An R-FCN Disconnector Detection Method Combining Local Features and Global Features

Abstract

The present invention provides an R-FCN disconnect switch detection method that combines local features and global features. This method is based on the deep learning network model R-FCN to detect the status of disconnect switches in a substation. By embedding network camera centralized control software in the substation auxiliary monitoring system, images of disconnect switches at multiple positions and angles are collected in different outdoor weather conditions and backgrounds, constructing diverse disconnect switch data. By integrating a global feature prediction module in parallel with the R-FCN output prediction network, the defect of insufficient receptive field generated by the original network that only predicts disconnect switches through local features is supplemented, the detection accuracy of partially occluded disconnect switches is improved, and the missed detection rate and false detection rate of disconnect switches in complex backgrounds are reduced. By regularizing the local feature prediction results and the global feature prediction results and then accumulating the prediction results, the global feature prediction results are used to supplement the local feature prediction results, meeting the requirements of remote detection and unmanned operation of substation disconnect switches.

Description

Technical Field

[0001] The present invention relates to the technical field of disconnector detection in a substation auxiliary monitoring system, and relates to an R-FCN disconnector detection method that combines local features and global features. Background Technique

[0002] A substation is an important transfer site in high-voltage power transmission, and various crucial on-off control devices are operating in the station. In the past, when a substation experienced an interruption in work, relevant operators needed to check each device one by one to find the fault. The length of the fault troubleshooting time determined when the substation would return to normal operation, seriously affecting the daily life of residents. As an important power control switch in the substation, the opening and closing state of the disconnector directly determines whether the overall electrified circuit is operating. For the traditional troubleshooting of the power outage state in the substation, operators first need to enter the high-voltage area of the substation to check whether the disconnector is disconnected, and then check other connected devices one by one, which is prone to electric shock danger. With the development demand of unmanned substation operation, researching the automatic detection of the disconnector state in the captured images will help quickly identify the current state of the disconnector, shorten the fault detection efficiency of the substation, ensure the safety of operators' lives, and effectively improve the safe operation of the power grid.

[0003] Regarding the automatic detection of disconnectors, the problems often encountered currently are as follows: The working environment of the disconnector is outdoors, and there are various devices made of metal that are the same as the disconnector, with similar colors and are not easy to distinguish; affected by the weather environment, the recognition difficulty of disconnectors in the same position is different in different environments; due to the existence of various voltage control devices, wires and the long strip rectangular shape of the disconnector outdoors, when deploying an auxiliary monitoring system in the substation, the captured disconnectors are often blocked by other target devices or the entire features cannot be completely captured.

[0004] For the switch detection method, it is divided into two types: 1) The method based on image processing; 2) The method based on deep learning. Affected by the switch data collection, most current methods adopt the method based on image processing. First, a switch template is established for the standard images collected at a fixed angular position, and then image processing algorithms such as feature point matching algorithms or template matching algorithms are used for positioning and status recognition. This method can obtain a high monitoring accuracy in a good weather environment in a short time. However, with the frequent change of weather conditions, the camera position will shift based on the angle of the pre-established template image, and thus the status of the switch cannot be recognized. Therefore, it is necessary to often manually check and adjust the camera position or re-establish the template. For the method based on deep learning, it needs to rely on a large number of data sets with different angles and different weathers. The lack of data sets will limit the progress of related work. In addition, most current deep learning network models predict through the global receptive field, such as: Faster-Rcnn series network models, Yolo series network models, etc. When detecting the switch, if the switch is blocked or the acquisition object is incomplete, it will lead to the lack of the overall feature information of the switch, resulting in missed detection or false detection; some deep learning network models predict through local features, such as: R-FCN network model, VIT network model, etc. When detecting the switch, although it can effectively detect the incomplete switch, when the switch fills the entire image, local feature prediction will also have missed detection situations. Summary of the Invention

[0005] Aiming at the above technical problems, the object of the present invention is: based on the R-FCN network model, to propose a region-based fully convolutional object detection network R-FCN (R-FCN: Object Detection via Region-based Fully Convolutional Networks) switch detection method that combines local features and global features, embed global feature prediction into the R-FCN network model, and improve the accuracy of switch detection by combining the advantages of local feature and global feature prediction, and improve the accuracy of the existing switch detection method.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is: An R-FCN switch detection that combines local features and global features, including the following steps:

[0007] Step 1: Build a substation auxiliary monitoring system to collect switch images;

[0008] Step 2: Divide, clean and label the image data set;

[0009] Step 3: Select the R-FCN backbone feature extraction network;

[0010] Step 4: Adjust the R-FCN backbone feature extraction network;

[0011] Step 5: Construct a local feature prediction branch;

[0012] Step 6: Construct a global feature prediction branch;

[0013] Step 7: Integrate the local prediction result and the global prediction result;

[0014] Step 8: Train and save the model.

[0015] Furthermore, the substation auxiliary monitoring system in Step 1 is an integration of all network monitoring cameras and related control devices in the substation. By adjusting the angles of each camera, an image dataset including images with and without disconnectors at different times, weather conditions, and backgrounds is collected, and the number of image datasets with disconnectors should evenly include two states of the disconnector switch.

[0016] Furthermore, the division of the dataset in Step 2 means dividing the collected images into images with disconnectors and images without disconnectors, and dividing the dataset of images with disconnectors into a training set and a test set with a ratio of 8:2; cleaning the dataset means removing the blurred datasets in the collected dataset; only the images containing disconnector objects are labeled.

[0017] Furthermore, in Step 3, the R-FCN backbone feature extraction network is selected, and the implementation method is as follows: The classification network ResNet101 is used as the backbone feature extraction network. After the fourth group of convolutional layers of ResNet101, the Region Proposal Network (RPN) operation is used to generate regions of interest (for the later detection of disconnectors), and the pooling layer and fully connected layer after the fifth group of convolutional layers of ResNet101 are discarded. The final output number of channels is 2048. The region of interest is the region where the target exists.

[0018] Furthermore, in Step 4, the backbone feature extraction network of R-FCN is adjusted, and the implementation method is as follows: After selecting the trimmed backbone feature extraction network ResNet101 in Step 3, a convolutional layer with a kernel size of 1×1 is added to reduce the number of channels to 1024, so as to reduce the dimension of the feature data without changing the size of the feature map and improve the calculation speed.

[0019] Furthermore, in Step 5, the local feature prediction branch is constructed, and the implementation method is as follows: The original network model R-FCN with local feature prediction is used to output the local prediction result, and the local feature prediction uses the RPN proposed regions divided into 7×7 local regions for prediction.

[0020] Furthermore, in step 6, a global feature prediction branch is constructed, and the implementation method is as follows: perform a pooling operation on the extracted semantic features to unify the size of the extracted semantic features; then, concatenate and use convolutional layers with a convolutional kernel size of 7×7 and 1×1 to output the global prediction result.

[0021] Furthermore, in step 7, the local prediction result and the global prediction result are fused. For the output prediction results in steps 5 and 6, L2 regularization is to be used, and the mathematical expression is: In the formula, x and y are vectors, x = (x0, x1, x2... x n ), representing the prediction output result, y = (y0, y1, y2... y n ), representing the prediction output result after regularization. Finally, the numerical values of the two prediction results are uniformly scaled to the interval from 0 to 1, and the two prediction output results are vector-added, and then a Softmax operation is performed to output the final prediction result. The mathematical expression of Softmax is: R i is the i-th output value in the output result vector.

[0022] Furthermore, in step 8, the model is trained. The training is carried out twice using the two divided data sets: the first training only uses the clear data set containing disconnectors, and the second training uses the fuzzy data set with a ratio of 1:1 for the data quantity of the data set containing disconnectors and the data set not containing disconnectors but similar to disconnectors.

[0023] Furthermore, in step 8, when training the model, the mathematical expression of the loss function used for training is: Among them, L(s, t) represents the total loss of the classification loss and the regression loss, s is the class prediction probability, represents the prediction probability of class c * , t represents the regression box predicted by the model, and L cls (s) is the classification loss, and the mathematical expression is: λ is the multi-task balance factor, c * is the true label of the class. When c * = 0, it represents the background class, and when c * ≠0, it represents the corresponding class. [c * > 0] constitutes the guiding factor, and its value is taken as 1, indicating that the regression box only adjusts the non-background class objects. L reg (t, t * ) is the regression loss, t represents the target box predicted by the model, and t * represents the true box manually annotated.

[0024] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: For the detection of disconnect switches with partial feature occlusion and large features, the R-FCN network based on combined local and global features can perform prediction and judgment compared with the prediction based on original local features. The proposed R-FCN disconnect switch detection method based on combined local and global features effectively reduces the false detection rate and missed detection rate of disconnect switch detection in complex substation scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] To more clearly illustrate the technical solution of the present invention, the following further describes the drawings required in the embodiments.

[0026] Figure 1 It is a schematic diagram of the operation process of the R-FCN disconnect switch detection method based on combined local and global features of the present invention.

[0027] Figure 2 It is a schematic diagram of the module of the R-FCN model based on combined local and global features of the present invention.

[0028] Figure 3 It is a schematic diagram of the module of local feature prediction of the present invention.

[0029] Figure 4 It is a schematic diagram of the module of global feature prediction of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0030] The following further details the implementation scheme of the present invention with reference to the attached Figures 1-4 drawings.

[0031] As Figure 1 shown, a R-FCN disconnect switch detection method based on combined local and global features comprises the following steps:

[0032] Step 1: When the substation auxiliary monitoring system collects disconnect switch images, collect them from natural factors such as different angles, different fields of view, and different backgrounds to ensure the diversity of the collected pictures.

[0033] Step 2: Clean the data set collected in Step 1, remove damaged image data, and divide it into two types: an image data set containing only disconnect switches and a data set not containing disconnect switches but similar to disconnect switches. The total number of images containing disconnect switches in the data set collected by the present invention is 800, and 800 images not containing disconnect switches but similar to disconnect switches. Divide the two types respectively, where the ratio of the training set to the test set is 8:2, and label the position and status of the disconnect switches in the data set containing only disconnect switch images. For the first training, the number of samples used is shown in Table 1, and for the second training, the number of samples used is shown in Table 2.

[0034] Step 3: Adjust the R-FCN backbone feature extraction network: Use the classification network ResNet101 as the backbone feature extraction network. After the fourth group of convolutional layers of ResNet101, use the RPN operation to generate regions of interest (for the later detection of disconnect switches), discard the pooling layer and fully connected layer after the fifth group of convolutional layers of ResNet101, and the final number of output channels is 2048, as Figure 2 shown.

[0035] Step 4: After selecting the trimmed backbone feature extraction network ResNet101 in Step 3, add a convolutional layer with a kernel size of 1×1 to reduce the number of channels to 1024, so as to reduce the dimensionality of the feature data without changing the size of the feature map, as Figure 2 shown.

[0036] Step 5: Construct the local feature prediction branch: Use the original network model R-FCN with local feature prediction to output the local prediction results, and the local feature prediction is divided into 7×7 local regions by the RPN proposed regions for prediction, as Figure 3 shown.

[0037] Step 6: Construct the global feature prediction branch: Perform a pooling operation on the extracted semantic features to unify the size of the extracted semantic features; then, concatenate convolutional layers with kernel sizes of 7×7 and 1×1 to output the global prediction results, as Figure 4 shown.

[0038] Step 7: Fuse the local prediction results and the global prediction results: Regularize the local prediction results and the global prediction results, uniformly scale them to the same numerical interval and add them to complete the information fusion prediction.

[0039] Furthermore, the mathematical expression is: In the formula, x and y are vectors, x = (x0, x1, x2... x n ), represents the prediction output result, y = (y0, y1, y2... y n ), represents the prediction output result after regularization. Finally, the two prediction results are uniformly scaled to the interval from 0 to 1, and the two prediction output results are added vectorially, and then the Softmax operation is performed to output the final prediction result. The mathematical expression of Softmax is: R i is the i-th output value in the output result vector.

[0040] Step 8: Training the model: When training, the initial parameters of the ResNet101 in the backbone feature extraction network of the model use the model weights trained on the ImageNet dataset. The loss function used for model training selects the same loss function as the R-FCN network, and two trainings are carried out: The first time is to train using only the data containing switch-disconnector images until the loss function converges, and save the network model; The second time is to train based on the model parameters of the first training, using the data containing switch-disconnectors and data similar to switch-disconnectors without switch-disconnectors to further improve the switch-disconnector discrimination ability of the feature network model. Finally, the test results of the model are shown in Table 3.

[0041] Furthermore, in Step 8 for training the model, the mathematical expression of the loss function used for training is: where \(L(s, t)\) represents the total loss of the classification loss and the regression loss, \(s\) is the class prediction probability, represents the prediction probability of class \(c\), * \(t\) represents the regression box predicted by the model, \(L\) cls (s) is the classification loss, and its mathematical expression is: \(\lambda\) is the multi-task balance factor, \(c\) * is the true label of the class. When \(c\) * = 0, it represents the background class, and when \(c\) * ≠ 0, it represents the corresponding class. \([c * > 0]\) constitutes the guiding factor, and its value is taken as 1, indicating that the regression box only adjusts the non-background class objects. \(L\) reg (t, t * ) is the regression loss, \(t\) represents the target box predicted by the model, and \(t * represents the true box manually labeled.

[0042] Table 1 is the statistical result of the sample quantity of the first training set.

[0043]

[0044] Table 2 is the statistical result of the sample quantity of the second training set.

[0045]

[0046] Table 3 is the comparison result of the R-FCN switch-disconnector detection accuracy combining local features and global features.

[0047]

[0048] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting disconnectors using R-FCN by combining local features and global features, characterized in that, The following steps are involved: Step 1: Build a substation auxiliary monitoring system to collect images of the target from different camera angles and weather conditions at different times; Step 2: Divide the images collected in step 1 into training set images and test set images, and clean and annotate the image data set; Step 3: Build an R-FCN knife switch detection model with joint local features and global features: Adjust the R-FCN backbone feature extraction network: Use the classification network ResNet101 as the backbone feature extraction network, use the region proposal network RPN operation after the fourth group of convolutional layers of ResNet101 to generate the region of interest, and discard the pooling layer and fully connected layer after the fifth group of convolutional layers of ResNet101; Step 4: Based on the backbone feature extraction network adjusted in step 3, the number of output channels is 2048, and a convolution layer with a convolution kernel size of 1×1 is added to it to reduce the number of channels to 1024; Step 5: Construct a local feature prediction branch: Use the original network model R-FCN with local feature prediction to output the local prediction results, and the local feature prediction uses the RPN recommended area to be divided into 7×7 local areas for prediction; Step 6: Construct a global feature prediction branch: The global feature prediction is based on the semantic features extracted in step 4. First, the extracted semantic features are pooled to unify the size of the extracted semantic features. Then, convolutional layers with kernel sizes of 7×7 and 1×1 are used in series to output the global prediction results. Step 7: Fusion of local prediction results and global prediction results: Regularization of local prediction results and global prediction results, uniform scaling to the same numerical range and addition, to complete information fusion prediction; Step 8: Training model: The loss function used for model training is the same as that of the R-FCN network to guide the optimization of model parameters; the network training parameters are updated until the loss function converges and the network model is saved.

2. The R-FCN knife switch detection method combining local features and global features according to claim 1, wherein: The auxiliary monitoring system in step 1 is composed of a combination of multiple network cameras to implement comprehensive coverage and real-time monitoring of all angles of the substation. The auxiliary monitoring system is embedded with a camera centralized control program, and realizes data collection from different angles, different backgrounds, and different weather factors through multiple network cameras.

3. The R-FCN disconnector detection method combining local features and global features according to claim 1, characterized in that: The cleaning data set in step 2 specifically includes: removing blurred images caused by camera shake, classifying images containing knife switches into one data set, and classifying images that do not contain knife switches but have knife switch-like images into one data set.

4. A method for detecting a knife switch using R-FCN that combines local features and global features according to claim 1, characterized in that: The local feature prediction in step 5 is to divide the RPN suggestion area into 7×7 local areas for prediction.

5. A method for detecting knife switches by combining local features and global features in R-FCN according to claim 1, characterized in that: The fusion of the local prediction results and the global prediction results in step 7 is specifically: regularizing the local prediction results and the global prediction results respectively using a regularization method, and using L2 regularization, the mathematical expression is: where x and y are vectors, x = (x0, x1, x2... x n ), y = (y0, y1, y2... y n ), representing the predicted output result. Or, finally, the numerical values of the two predicted results are uniformly scaled to the interval from 0 to 1.

6. The R-FCN switch-disconnector detection method combining local features and global features according to claim 5, characterized in that The fusion method used is: adding vectors of the same dimension.

7. A method for detecting knife switches of R-FCN by combining local features and global features according to claim 1, characterized in that: In the step 8, the model is trained twice using two divided data sets: for the first training, only the clear data set containing disconnect switches is used; for the second training, a fuzzy data set with a ratio of 1:1 of the number of data containing disconnect switches to the number of data not containing disconnect switches but similar to disconnect switches is used.

8. A method for detecting knife switches using R-FCN that combines local features and global features according to claim 7, characterized in that: The specific training model is as follows: when training, the idea of transfer learning is used, and the model weights trained on the ImageNet data set are adopted for the backbone feature extraction network; the clear data set refers to that all training images contain at least one disconnect switch image, and the fuzzy data set refers to that the images containing disconnect switch images and the images not containing disconnect switches but similar to disconnect switches each account for 50%.

9. A method for detecting knife switches of R-FCN combining local features and global features according to claim 7, characterized in that: The loss function mathematical expression used for training the model is: Among them, L(s, t) represents the total loss of classification loss and regression loss, s is the class prediction probability, represents the prediction probability of class c * of, t represents the regression box predicted by the model, L cls (s) is the classification loss, and its mathematical expression is: λ is the multi-task balance factor, c * is the true class label. When c * = 0 represents the background class, c * ≠ 0 represents the corresponding pair, [c * > 0] constitutes the guiding factor, and its value is taken as 1, indicating that the regression box only adjusts non-background class objects. L reg (t, t * ) is the regression loss, t represents the target box predicted by the model, t * represents the true box manually annotated.

Images