TL;DR
This paper presents a novel, real-time system for monitoring the deformation of Gas-Insulated Transmission Lines (GILs) using a single camera. By combining deep learning for object detection (YOLOv10) with advanced image processing techniques like adaptive corner detection and subpixel refinement, the system achieves high accuracy and speed, crucial for ensuring the safety and reliability of power transmission.
What Problem Does This Paper Address?
Gas-Insulated Transmission Lines (GILs) are critical infrastructure for high-voltage power transmission, offering advantages over traditional lines. However, their structural health, particularly deformation, needs constant monitoring to prevent failures and ensure safe operation. Traditional methods are often inefficient, complex, and susceptible to electromagnetic interference. While machine vision has emerged as a promising alternative, monitoring GIL pipelines specifically remains a nascent area. Existing vision-based methods can be labor-intensive, require offline processing, or struggle with accuracy and speed, especially in real-time applications. This paper aims to bridge this gap by developing an automated, fast, and accurate monocular measurement system for GIL deformation.
Methods
The proposed system integrates several key techniques:
- Deep Learning for Region of Interest (ROI) Extraction: A YOLOv10 model is employed to rapidly and automatically detect and localize artificial markers placed on the GIL pipeline. This provides the initial bounding boxes for potential deformation areas.
- Adaptive Corner Detection: Within the detected ROIs, a novel approach is used to dynamically set thresholds for the FAST corner detection algorithm. This is achieved by analyzing the grayscale mean and standard deviation of the image region, making the detection more robust to varying lighting conditions.
- Redundant Corner Filtering: The Shi-Tomasi algorithm is then applied to the FAST-detected corners to filter out redundant points and precisely identify unique feature points within the markers.
- Subpixel Refinement: To further enhance accuracy, a subpixel repositioning technique is used to extract more precise corner coordinates, moving beyond pixel-level precision.
- Image Tilt Correction and Scaling: ArUco markers are utilized to correct for image tilt caused by non-perpendicular camera angles. The known physical dimensions of these markers are used to compute a scaling factor, enabling the conversion of pixel displacements into real-world measurements. This process involves calculating a projection transformation matrix.
Datasets
The paper mentions the creation of a dataset for training the YOLOv10 model. This dataset consists of 930 images of GILs, augmented from a smaller initial set. The markers within these images were annotated with bounding rectangles. The dataset was split into training, validation, and testing sets in an 8:1:1 ratio.
Key Insights and Ideas
- Adaptive Thresholding for Corner Detection: Dynamically adjusting corner detection thresholds based on image statistics (mean and standard deviation of grayscale) is a clever way to improve robustness against varying illumination, a common challenge in real-world vision systems.
- Synergy of YOLO and Traditional Vision: The paper effectively combines the strengths of deep learning (fast, automated ROI detection) with established computer vision techniques (FAST, Shi-Tomasi, subpixel refinement) for a more comprehensive and accurate measurement pipeline.
- ArUco Markers for Robust Calibration: Using ArUco markers not only for localization but also for geometric correction (tilt and scaling) is a practical approach to mitigate the inherent challenges of monocular measurements and camera placement.
- Subpixel Refinement's Impact: The experimental results clearly demonstrate that subpixel-level measurements significantly improve accuracy and data stability compared to pixel-level measurements, justifying the added computational complexity.
Key Results
- Accuracy: The proposed method achieves a mean absolute error (MAE) of no more than 1.337 mm across multiple measurements, with the minimum average absolute error being 0.265 mm. The subpixel-level measurement method significantly improved MAE to 0.791 mm and reduced the standard deviation to 0.2953 mm compared to pixel-level measurements.
- Speed: The system achieves a processing speed of approximately 0.024 seconds per frame, enabling real-time monitoring.
- Robustness: Experimental results show that the system is insensitive to variations in applied forces, angles, and distances, and is robust to changes in camera orientation.
- Torsional Deformation: The method also demonstrated effectiveness in measuring torsional deformation angles, with a maximum average absolute error of 0.31°.
Limitations and Future Work
- Intrinsic Camera Calibration: The current work acknowledges that it does not delve deeply into intrinsic camera calibration. Future research will explore comprehensive calibration methods to further enhance measurement accuracy without increasing manual intervention.
- YOLOv10 Optimization: Further optimization of the YOLO-based model is planned to enhance its specificity for GIL deformation, improve ROI extraction speed, and reduce computational overhead.
- Multi-Camera Systems and AI: The authors suggest exploring multi-camera systems and alternative AI techniques to more effectively capture three-dimensional deformations.
- Applicable Measurement Range for Torsional Deformation: While the method shows promise for torsional deformation, the specific applicable measurement range requires further investigation.
Summary updated