A Method for Fault Diagnosis of Capacitive Equipment by Improving the YOLO Model

Abstract

The present invention proposes a capacitive equipment fault diagnosis method based on an improved YOLO model, which relates to the technical field of digital data processing in electric power, and includes collecting infrared images of capacitive equipment; denoising the images using a median filtering algorithm; using the YOLOv7 model to identify and extract targets from the denoised images; using the FCM algorithm to cluster the processed infrared images; and using the F-measure method to evaluate the clustering results. Through image denoising and target extraction, the present invention realizes the efficient identification of capacitive equipment, and through the FCM algorithm, clusters the infrared images of capacitive equipment to diagnose the fault types, which is convenient for realizing the intelligent management and control of the power grid, improves the efficiency of fault diagnosis, and reduces the danger of manual diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical digital data processing, and particularly to a capacitive equipment fault diagnosis method based on an improved YOLO model. Background Art

[0002] With the increasing complexity of the power system, the importance of capacitive equipment has been continuously enhanced, and its insulation state has a decisive impact on the stable operation of the system. Traditional monitoring methods are no longer suitable for the dynamically changing operation environment, and often cannot detect insulation defects in time, increasing system risks. Capacitive equipment accounts for most of the substation equipment, and its insulating medium may be deteriorated by multiple factors. In severe cases, it can lead to functional failure, threatening safe operation and personnel safety. Therefore, it is particularly important to monitor the insulation status of capacitive equipment in real time.

[0003] In the face of the growth of the scale and complexity of the power system, more intelligent and accurate monitoring methods are needed to improve the operation efficiency and safety of equipment. Deep learning and machine learning technologies provide new ways for intelligent monitoring. The present invention proposes an online monitoring method for all-station capacitive equipment based on the YOLOv7 algorithm, in order to ensure the safe and stable operation of the power system. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a capacitive equipment fault diagnosis method for improving the YOLO model, which can solve the problems of low efficiency and low accuracy of the existing methods for diagnosing capacitive equipment faults.

[0006] To solve the above technical problems, the present invention provides the following technical solution. A capacitive equipment fault diagnosis method for improving the YOLO model includes:

[0007] Collect infrared images of capacitive equipment; preprocess the infrared images, and use the median filtering algorithm to remove noise; apply the YOLOv7 model to identify and extract the equipment in the preprocessed infrared images; apply the FCM mean clustering method to cluster the infrared images processed by the YOLOv7 model; introduce the F-measure evaluation index to evaluate the clustering results.

[0008] As a preferred scheme of the capacitive equipment fault diagnosis method for improving the YOLO model according to the present invention, wherein: collecting infrared images of capacitive equipment includes the following steps: automatically scanning and collecting data of capacitive equipment by using an inspection robot with an infrared thermal imager.

[0009] As a preferred solution of the user-preference-based adaptive user interface generation method of the present invention, the preprocessing of the infrared image using the median filtering algorithm includes the following steps: The preliminary noise reduction of the image using the median filtering algorithm is specifically operated as follows: According to the shape of the picture taken by the infrared imager, a square filtering window is selected; A 3×3 filtering window is selected to ensure that more details are retained in the filtered image; All pixel values within the window are sorted to generate a monotonic two-dimensional data sequence; The median value is taken from the sorted pixel values, and the pixel count of the current image is replaced with this median value.

[0010] As a preferred solution of the capacitive device fault diagnosis method using the improved YOLO model of the present invention, the expression for selecting the median pixel value is:

[0011] ,

[0012] where, represents the selected median value, Med represents the median operation, represents the pixel value of the block diagram, N represents a natural number, m represents the window length, and v represents the number of samples included on both sides of the window.

[0013] As a preferred solution of the capacitive device fault diagnosis method using the improved YOLO model of the present invention, the identification and extraction of devices in the preprocessed infrared image using the YOLOv7 model include the following steps: The resize_image function is used to scale the input picture, and the input picture is adjusted to a size of 640×640.

[0014] As a preferred solution of the capacitive device fault diagnosis method using the improved YOLO model of the present invention, after the convolutional layer of the traditional backbone network, an SE module is added. This module consists of two main parts: The Squeeze stage: The spatial information of the feature map is compressed into channel-level global information through global average pooling. The purpose of the Squeeze stage is to compress the information in the spatial dimension into the information in the channel dimension. Given the input feature map , where H and W are the spatial dimensions of the feature map, and C is the number of channels. The global average pooling operation GAP compresses the dimension of each channel into a scalar:

[0015] ,

[0016] where, is the value of the input feature map X at position and channel c, It is the global descriptor of channel c. In this way, a C-dimensional vector can be obtained. It represents the global information of each channel.

[0017] Excitation stage: Use a fully connected layer to generate the weight coefficients of each channel, then normalize the weight values through the Sigmoid activation function, and finally perform weighted adjustment on each channel of the input feature map. This enables the model to automatically focus on those feature channels with important diagnostic information, such as temperature gradients and texture features.

[0018] First, reduce the dimension of the global descriptor z through a fully connected layer (through an activation function, such as ReLU). Assume it is reduced to an intermediate dimension. (r is a scaling factor, usually set to 16 or 32):

[0019] ,

[0020] Among them, is the weight matrix for dimensionality reduction, is the bias term, is the intermediate activation vector.

[0021] Then, increase the dimension back to channel C through another fully connected layer and normalize the result to the range [0,1] through the Sigmoid activation function:

[0022] ,

[0023] Among them, is the weight matrix for dimensionality increase, is the bias term, is a C-dimensional vector representing the weight coefficients of each channel.

[0024] The weight coefficients generated through the above Excitation stage will act on the input feature map. Assume the input feature map is X, and multiply the weight coefficients by the feature map of each channel to obtain the weighted feature map :

[0025] ,

[0026] Among them, is the value of the original feature map at position and channel c, is the channel weight obtained in the Excitation stage.

[0027] As a preferred solution of the capacitive device fault diagnosis method using the improved YOLO model of the present invention, the following steps are included: The picture passes through 4 CBS modules. By introducing the Transformer module to optimize the feature extraction process, the Transformer module is combined with the traditional convolutional layer. Position encoding is used to enhance the model's sensitivity to local features while retaining its advantage of global context modeling. Through the self-attention mechanism, the model can establish long-range dependencies between different positions of the input image. Especially when the relationship between the capacitive device area and other parts is relatively complex, the Transformer can effectively extract this global information. The multi-head self-attention mechanism of the Transformer enables the model to focus more on the heat map information of the target area, thereby improving the accuracy of image recognition. And through the ELAN module, the input and output of the convolutional layer are connected to generate more gradient flows. The feature map passes through the Head part, where feature maps at different levels are fused and the structure is predicted. The specific operations are as follows: C5 passes through the SPPCSP module, and the number of channels is reduced to 512; it is fused with the C4 and C3 feature maps in the top-down order to obtain the P3, P4, and P5 feature maps; it is fused with the P4 and P5 feature maps in the bottom-up order; the fused feature map is mapped to different target variables through the convolutional layer, including the category of the capacitive device, the bounding box coordinates, and the probability of the object's existence; finally, the infrared feature map enters the Head module to convert the feature map into the target detection result.

[0028] As a preferred solution of the capacitive device fault diagnosis method using the improved YOLO model of the present invention, the following steps are included: The YOLO model needs to be trained with a database. Based on the infrared images of capacitive devices obtained through network search, the image data for training the model is obtained. At the same time, infrared images are taken from multiple angles on all sides of various capacitive devices through an infrared imager to establish a dataset for the subsequent training of the model.

[0029] As a preferred solution of the capacitive device fault diagnosis method using the improved YOLO model of the present invention, the following steps are included: The YOLO model needs to be trained with a database. Based on the infrared images of capacitive devices obtained through network search, the image data for training the model is obtained. At the same time, infrared images are taken from multiple angles on all sides of various capacitive devices through an infrared imager to establish a dataset for the subsequent training of the model.

[0030] As a preferred solution of the capacitive device fault diagnosis method using the improved YOLO model of the present invention, the following steps are included: For the FCM mean clustering method, the constraint conditions of its objective function are as follows:

[0031] ,

[0032] Among them, J represents the clustering error, C1 represents the number of clusters, n represents the number of data points, represents the membership degree that data point j belongs to category i, represents the feature vector of data point j, p represents the fuzzy factor, represents the number of categories. After clustering is completed, it is necessary to analyze the features and fault modes represented by each cluster center in the clustering result. Each cluster center corresponds to the feature mode in the infrared image of the capacitive equipment in the substation.

[0033] After FCM clustering is completed, it is first necessary to analyze the features and possible fault modes represented by each cluster center in the clustering result. Each cluster center corresponds to the feature mode in the infrared image of the capacitive equipment in the substation, and these modes can reflect different states of the equipment during operation.

[0034] According to the clustering result, map the features of each cluster to the known fault modes of capacitive equipment. According to the working principle and empirical knowledge of the substation capacitive equipment, judge which clusters represent the fault areas. Among them, the high-temperature area (the area with higher temperature in the clustering result) represents the overheating fault, and the change of the hot spot area indicates that there is a fault in a certain component inside the equipment (such as capacitor failure, cable overheating, poor contact of switchgear, etc.). Using the membership degree information of each sample in the FCM clustering result, the potential fault area can be located. After clustering, the area with a higher membership degree usually indicates that this area has stronger fault characteristics and may be the hot spot area of the equipment.

[0035] As a preferred scheme of the capacitive equipment fault diagnosis method using the improved YOLO model described in the present invention, among them: the specific calculation formula for evaluating the clustering result using the F-measure evaluation index is as follows:

[0036] ,

[0037] ,

[0038] ,

[0039] ,

[0040] Among them, P represents the accuracy rate, R represents the recall rate, AP represents the average precision, rAP represents the mean average precision, TP represents the true positive, FP represents the false positive, TN represents the true negative, FN represents the false negative, and t represents the number of categories in the dataset.

[0041] Advantages of the present invention: The present invention uses a median filtering algorithm to filter the obtained infrared image. On the basis of preserving the basic information of the image, it can effectively suppress image noise and improve the accuracy of image recognition. The present invention uses the YOLOv7 model to perform target recognition and extraction on the filtered image. By adding auxiliary heads during the training process, this algorithm helps the model learn more complete feature representations and quickly and efficiently identify and extract the device in real time. The present invention uses the FCM mean clustering method to cluster the infrared image processed by the YOLOv7 algorithm. This method improves the stability of clustering through the effective allocation of inter-cluster boundary points and the excellent representation of covered data. Using the membership information of each sample in the FCM clustering result, potential fault areas can be located, and combined with the working conditions of the device, the nature of the fault can be further determined, thereby effectively improving the fault detection and maintenance efficiency of capacitive devices in substations. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0043] Figure 1 Schematic flow chart of a capacitive device fault diagnosis method for improving the YOLO model provided by an embodiment of the present invention.

[0044] Figure 2 Schematic flow chart of the FCM mean clustering method of a capacitive device fault diagnosis method for improving the YOLO model provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will give a detailed description of the specific embodiments of the present invention in conjunction with the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0046] In the following description, many specific details are set forth to facilitate a thorough understanding of the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0047] Secondly, the so-called "one embodiment" or "embodiment" herein refers to specific features, structures or characteristics that may be included in at least one implementation manner of the present invention. The appearances of "in one embodiment" in different places in this specification do not all refer to the same embodiment, nor are they separate or alternative embodiments that exclude each other with other embodiments.

[0048] The present invention is described in detail in conjunction with schematic diagrams. When detailing the embodiments of the present invention, for the convenience of explanation, the cross-sectional views showing the device structure will be locally enlarged in an ungeneralized scale, and the schematic diagrams are only examples and should not limit the scope of protection of the present invention herein. In addition, in actual production, three-dimensional spatial dimensions of length, width and depth should be included.

[0049] At the same time, in the description of the present invention, it should be noted that the orientation or positional relationship indicated by terms such as "upper, lower, inner and outer" is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation to the present invention. In addition, the terms "first, second or third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0050] Unless otherwise clearly defined and limited in the present invention, the terms "installed, connected, connected" should be understood in a broad sense. For example: it can be a fixed connection, a detachable connection or an integral connection; it can also be a mechanical connection, an electrical connection or a direct connection, and can also be indirectly connected through an intermediate medium, or can be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0051] Embodiment 1

[0052] Referring to Figure 1 - Figure 2 , which is the first embodiment of the present invention. This embodiment provides a capacitive device fault diagnosis method for improving the YOLO model, including:

[0053] S1: Collect infrared images of capacitive devices.

[0054] Specifically, use an inspection robot equipped with an infrared imager to photograph the capacitive device along a set route and generate its infrared image.

[0055] S2: Preprocess the infrared image and use the median filtering algorithm to remove noise.

[0056] Specifically, to save computing resources, for each pixel point in the infrared image, a 5×5 neighborhood window is selected; all the pixels within the selected 5×5 neighborhood are sorted in ascending order according to the pixel values to generate a monotonic two-dimensional data sequence; the median value is taken from the sorted pixel values and used to replace the original pixel value of the current pixel point. The noise points in the original image are replaced by the median values of the surrounding pixels, thereby achieving the purpose of smoothing the image; the above operations are repeated until the neighborhood window has traversed all the pixel points in the image.

[0057] Specifically, the expression is:

[0058] ,

[0059] where, represents the selected median value, Med represents the median operation, represents the pixel value of the block diagram, N represents a natural number, m represents the window length, here m takes 3, and v represents the number of samples included on both sides of the window.

[0060] S3: Apply the YOLOv7 model to identify and extract the devices in the preprocessed infrared image.

[0061] Specifically, use the resize_image function to scale the input image and adjust the input image to a size of 640×640; the adjusted image is input into the backbone network for further processing; after the convolutional layer of the traditional backbone network, an SE module is added. This module consists of two main parts: Squeeze stage: The spatial information of the feature map is compressed into channel-level global information through global average pooling (GlobalAverage Pooling). The purpose of the Squeeze stage is to compress the information in the spatial dimension into the information in the channel dimension. Given the input feature map , where H and W are the spatial dimensions of the feature map, and C is the number of channels. The global average pooling operation GAP compresses each channel's dimension into a scalar:

[0062] ,

[0063] where, is the value of the input feature map X at position and channel c, is the global descriptor of channel c. In this way, a C-dimensional vector can be obtained, which represents the global information of each channel.

[0064] Excitation Phase: Use a fully connected layer to generate the weight coefficients for each channel. Then, normalize the weight values through the Sigmoid activation function. Finally, perform weighted adjustment on each channel of the input feature map. This enables the model to automatically focus on those feature channels with important diagnostic information, such as temperature gradients and texture features.

[0065] First, reduce the dimension of the global descriptor z through a fully connected layer (using an activation function such as ReLU). Assume it is reduced to an intermediate dimension (r is a scaling factor, usually set to 16 or 32):

[0066] ,

[0067] where, is the weight matrix for dimensionality reduction, is the bias term, is the intermediate activation vector.

[0068] Then, increase the dimension back to channel C through another fully connected layer and normalize the result to the range [0, 1] through the Sigmoid activation function:

[0069] ,

[0070] where, is the weight matrix for dimensionality increase, is the bias term, is a vector of dimension C, representing the weight coefficients for each channel.

[0071] The weight coefficients generated through the above Excitation Phase will act on the input feature map. Assume the input feature map is X, multiply the weight coefficients by the feature map of each channel to obtain the weighted feature map :

[0072] ,

[0073] where, is the value of the original feature map at position and channel c, is the channel weight obtained in the Excitation Phase.

[0074] The image passes through 4 CBS modules. By introducing a Transformer module to optimize the feature extraction process, the Transformer module is combined with traditional convolutional layers. Position encoding is used to enhance the model's sensitivity to local features while retaining its advantage in global context modeling. Through the self-attention mechanism, the model can establish long-range dependencies between different positions in the input image. Especially when the relationship between the capacitive device area and other parts is complex, the Transformer can effectively extract this global information. The multi-head self-attention mechanism of the Transformer enables the model to focus more on the heatmap information of the target area, thereby improving the accuracy of image recognition; and through the ELAN module, the input and output of the convolutional layer are connected to generate more gradient flows.

[0075] Specifically, the feature map passes through the Head part, where feature maps at different levels are fused and the structure is predicted. The specific operations are as follows: C5 passes through the SPPCSP module, and the number of channels is reduced to 512; it is fused with the C4 and C3 feature maps in the top-down order to obtain the P3, P4, and P5 feature maps; it is fused with the P4 and P5 feature maps in the bottom-up order; the fused feature map is mapped to different target variables through a convolutional layer, including the category of the capacitive device, the bounding box coordinates, and the probability of the object's existence.

[0076] Furthermore, the infrared feature map enters the Head module to convert the feature map into an object detection result. The specific operations are as follows: Through the anchor box mechanism, the position and size of the prediction box of the capacitive device in the infrared image are partially defined; using the classification head, the category of the capacitive device in the image is identified by analyzing the feature map; using the regression head, the offset of the prediction box is predicted to more accurately locate the target object.

[0077] S4: Apply the FCM mean clustering method to cluster the infrared images processed by the YOLOv7 model.

[0078] Specifically, adopt the FCM mean clustering method, as Figure 2 shown, to classify the feature images extracted by the YOLO model: Calculate the cluster centers according to the current membership degrees (the cluster centers are the weighted averages of the data points, and the weights are represented by the membership degrees); update the membership degrees according to the current cluster centers; repeat the above operations until the objective function meets the conditions.

[0079] Specifically, the calculation formula for the cluster center C(k,d) is:

[0080] ,

[0081] where, represents the membership degree of sample j belonging to category i, Denote the value of data point \(i\) in feature dimension \(d\), and \(g\) represents the fuzzy factor, usually taking a real number greater than 1, and here it takes 1.5.

[0082] The updated membership degree calculation formula is as follows:

[0083] ,

[0084] where \(X(i)\) represents the feature vector of data point \(i\), and \(C(k)\) represents the feature vector of cluster center \(k\).

[0085] The constraint conditions of its objective function are as follows:

[0086] ,

[0087] where \(J\) represents the clustering error, \(C1\) represents the number of clusters, \(n\) represents the number of data points, represents the membership degree that data point \(j\) belongs to category \(i\), represents the feature vector of data point \(j\), and \(p\) represents the fuzzy factor, represents the number of categories. After clustering is completed, it is necessary to analyze the features and fault modes represented by each cluster center in the clustering results. Each cluster center corresponds to a feature pattern in the infrared image of substation capacitive equipment.

[0088] Specifically, after the FCM clustering is completed, it is first necessary to analyze the features and fault modes represented by each cluster center in the clustering results. Each cluster center corresponds to a feature pattern in the infrared image of substation capacitive equipment, and these patterns can reflect different states of the equipment during operation. According to the clustering results, map the features of each cluster to the known fault modes of capacitive equipment. Based on the working principle and empirical knowledge of substation capacitive equipment, judge which clusters may represent the fault areas. Among them, the high-temperature area (the area with a higher temperature in the clustering results) represents an overheating fault, and the change in the hot spot area indicates that there is a fault in a certain component inside the equipment (such as capacitor failure, cable overheating, poor contact of switchgear, etc.). Using the membership degree information of each sample in the FCM clustering results, the potential fault areas can be located. After clustering, the areas with higher membership degrees usually indicate that these areas have stronger fault characteristics and may be the hot spots of the equipment.

[0089] S5: Introduce evaluation indicators to evaluate the clustering results.

[0090] Specifically, for the method of evaluating the clustering results, among the evaluation metrics, P (Precision), R (Recall), AP (Average Precision), and mAP (mean Average Precision) are selected as references. The specific calculation formulas are as follows:

[0091] ,

[0092] ,

[0093] ,

[0094] ,

[0095] Among them, P represents precision, R represents recall, AP represents average precision, rAP represents mean average precision, TP represents true positive, FP represents false positive, TN represents true negative, FN represents false negative, and t represents the number of categories in the dataset.

[0096] The terms used to describe the positional relationship in the drawings are for illustrative purposes only and should not be construed as a limitation of this patent.

[0097] Obviously, the above embodiments of the present invention are only examples for clearly explaining the present invention and are not limitations on the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all the implementation manners here. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. An improved YOLO model-based capacitive equipment fault diagnosis method, characterized in that: Including, Collecting infrared images of capacitive devices; Preprocessing the infrared images, using the median filtering algorithm for denoising; Applying the YOLOv7 model to identify and extract the devices in the preprocessed infrared images; Applying the FCM mean clustering method to cluster the infrared images processed by the YOLOv7 model; Introducing the F-measure evaluation index to evaluate the clustering results; The YOLOv7 model uses the resize_image function to scale the input image and adjusts the input image to a size of 640×640; The adjusted image is input into the backbone network for processing; after the convolutional layer of the backbone network, an SE module is added; In the Squeeze stage: The spatial information of the feature map is compressed into channel-level global information through global average pooling. Given the input feature map X ∈ R H×W×C , where H and W are the spatial dimensions of the feature map and C is the number of channels, the global average pooling operation GAP compresses the H×W dimension of each channel into a scalar: where x ijc is the value of the input feature map X at position (i, j) and channel c, and z c is the global descriptor of channel c, and at the same time, a C-dimensional vector z = [z1, z2,..., z C is obtained, representing the global information of each channel; In the Excitation stage, a fully connected layer is used to generate the weight coefficients for each channel, the weight values are normalized through the Sigmoid activation function, and each channel of the input feature map is weighted and adjusted: s1 = ReLU(W1z + b1) s2 = Sigmoid(W2s1 + b2) X′ ijc = X ijc ·s 2c , c = 1, 2, ..., C Among them, W1 and W2 are weight matrices for dimensionality reduction, b1 and b2 are bias terms, s1 is an intermediate activation vector, s2 is a vector of C dimensions representing the weight coefficients for each channel, and X ijc is the value of the original feature map at position (i, j) and channel c, and s 2c is the channel weight obtained in the Excitation stage; The median filtering algorithm performs preliminary denoising on the image. The specific operation is as follows: According to the shape of the image taken by the infrared imager, a square filtering window is selected; A 3×3 filtering window is selected, and all pixel values within the window are sorted to generate a monotonic two-dimensional data sequence; The median value is taken from the sorted pixel values, and the pixel points of the current image are replaced with the median value; The expression is: Among them, y i represents the selected intermediate value, Med represents the median operation, f i represents the pixel value of the block diagram, N represents a natural number, m represents the window length, and v represents the number of samples included on both sides of the window; After the image passes through 4 CBS modules, the self-attention mechanism of the Transformer module is introduced to optimize the feature extraction process. The Transformer module is combined with the convolutional layer to establish a dependency relationship between different positions of the input image through the self-attention mechanism, and the input and output of the convolutional layer are connected through the ELAN module; The feature map passes through the Head part, and the feature maps at different levels are fused and the structure is predicted. The specific operation is as follows: The C5 feature map is processed by the SPPCSP module, and the number of channels is reduced to 512. Then, the C5 feature map is fused with the C4 and C3 feature maps in the top-down order. At the same time, the feature maps are gradually fused from high level to low level to form three different levels of feature maps, namely P3, P4, and P5; In the bottom-up order, the P4 and P5 feature maps are fused again to ensure the complementarity of multi-level features; After the feature map fusion, these fused feature maps are mapped to different target variables through the convolutional layer, including the category of the capacitive device, the bounding box coordinates, and the probability of the object's existence; Finally, the infrared feature map is input into the Head module to convert all feature maps into the final object detection results; The YOLOv7 model needs to be trained with a database. Based on the infrared images of capacitive devices obtained through network search, the image data for training the model is obtained. At the same time, infrared images of various capacitive devices are taken through the infrared imager to establish a dataset for model training.

2. The capacitive device fault diagnosis method for improving the YOLO model according to claim 1, wherein: By using an inspection robot equipped with an infrared thermal imager, the capacitive devices are automatically scanned and data is collected.

3. The capacitive device fault diagnosis method for improving the YOLO model according to claim 2, characterized in that: The objective function constraint conditions of the FCM mean clustering method are as follows: where J represents the clustering error, C represents the number of clusters, n represents the number of data points, represents the membership degree of data point j belonging to category i, x j represents the feature vector of data point j, p represents the fuzzy factor, C i represents the number of categories. After clustering, it is necessary to analyze the features and possible fault modes represented by each cluster center in the clustering result. Each cluster center corresponds to the feature pattern in the infrared image of the capacitive equipment in the substation; According to the clustering results, map the features of each cluster to the known capacitive device fault modes and determine the possible fault areas; Among them, the area with a temperature higher than the set threshold in the clustering result is the high-temperature area, and the high-temperature area represents an overheating fault. The change in the hot spot area indicates that there is a fault in the internal components of the device; Using the membership information of each sample in the FCM clustering result, locate the potential fault areas.

4. The capacitive equipment fault diagnosis method for improving the YOLO model according to claim 3, characterized in that: Evaluate the clustering results, and the specific calculation formula for evaluation is as follows: Among them, P represents the accuracy rate, R represents the recall rate, AP represents the average precision, rAP represents the mean of the average precision, TP represents the true positive, FP represents the false positive, TN represents the true negative, FN represents the false negative, and t represents the number of categories in the dataset.

Images