A substation cubicle equipment state detection method and device

Abstract

The present application belongs to the technical field of target detection, and more particularly relates to a substation screen cabinet equipment state detection method and device. The method comprises: acquiring a substation screen cabinet secondary equipment image to construct a sample data set, labeling the sample data set and dividing it into a training set and a validation set; constructing a target detection model and an image classification model, training the target detection model and the image classification model respectively using the training set; using the trained target detection model to identify the state of the screen cabinet secondary equipment, using the trained image classification model to detect whether the target detection model has target missed detection, and combining an OCR model to identify the screen cabinet name, thereby indexing the screen cabinet secondary equipment state comparison, and labeling the final comparison difference area. The present application solves the problems of difficult confirmation of secondary equipment state in the intelligent inspection of substation screen cabinets and large personnel workload and low inspection efficiency of the existing inspection method.

Description

Technical Field

[0001] This invention belongs to the technical field of target detection, and more specifically, relates to a method and device for detecting the status of substation cabinet equipment. Background Technology

[0002] Conventional power stations face a large amount of repetitive work involving manual inspections and manual reviews. Taking a substation in Shanghai as an example, the substation operation and maintenance center is responsible for the inspection and recording of more than 100,000 panels. Routine inspections of the panels are required every quarter. However, the existing inspection methods have the following problems: (1) There are many panels and the operating conditions are complex. The inspection personnel need to manually check a large number of points in each station, resulting in low inspection efficiency. At the same time, the inspection results vary due to differences in the individual abilities and sense of responsibility of the inspection personnel. (2) After the operation of equipment such as panel pressure plates and circuit breakers, the on-site operators need to compare them one by one and reprint the current status, which is quite tedious. (3) The panel doors will block some equipment, and most of the circuit breakers are installed behind the panel. During the inspection, the back door of the panel needs to be opened to see the specific status. Conventional cameras or track robots cannot monitor the status of the above equipment. The above inspection methods have problems such as large workload of personnel and low inspection efficiency.

[0003] In summary, maintenance personnel hope to use a more convenient method to intelligently detect and identify cabinet pressure plates, indicator lights, circuit breakers, etc., in order to achieve intelligent judgment of equipment status, reduce the workload of maintenance personnel, and create digital maintenance, establishing a new maintenance management model driven by data.

[0004] Chinese patent document CN115409816A proposes a method, device, equipment, and storage medium for identifying the status of pressure plates in substation cabinets. The method involves determining the pressure plate detection area of ​​the current cabinet image based on a pre-trained pressure plate area detection model, and then obtaining the cabinet label image and the pressure plate area image of the pressure plate to be identified based on the pre-trained target pressure plate detection model. Finally, the method determines the engagement / disengagement status of the pressure plate to be identified based on the cabinet label image and the pressure plate area image.

[0005] Chinese patent document CN113221687A proposes a training method and a method for recognizing the state of pressure plates. By replacing the official CSPDarknet53 network with a MobilenetV3 network in the feature extraction backbone network of the standard YOLOv5s algorithm, and by performing sparsity processing and model channel pruning, the network parameters and model size can be simplified, reducing the algorithm's consumption of computer hardware resources. This makes the model suitable for recognizing pressure plates in edge terminal devices with general computing performance. At the same time, the model distillation technique is used to ensure that the final model file has high recognition accuracy, which can meet the application requirements for intelligent recognition of the pressure plate state of cabinets on edge terminals.

[0006] The above method only achieves automated identification of the status of the substation cabinet pressure plate, but cannot identify the status of other secondary equipment such as cabinet switches and indicator lights, cannot identify the changes that have occurred in the cabinet from the previous moment, and cannot detect whether the model has missed detection or reporting of targets.

[0007] Therefore, in order to solve the above-mentioned problems, there is an urgent need for a method or device that can intelligently detect and identify the status of substation cabinet equipment, so as to realize intelligent judgment of equipment status, reduce the workload of operation and maintenance personnel, and improve the efficiency of inspection work. Summary of the Invention

[0008] The present invention aims to overcome at least one of the defects of the prior art and provide a method for detecting the status of substation cabinet equipment, so as to solve the problems of difficulty in confirming the status of secondary equipment during intelligent inspection of substation cabinets and the problems of large workload and low efficiency of existing inspection methods.

[0009] The present invention also discloses an apparatus loaded with a method for detecting the status of substation cabinet equipment.

[0010] The detailed technical solution of this invention is as follows:

[0011] A method for detecting the status of substation switchgear equipment, the method comprising:

[0012] S1. Obtain images of secondary equipment in the substation cabinet to construct a sample dataset, label the sample dataset, and divide the labeled sample dataset into a training set and a validation set.

[0013] S2. Construct an object detection model and an image classification model, and train the object detection model and the image classification model respectively using the training set;

[0014] S3. Use the trained target detection model to identify the status of the secondary equipment in the cabinet, and use the trained image classification model to detect whether the target detection model has missed any targets. At the same time, combine the OCR model to identify the cabinet name, use this as an index to compare the status of the secondary equipment in the cabinet, and mark the difference areas in the final comparison.

[0015] According to a preferred embodiment of the present invention, in step S2, a target detection model is constructed based on an improved YOLOv5 network and combined with the CBAM attention mechanism; wherein, the improved YOLOv5 network is an upsampling branch added to the output head of the initial YOLOv5 network to increase the feature map output after upsampling, specifically including:

[0016] Extract the first feature map from the initial YOLOv5 network after 4x downsampling;

[0017] The feature map in the initial YOLOv5 network, which was downsampled by 8 times, is upsampled by 1 times to obtain the second feature map.

[0018] The first feature map and the second feature map are fused to obtain a new feature map, and the size of the new feature map is larger than the size of the first feature map and the second feature map.

[0019] According to a preferred embodiment of the present invention, the CBAM attention mechanism includes a channel attention map M. c (F) and spatial attention map M s (F), where:

[0020] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) (1)

[0021] M s (F)=σ(f 7×7 ([AvgPool(F));MaxPool(F)])) (2)

[0022] In equations (1) and (2), σ(·) represents the sigmoid function, MLP represents a multilayer perceptron, AvgPool(F) and MaxPool(F) represent the average pooling feature and the max pooling feature, respectively, and f 7×7 This represents a convolution operation with a filter size of 7×7; and, in equation (1), AvgPool(F) and MaxPool(F) are forwarded to a shared network composed of a multilayer perceptron (MLP), and pass through the sigmoid function σ to generate a channel attention map M. c(F); Equation (2) indicates that AvgPoOl(F) and MaxPool(F) are aggregated to generate two 2D graphs, and then the spatial attention graph M is generated by the sigmoid function σ. s (F).

[0023] According to a preferred embodiment of the present invention, in step S2, before training the target detection model using the training set, the K-means clustering algorithm is first used to perform anchor box clustering on the training set, and the obtained anchor boxes are used to train the target detection model; wherein, the step of using the K-means clustering algorithm to perform anchor box clustering on the training set specifically includes:

[0024] S21. Extract the coordinate values ​​of all bounding boxes for each image in the training set, and convert the coordinate values ​​of all bounding boxes into the height and width values ​​of the bounding boxes, wherein the bounding boxes are obtained by labeling the sample dataset;

[0025] S22. Randomly select M coordinates as the center points of the Kmeans clustering algorithm, calculate the intersection-union ratio (IUR) between each bounding box of each image in the training set and each center point, and assign each bounding box to the cluster to which the center point with the largest IUR belongs;

[0026] S23. Update the center point of each cluster based on the average coordinates of all bounding boxes in each cluster;

[0027] S24. Repeat steps S22 to S23 until the center point of each cluster is stable or the preset maximum number of iterations is reached, then stop the operation and obtain M anchor boxes for training the target detection model.

[0028] According to a preferred embodiment of the present invention, the image classification model constructed in step S2 sequentially includes a 5*5 convolutional layer, 7 groups of depthwise separable convolutional modules and a fully connected layer, wherein each group of depthwise separable convolutional modules sequentially includes a 3*3 depthwise convolutional layer, a 1*1 point convolutional layer and a 2*2 pooling layer.

[0029] According to a preferred embodiment of the present invention, in step S2, training the image classification model using the training set specifically includes:

[0030] The training set is input into the image classification model, and the number of channels in the input feature map is N;

[0031] Perform convolution operations on each of the N channels using a single convolution kernel to obtain a feature map with N channels equal to 1.

[0032] The N feature maps with 1 channel are concatenated sequentially to obtain an output feature map with N channels.

[0033] The output feature maps with N channels are weighted and combined in the depth direction to obtain a new high-dimensional feature map.

[0034] The new high-dimensional feature map is output using the Softmax activation function, i.e.:

[0035]

[0036] In equation (3), y represents the category of the current sample, e represents the base of the natural logarithm, and Z... i z represents the feature value extracted by the network that belongs to the i-th class. y is the feature value of the current class, and C represents the total number of classes.

[0037] According to a preferred embodiment of the present invention, in step S2, training the image classification model using the training set further includes: using Softmax cross-entropy loss and iteratively applying gradient descent to gradually decrease the value of the Softmax cross-entropy loss until convergence, thereby completing the model training, wherein:

[0038]

[0039] In equation (4), L softmax This represents the Softmax cross-entropy loss.

[0040] According to a preferred embodiment of the present invention, in step S3, the state of the secondary equipment of the cabinet is identified using the trained target detection model, specifically including: inputting the image of the secondary equipment of the cabinet to be identified into the trained target detection model; the target detection model first performs horizontal correction on the input image; the correction operation includes extracting straight lines in the image and removing straight lines that are close to vertical; then calculating the average tilt angle of the remaining straight lines; and rotating and correcting the image according to the average tilt angle; wherein, the straight lines in the image are given by Hough line detection, which here refers to the edge lines on the cabinet.

[0041] According to a preferred embodiment of the present invention, step S3, which involves using the trained image classification model to detect whether the target detection model has missed any targets, specifically includes:

[0042] The system detects the distribution of targets in the image to be identified and finds targets with abnormal distribution. Targets with abnormal distribution are identified by statistically analyzing the distance between every two adjacent secondary devices (pressure plates, indicator lights, etc.) in the cabinet. If the distance is greater than the average distance between all secondary devices, then these two secondary devices are determined to be targets with abnormal distribution.

[0043] Centered on the abnormally distributed target, images of the four regions above, below, left, and right of the target are cropped and fed into the trained image classification model. The output of the image classification model is used to determine whether the target detection model has missed any targets.

[0044] In another aspect of the present invention, an apparatus for implementing a substation cabinet equipment status detection method is also provided. The apparatus stores an executable computer program, which, when executed, performs the substation cabinet equipment status detection method as described above.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] (1) The present invention provides a substation cabinet equipment status detection method, which improves the target detection network by using deep learning technology and builds a lightweight image recognition network. Through the classification learning of image data, it has a stable status recognition capability for substation cabinet equipment.

[0047] (2) This invention utilizes the uniqueness of the inspection cabinet name, uses an OCR model to identify the cabinet name, and combines a target detection model and an image classification model to identify the status of secondary equipment such as cabinet switches and pressure plates. It automatically establishes a cabinet database and performs row and column processing on the identification results. Using the cabinet name identified by the OCR model as an index, it finds the historical status of the current cabinet in the cabinet database. By comparing the current identified status data with the corresponding historical status data, it identifies the status changes of the secondary equipment in the cabinet. Finally, it outputs the row and column position of the changed target and issues an alarm.

[0048] (3) The present invention systematically solves the key problems of substation panel inspection and abnormal alarm, greatly reduces the workload of personnel and manual operation, and significantly improves the efficiency of inspection work. Attached Figure Description

[0049] Figure 1 This is a flowchart of the substation panel equipment status detection method described in this invention.

[0050] Figure 2 This is a network structure diagram of the target detection model in Embodiment 1 of the present invention.

[0051] Figure 3 This is a network structure diagram of the CBAM attention mechanism in Embodiment 1 of the present invention.

[0052] Figure 4 This is a schematic diagram illustrating the evaluation performance of the trained target detection model in Embodiment 1 of the present invention.

[0053] Figure 5 This is a network structure diagram of the image classification model in Embodiment 1 of the present invention.

[0054] Figure 6 This is a schematic diagram illustrating the principle of depthwise convolution in the image classification model of Embodiment 1 of the present invention.

[0055] Figure 7 It is the effect diagram of the state recognition of the substation panel cabinet equipment in Embodiment 1 of the present invention. Specific implementation manners

[0056] The following further describes the present disclosure in conjunction with the accompanying drawings and embodiments.

[0057] It should be noted that the following detailed descriptions are all exemplary and intended to provide further descriptions of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present disclosure belongs.

[0058] It should be noted that the terms used herein are only for describing specific implementation manners and are not intended to limit the exemplary implementation manners according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should also be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0059] Without conflict, the embodiments in the present disclosure and the features in the embodiments may be combined with each other.

[0060] The present invention provides a method for detecting the state of substation panel cabinet equipment to solve the problems of large workload of personnel and low inspection efficiency in the existing method for inspecting the state of substation panel cabinet equipment. The method mainly includes two stages: model training and model application. In the model training stage, the constructed object detection model and image classification model are trained using the images of the secondary equipment of the panel cabinets in the substation; in the model application stage, the trained object detection model is used to detect and identify the state of the secondary equipment of the panel cabinets, and at the same time, the image classification model is used to detect whether there are missed detections and false alarms of the object detection model, and then the OCR model is combined to identify the panel cabinet name, and the state of the secondary equipment of the panel cabinets is compared with this as an index, and the finally compared difference area is marked in the form of a rectangular box, so as to accurately detect the change of the state of the secondary equipment of the panel cabinets.

[0061] For convenient use, the above functions can be encapsulated into a mobile (such as mobile phone) APP. The operation and inspection personnel can hold the mobile device to take pictures and identify, and the recognition result automatically generates an electronic report, greatly reducing the personnel workload and manual operation, and significantly improving the inspection efficiency.

[0062] The following further describes the method and device for detecting the state of substation panel cabinet equipment of the present invention in detail with specific embodiments.

[0063] Embodiment 1

[0064] Refer Figure 1 , this embodiment provides a method for detecting the status of substation panel cabinet equipment, and the method includes:

[0065] Step S1, obtain images of secondary equipment of panel cabinets in the substation to construct a sample data set, annotate the sample data set, and divide the annotated sample data set into a training set and a validation set.

[0066] Specifically, devices such as mobile phones or industrial cameras can be used to collect images of all secondary equipment of panel cabinets in the substation. Since there may be large light when collecting images, resulting in brightness differences between the collected images, in order to ensure the anti-interference ability of the model at the data level, the collected images need to be preprocessed. In this embodiment, conventional image augmentation means can be used, including operations such as brightness adjustment, Gaussian noise removal, horizontal flipping, and vertical flipping, to standardize the collected images, and then construct the standardized image data into a sample data set.

[0067] Then, all target detection points in each image in the constructed sample data set are marked in the form of a rectangular box, and then the annotated sample data set is divided into a training set and a validation set, where the training set is used for subsequent model training, and the validation set is used to evaluate the performance of the trained model.

[0068] Step S2, construct a target detection model and an image classification model, and use the training set to train the target detection model and the image classification model respectively.

[0069] In this embodiment, a target detection model can be constructed based on the YOLO network as the basic architecture, and preferably, a target detection model is constructed based on the improved YOLOv5 network architecture and combined with the CBAM attention mechanism.

[0070] Since the target detection object in this embodiment is the secondary equipment of the panel cabinet, considering that there are many small targets in the secondary equipment of the panel cabinet, it is therefore preferred to improve the output head of the target detection network. By adding an upsampling branch to the network output head, the detection effect of small targets is improved. Specifically, refer to Figure 2 As shown, the first step is to extract the first feature map obtained after 4-fold downsampling in the backbone network; the second step is to perform 1-fold upsampling operation on the feature map with 8-fold downsampling in the original network to obtain the second feature map; the third step is to fuse the first feature map and the second feature map in the previous two steps to obtain a new feature map with a size of 160*160. The advantage of such improvement is that the feature map becomes larger after upsampling, and small targets will become obvious, which is beneficial to the detection of small targets.

[0071] Refer to Figure 3 , the CBAM attention mechanism is divided into two parts: spatial attention and channel attention, and can be directly inserted into the target detection network, which is a plug-and-play module. The CBAM attention mechanism contains the channel attention map Mc (F) and spatial attention map M s (F), where:

[0072] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) (1)

[0073] M s (F)=σ(f 7×7 ([Avgpool(F));MaxPool(F)])) (2)

[0074] In equations (1) and (2), σ(·) represents the sigmoid function, MLP represents a multilayer perceptron, AvgPool(F) and MaxPool(F) represent the average pooling feature and the max pooling feature, respectively, and f 7×7 This represents a convolution operation with a filter size of 7×7. Furthermore, in equation (1), AvgPool(F) and MaxPool(F) are forwarded to a shared network composed of a multilayer perceptron (MLP), and pass through a sigmoid function σ to generate a channel attention map M. c (F); Equation (2) indicates that AvgPool(F) and MaxPool(F) are aggregated to generate two 2D graphs, and then the spatial attention graph M is generated by the sigmoid function σ. s (F).

[0075] The input feature map F is compared with the channel attention map M respectively. c (F) and spatial attention map M s (F) Performing element-wise multiplication yields a new output feature map F′, which possesses stronger feature representation capabilities. This can be viewed as extracting spatial and channel dimensions of features from the original two feature layers, resulting in a new output layer with enhanced feature representation capabilities compared to the original output layer. Therefore, by incorporating the CBAM attention mechanism, the overall recognition rate of the model can be significantly improved.

[0076] Furthermore, since the acquired images of the secondary equipment in the substation cabinets contain cabinet names and secondary equipment, and considering the large size difference between cabinet names and secondary equipment, in this embodiment, before training the target detection model using the training set, it is preferable to first use the Kmeans clustering algorithm to perform anchor box clustering on the training set for target detection, and then use the obtained anchor boxes to train the target detection model.

[0077] Specifically, the steps for generating anchor boxes on the training set using the K-means clustering algorithm include:

[0078] S21. Extract the coordinate values of all bounding boxes in each image in the training set, and convert the coordinate values of all bounding boxes into the height values and width values of the bounding boxes; wherein, the bounding box is the rectangular box used to label the target detection points in step S1.

[0079] S22. Randomly select M coordinates as the center points of the Kmeans clustering algorithm, calculate the IoU (Intersection over Union) values between each bounding box in each image in the training set and each center point, and assign each bounding box to the cluster where the center point with the largest IoU value belongs.

[0080] S23. Update the center points of the clusters according to the average values of the coordinates of all bounding boxes in each cluster.

[0081] S24. Repeat steps S22 - S23 until the center points of each cluster are stable or the preset maximum number of iterations is reached, and then stop the operation, and obtain M center points for the anchor boxes used in object detection.

[0082] Then, use the obtained anchor boxes to train the constructed object detection model. Continue to refer Figure 2 In this embodiment, the improved object detection model contains 4 branches, and each branch has 3 anchor boxes. Therefore, 12 coordinates are randomly selected as the center points of the Kmeans clustering algorithm in the above steps, and finally 12 anchor boxes are obtained for model training.

[0083] In this embodiment, 150 rounds of object detection model training cycles are preset. During the training process, the performance of the trained model is evaluated in combination with the divided validation set. The evaluation metrics include the detection rate Precision, the recall rate Recall, and the accuracy mAP50. The evaluation results are as Figure 4 shown. It can be seen that the model has a detection rate and a recall rate of more than 98%, and the average accuracy (mAP50) reaches more than 99%, which significantly demonstrates the effectiveness of the model improvement scheme.

[0084] After training is completed, the final object detection model can be obtained to identify the status of the secondary equipment in the switch cabinet.

[0085] Refer Figure 5 In this embodiment, the image classification model sequentially includes a 5*5 convolutional layer, 7 groups of depthwise separable convolutional modules, and a fully connected layer. Each group of depthwise separable convolutional modules sequentially includes a 3*3 depth convolutional layer, a 1*1 point convolutional layer, and a 2*2 pooling layer. The image classification model receives an input image of 224*224. The image first passes through a convolution with a convolutional kernel size of 5*5, and then the features are extracted through 7 groups of depthwise separable convolutional modules to obtain a feature map. Finally, the feature map is compressed into a one-dimensional vector, and the output is predicted through the fully connected layer using the softmax activation function.

[0086] Furthermore, this embodiment employs depthwise separable convolution to reduce model size and utilizes the training set to train the image classification model. See details... Figure 6 As shown, the training set is input into the image classification model. The number of channels in the input feature map is N. After applying a convolution kernel to each of the N channels individually, N feature maps with 1 channel are obtained. These N feature maps are then concatenated sequentially to obtain an output feature map with N channels. The kernel size of the point convolution is 1×1×N. The feature maps from the previous step are weighted and combined in the depth direction to generate a new high-dimensional feature map. Finally, after multiple convolution and pooling operations, the extracted high-dimensional features are output using the Softmax activation function, i.e.:

[0087]

[0088] In equation (3), y represents the category of the current sample, e represents the base of the natural logarithm, and Z... i z represents the feature value extracted by the network that belongs to the i-th class. y is the feature value of the current class, and C represents the total number of classes.

[0089] Finally, the Softmax cross-entropy loss L is used. softmax And use gradient descent to iterate, making L softmax The loss value gradually decreases until it converges, completing the model training process. Wherein:

[0090]

[0091] After training, the final image classification model is obtained, primarily used to detect whether the target detection model has missed detections or false alarms. Furthermore, the trained image classification model can also identify whether an image belongs to the foreground or background category. It utilizes prior knowledge of the target distribution in the image to mine potential missed detection samples, crops the potential sample regions, and feeds them into the trained image classification model. The image classification model provides confidence scores for different categories to determine whether the input image belongs to the foreground or background category. The category confidence scores are obtained through Softmax output, specifically a 1*C array (where C is the number of categories), with the category having the largest number being the current foreground class.

[0092] Step S3: Use the trained target detection model to identify the status of the secondary equipment in the cabinet, and use the trained image classification model to detect whether the target detection model has missed any targets. At the same time, combine the OCR model to identify the cabinet name, use this as an index to compare the status of the secondary equipment in the cabinet, and mark the difference areas in the final comparison.

[0093] Specifically, the image of the secondary equipment in the control cabinet to be identified is input into the trained object detection model. The object detection model first performs horizontal correction on the input image. The specific correction method is as follows: extract straight lines from the image and remove lines that are close to vertical, then calculate the average tilt angle of the remaining lines for rotation correction of the image; the straight lines in the image are obtained through Hough line detection, which in this case refers to the edge lines on the control cabinet. Afterwards, the object detection model performs secondary equipment status recognition on the corrected image.

[0094] To improve the accuracy of target detection and prevent missed detections, a trained image classification model is run simultaneously with the trained target detection model. Based on the target distribution in the image to be identified, a four-neighbor search is performed on abnormally distributed targets. Specifically: First, the target distribution in the image to be identified is detected, and abnormally distributed targets are identified. Abnormally distributed targets are defined as those identified by statistically analyzing the distance between any two adjacent devices (pressure plates, indicator lights, etc.) in the control panel. If the distance is greater than the average distance between all devices, these two devices are considered abnormally distributed targets, indicating a potential missed detection. Then, images of the four regions above, below, left, and right of the abnormally distributed target are cropped and fed into the trained image classification model. The output of the image classification model is then used to determine whether the target detection model has missed any targets.

[0095] Thus, by utilizing deep learning technology to improve the target detection network and building a lightweight image recognition network, a stable state recognition capability for substation cabinet equipment can be achieved through the classification and learning of image data.

[0096] Furthermore, this embodiment combines OCR (Optical Character Recognition) model to identify cabinet names. The main algorithm models of OCR include neural network models, Hidden Markov Models (HMMs), Conditional Random Field (CRF) models, and Maximum Entropy models. Among these, neural network models have high recognition accuracy and strong ability to handle different fonts, sizes, and complex background text; HMMs are suitable for processing sequence data and can be used in applications such as speech recognition and handwriting recognition; CRF models are commonly used for tasks such as text classification and part-of-speech tagging; and Maximum Entropy models are commonly used for tasks such as text segmentation and named entity recognition. This embodiment preferably uses an open-source neural network model to identify cabinet names.

[0097] In this embodiment, the uniqueness of the inspection cabinet name is utilized. An OCR model is used to identify the cabinet name, and a target detection model and an image classification model are combined to identify the status of secondary equipment such as cabinet switches and pressure plates. A cabinet database is automatically established to store the identified cabinet names and the status of secondary equipment such as cabinet switches and pressure plates. Using the cabinet name identified by the OCR model as an index, the historical status of the current cabinet is found in the cabinet database. Since the target detection results are unordered, the target detection results must be sorted by row and column before detecting status changes, so that the results are stored in the cabinet database in a fixed order. Then, by comparing the currently identified status data with the corresponding historical status data, the status change of the secondary equipment in the cabinet is identified. Finally, the row and column position of the changed target is output and an alarm is triggered.

[0098] Example 2

[0099] This embodiment provides an apparatus for implementing a substation cabinet equipment status detection method. The apparatus stores an executable computer program, which, when executed, performs the substation cabinet equipment status detection method as described above.

[0100] In actual operation, the working process of this device includes:

[0101] ① Capture images of the target sample;

[0102] ② Perform horizontal correction on the image;

[0103] ③ Call the object detection model to obtain the location of the cabinet name in the image and crop it;

[0104] ④ Use the OCR model to identify the name of the cropped screen cabinet;

[0105] ⑤ Obtain the position and status of secondary equipment in the cabinet, such as switches and pressure plates, based on the target detection model;

[0106] ⑥ Call the image classification model to determine if there are any missed detections of the target;

[0107] ⑦ Merge the identified cabinet name and secondary equipment location status results, and choose to save the result as a template or perform change detection; if you choose to save the template, you will create information in the cabinet database using the cabinet name as the index; if you choose to perform change detection, you will find the historical data of equipment location status stored in the cabinet database using the cabinet name as the index for comparison; generally, if the cabinet is being photographed for the first time, you will choose to save it as a template.

[0108] ⑧ Write the identified secondary device location status information into the current index;

[0109] ⑨ Sort the historical data and current data by position, and compare the device status for each position to see if they are consistent;

[0110] ⑩ Output results include information such as the name of the identified cabinet, the location status of the secondary equipment, and the areas where changes have occurred.

[0111] The above functions can be packaged into a mobile app (such as a mobile phone). Maintenance personnel can use handheld mobile devices to take photos for recognition, and the recognition results will automatically generate electronic reports, which will greatly reduce the workload of personnel and manual operation, and significantly improve the efficiency of inspection work.

[0112] Figure 7 This is a demonstration diagram illustrating the status identification effect of substation cabinet equipment. Combined with... Figure 7 As shown, in actual operation: First, take a picture with the name of the screen cabinet as prompted on the page, such as... Figure 7 As shown in the left image, clicking "Recognize Cabinet Name" automatically uses an OCR algorithm to identify the current cabinet name. The page then prompts for image collection. After collection, an object detection algorithm is used to identify the position and status of secondary equipment such as switches, pressure plates, and indicator lights within the cabinet. The backend searches the cabinet database using the cabinet name as an index. If the image does not exist, it is created as a comparison image. If the image exists, the page prompts whether to perform a status comparison. If a status comparison is selected, as shown... Figure 7 As shown in the right figure, the device status in the currently captured image is compared with that in the historical comparison images. The changed areas are visualized in the "Recognition Results" and accompanied by text descriptions indicating what kind of status change has occurred in a certain device in a certain row and column.

[0113] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A method for detecting the status of substation switchgear equipment, characterized in that, The method includes: S1: Obtain images of secondary equipment in the substation cabinet to construct a sample dataset, label the sample dataset, and divide the labeled sample dataset into a training set and a validation set; S2: Construct an object detection model and an image classification model, and train the object detection model and the image classification model respectively using the training set; S3: Use the trained target detection model to identify the status of the secondary equipment in the cabinet, and use the trained image classification model to detect whether the target detection model has missed any targets. At the same time, combine the OCR model to identify the cabinet name, use this as an index to compare the status of the secondary equipment in the cabinet, and mark the difference areas in the final comparison. In step S2, an object detection model is constructed based on an improved YOLOv5 network and combined with the CBAM attention mechanism; wherein, the improved YOLOv5 network adds an upsampling branch to the output head of the initial YOLOv5 network to increase the feature map output after upsampling, specifically including: Extract the first feature map from the initial YOLOv5 network after 4x downsampling; The feature map in the initial YOLOv5 network, which was downsampled by 8 times, is upsampled by 1 times to obtain the second feature map. The first feature map and the second feature map are fused to obtain a new feature map, and the size of the new feature map is larger than the size of the first feature map and the second feature map. In step S3, the trained image classification model is used to detect whether the target detection model has missed any targets. Specifically, this includes: The distribution of targets in the image to be identified is detected and targets with abnormal distribution are found. The abnormally distributed targets are those in the cabinet whose interval between any two adjacent secondary devices is greater than the average interval between all secondary devices. If the interval between any two secondary devices is greater than the average interval between all secondary devices, then these two secondary devices are determined to be abnormally distributed targets. Centered on the abnormally distributed target, images of the four regions above, below, left, and right of the target are cropped and fed into the trained image classification model. The output of the image classification model is used to determine whether the target detection model has missed any targets.

2. The substation panel equipment status detection method according to claim 1, characterized in that, The CBAM attention mechanism includes channel attention graphs. Spatial attention map ,in: (1) (2) In equations (1) and (2): and These represent the average pooling feature and the max pooling feature, respectively. This represents a convolution operation with a filter size of 7×7; and, in equation (1) it represents and Forwarded to a multilayer perceptron In the shared network formed, after To generate channel attention maps In equation (2), it is indicated that... and Aggregate to generate two 2D graphs, then... Generate spatial attention map .

3. The substation panel equipment status detection method according to claim 1, characterized in that, In step S2, before training the object detection model using the training set, the K-means clustering algorithm is first used to cluster anchor boxes in the training set, and the obtained anchor boxes are then used to train the object detection model; wherein, the K-means clustering algorithm is used to cluster anchor boxes in the training set, specifically including: S21: Extract the coordinate values ​​of all bounding boxes for each image in the training set, and convert the coordinate values ​​of all bounding boxes into the height and width values ​​of the bounding boxes, wherein the bounding boxes are obtained by labeling the sample dataset; S22: Randomly select M coordinates as the center points of the Kmeans clustering algorithm, calculate the intersection-union ratio (IUR) between each bounding box of each image in the training set and each center point, and assign each bounding box to the cluster to which the center point with the largest IUR belongs; S23: Update the center point of each cluster based on the average coordinates of all bounding boxes in each cluster; S24: Repeat steps S22~S23 until the center point of each cluster is stable or the preset maximum number of iterations is reached, then stop the operation and obtain M anchor boxes for training the target detection model.

4. The substation panel equipment status detection method according to claim 1, characterized in that, The image classification model constructed in step S2 includes a 5*5 convolutional layer, 7 groups of depthwise separable convolutional modules, and a fully connected layer. Each group of depthwise separable convolutional modules includes a 3*3 depthwise convolutional layer, a 1*1 point convolutional layer, and a 2*2 pooling layer.

5. The substation panel equipment status detection method according to claim 4, characterized in that, In step S2, the image classification model is trained using the training set, specifically including: The training set is input into the image classification model, and the number of channels in the input feature map is N; Perform convolution operations on each of the N channels using a single convolution kernel to obtain a feature map with N channels equal to 1. The N feature maps with 1 channel are concatenated sequentially to obtain an output feature map with N channels. The output feature maps with N channels are weighted and combined in the depth direction to obtain a new high-dimensional feature map. The new high-dimensional feature map is output using the Softmax activation function, i.e.: (3) In equation (3): y represents the category of the current sample, and e represents the base of the natural logarithm. This represents the feature value extracted by the network that belongs to the i-th class. is the feature value of the current class, and C represents the total number of classes.

6. The substation panel equipment status detection method according to claim 5, characterized in that, In step S2, the image classification model is trained using the training set. Specifically, this includes: using Softmax cross-entropy loss and iteratively applying gradient descent to gradually decrease the value of the Softmax cross-entropy loss until convergence, thus completing model training. (4) In equation (4): This represents the Softmax cross-entropy loss.

7. The substation panel equipment status detection method according to claim 1, characterized in that, In step S3, the trained target detection model is used to identify the status of the secondary equipment in the cabinet, specifically including: First, the input image of the secondary equipment of the cabinet is horizontally corrected to obtain the corrected image. The correction operation includes extracting straight lines from the image and removing straight lines that are close to vertical. Then, the average tilt angle of the remaining straight lines is calculated, and the image is rotated and corrected according to the average tilt angle. The straight lines in the image are obtained through Hough line detection, and these straight lines refer to the edge lines on the cabinet. The target detection model then performs secondary equipment status recognition on the corrected image.

8. An apparatus for implementing a method for detecting the status of substation cabinet equipment, characterized in that, The device stores an executable computer program, which, when executed, performs the substation cabinet equipment status detection method as described in any one of claims 1-7.

Images