Automatic detection platform for distribution automation terminal

Abstract

The invention relates to the technical field of power distribution terminal monitoring, and discloses a power distribution automation terminal automatic detection platform, which comprises a monitoring area division module for dividing a ring main unit into M monitoring areas; the multi-modal data acquisition module is used for acquiring a discharge signal and a thermal image of each monitoring area; the feature sequence construction module is used for constructing a feature sequence; the feature sequence updating module is used for updating the feature sequence through a first neural network model; the graph network data construction module is used for constructing graph network data; the discharge fault monitoring module obtains fault types of M monitoring areas through a second neural network model; according to the invention, the change features of the discharge signal and the thermal image of each monitoring area in the time dimension are extracted through the first neural network model, and the correlation features of the discharge signals and the thermal images of different monitoring areas in the space dimension are extracted through the second neural network model. Therefore, automatic and real-time positioning and classification of partial discharge of the ring main unit are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution terminal monitoring, and more specifically, it relates to an automatic detection platform for distribution automation terminals. Background Art

[0002] The distribution terminal unit (DTU) is an important part of the power system for realizing the automation of the distribution network. It is mainly used to collect the operating status of distribution equipment and transmit the data to the master station system, so as to realize the remote monitoring and automatic operation of the distribution network. As a key node in the distribution network, the ring main unit is mainly used to distribute electric energy and protect the line to ensure the safe and stable operation of the power system. Since the ring main unit works in a high-voltage environment, the change of the internal insulation performance of the equipment directly affects its operation safety. Partial discharge is a precursor to the degradation of the internal insulation performance of the ring main unit and an important cause of ring main unit failures. Partial discharge refers to the electrical discharge phenomenon occurring inside or on the surface of high-voltage insulating materials. Long-term partial discharge may lead to equipment failures and even cause serious power accidents. Therefore, the detection of partial discharge in the ring main unit is particularly important.

[0003] The existing methods for detecting partial discharge in ring main units mainly include: 1. The ultrasonic detection method uses an ultrasonic sensor to collect the ultrasonic waves (20 kHz - 40 kHz) generated during partial discharge. However, this method is easily interfered by external noise and mechanical vibration, with low detection accuracy and difficult to achieve real-time online detection; 2. The ultra-high frequency (UHF) detection method uses a UHF antenna or coupler to capture the ultra-high frequency electromagnetic waves (300 MHz - 3 GHz) generated by partial discharge to locate the discharge source. This method has high detection accuracy, but the equipment cost is high and the installation and commissioning are complex; 3. The high-frequency current detection method uses a high-frequency current sensor to capture the high-frequency pulse current (1 MHz - 50 MHz) generated by partial discharge on the grounding wire. This method has low cost and convenient installation and commissioning, but it cannot directly locate the fault and can only judge the strength of partial discharge; 4. The infrared thermal imaging detection method uses an infrared thermal imager to detect the temperature abnormal area caused by partial discharge. However, it is difficult to detect early partial discharge by this method, and the infrared penetration is limited, making it difficult to detect partial discharge inside the ring main unit.

[0004] Therefore, there is an urgent need for a real-time detection platform for ring main units to solve the above problems. Summary of the Invention

[0005] The present invention provides an automatic detection platform for distribution automation terminals to solve the technical problems in the above background art.

[0006] The present invention provides an automatic detection platform for distribution automation terminals, including: A monitoring area division module, which is used to divide the ring main unit into M monitoring areas according to the locations of key components inside the ring main unit; A multi-modal data acquisition module, which is used to collect the discharge signals of each monitoring area through a high-frequency current sensor and collect the thermal imaging maps of each monitoring area through an infrared thermal imager at a preset time interval t within a preset time period T; The discharge signal is represented by a time-domain waveform diagram, where the horizontal axis of the time-domain waveform diagram represents the acquisition time point and the vertical axis represents the voltage value; A feature sequence construction module, which is used to construct corresponding feature sequences according to the discharge signals and thermal imaging maps of M monitoring areas; The feature sequence corresponding to each monitoring area includes N sequence units, and the nth sequence unit is represented by the time-domain feature obtained by preprocessing the discharge signal at the nth time point and the preprocessed thermal imaging map at the nth time point, where 1 ≤ n ≤ N and N = T / t; A feature sequence update module, which is used to input the feature sequence corresponding to each monitoring area into a first neural network model for updating and output a first update vector; A graph network data construction module, which is used to construct graph network data according to the first update vectors corresponding to M monitoring areas; The graph network data consists of M vertices and the edges between the vertices; Each vertex establishes a mapping relationship with a monitoring area, and the initial feature of each vertex is represented by the first update vector of the monitoring area with which it establishes a mapping relationship; A discharge fault monitoring module, which is used to input the graph network data into a second neural network model for updating and output the fault types of M monitoring areas; The fault types include: normal, minor partial discharge, and severe partial discharge.

[0007] Furthermore, both the preset time period T and the preset time interval t are user-defined parameters. Dividing the ring main unit into M monitoring areas includes the following steps: Step S201, collect the visible light image of the ring main unit at the same position where the thermal imaging map is collected, and manually mark the locations of key components inside the ring main unit in the visible light image; The key components inside the ring main unit include: busbar connection points, disconnector switches, cable joints, insulators, and grounding buses; Step S202, obtain the edge contour of the ring main unit in the visible light image through an object detection model; Step S203, calculate the minimum Euclidean distance between the locations of key components, and generate rectangular regions with a side length of 1.5 times this distance, with each location of a key component as the center; Step S204: If the overlapping area between two rectangular regions is greater than or equal to a preset overlapping area threshold, recalculate the side lengths of these two rectangular regions and regenerate the rectangular regions until the overlapping area between these two rectangular regions is less than the preset overlapping area threshold; The side lengths of the rectangular region after recalculation The calculation formula is as follows: ; where represents the side lengths of the rectangular region before recalculation, represents the preset overlapping area threshold, represents the overlapping area between two rectangular regions, and the preset overlapping area threshold is a custom parameter; Step S205: Take all rectangular regions as monitoring regions and count the number M of monitoring regions obtained.

[0008] Furthermore, preprocess the discharge signals and thermal images at N time points to construct a feature sequence, including the following steps: Step S301: When the discharge signal or thermal image at a time point is missing, randomly select the discharge signal or thermal image at an adjacent time point for supplementation; Step S302: Perform noise reduction processing on the discharge signal through wavelet transform and perform smoothing processing on the thermal image through Gaussian filtering; Step S303: Extract the time-domain features of the time-domain waveform diagrams corresponding to the discharge signals at N time points; The time-domain features include: the maximum voltage value, the average voltage value, the root mean square value, the difference between the maximum voltage value and the minimum voltage value, the number of times exceeding the preset voltage value, and the duration of exceeding the preset voltage value, where the preset voltage value is a custom parameter; Step S304: Perform normalization processing on the time-domain features and thermal images at N time points through the Min-Max method to obtain a feature sequence.

[0009] Furthermore, the edges between vertices include: if there is a direct electrical connection relationship between key components in the monitoring region, an edge is constructed between the corresponding vertices; if the Euclidean distance between the locations of key components in the monitoring region is less than the preset distance threshold, an edge is constructed between the corresponding vertices, and the preset distance threshold The calculation formula is as follows: ; where A, B, and C respectively represent the length, width, and height of the ring main unit, represents a custom proportionality coefficient with a value range between 0.1 and 0.3.

[0010] Further, the first neural network model includes: a first hidden layer, a second hidden layer, a feature fusion layer, and a first classifier; The first hidden layer includes: N first update units, an unfolding unit, a splicing unit, and a second update unit; The u-th first update unit inputs the preprocessed thermal imaging map of the u-th sequence unit of the feature sequence and outputs a feature matrix, where 1 ≤ u ≤ N; The u-th unfolding unit is used to unfold the feature matrix output by the u-th first update unit into a vector representation and outputs a feature vector; The u-th splicing unit is used to splice the feature vector output by the u-th unfolding unit with the time-domain feature of the u-th sequence unit of the feature sequence and outputs a combined vector; The u-th second update unit inputs the combined vector output by the u-th splicing unit and outputs a second update vector; The second update unit is constructed based on the GRU model, and the dimension number of the second update vector is a custom parameter; The second hidden layer inputs the feature sequence and outputs a third update vector; The second hidden layer is constructed based on the Transformer model, and the dimension number of the third update vector is a custom parameter; The feature fusion layer is used to splice the second update vector output by the N-th second update unit and the third update vector output by the second hidden layer to obtain a first update vector; Input the first update vector into the first classifier, and the classification space of the first classifier represents whether partial discharge occurs in the monitoring area.

[0011] Further, the feature matrix output by the u-th first update unit The calculation formula is as follows: ; Where represents the preprocessed thermal imaging map of the u-th sequence unit of the feature sequence input by the u-th first update unit, represents scaling the thermal imaging map to a size of 128×128, represents a convolution operation with the number of convolution kernels being K1, the convolution kernel size being 5×5, the stride being 1, and the padding method being Valid, represents a max pooling operation with the pooling window size being 5×5 and the stride being 5, represents a convolution operation with the number of convolution kernels being K2, the convolution kernel size being 3×3, the stride being 1, and the padding method being Valid, It represents a max pooling operation with a pooling window size of 3×3 and a stride of 3. ChannelAvgPool represents a channel average pooling operation. Swish represents the Swish activation function. Both K1 and K2 are custom parameters.

[0012] Further, the ultrasonic detection method is used to determine whether partial discharge occurs in the monitoring area, serving as the sample label for the training samples used to train the first neural network model.

[0013] Further, the second neural network model includes: a third hidden layer and a second classifier; The third hidden layer inputs graph network data and outputs an updated matrix. The updated matrix includes M row vectors, and each row vector corresponds to a fourth update vector of a vertex. Each row vector of the updated matrix is input into the second classifier, and the classification space of the second classifier represents the fault type of the monitoring area.

[0014] Further, the calculation formula for the updated matrix S output by the third hidden layer includes: ; ; ; where 1≤i≤M, represents the fourth update vector of the i-th vertex of the updated matrix, and respectively represent the initial features of the i-th vertex and the j-th vertex, represents the set of vertices that have an edge connection with the i-th vertex, represents the association vector between the i-th vertex and the j-th vertex. The dimensionality of the association vector is the same as the dimensionality of the initial features of the vertex, represents the number of edges between the i-th vertex and the j-th vertex, 、 and respectively represent the first weight parameter, the second weight parameter, and the third weight parameter, and respectively represent the first bias parameter and the second bias parameter. Concat represents the concatenation function, sigmoid represents the sigmoid activation function, Swish represents the Swish activation function, MLP represents a multi-layer perceptron, represents the stacking operation of the fourth update vectors of M vertices.

[0015] Further, the fault type of the monitoring area is obtained through the ultra-high frequency detection method, serving as the sample label for the training samples used to train the second neural network model.

[0016] The beneficial effects of the present invention are as follows: The present invention divides the ring main unit into multiple monitoring areas, extracts the change characteristics of the discharge signals and thermal images of each monitoring area in the time dimension through the first neural network model, extracts the correlation characteristics of the discharge signals and thermal images of different monitoring areas in the spatial dimension through the second neural network model, and obtains the sample labels of the above models through the ultrasonic detection method and the ultra-high frequency detection method, so as to realize the automatic and real-time positioning and classification of partial discharges in the ring main unit. Description of the Drawings

[0017] Figure 1 is a schematic diagram of an automatic detection platform for a distribution automation terminal of the present invention; Figure 2 is a flowchart of dividing the ring main unit into M monitoring areas according to the present invention; Figure 3 is a flowchart of preprocessing to construct a feature sequence according to the present invention.

[0018] In the figure: monitoring area division module 101, multi-modal data acquisition module 102, feature sequence construction module 103, feature sequence update module 104, graph network data construction module 105, discharge fault monitoring module 106. Detailed Embodiments

[0019] Now, the subject matter described herein will be discussed with reference to exemplary embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein, and the functions and arrangements of the elements discussed can be changed without departing from the scope of protection of the content of this specification. Each example can omit, substitute or add various processes or components as needed. Additionally, the features described relative to some examples can also be combined in other examples.

[0020] It should be noted that unless otherwise defined, the technical terms or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by those of ordinary skill in the art to which the present invention pertains. The "first", "second" and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. The terms such as "including" or "comprising" mean that the elements or objects appearing before the term cover the elements or objects listed after the term and their equivalents, without excluding other elements or objects. The terms such as "connected" or "coupled" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. The terms such as "upper", "lower", "left", "right" are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0021] As Figures 1 to 3 shown, a distribution automation terminal automatic detection platform includes: A monitoring area division module 101, which is used to divide the ring main unit into M monitoring areas according to the positions of key components inside the ring main unit; A multimodal data acquisition module 102, which is used to collect the discharge signals of each monitoring area through a high-frequency current sensor and collect the thermal imaging maps of each monitoring area through an infrared thermal imager at a preset time interval t within a preset time period T; The discharge signal is represented by a time-domain waveform diagram, where the horizontal axis of the time-domain waveform diagram represents the acquisition time point and the vertical axis represents the voltage value; A feature sequence construction module 103, which is used to construct corresponding feature sequences according to the discharge signals and thermal imaging maps of M monitoring areas; The feature sequence corresponding to each monitoring area includes N sequence units, and the nth sequence unit is represented by the time-domain feature obtained by preprocessing the discharge signal at the nth time point and the thermal imaging map at the preprocessed nth time point, where 1 ≤ n ≤ N and N = T / t; A feature sequence update module 104, which is used to input the feature sequence corresponding to each monitoring area into the first neural network model for updating and output the first update vector; A graph network data construction module 105, which is used to construct graph network data according to the first update vectors corresponding to M monitoring areas; The graph network data consists of M vertices and the edges between the vertices; Each vertex is mapped to a monitoring area, and the initial feature of each vertex is represented by the first update vector of the monitoring area mapped to it; A discharge fault monitoring module 106, which is used to input the graph network data into the second neural network model for updating and output the fault types of M monitoring areas; The fault types include: normal, slight partial discharge, and severe partial discharge.

[0022] In an embodiment of the present invention, both the preset time period T and the preset time interval t are user-defined parameters. Preferably, the preset time period T is set to 5 minutes and the preset time interval t is set to 30 seconds, then N = T / t = 10; As Figure 2 shown, dividing the ring main unit into M monitoring areas includes the following steps: Step S201, collect the visible light image of the ring main unit at the same position where the thermal imaging map is collected, and manually mark the positions of key components inside the ring main unit in the visible light image; The key components inside the ring main unit include: busbar connection points, circuit breakers, cable joints, insulators, and grounding buses; Step S202: Obtain the edge contour of the ring main unit in the visible light image through the target detection model. Step S203: Calculate the minimum Euclidean distance between the positions of the key components, and generate rectangular regions with each position of the key components as the center and 1.5 times of this distance as the side length. Step S204: Determine whether the overlapping area between two rectangular regions is greater than or equal to the preset overlapping area threshold. If so, recalculate the side lengths of these two rectangular regions and regenerate the rectangular regions until the overlapping area between these two rectangular regions is less than the preset overlapping area threshold. The side length of the rectangular region after recalculation The calculation formula is as follows: ; Where represents the side length of the rectangular region before recalculation, represents the preset overlapping area threshold, represents the overlapping area between two rectangular regions. The preset overlapping area threshold is a custom parameter. Preferably, the preset overlapping area threshold is set to 30% of the area of the rectangular region generated for the first time. Step S205: Take all the rectangular regions as monitoring regions, and count to obtain the number M of the monitoring regions.

[0023] It should be noted that the busbar connection point is responsible for power transmission. The reasons for easy discharge include: poor electrical connection, overload, aging, and surface contamination, etc.; the circuit breaker is used to control and protect the circuit. The reasons for easy discharge include: poor contact of the contacts, insulation aging and damage, etc.; the reasons for easy discharge of the cable joint include: loose wiring of the cable head, insulation aging and dampness, etc.; the insulator is used for the support and insulation of the cable entering and leaving the ring main unit. The reasons for easy discharge include: insulation breakdown, surface contamination and dampness, etc.; the reasons for easy discharge of the grounding busbar include: poor grounding and loose welding points, etc.; the above components are all high-incidence areas of partial discharge; in addition, the annotation tools for manual annotation can be LabelImg, VIA, etc., and the target detection model can be YOLO, Faster R-CNN, etc., which will not be elaborated here.

[0024] In an embodiment of the present invention, as Figure 3 shown, preprocess the discharge signals and thermal imaging maps at N time points to construct a feature sequence, including the following steps: Step S301: When the discharge signal or thermal imaging map at a time point is missing, randomly select the discharge signal or thermal imaging map at an adjacent time point for supplementation. Step S302, perform noise reduction processing on the discharge signal through wavelet transform, and perform smoothing processing on the thermal imaging map through Gaussian filtering; Step S303, extract the time-domain features of the time-domain waveform diagrams corresponding to the discharge signals at N time points; The time-domain features include: the maximum voltage value, the average voltage value, the root mean square value, the difference between the maximum voltage value and the minimum voltage value, the number of times exceeding the preset voltage value, and the duration of exceeding the preset voltage value, where the preset voltage value is a user-defined parameter. Preferably, the preset voltage value is set to the average value of all voltage values; Step S304, perform normalization processing on the time-domain features and the thermal imaging map at N time points through the Min-Max method to obtain a feature sequence.

[0025] In an embodiment of the present invention, the edges between vertices include: if there is a direct electrical connection relationship between key components in the monitoring area, an edge is constructed between the corresponding vertices; if the Euclidean distance between the locations of key components in the monitoring area is less than a preset distance threshold, an edge is constructed between the corresponding vertices. The preset distance threshold has the following calculation formula: ; where A, B, and C respectively represent the length, width, and height of the ring main unit, represents a user-defined proportionality coefficient with a value range between 0.1 and 0.3. Preferably, it is set to 0.2. For example, if the length, width, and height of the ring main unit are 2m, 1.5m, and 1m respectively, the preset distance threshold obtained according to the above calculation formula is 0.5m.

[0026] In an embodiment of the present invention, the first neural network model includes: a first hidden layer, a second hidden layer, a feature fusion layer, and a first classifier; The first hidden layer includes: N first update units, an unfolding unit, a splicing unit, and a second update unit; The u-th first update unit inputs the preprocessed thermal imaging map of the u-th sequence unit of the feature sequence and outputs a feature matrix, where 1 ≤ u ≤ N; The u-th unfolding unit is used to unfold the feature matrix output by the u-th first update unit into a vector representation and output a feature vector; The u-th splicing unit is used to splice the feature vector output by the u-th unfolding unit with the time-domain feature of the u-th sequence unit of the feature sequence and output a combined vector; The u-th second update unit inputs the combined vector output by the u-th splicing unit and outputs a second update vector; The second update unit is constructed based on the GRU model. The dimension number of the second update vector is a custom parameter. Preferably, the dimension number of the second update vector is set to 32; The second hidden layer inputs the feature sequence and outputs a third update vector; The second hidden layer is constructed based on the Transformer model. The dimension number of the third update vector is a custom parameter. The dimension number of the third update vector is set to 32; The feature fusion layer is used to splice the second update vector output by the Nth second update unit and the third update vector output by the second hidden layer to obtain a first update vector; The first update vector is input into the first classifier, and the classification space of the first classifier represents whether partial discharge occurs in the monitoring area.

[0027] In an embodiment of the present invention, the feature matrix output by the u-th first update unit has the following calculation formula: ; where represents the preprocessed thermal imaging map of the u-th sequence unit of the feature sequence input to the u-th first update unit, represents scaling the thermal imaging map to a size of 128×128, represents a convolution operation with the number of convolution kernels being K1, the convolution kernel size being 5×5, the stride being 1, and the padding method being Valid, represents a max-pooling operation with the pooling window size being 5×5 and the stride being 5, represents a convolution operation with the number of convolution kernels being K2, the convolution kernel size being 3×3, the stride being 1, and the padding method being Valid, represents a max-pooling operation with the pooling window size being 3×3 and the stride being 3, ChannelAvgPool represents a channel average pooling operation, Swish represents a Swish activation function, and both K1 and K2 are custom parameters. Preferably, K1 is set to 32 and K2 is set to 64.

[0028] It should be noted that the size of the feature map after processing is 124×124×K1, the size of the feature map after processing is 24×24×K1, the size of the feature map after processing is 22×22×K2, the size of the feature map after processing is 7×7×K2, and the size of the feature matrix obtained after ChannelAvgPool processing is 7×7. Then the dimension number of the combined vector is equal to 7×7 + 6 = 55.

[0029] In one embodiment of the present invention, the ultrasonic detection method is used to determine whether partial discharge occurs in the monitoring area, serving as the sample label for the training samples used to train the first neural network model.

[0030] In one embodiment of the present invention, the second neural network model includes: a third hidden layer and a second classifier; The third hidden layer inputs graph network data and outputs an updated matrix. The updated matrix includes M row vectors, and each row vector corresponds to a fourth update vector of a vertex. Each row vector of the updated matrix is input into the second classifier, and the classification space of the second classifier represents the fault type of the monitoring area.

[0031] In one embodiment of the present invention, the calculation formula of the updated matrix S output by the third hidden layer includes: ; ; ; where 1 ≤ i ≤ M, represents the fourth update vector of the i-th vertex of the updated matrix, and respectively represent the initial features of the i-th vertex and the j-th vertex, represents the set of vertices connected to the i-th vertex by edges, represents the association vector between the i-th vertex and the j-th vertex. The dimension number of the association vector is the same as the dimension number of the initial features of the vertices, represents the number of edges between the i-th vertex and the j-th vertex, , and respectively represent the first weight parameter, the second weight parameter, and the third weight parameter, and respectively represent the first bias parameter and the second bias parameter. Concat represents the concatenation function, sigmoid represents the sigmoid activation function, Swish represents the Swish activation function, and MLP represents the multi-layer perceptron, represents the stacking operation of the fourth update vectors of M vertices.

[0032] For example, if the size of the initial feature (the first update vector) of the vertex is 1×64, then the size of the association vector obtained after being processed by the MLP is also 1×64. The second weight parameter and the third weight parameter can be designed as matrices with a size of 64×32, and then the size of the fourth update vector is 1×32. The size of [it] is M×32. The first weight parameter can be designed as a vector of size 32×16. Then the size of the finally output updated matrix S is M×16.

[0033] In an embodiment of the present invention, the fault type of the monitoring area is obtained by the ultra-high frequency detection method and used as the sample label of the training sample for training the second neural network model.

[0034] It should be noted that the parameters in the first neural network model and the second neural network model are all learnable hyperparameters. The parameters in the first neural network model and the second neural network model are updated by the chain rule and the gradient descent algorithm. The gradient descent algorithm can be Adam, RMSProp, etc., which will not be elaborated here.

[0035] The above has described the embodiments of this embodiment, but this embodiment is not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative rather than restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make many forms, all of which fall within the protection scope of this embodiment.

Claims

1. A distribution automation terminal automation detection platform, characterized in that: include: A monitoring area division module, which is used to divide the ring main unit into M monitoring areas according to the locations of key components inside the ring main unit; A multimodal data acquisition module, which is used to collect the discharge signal of each monitoring area through a high-frequency current sensor and collect the thermal image of each monitoring area through an infrared thermal imager within a preset time period T and at a preset time interval t; The discharge signal is represented by a time domain waveform diagram, the horizontal axis of which represents the acquisition time point, and the vertical axis represents the voltage value; A feature sequence construction module, which is used to construct a corresponding feature sequence according to the discharge signals and thermal imaging images of M monitoring areas; The characteristic sequence corresponding to each monitoring area includes N sequence units, and the nth sequence unit is represented by the time domain feature obtained by preprocessing the discharge signal at the nth time point and the thermal image at the nth time point after preprocessing, where 1≤n≤N, N=T / t; A feature sequence updating module, which is used to input the feature sequence corresponding to each monitoring area into the first neural network model for updating, and output a first update vector; A graph network data construction module, which is used to construct graph network data according to the first update vectors corresponding to the M monitoring areas; Graph network data consists of M vertices and edges between vertices; A mapping relationship is established between each vertex and a monitoring area, and the initial feature of each vertex is represented by the first update vector of the monitoring area with which the mapping relationship is established; A discharge fault monitoring module, which is used to input the graph network data into the second neural network model for updating, and output the fault types of M monitoring areas; Fault types include: normal, minor partial discharge and severe partial discharge.

2. A distribution automation terminal automation detection platform according to claim 1, characterized in that: The preset time period T and the preset time interval t are both custom parameters. The ring main unit is divided into M monitoring areas, including the following steps: Step S201, collecting a visible light image of the ring main unit at the same position as the thermal imaging image, and manually marking the positions of key components inside the ring main unit in the visible light image; Key components inside the ring main unit include: busbar connection points, circuit breakers, cable connectors, insulators and grounding busbars; Step S202, obtaining the edge contour of the ring main unit in the visible light image through the target detection model; Step S203, calculating the minimum Euclidean distance between the locations of the key components, and generating a rectangular area with a side length of 1.5 times the distance and the location of each key component as the center; Step S204, if it is determined that the overlapping area between the two rectangular regions is greater than or equal to a preset overlapping area threshold, the side lengths of the two rectangular regions are recalculated and the rectangular regions are regenerated until the overlapping area between the two rectangular regions is less than the preset overlapping area threshold; Recalculate the side length of the rectangular area The calculation formula is as follows: ； in Indicates the side length of the rectangular area before recalculation. Indicates the preset overlap area threshold. Represents the overlapping area between two rectangular areas, where the preset overlapping area threshold is a custom parameter; Step S205: All rectangular areas are used as monitoring areas, and the number M of monitoring areas is obtained by counting.

3. A distribution automation terminal automation detection platform according to claim 1, characterized in that: Preprocessing the discharge signals and thermal images at N time points to construct a feature sequence includes the following steps: Step S301, when a discharge signal or thermal imaging image at a time point is missing, a discharge signal or thermal imaging image at an adjacent time point is randomly taken for supplementation; Step S302, performing noise reduction processing on the discharge signal by wavelet transform, and performing smoothing processing on the thermal image by Gaussian filtering; Step S303, extracting the time domain features of the time domain waveform graphs corresponding to the discharge signals at N time points; The time domain characteristics include: maximum voltage value, average voltage value, root mean square value, difference between maximum voltage value and minimum voltage value, number of times exceeding preset voltage value and duration of exceeding preset voltage value, wherein preset voltage value is a custom parameter; Step S304, normalizing the time domain features and thermal imaging images of N time points by using the Min-Max method to obtain a feature sequence.

4. A distribution automation terminal automation detection platform according to claim 1, characterized in that: The edges between vertices include: if there is a direct electrical connection between the key components in the monitoring area, then the corresponding vertices are constructed with edges; if the Euclidean distance between the locations of the key components in the monitoring area is less than the preset distance threshold, then the corresponding vertices are constructed with edges, and the preset distance threshold The calculation formula is as follows: ； A, B and C represent the length, width and height of the ring main unit respectively. Indicates a custom scale factor ranging from 0.1 to 0.

3.

5. A distribution automation terminal automation detection platform according to claim 1, characterized in that: The first neural network model includes: a first hidden layer, a second hidden layer, a feature fusion layer and a first classifier; The first hidden layer includes: N first update units, an expansion unit, a splicing unit, and a second update unit; The u-th first updating unit inputs the preprocessed thermal image of the u-th sequence unit of the feature sequence and outputs a feature matrix, where 1≤u≤N; The u-th expansion unit is used to expand the feature matrix output by the u-th first update unit into a vector representation and output a feature vector; The u-th concatenation unit is used to concatenate the feature vector output by the u-th expansion unit with the time domain feature of the u-th sequence unit of the feature sequence, and output a combined vector; The u-th second updating unit inputs the combined vector output by the u-th concatenation unit, and outputs a second updating vector; The second update unit is built based on the GRU model, and the number of dimensions of the second update vector is a custom parameter; The second hidden layer inputs the feature sequence and outputs the third update vector; The second hidden layer is built based on the Transformer model, and the number of dimensions of the third update vector is a custom parameter; The feature fusion layer is used to concatenate the second update vector output by the Nth second update unit and the third update vector output by the second hidden layer to obtain a first update vector; The first update vector is input into the first classifier, and the classification space of the first classifier indicates whether partial discharge occurs in the monitoring area.

6. A distribution automation terminal automation detection platform according to claim 5, characterized in that: The feature matrix output by the u-th first update unit The calculation formula is as follows: ； in represents the preprocessed thermal image of the u-th sequence unit of the feature sequence input by the u-th first update unit, Indicates scaling the thermal image to 128×128 size. It indicates a convolution operation with K1 convolution kernels, 5×5 convolution kernel size, 1 step size and Valid padding. represents a maximum pooling operation with a pooling window size of 5×5 and a step size of 5. It indicates a convolution operation with K2 kernels, 3×3 kernel size, 1 step size and Valid padding. It represents the maximum pooling operation with a pooling window size of 3×3 and a step size of 3. ChannelAvgPool represents the channel average pooling operation. Swish represents the Swish activation function. K1 and K2 are both custom parameters.

7. A distribution automation terminal automation detection platform according to claim 1, characterized in that: Whether local discharge occurs in the monitoring area is determined by ultrasonic detection method, and the results are used as sample labels of training samples for training the first neural network model.

8. A distribution automation terminal automation detection platform according to claim 1, characterized in that: The second neural network model includes: a third hidden layer and a second classifier; The third hidden layer inputs the graph network data and outputs an update matrix, the update matrix includes M row vectors, each row vector corresponds to the fourth update vector of a vertex; Each row vector of the update matrix is ​​input to the second classifier, and the classification space of the second classifier represents the fault type of the monitoring area.

9. A distribution automation terminal automation detection platform according to claim 8, characterized in that: The calculation formula of the update matrix S output by the third hidden layer includes: ； ； ； Where 1≤i≤M, represents the fourth update vector for the ith vertex of the update matrix, and Represent the initial features of the i-th vertex and the j-th vertex respectively, represents the set of vertices connected to the i-th vertex by edges. represents the association vector between the i-th vertex and the j-th vertex. The number of dimensions of the association vector is the same as the number of dimensions of the initial features of the vertex. represents the number of edges between the i-th vertex and the j-th vertex, , and represent the first weight parameter, the second weight parameter and the third weight parameter respectively, and Respectively represent the first bias parameter and the second bias parameter, Concat represents the concatenation function, sigmoid represents the sigmoid activation function, Swish represents the Swish activation function, MLP represents the multi-layer perceptron, Indicates that the fourth update vectors of M vertices are stacked.

10. A distribution automation terminal automation detection platform according to claim 1, characterized in that: The fault types in the monitoring area are obtained through the ultra-high frequency detection method as sample labels of training samples for training the second neural network model.

Images