A method and system for detecting abnormality of a box-type substation

Abstract

The application discloses a kind of box-type substation anomaly detection method and system, belong to the technical field of computer vision.The present application comprehensively traditional method and deep learning method, in combination with target detection and salient region detection, coarse-grained positioning of abnormal region is carried out, then potential abnormal region is positioned in fine granularity using feature point matching technology, finally, potential abnormal region is filtered and confirmed in combination with salient region map, and finally the accurate identification of abnormal region is realized.The present application can realize the accurate monitoring of the state of box-type substation, greatly reduce the false alarm problem caused by environmental change in traditional method, can be applied to power inspection task, save the human consumption in power inspection, improve the efficiency of inspection.

Description

Technical Field

[0001] This invention discloses a method and system for detecting anomalies in a prefabricated substation, belonging to the technical field of computer vision. Background Technology

[0002] A prefabricated substation, or simply prefabricated substation, is a compact, factory-prefabricated indoor or outdoor power distribution unit that integrates high-voltage switchgear, distribution transformers, and low-voltage power distribution equipment according to a specific wiring scheme. It organically combines transformer voltage reduction and low-voltage power distribution functions, all housed in a moisture-proof, rust-proof, dust-proof, rodent-proof, fire-proof, theft-proof, heat-insulated, fully enclosed, and movable steel structure box. It is particularly suitable for urban power grid construction and renovation, and represents a new type of substation that has emerged after civil engineering substations. However, in recent years, safety incidents such as personnel and animals entering prefabricated substations and maintenance personnel forgetting to close the box doors have frequently occurred. Therefore, monitoring the status of prefabricated substations is a pressing technical problem that needs to be solved.

[0003] Current image monitoring technologies largely rely on traditional image comparison methods, such as background subtraction and template matching. While these methods are fast and easy to deploy, they also have several problems: First, the recognition rate is low, making it easy to miss or misdetect. Second, the algorithms are not robust enough; traditional methods rely on manually set features, making them poorly adaptable to environmental changes and unable to cope with various abnormal situations, thus often failing to achieve the desired results. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes an anomaly detection method for box-type substations that combines target detection with salient region feature point matching.

[0005] The present invention also discloses a system for implementing the above-described detection method. Invention Overview:

[0007] An anomaly detection method and system for prefabricated substations integrates traditional and deep learning methods. It combines target detection and salient region detection for coarse-grained localization of anomaly areas, then uses feature point matching technology for fine-grained localization of potential anomaly areas, and finally combines a salient region map to filter and confirm potential anomaly areas, thereby achieving accurate identification of anomaly areas.

[0008] This invention enables accurate monitoring of the status of prefabricated substations, significantly reducing false alarms caused by environmental changes in traditional methods. It can be applied to power inspection tasks, saving manpower and improving inspection efficiency.

[0009] The detailed technical solution of this invention is as follows:

[0010] A method for detecting anomalies in a prefabricated substation, characterized by comprising:

[0011] 1) Collect data on the scene to be inspected and mark the transformer substation areas in the scene data;

[0012] 2) Establish a detection model for the transformer substation area:

[0013] The substation detection model is trained using labeled scene data to obtain a trained substation area detection model.

[0014] The image to be tested is input into the transformer substation area detection model to obtain the coordinates of the transformer substation area in the image to be tested. The transformer substation area is cropped and saved as a template image according to the coordinates of the transformer substation area. Preferably, the transformer substation detection model is a target detection framework model, which adopts any one of the first or second order target detection networks of YOLOV3 / V4 / V5, SSD, Faster R-CNN, and Cascade R-CNN.

[0015] Obtain the backbone network based on ImageNet pretraining, taking ResNet-18 as an example;

[0016] After removing the fully connected layers from ResNet-18, we obtain the feature map before classification, which is the output feature map of layer 4 here;

[0017] The tensors of the feature map are summed by channel-weighted features, and the summation result is normalized.

[0018] The normalized feature map is resized to the same size as the image under test using interpolation. This feature map is the salient region distribution map of the image under test.

[0019] 3) Calculate the homography matrix:

[0020] SIFT / ORB feature points are extracted from the subsequent inspection images and the template images. Then, the feature points of the template images are matched with the feature points of the inspection images, and the homography matrix is ​​calculated based on the matched feature points.

[0021] 4) Calculate the vector Euclidean distance between the template image monitoring points and the inspection image monitoring points:

[0022] Monitoring points are set within the template image, and homography transformation is performed using the homography matrix to map the monitoring points of the template image to the corresponding positions in the inspection image;

[0023] Extract the position and orientation attribute vectors of the monitoring points in the template image and the inspection image, and calculate the Euclidean distance between the corresponding monitoring point vectors in the two images;

[0024] 5) Abnormal detection:

[0025] A preset abnormal threshold for the prefabricated substation is established, and abnormalities are determined based on this threshold.

[0026] When the vector Euclidean distance between the monitoring points corresponding to the template image and the inspection image is greater than the threshold, the monitoring point in the inspection image is determined to be a potential anomaly.

[0027] When the vector Euclidean distance between the monitoring points corresponding to the template image and the inspection image is less than or equal to the threshold, the monitoring point in the inspection image is determined to be a normal point.

[0028] According to a preferred embodiment of the present invention, in step 5) detecting an anomaly, the method further includes:

[0029] Calculate the relationship between potential outliers and salient regions:

[0030] If a potential outlier exceeds the salient region, it is discarded and will no longer be considered an outlier.

[0031] Otherwise, treat potential outliers as outliers.

[0032] According to a preferred embodiment of the present invention, in step 5) detecting an anomaly, the method further includes:

[0033] Perform K-means clustering on the anomalies, visualize the anomaly regions, and issue alerts.

[0034] A system for detecting anomalies in prefabricated substations is characterized by comprising: a data acquisition module, a prefabricated substation area coordinate acquisition module, a significant area acquisition module, an anomaly capture module, and an anomaly confirmation module;

[0035] The data acquisition module is used to perform step 1) and is configured to: acquire scene data and mark the transformer substation area in the scene data;

[0036] The transformer substation area coordinate acquisition module is used to perform step 2) of inputting the image to be tested into the transformer substation detection model to obtain the transformer substation area coordinates of the image to be tested. It is configured to: establish a transformer substation detection model, train the transformer substation detection model with labeled scene data to obtain the trained transformer substation detection model; input the image to be tested into the transformer substation detection model to obtain the transformer substation area coordinates of the image to be tested.

[0037] The salient region acquisition module, used to perform step 2) to obtain the distribution of salient regions in the image, is configured to: use a pre-trained deep learning neural network to perform feature weighted summation of the tensor of the extracted high-dimensional information feature map by channel, and after normalization of the summation result, resize it to the same size as the input image by interpolation. This feature map represents the distribution of salient regions in the image.

[0038] The anomaly capture module, used to execute steps 3), 4), and 5), is configured to: save a template image based on the transformer substation area, extract feature points from the template image and the inspection image, obtain the homography matrix, set image monitoring points in the template image, map the position of the monitoring points in the template image to the inspection image through homography transformation, further calculate the feature distance between the two image monitoring points, and if it exceeds the threshold, it is an anomaly point.

[0039] According to a preferred embodiment of the present invention, the system further includes an anomaly confirmation module, which is used to determine whether a potential anomaly is an anomaly by calculating the relationship between the potential anomaly and the salient region. The module is configured to: if the anomaly exceeds the salient region by calculating the relationship between the anomaly and the salient region, the anomaly is considered to be a misjudgment caused by background interference, and the point is discarded.

[0040] According to a preferred embodiment of the present invention, the anomaly confirmation module further includes: performing K-means clustering on the anomaly points, visualizing the anomaly region and issuing an alarm.

[0041] The present invention has the following beneficial effects:

[0042] 1. This invention constructs a detection dataset for prefabricated substations. It utilizes target detection algorithms to locate the substation positions in any monitoring image. First, it uses a salient area map for coarse-grained localization of abnormal areas. Then, it employs a key feature matching algorithm to locate potential abnormal areas, and further confirms these areas by matching them with the salient area map. This achieves accurate monitoring of abnormal states in prefabricated substations and can be applied to power line inspection tasks, saving manpower and improving inspection efficiency.

[0043] 2. The application of target detection algorithm, salient region extraction algorithm and anomaly detection algorithm in this invention ensures the robustness of the system in actual use. The recognition result integrates traditional algorithm and deep learning algorithm. By complementing the advantages of traditional method and deep learning method, and combining secondary confirmation of salient region map, accurate recognition from coarse granularity to fine granularity is achieved, which greatly reduces false alarms.

[0044] 3. The anomaly detection method for prefabricated substations that combines target detection and salient area feature point matching provided by this invention is the first to apply detection technology and feature point matching technology to anomaly detection in prefabricated substations. It is innovative, and its effectiveness has been verified by experiments and applications in real-world scenarios. It can be widely promoted. Attached Figure Description

[0045] Figure 1 This is a flowchart of the detection method described in this invention;

[0046] Figure 2 This is a schematic diagram of the area detection network of the box-type substation according to Embodiment 1 of the present invention;

[0047] Figure 3 This is a demonstration diagram of the detection system of Embodiment 1 of the present invention. Detailed implementation method:

[0048] The present invention will be further described below with reference to the accompanying drawings and embodiments, but is not limited thereto.

[0049] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0050] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should 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.

[0051] The purpose of this invention is to address the challenges encountered in anomaly detection in prefabricated substation scenarios. It introduces a target detection algorithm to determine the substation area, designs a salient region extraction algorithm to obtain coarse-grained anomaly regions, and processes and analyzes visual data within the target scene using key point feature extraction technology. By defining monitoring points and calculating feature distances between them, potential anomaly regions are identified. Furthermore, the relationship between potential anomaly regions and salient regions is determined, discarding anomalies outside the salient regions to achieve fine-grained anomaly region localization and improve recognition accuracy. First, a target detection algorithm is needed to locate the target area during use. Then, a salient region extraction algorithm is designed to obtain coarse-grained anomaly regions. Next, a key feature matching algorithm is designed to determine potential anomalies. Then, the results of potential and salient anomaly region matching are combined to achieve precise localization of actual anomalies. Finally, this invention encapsulates the algorithm into a prefabricated substation anomaly detection system for practical testing and deployment.

[0052] Example 1

[0053] A method for detecting anomalies in a prefabricated substation includes:

[0054] 1) Collect the scene data to be inspected and mark the transformer substation area in the scene data; further, the step of preprocessing the inspection data includes performing Gaussian filtering on the normal inspection data to reduce noise interference and obtain the image to be tested.

[0055] 2) Establish a detection model for the transformer substation area:

[0056] The substation detection model is trained using labeled scene data to obtain a trained substation area detection model.

[0057] Input the image to be tested into the transformer substation area detection model to obtain the coordinates of the transformer substation area in the image to be tested, and then crop and save the transformer substation area as a template image according to the coordinates of the transformer substation area.

[0058] The transformer substation detection model is a target detection framework model, which adopts any one of the first- or second-order target detection networks from YoloV3 / V4 / V5, SSD, Faster R-CNN, and Cascade R-CNN.

[0059] The classification network can be any one of AlexNet, ResNet, VGG, GoogLeNet, or MobileNet. For example... Figure 2 As shown, the transformer substation anomaly detection system uses outdoor transformer substation monitoring data as input. Then, a target detection algorithm is used to locate the transformer substation in the monitoring data, narrowing the detection range. Next, by modifying the classification network structure, removing the fully connected layers, and superimposing and normalizing the outputs of the convolutional layers by channel, a salient region map is obtained. This salient region map is then used for coarse-grained localization of anomaly areas.

[0060] That is, to obtain the backbone network based on ImageNet pretraining, taking ResNet-18 as an example;

[0061] After removing the fully connected layers from ResNet-18, we obtain the feature map before classification, which is the output feature map of layer 4 here;

[0062] The tensors of the feature map are summed by channel-weighted features, and the summation result is normalized.

[0063] The normalized feature map is resized to the same size as the image under test using interpolation. This feature map is the salient region distribution map of the image under test.

[0064] 3) Calculate the homography matrix:

[0065] SIFT / ORB feature points are extracted from the subsequent inspection images and the template images. Then, the feature points of the template images are matched with the feature points of the inspection images, and the homography matrix is ​​calculated based on the matched feature points.

[0066] 4) Calculate the vector Euclidean distance between the template image monitoring points and the inspection image monitoring points:

[0067] Monitoring points are set within the template image, and homography transformation is performed using the homography matrix to map the monitoring points of the template image to the corresponding positions in the inspection image;

[0068] The position and orientation attribute vectors of the monitoring points in the template image and the inspection image are extracted, and the Euclidean distance between the corresponding monitoring points in the two images is calculated.

[0069] 5) Abnormal detection:

[0070] A preset abnormal threshold for the prefabricated substation is established, and abnormalities are determined based on this threshold.

[0071] When the vector Euclidean distance between the monitoring points corresponding to the template image and the inspection image is greater than the threshold, the monitoring point in the inspection image is determined to be a potential anomaly.

[0072] When the Euclidean distance between the corresponding monitoring points of the template image and the inspection image is less than or equal to the threshold, the inspection image is determined to be a normal point.

[0073] The detected transformer substation is used as a template image. Feature points are extracted from the template image and subsequent inspection images, and the homography transformation matrix H is calculated. Matrix H is used to map the monitoring points in the template image to those in the inspection image. Features of the monitoring points before and after are calculated; points with a feature distance exceeding a threshold are defined as potential anomalies. Then, a salient region map is used to filter potential anomalies, and secondary confirmation of anomalies through the salient region map greatly reduces false alarms caused by environmental interference. Finally, K-means clustering is performed on the anomalies to achieve fine-grained localization of anomaly regions. The combined application of the target detection algorithm, salient region detection algorithm, and keypoint detection algorithm ensures the robustness of the system in practical use.

[0074] Example 2

[0075] The detection method as described in Example 1, in step 5) detecting an abnormality, further includes:

[0076] Calculate the relationship between potential outliers and salient regions:

[0077] If a potential outlier exceeds the salient region, it is discarded and will no longer be considered an outlier.

[0078] Otherwise, treat potential outliers as outliers.

[0079] Example 3

[0080] The detection method as described in Example 2, in step 5) detecting an abnormality, further includes:

[0081] Perform K-means clustering on the anomalies, visualize the anomaly regions, and issue alerts.

[0082] The detection method described in Examples 1-3 above specifically includes the following steps:

[0083] like Figure 1 , 2As shown,

[0084] Step S0: Collect data on the inspection scene of the transformer substation, label the transformer substation area, and train the transformer substation detection model;

[0085] Step S1: Input the image to be tested into the transformer substation area detection model to obtain the transformer substation area coordinates of the image to be tested, and crop the transformer substation area according to the transformer substation area coordinates and save it as a template image;

[0086] Step S2: Modify the model structure of the deep neural network and extract the network layer output of the feature map to be visualized. For example, taking ResNet-18 as an example, this is a tensor of size 512*7*7. Obtain the weights of each channel of this tensor, which are vectors of length 512. Through linear fusion, weighted summation of this tensor along the channel dimension is performed to obtain a feature map of size 7*7. Normalize this feature map and resize it to the original size through interpolation. Figure 1 The size of the sample is used to obtain a salient region map;

[0087] Step S3: Extract key point features from the template image and the inspection image. Taking the SIFT feature extraction algorithm as an example, firstly, the original image is continuously reduced and Gaussian filtered to generate a difference-of-gaussian image pyramid. Then, by detecting the extreme points of the image between different layers, the Hessian matrix is ​​used for filtering, as shown in formula (I), to obtain the location of the feature points, where D... xx D xy D yy It is obtained by the difference between the corresponding positions in the neighborhood of the candidate point:

[0088]

[0089] After obtaining the location of the feature points, their amplitude m and direction θ are calculated, where L is the scale space value of the keypoint:

[0090]

[0091]

[0092] Finally, feature matching points are generated. The neighborhood of each feature point is divided into several blocks, and the gradient direction histograms in eight directions are calculated. There are 16 regions, so 16×8=128 dimensions of data are generated.

[0093] Step S4: By matching the feature point descriptors of the template image and the inspection image, find the matching feature point pairs in the two images, and calculate the homography transformation matrix H;

[0094]

[0095] Step S5: Densely define monitoring points in the template image;

[0096] Step S6: Use the homography matrix obtained in S4 to map the monitoring points of the template image to the inspection image;

[0097] Step S7: Calculate the Euclidean distance between the feature vectors of the corresponding monitoring points. If the distance exceeds the threshold T, it is defined as a potential anomaly.

[0098] Step S8: Calculate the relationship between potential outliers and the salient regions in S2. If an outlier exceeds the salient region, it is considered to be interfering data and is discarded. Points that remain within the salient region are considered outliers.

[0099] Step S9: Perform K-means clustering on the retained outliers to determine the number and specific locations of outlier regions;

[0100] Step S10: Encapsulate the entire algorithm into a transformer substation anomaly detection system.

[0101] Example 4

[0102] like Figure 3 As shown,

[0103] A system for detecting anomalies in a prefabricated substation as described in Embodiment 1, characterized in that it includes: a data acquisition module, a prefabricated substation area coordinate acquisition module, a significant area acquisition module, an anomaly capture module, and an anomaly confirmation module;

[0104] The data acquisition module is used to perform step 1) and is configured to: acquire scene data and mark the transformer substation area in the scene data;

[0105] The transformer substation area coordinate acquisition module is used to perform step 2) of inputting the image to be tested into the transformer substation detection model to obtain the transformer substation area coordinates of the image to be tested. It is configured to: establish a transformer substation detection model, train the transformer substation detection model with labeled scene data to obtain the trained transformer substation detection model; input the image to be tested into the transformer substation detection model to obtain the transformer substation area coordinates of the image to be tested.

[0106] The salient region acquisition module, used to perform step 2) to obtain the distribution of salient regions in the image, is configured to: use a pre-trained deep learning neural network to perform feature weighted summation of the tensor of the extracted high-dimensional information feature map by channel, and after normalization of the summation result, resize it to the same size as the input image by interpolation. This feature map represents the distribution of salient regions in the image.

[0107] The anomaly capture module, used to execute steps 3), 4), and 5), is configured to: save a template image based on the transformer substation area, extract feature points from the template image and the inspection image, obtain the homography matrix, set image monitoring points in the template image, map the position of the monitoring points in the template image to the inspection image through homography transformation, further calculate the feature distance between the two image monitoring points, and if it exceeds the threshold, it is an anomaly point.

[0108] Example 5

[0109] As described in Examples 4 and 2, the system further includes an anomaly confirmation module, which is used to determine whether a potential anomaly is an anomaly by calculating the relationship between the potential anomaly and the salient region. It is configured to: if the anomaly exceeds the salient region by calculating the relationship between the anomaly and the salient region, the anomaly is considered to be a misjudgment caused by background interference, and the point is discarded.

[0110] Example 6

[0111] The anomaly confirmation module of the detection system described in Example 5 further includes: performing K-means clustering on the anomaly points, visualizing the anomaly region and issuing an alarm.

[0112] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for detecting anomalies in a prefabricated substation, characterized in that, include: 1) Collect data on the scene to be inspected and mark the transformer substation areas in the scene data; 2) Establish a detection model for the transformer substation area: The substation detection model is trained using labeled scene data to obtain a trained substation area detection model. Input the image to be tested into the transformer substation area detection model to obtain the coordinates of the transformer substation area in the image to be tested, and then crop and save the transformer substation area as a template image according to the coordinates of the transformer substation area. Obtain the backbone network based on ImageNet pretraining; After removing the fully connected layer, the feature map before classification is obtained; The tensors of the feature map are summed by channel-weighted features, and the summation result is normalized. The normalized feature map is resized to the same size as the image under test using interpolation. This feature map is the salient region map of the image under test. 3) Calculate the homography matrix: SIFT / ORB feature points are extracted from the subsequent inspection images and the template images. Then, the feature points of the template images are matched with the feature points of the inspection images, and the homography matrix is ​​calculated based on the matched feature points. 4) Calculate the vector Euclidean distance between the template image monitoring points and the inspection image monitoring points: Monitoring points are set within the template image, and homography transformation is performed using the homography matrix to map the monitoring points of the template image to the corresponding positions in the inspection image; Calculate the Euclidean distance between the corresponding monitoring point vectors of the two images; 5) Abnormal detection: A preset abnormal threshold for the prefabricated substation is established, and abnormalities are determined based on this threshold. When the vector Euclidean distance between the template image monitoring point and the inspection image monitoring point is greater than the threshold, the inspection image monitoring point is determined to be a potential anomaly point. When the vector Euclidean distance between the template image monitoring point and the inspection image monitoring point is less than or equal to the threshold, the inspection image monitoring point is determined to be a normal point. Step 5) of detecting anomalies also includes: Calculate the relationship between potential outliers and salient regions: If a potential outlier exceeds the salient region, it is discarded and will no longer be considered an outlier. Otherwise, treat potential outliers as outliers.

2. The method for detecting anomalies in a prefabricated substation according to claim 1, characterized in that, The transformer substation detection model is a target detection framework model, which adopts any one of the first or second order target detection networks from YOLOv3 / V4 / V5, SSD, Faster R-CNN, and Cascade R-CNN.

3. The method for detecting anomalies in a prefabricated substation according to claim 1, characterized in that, Step 5) of detecting anomalies also includes: Perform K-means clustering on the anomalies, visualize the anomaly regions, and issue alerts.

4. A detection system for implementing the method of any one of claims 1-3, characterized in that, include: The system includes a data acquisition module, a transformer substation area coordinate acquisition module, a significant area acquisition module, an anomaly capture module, and an anomaly confirmation module. The data acquisition module is used to perform step 1). The transformer substation area coordinate acquisition module is used to perform step 2) of inputting the image to be tested into the transformer substation detection model to obtain the transformer substation area coordinates of the image to be tested; The salient region acquisition module is used to perform step 2) to obtain the distribution of salient regions in the image; The exception capture module is used to execute steps 3), 4), and 5).

5. The detection system according to claim 4, the system further includes an anomaly confirmation module, used to determine whether a potential anomaly is an anomaly by calculating the relationship between the potential anomaly and the significant region.

6. The detection system according to claim 5, wherein the anomaly confirmation module further comprises: Perform K-means clustering on the anomalies, visualize the anomaly regions, and issue alerts.

Images