matrix and the input feature matrix, which is used to dynamically adjust the features of the nodes according to the neighbor relationships of the nodes; S53. Introduce the attention mechanism so that each node can adaptively focus on the most relevant neighbor node features when updating. By calculating the attention weights between neighbor nodes, adjust the information transfer between nodes, and update the node features: Among them, H k+1 represents the feature matrix of node k+1, represents the standardized adjacency matrix, N(k) represents the set of neighbor nodes of node k, ReLU represents the nonlinear activation function, W k represents the weight matrix of node k, ε represents the learning parameter of the attention mechanism, which is used to adjust the attention weight and describe the influence of node i on node k, H k and H i Represents the feature matrix of node k and node i respectively, Attention(H k ,H i ) represents the attention score between node k and its neighbor node i: Among them, exp represents the natural exponential function, LeakyReLU represents the nonlinear activation function, a represents the parameter for calculating the attention weight, T represents the transpose operation, || represents the operation of splicing two vectors together, and W represents the weight matrix; S54. After several layers of graph convolution and attention mechanism, the node features are updated, the graph size is reduced through graph pooling operation, and the node information is aggregated: H pooled =Pooling(∑ i∈V α i ·H k ); Among them, H pooled represents the node feature matrix after graph pooling, which is the graph information compressed after pooling. Pooling represents the graph pooling operation, V represents the node set in the graph, and α i represents the weight applied to node i during the pooling process, H k Represents the feature matrix of node k; S55. The pooled feature matrix is ​​used for final fault prediction through the fully connected layer: Y output =Softmax(W out H pooled +b out ): Among them, Y output Represents the predicted probability of the fault category, which is used to predict the fault type of the cable branch box, W out represents the weight matrix of the output layer, mapping the pooled node features to the fault category space, Softmax represents normalization, and b out Represents the bias term of the output layer.

Images