Electrical equipment corrosion cascade detection method based on improved deformable convolution

Abstract

The invention relates to the technical field of distribution network equipment detection, and provides an improved deformable convolution-based electrical equipment corrosion cascade detection method, which comprises the following steps of S1, collecting high-definition image data of equipment in a distribution network station through an intelligent inspection robot and a fixed high-definition camera; s2, through image enhancement, rotation, overturning and brightness and format adjustment, preprocessing the acquired image data, and then using an improved YOLOv7 detection model to identify the acquired image data; and S3, marking a corrosion area in the preliminary detection result. Through a special detection framework for corrosion of the distribution network equipment, YOLOv7 is taken as a target detection basic model, and an improved deformable convolution module is combined, so that the framework specially used for corrosion detection of the distribution network equipment is constructed, rapid and accurate recognition of image data acquired by equipment in a distribution network station is facilitated in the later period, the probability of false detection is reduced, and the labor cost is saved.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network equipment detection, and particularly relates to a method for cascaded detection of rust on power equipment based on improved deformable convolution. Background Art

[0002] As an important part of the power system, distribution network equipment is exposed to the outdoor environment for a long time and is affected by natural environments (such as sunlight, rain, temperature and humidity changes, salt spray corrosion, etc.) and human factors (such as industrial pollution, mechanical wear, etc.), making it extremely prone to metal rust. Rust not only reduces the mechanical strength and electrical performance of equipment, but may also lead to equipment failures, thereby affecting the safe and stable operation of the entire power system. For example, rust on distribution network equipment may cause poor contact, decreased insulation performance, and even short-circuit accidents, seriously affecting production and life. Under the strategic background of "carbon neutrality and carbon peak", as an important platform for energy conversion and resource allocation, the stable operation of the power grid system is crucial for promoting new energy supply and ensuring energy security. Therefore, timely and accurately detecting and dealing with rust problems on distribution network equipment is of great significance for ensuring the safe operation of the power system.

[0003] Traditional manual inspection methods rely on the experience of inspectors, suffering from problems such as strong subjectivity, low efficiency, and potential safety hazards. Manual inspection not only involves a large workload, increasing the operating costs of distribution substations, but also is prone to misdetection. This not only further increases the workload, but also some equipment may experience initial damage without being detected, resulting in further damage to the equipment in distribution substations.

[0004] Therefore, we have made improvements in this regard and proposed a method for cascaded detection of rust on power equipment based on improved deformable convolution. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to address the problem of insufficient accuracy in detecting rust on distribution network equipment in complex scenarios, which is mainly reflected in the irregular, complex and variable shapes of rust areas in the detection of rust on distribution network equipment, background interference, misdetection and low-confidence problems, and data insufficiency problems.

[0006] To achieve the above-mentioned invention objective, the present invention provides a method for cascaded detection of rust on power equipment based on improved deformable convolution, including the following steps: S1. Collect high-definition image data of equipment in the distribution substation through intelligent inspection robots and fixed high-definition cameras. The image data includes the front, side, and views at different angles of the equipment for subsequent rust detection; S2. Preprocess the collected image data through image enhancement, rotation, flipping, brightness, and format adjustment, and then use the improved YOLOv7 detection model to identify the collected image data for rapid detection of rust areas; S3. Mark the rust areas in the preliminary detection results, add detailed information of the distribution network equipment to the marking information, including equipment type, location coordinates, and rust degree, and use the marked data as training samples for subsequent classification model training; S4. Use the marked rust data to fine-tune and train the Swin Transformer V2 classification model, build a dedicated classification model for the rust detection of distribution network equipment, and perform refined classification on the areas within the detection frame to distinguish rust areas and background areas; S5. Use the trained Swin Transformer V2 classification model to perform secondary identification on the rust areas in the preliminary detection results; S6. Display the recognition results in the form of charts, histograms, and pie charts, and send them to the service terminal and user terminal.

[0007] Preferably, in step S2, the improved YOLOv7 detection model also integrates the D-DCNv3 module, which can adapt to the irregular shape of the rust area and output a detection frame to locate the rust area.

[0008] Preferably, in step S3, the marking work is completed by power system experts, and only the rust areas that pose a potential threat to the safe operation of the equipment are marked, including the rust at equipment joints and key parts of insulators.

[0009] Preferably, in step S5, the recognition method includes the following steps: S51. Process the position coordinates of the detection frame to expand the recognition area and improve the recognition and understanding ability of the classification model; S52. Input the processed detection frame area into the Swin Transformer V2 classification model for recognition; S53. According to the recognition results of the classification model, combined with the equipment type and location coordinate parameters in the marking information, output the final rust detection results.

[0010] Preferably, in step S51, the recognition area is expanded by increasing the number of pixels. The width and height of the detection frame are each increased by at least 50 pixels, and if the number of pixels is less than 50, it is expanded to the image boundary points.

[0011] A power equipment rust cascade detection system based on improved deformable convolution is applied to the power equipment rust cascade detection method described in any one of the above, and includes: A data acquisition module for acquiring high-definition image data of the internal equipment of the distribution network site; A data preprocessing module for preprocessing the acquired data, unifying the data format, structure, and encoding method, and eliminating inconsistencies and conflicts in the data; A YOLOv7 detection module for quickly detecting rust areas; A data annotation module for adding detailed information of the distribution network equipment to the annotation information; A Swin Transformer V2 classification module for constructing a dedicated classification model for the rust detection of distribution network equipment; A display and sending module for displaying the detection results and sending them to the service terminal and the user terminal.

[0012] A power equipment rust cascade detection method based on improved deformable convolution provided by the present invention has the following beneficial effects: 1. Through a dedicated detection framework for the rust of distribution network equipment, with YOLOv7 as the target detection basic model and combined with an improved deformable convolution module, a framework dedicated to the rust detection of distribution network equipment is constructed. Through multi-scale feature fusion and context information capture, and by means of improved deformable convolution, the model can capture multi-scale features and rich context information, which is beneficial to improving the fast and accurate recognition of the image data collected from the equipment in the distribution network site later, reducing the probability of false detection, and saving labor costs.

[0013] 2. By combining the method of deformable convolution with the method of dilated convolution to form an improved deformable convolution method, and through a cascade detection method, after separating the area within the detection frame from the background, it is sent to the Swin Transformer V2 classification network for reclassification, which improves the efficiency of detecting the images collected from the distribution network site equipment, avoids the situation where the background environment is complex and diverse, the clarity of the rust area is poor, the rust edge shape is irregular, and the morphology and position are random, and there will be no misdetection, missed detection, or the situation where the detection frame contains too much background area. The detection accuracy is high, which is beneficial to the maintenance of the equipment in the distribution network site and makes the maintenance cost of the equipment in the distribution network site low. Description of the Drawings

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0015] Figure 1Schematic flow chart of a power equipment rust cascade detection method based on improved deformable convolution provided by this application; Figure 2 Schematic diagram of the Swin Transformer V2 structure of a power equipment rust cascade detection method based on improved deformable convolution provided by this application; Figure 3 Schematic diagram of the system module of a power equipment rust cascade detection method based on improved deformable convolution provided by this application. Specific implementation manners

[0016] The following combines the description of the drawings in the specification and the embodiments to further describe in detail the specific implementation manners of the present invention. The following embodiments are only used to illustrate the present invention, but cannot be used to limit the scope of the present invention.

[0017] As Figures 1 - 3 shown, this implementation manner proposes a power equipment rust cascade detection method based on improved deformable convolution, including the following steps: S1. Collect high-definition image data of the equipment in the distribution substation through intelligent inspection robots and fixed high-definition cameras. The image data includes the front, side, and views at different angles of the equipment for subsequent rust detection; S2. Preprocess the collected image data through image enhancement, rotation, flipping, brightness, and format adjustment, and then use the improved YOLOv7 detection model to identify the collected image data for quickly detecting rust areas. The main advantage of the improved YOLOv7 model is that it can quickly detect rust areas. When traditional detection methods process these images, due to complex backgrounds, lighting changes, and irregular shapes of rust areas, missed detections and misdetections are likely to occur; S3. Mark the rust areas in the preliminary detection results, and add detailed information of the distribution network equipment to the marking information, including equipment type, position coordinates, and rust degree, and use the marked data as training samples for subsequent classification model training; S4. Use the marked rust data to fine-tune and train the Swin Transformer V2 classification model, construct a dedicated classification model for power distribution equipment rust detection, and perform refined classification on the areas within the detection frames to distinguish between rust areas and background areas; S5. Use the trained Swin Transformer V2 classification model to perform secondary identification on the rust areas in the preliminary detection results. Using the Swin Transformer V2 model alone will increase the detection cost. Therefore, cascaded detection in combination with the preliminary detection model can effectively reduce the consumption of computing resources; S6. Display the recognition results in the form of charts, histograms, and pie charts, and send them to the service terminal and the user terminal.

[0018] Specifically, through a dedicated detection framework for the corrosion of distribution network equipment, with YOLOv7 as the target detection basic model and combined with an improved deformable convolution module, a framework dedicated to the corrosion detection of distribution network equipment is constructed. Through multi-scale feature fusion and context information capture, and in an improved deformable convolution manner, the model can capture multi-scale features and rich context information, which is conducive to improving the rapid and accurate recognition of the image data collected from the equipment in the distribution substation in the later stage, reducing the probability of false detection, and saving labor costs.

[0019] In this embodiment, in step S2, the improved YOLOv7 detection model also integrates the D-DCNv3 module, which can adapt to the irregular shape of the corrosion area and output a detection box to locate the corrosion area.

[0020] In this embodiment, in step S3, the annotation work is completed by power system experts, and only the corrosion areas that pose a potential threat to the safe operation of the equipment are annotated, including the corrosion at the equipment connection points and the key parts of the insulators.

[0021] In this embodiment, in step S5, the recognition method includes the following steps: S51. Process the position coordinates of the detection box to expand the recognition area and improve the recognition and understanding ability of the classification model; S52. Input the processed detection box area into the Swin Transformer V2 classification model for recognition; S53. According to the recognition results of the classification model, combined with the equipment type and position coordinate parameters in the annotation information, output the final corrosion detection results.

[0022] In this embodiment, in step S51, the recognition area is expanded by increasing the number of pixels. The width and height of the detection box are each increased by at least 50 pixels, and if the number of pixels is less than 50, it is expanded to the image boundary points.

[0023] Specifically, by combining the deformable convolution method with the dilation convolution method, an improved deformable convolution method is formed. After separating the area inside the detection box from the background through the cascade detection method, it is sent to the Swin Transformer V2 classification network for reclassification, which improves the efficiency of detecting the images collected from the equipment in the distribution substation, avoids the situation where the background environment is complex and diverse, the clarity of the corrosion area is poor, the shape of the corrosion edge is irregular, and the morphology and position are random, and there will be no misdetection, missed detection, or the situation where the detection box contains too much background area. The detection accuracy is high, which is conducive to the maintenance of the equipment in the distribution substation and makes the maintenance cost of the equipment in the distribution substation low.

[0024] An electric power equipment rust cascade detection system based on improved deformable convolution, which is applied to an electric power equipment rust cascade detection method according to any one of the above, includes: A data acquisition module, which is used to acquire high-definition image data of the internal equipment of the distribution network substation; A data preprocessing module, which is used to preprocess the acquired data, unify the data format, structure and encoding method, and eliminate the inconsistencies and conflicts in the data; A YOLOv7 detection module, which is used to quickly detect the rust area; A data annotation module, which is used to add detailed information of the distribution network equipment to the annotation information; A Swin Transformer V2 classification module, which is used to build a special classification model for the rust detection of distribution network equipment; A display and sending module, which is used to display the detection results and send them to the service terminal and the user terminal.

[0025] Specifically, compared with the traditional single-stage detection model, the detection method adopted in the present invention can reduce about 30% of the misdetected pictures per day on average after cascade detection, significantly improving the accuracy and efficiency of rust detection.

[0026] The above embodiments are only used to illustrate the present invention, rather than to limit the present invention. Although the present invention has been described in detail with reference to the embodiments, those of ordinary skill in the art should understand that various combinations, modifications or equivalent replacements of the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and should all be covered within the scope of the claims of the present invention.

Claims

1. A method for cascaded detection of corrosion of power equipment based on improved deformable convolution, characterized in that, It includes the following steps: S1. Collect high-definition image data of the equipment in the distribution network substation through intelligent inspection robots and fixed high-definition cameras. The image data includes the front, side, and views at different angles of the equipment for subsequent rust detection; S2. Preprocess the collected image data through image enhancement, rotation, flipping, brightness, and format adjustment, and then use the improved YOLOv7 detection model to identify the collected image data for quickly detecting rust areas; S3. Mark the rust areas in the preliminary detection results, add detailed information of the distribution network equipment to the marking information, including equipment type, location coordinates, and rust degree, and use the marked data as training samples for subsequent classification model training; S4. Use the marked rust data to fine-tune and train the Swin Transformer V2 classification model, construct a dedicated classification model for rust detection of distribution network equipment, and perform refined classification on the areas within the detection frame to distinguish between rust areas and background areas; S5. Use the trained Swin Transformer V2 classification model to perform secondary identification on the rust areas in the preliminary detection results; S6. Display the recognition results in the form of charts, histograms, and pie charts and send them to the service terminal and user terminal.

2. The cascade detection method for corrosion of power equipment based on improved deformable convolution according to claim 1, characterized in that: In step S2, the improved YOLOv7 detection model also integrates the D-DCNv3 module, which can adapt to the irregular shape of the rust area and output a detection frame to locate the rust area.

3. The method for cascaded detection of corrosion of power equipment based on improved deformable convolution according to claim 1, characterized in that: In step S3, the marking work is completed by power system experts, and only the rust areas that pose a potential threat to the safe operation of the equipment are marked, including the rust at equipment joints and key parts of insulators.

4. A method for cascaded detection of corrosion of power equipment based on improved deformable convolution according to claim 1, characterized in that: In step S5, the recognition method includes the following steps: S51. Process the position coordinates of the detection frame to expand the recognition area and improve the recognition and understanding ability of the classification model; S52. Input the processed detection frame area into the Swin Transformer V2 classification model for recognition; S53. According to the recognition results of the classification model, combined with the equipment type and location coordinate parameters in the marking information, output the final rust detection results.

5. A method for cascaded detection of corrosion of power equipment based on improved deformable convolution according to claim 4, characterized in that: In step S51, the recognition area is expanded by adding pixel points. The width and height of the detection frame are each increased by at least 50 pixel points, and if it is less than 50 pixel points, it is expanded to the image boundary points.

6. A power equipment rust cascade detection system based on improved deformable convolution is applied to a power equipment rust cascade detection method based on improved deformable convolution according to any one of claims 1-5, and is characterized in that It includes: A data acquisition module for collecting high-definition image data of the equipment inside the distribution network substation; A data preprocessing module for preprocessing the collected data, unifying the data format, structure, and encoding method, and eliminating inconsistencies and conflicts in the data; A YOLOv7 detection module for quickly detecting rust areas; A data marking module for adding detailed information of the distribution network equipment to the marking information; A Swin Transformer V2 classification module for constructing a dedicated classification model for rust detection of distribution network equipment; A display and sending module for displaying the detection results and sending them to the service terminal and user terminal.

Images