Power transmission equipment fault diagnosis method based on image analysis technology

Abstract

The invention discloses a power transmission equipment fault diagnosis method based on an image analysis technology. The method comprises the following steps: S1, automatically collecting high-resolution visible light and infrared thermal image pictures / video streams of power transmission equipment; s2, normalizing, denoising and enhancing the acquired data, and positioning an equipment area; s3, a multi-task deep learning network based on an encoder-decoder architecture is deployed at the cloud end, and the network shares a feature extraction backbone network, but has a plurality of branches for executing different parallel diagnosis tasks; s4, fusing the output results of the multiple branches, matching the fused diagnosis result with a knowledge base, and automatically generating a diagnosis report; s5, the operation and maintenance personnel check and correct the diagnosis report generated by the system, and data with a correction label is returned to the training data set of the system; the method has the advantages of being high in diagnosis efficiency, high in diagnosis precision and reliability, capable of achieving early warning of faults and scientific and accurate in management decision.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power transmission equipment fault diagnosis, and particularly relates to a power transmission equipment fault diagnosis method based on image analysis technology. BACKGROUND

[0002] Power transmission equipment plays a vital role in the power system. Through the coordinated work of primary equipment (such as transformers, circuit breakers, arresters, etc.) and secondary equipment (protection, monitoring system), power transmission equipment ensures the safe and reliable transmission of electric energy from power plants to users, avoiding chain failures and large-scale power outages.

[0003] At present, the fault diagnosis of power transmission equipment still mainly relies on the traditional mode of manual experience, which mainly has the following problems: 1. Low inspection efficiency and coverage: the traditional manual inspection method is time-consuming and labor-intensive, and it is difficult to cover all equipment in a wide area, and is subject to environmental factors such as weather and terrain; 2. Defect identification accuracy and consistency are insufficient: relying on the human eye to judge the equipment state is easily affected by subjective experience and fatigue, and is difficult to standardize, and has a high missed detection rate for early and small defects (such as fine cracks and light rust); 3. Single fault type, lack of comprehensive diagnosis: existing methods often only identify a single fault (such as only identifying broken insulators), and cannot simultaneously analyze multiple potential defects of the same equipment or the same picture; 4. Slow response speed, unable to actively warn: from discovering the problem to manual reporting and then to maintenance decision-making, the process is long, and it is impossible to predict and early warn the development trend of the fault.

[0004] In order to solve the above problems, it is necessary to develop a power transmission equipment fault diagnosis method based on image analysis technology. SUMMARY

[0005] The purpose of the present application is to overcome the shortcomings of the prior art and provide a power transmission equipment fault diagnosis method based on image analysis technology, which has high diagnosis efficiency, high diagnosis accuracy and reliability, early fault warning, and scientific and accurate management decision-making, and provides a comprehensive, scientific and reasonable intelligent and fine diagnosis method for power transmission equipment fault diagnosis.

[0006] The purpose of the present application is achieved by a power transmission equipment fault diagnosis method based on image analysis technology, comprising the following steps: S1, data acquisition and transmission: using a patrol unmanned aerial vehicle and / or a high-definition camera deployed on a tower, automatically collecting high-resolution visible light and infrared thermal image pictures / video streams of power transmission equipment according to a predetermined flight route or triggering rule; S2, image preprocessing and enhancement: normalize, denoise and enhance the collected data, and perform device region positioning processing; S3, multi-task fine fault analysis and diagnosis: deploy a multi-task deep learning network based on an encoder-decoder architecture on the cloud, which shares a feature extraction backbone network (Encoder) but has multiple branches (Decoder) for performing different parallel diagnostic tasks; S4, diagnosis result fusion and decision support: fuse the output results of multiple branches, match the fused diagnosis results with the knowledge base, and automatically generate a diagnosis report; S5, feedback and model optimization: the operation and maintenance personnel review and correct the diagnosis report generated by the system, and the data with correction labels will be returned to the training data set of the system.

[0007] Preferably, in step S1, the high-definition camera includes a visible light high-definition camera and an infrared high-definition camera, and the power transmission equipment includes insulators, conductors, fittings, and tower bodies.

[0008] Preferably, step S2 includes the following steps: S21, data normalization: size standardization and color correction of images; S22, denoising and enhancement: using a filtering algorithm to remove image noise and using image enhancement techniques to improve the contrast of key areas; S23, device region positioning: using a lightweight target detection model to quickly frame each type of device in the image, and defining the region for subsequent fine analysis.

[0009] Preferably, in step S22, the filtering algorithm is median filtering or Gaussian filtering, and the image enhancement technique is histogram equalization; in step S23, the target detection model is the YOLO series.

[0010] Preferably, in step S3, the parallel diagnostic tasks include a semantic segmentation branch, a target detection and classification branch, an anomaly detection branch, and an infrared thermal image analysis branch.

[0011] Preferably, the semantic segmentation branch uses a U-Net or high-resolution HRNet model to perform pixel-level classification on the preprocessed image, generating a segmentation mask image to accurately locate the specific position and number of damaged and exploded insulators.

[0012] Preferably, the target detection and classification branch uses a Faster R-CNN or DETR model to identify and locate specific fault types in the preprocessed image and directly frame and classify them.

[0013] Preferably, the anomaly detection branch adopts an autoencoder or a generative adversarial network to learn the reconstruction of normal equipment images, and when a defect image is input, the reconstruction error will significantly increase, thereby triggering an alarm, discovering unknown or insufficient sample training defect types, and giving an anomaly score to indicate the possibility of unknown anomalies in the abnormal area.

[0014] Preferably, the infrared thermal image analysis branch adopts a special model or is fused with a visible light model to identify temperature anomaly areas, analyze infrared images, detect equipment overheating faults, locate overheating points, and label their temperature levels.

[0015] Preferably, the step S4 comprises the following steps: S41, result fusion: the outputs of multiple branches in step S3 are fused to confirm each other and improve the diagnostic confidence; S42, knowledge base matching: the fused diagnostic results are associated with equipment historical archives and maintenance records to evaluate the emergency degree and potential impact of the fault; S43, generating a diagnostic report: automatically generating a structured report containing fault location, fault type, severity level, confidence, and processing suggestions.

[0016] Due to the adoption of the above technical solutions, the present application has the following advantages: 1. The present application can realize 7x24 hour automatic inspection and diagnosis, can shorten the time for analyzing one picture from minutes to seconds or even milliseconds, greatly liberates manpower, and greatly improves the diagnostic efficiency; 2. The present application adopts a multi-task deep learning model, the detection rate of tiny defects is far higher than that of the human eye, effectively reduces the missed detection / misjudgment rate, and the multi-branch result fusion mechanism further ensures the reliability of the diagnostic conclusion, and the diagnostic precision and reliability are significantly enhanced; 3. The present application can analyze the trend of defects of the same equipment through periodic collection and analysis, thereby giving a warning before the fault occurs, guiding the development of predictive maintenance, avoiding shutdown accidents, and facilitating the transformation of the operation and maintenance mode to prediction; 4. The full-quantity and structured diagnostic report provided by the present application provides strong data support for power grid asset management, makes the configuration of operation and maintenance resources and the formulation of overhaul plans more scientific and accurate, thereby making the management decision more scientific and digitalized; Overall, the present application has the advantages of high diagnostic efficiency, strong diagnostic precision and reliability, early warning of faults, and scientific and accurate management decisions, and provides a comprehensive, scientific, and reasonable intelligent and fine diagnostic method for fault diagnosis of power transmission equipment. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a method step flowchart of the present application. Detailed Implementation

[0018] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0019] like Figure 1 As shown, the present invention provides a method for fault diagnosis of power transmission equipment based on image analysis technology, comprising the following steps: The first step, data acquisition and transmission: This step is used to achieve edge-side sensing and is responsible for acquiring the most basic state data. It is set to be a periodic and automated acquisition method.

[0020] Using inspection drones or high-definition cameras deployed on power transmission towers, including visible light high-definition cameras and infrared high-definition cameras, high-resolution visible light and infrared thermal images / video streams of power transmission equipment (such as insulators, conductors, fittings, and tower bodies) are automatically collected according to predetermined routes or triggering rules.

[0021] The second step, image preprocessing and enhancement, is used to achieve rapid edge processing, aiming to improve data quality and initially identify the analysis target, providing clean and well-organized data input for complex deep learning models in the cloud, while reducing the computing pressure on the cloud.

[0022] 1. Data normalization: Standardize the size and correct the color of the image to eliminate interference from changes in lighting and weather.

[0023] 2. Denoising and Enhancement: Filtering algorithms (such as median filtering and Gaussian filtering) are used to remove image noise, and image enhancement techniques (such as histogram equalization) are used to improve the contrast of key areas, making defect features more prominent.

[0024] 3. Equipment area localization: Using a lightweight target detection model (such as the YOLO series), various types of equipment (such as insulator strings, vibration dampers, and wire clamps) in the image are quickly selected to define the area for subsequent detailed analysis.

[0025] The third step is multi-task refined fault analysis and diagnosis: used in the cloud-based intelligent core.

[0026] Deploy a multi-task deep learning network based on an encoder-decoder architecture in the cloud. This network shares a feature extraction backbone (Encoder) but has multiple branches (Decoders) for performing different parallel diagnostic tasks.

[0027] 1. Semantic Segmentation Branch: Responsible for classifying each pixel in the image and assigning it to a specific part or state. The workflow is as follows: (1) Input: Image of the insulator string region after preprocessing and positioning; (2) Network structure: encoder-decoder structure similar to U-Net or high-resolution HRNet with skip connections; Encoder (main trunk): shared with other branches, responsible for extracting multi-level feature maps from images (from low-level features such as edges and textures to high-level semantic features such as insulators and damage); Decoder: through upsampling operations, the compressed feature maps are gradually restored to the original image size, and the skip connections fuse the encoder feature maps with the decoder feature maps, thus preserving rich spatial detail information when restoring the size; (3) Output layer: a Softmax classifier after the last convolutional layer calculates the probability of each pixel belonging to each class (such as intact porcelain, damaged porcelain, metal cap, and background); (4) Final output: a segmentation mask image of the same size as the input image, where each pixel is assigned a color label representing its class; Sign of completing the task: the generated segmentation mask can clearly show, for example, that the 3rd, 4th, and 5th insulators are marked as damaged, thus accurately counting the specific location and number of damaged insulators.

[0028] 2. Target detection and classification branch: use Faster R-CNN or DETR model to identify and locate specific fault types in the preprocessed image, directly frame and classify, quickly locate all fault targets in the image, and enclose them with a bounding box, while giving the fault class and confidence.

[0029] The workflow of Faster R-CNN is as follows: (1) Input: the same preprocessed image as the segmentation branch; (2) Feature extraction: use the shared main network to extract the deep feature map of the image; (3) Region Proposal Network (RPN): slide a small network over the feature map to quickly generate a large number of candidate regions (candidate boxes) that may contain targets; (4) Region of Interest (RoI) Pooling: map different size candidate boxes to the feature map and pool them into fixed size feature vectors; (5) Classification and regression: send the fixed size feature vectors to the fully connected layer and perform two tasks in parallel: classification branch: determine which specific fault class (such as bird nest, insulator explosion, and foreign object hanging) the target in the candidate box belongs to; regression branch: fine-tune the coordinates (x, y, width, height) of the candidate box to make it fit the target more closely; (6) Final output: a series of bounding boxes with class labels and confidence; Signs of successful completion: Accurately mark the bird's nest with a box and label it in the image, or outline the area where the insulator exploded.

[0030] 3. Anomaly Detection Branch: Using an autoencoder or generative adversarial network, it learns to reconstruct images of normal devices. When a defective image is input, its reconstruction error will increase significantly, thereby triggering an alarm. This allows it to discover unknown defect types or those for which there are not enough samples for training, and gives an anomaly score to indicate the possibility of unknown anomalies in the abnormal area.

[0031] The workflow of a self-encoder is as follows: (1) Training phase: Input: A large number of normal, defect-free images of power transmission equipment; Learning objective: Train an autoencoder network that can compress (encode) an input image into a low-dimensional latent space, and then decompress (decode) it to reconstruct an image that is as similar as possible to the original image; Core hypothesis: Through learning, the network only learns the patterns of normal images; (2) Reasoning (diagnosis) stage: Input: An image to be detected; Reconstruction and Comparison: Input the image into the trained autoencoder to obtain the reconstructed image; Computational error: The difference between the original image and the reconstructed image is calculated pixel by pixel (reconstruction error). On normal images, the error will be very small; however, on defective images, the error in defective areas will be significantly larger because the model cannot reconstruct defect patterns it has not seen well. Generate anomaly map: Visualize the reconstruction error as a heatmap (anomaly score map), where brighter areas indicate a higher probability of an anomaly; (3) Final output: an anomaly score graph and an overall anomaly score value; The sign that the task is completed: The system alarms to detect an unknown anomaly and highlights an unusual texture or structure (such as a crack or corrosion pattern never seen before) on the anomaly score graph.

[0032] 4. Infrared Thermal Imaging Analysis Branch: Utilizing a dedicated model or fusion with a visible light model, this branch identifies areas of abnormal temperature, analyzes infrared images, detects overheating faults in equipment, locates hot spots, and labels their temperature levels. It converts pixel values ​​(temperature) in infrared thermal images into fault information, diagnoses defects related to heat generation, and follows the workflow: (1) Input: Preprocessed infrared thermal image (can be input alone, or can be fused with visible light image at the pixel level or feature level); (2) Temperature calibration and feature extraction: Convert image pixel values to temperature values and extract temperature-related features such as maximum temperature, average temperature, relative temperature difference, heat field distribution, etc. (3) Fault diagnosis logic: Absolute temperature method: directly judge whether the temperature of a certain point exceeds the safety threshold (such as the temperature of the wire clamp exceeding 90℃); Relative temperature difference method: compare the temperature difference between the fault point and the normal reference point, such as calculating the relative temperature difference = (fault point temperature - normal corresponding point temperature) / (fault point temperature - ambient temperature) x 100%, if the calculation result is greater than the standard (such as 95%), it is diagnosed as a serious overheating fault; Heat pattern recognition: use a classification network to identify specific heat patterns such as local overheating points, overall heating, uneven heat field distribution, etc. (4) Final output: Thermal fault diagnosis report, indicating the overheating point position, temperature value, relative temperature difference and corresponding fault type (such as poor contact of wire joint, current leakage heating due to contamination of insulator).

[0033] At this point, through the cooperative work of the four branches, the system can not only identify known and common faults, but also discover unknown abnormalities, and provide more comprehensive equipment health status evaluation combined with thermal image data, thereby truly realizing intelligent and refined diagnosis of power transmission equipment faults.

[0034] Fourth step, diagnosis result fusion and decision support: used for comprehensive research and judgment, through cross-validation of multi-task results, to improve the accuracy and reliability of diagnosis, and to convert abstract recognition results into decision information that can be directly used by operation and maintenance personnel.

[0035] 1. Result fusion: fuse the outputs of multiple branches in the third step, such as when the semantic segmentation branch identifies insulator disc loss and the target detection branch also reports insulator damage, they can be mutually confirmed to improve the diagnosis confidence.

[0036] 2. Knowledge base matching: associate the fused diagnosis results with device historical archives, maintenance records, etc. to assess the urgency and potential impact of the fault.

[0037] 3. Generate diagnosis report: automatically generate a structured report containing fault location (which can be marked on the map), fault type, severity level, confidence and handling suggestions.

[0038] Fifth step, feedback and model optimization: used for system self-evolution, through continuous learning or online learning mechanism, periodically retrain the model with new data, so that the system can continuously adapt to new fault patterns, realize the continuous improvement of performance and self-evolution of knowledge.

[0039] Specifically, the operation and maintenance personnel review and correct the diagnosis report generated by the system, and these data with correction tags will be returned to the training data set of the system.

[0040] To sum up, the application constructs a cloud-edge-end collaborative power transmission equipment image intelligent analysis and fault diagnosis system, which automatically collects images by using unmanned aerial vehicles, fixed cameras and other end-side equipment, then carries out real-time preprocessing and preliminary analysis through edge-side computing nodes, finally deploys an advanced multi-task deep learning model on the cloud side to finely analyze and comprehensively diagnose the images, and realizes the operation and maintenance mode change from passive response to active early warning by deeply integrating computer vision, deep learning and power equipment knowledge.

[0041] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application, and are not limiting. Although the present application has been described in detail with reference to the above examples, those skilled in the art should understand that the specific embodiments of the present application can still be modified or replaced equivalently without departing from the spirit and scope of the present application, and any modification or equivalent replacement without departing from the spirit and scope of the present application should be covered in the scope of the claims of the present application.

Claims

1. A method for fault diagnosis of power transmission equipment based on image analysis technology, characterized in that, Includes the following steps: S1, Data Acquisition and Transmission: Using inspection drones and / or high-definition cameras deployed on towers, high-resolution visible light and infrared thermal images / video streams of power transmission equipment are automatically acquired according to predetermined routes or triggering rules. S2, Image Preprocessing and Enhancement: Normalize, denoise and enhance, and perform device area localization processing on the acquired data; S3, Multi-task refined fault analysis and diagnosis: Deploy a multi-task deep learning network based on an encoder-decoder architecture in the cloud. This network shares a feature extraction backbone network (Encoder) but has multiple branches (Decoder) to perform different parallel diagnostic tasks; S4, Diagnostic Result Fusion and Decision Support: The output results of multiple branches are fused, the fused diagnostic results are matched with the knowledge base, and a diagnostic report is automatically generated; S5, Feedback and Model Optimization: Operations personnel review and correct the diagnostic reports generated by the system, and data with correction labels will be fed back into the system's training dataset.

2. The fault diagnosis method for power transmission equipment based on image analysis technology according to claim 1, characterized in that: In step S1, the high-definition camera includes a visible light high-definition camera and an infrared high-definition camera, and the power transmission equipment includes insulators, conductors, fittings, and tower bodies.

3. The method for fault diagnosis of power transmission equipment based on image analysis technology according to claim 1, characterized in that, Step S2 includes the following steps: S21, Data Normalization: Standardize the size and correct the color of the image; S22, Denoising and Enhancement: Image noise is removed using a filtering algorithm, and the contrast of key areas is improved using image enhancement technology; S23, Device Area Localization: Using a lightweight target detection model, various devices in the image are quickly selected to define the area for subsequent detailed analysis.

4. The fault diagnosis method for power transmission equipment based on image analysis technology according to claim 3, characterized in that: In step S22, the filtering algorithm is median filtering and Gaussian filtering, and the image enhancement technique is histogram equalization; in step S23, the target detection model is the YOLO series.

5. The fault diagnosis method for power transmission equipment based on image analysis technology according to claim 1, characterized in that: In step S3, the parallel diagnostic task includes a semantic segmentation branch, a target detection and classification branch, an anomaly detection branch, and an infrared thermal imaging analysis branch.

6. The fault diagnosis method for power transmission equipment based on image analysis technology according to claim 5, characterized in that: The semantic segmentation branch uses U-Net or high-resolution HRNet models to perform pixel-level classification on the preprocessed image, generate a segmentation mask map, and accurately locate the specific location and number of damaged and burst insulators.

7. The fault diagnosis method for power transmission equipment based on image analysis technology according to claim 5, characterized in that: The target detection and classification branch uses Faster R-CNN or DETR models to identify and locate specific fault types in the preprocessed image, directly selecting and classifying them.

8. The fault diagnosis method for power transmission equipment based on image analysis technology according to claim 5, characterized in that: The anomaly detection branch uses an autoencoder or generative adversarial network to learn the reconstruction of normal device images. When a defective image is input, its reconstruction error will increase significantly, thereby triggering an alarm. This allows the branch to discover unknown defect types or those for which there are not enough samples for training, and to give an anomaly score to indicate the possibility of unknown anomalies in the abnormal area.

9. The fault diagnosis method for power transmission equipment based on image analysis technology according to claim 5, characterized in that: The infrared thermal imaging analysis branch uses a dedicated model or is fused with a visible light model to identify abnormal temperature areas, analyze infrared images, detect equipment overheating faults, locate hot spots, and label their temperature levels.

10. The method for fault diagnosis of power transmission equipment based on image analysis technology according to claim 1, characterized in that, Step S4 includes the following steps: S41, Result Fusion: The outputs of multiple branches in step S3 are fused together to corroborate each other and improve the diagnostic confidence. S42, Knowledge Base Matching: The merged diagnostic results are linked with equipment history files and maintenance records to assess the urgency and potential impact of the fault; S43, Generate Diagnostic Report: Automatically generate a structured report containing fault location, fault type, severity level, confidence level, and handling recommendations.

Images