In power grid systems, power lines play a vital role. Detecting strand breakage is challenging due to the slim design of power lines and limited sample data. In a recent publication in the journal Drones, researchers introduced a novel method for identifying broken strands in unmanned aerial vehicle (UAV)-captured images.
Study: Elevating Power Line Inspections: UAV-Powered Strand Breakage Detection. Image credit: abriendomundo/Shutterstock
Background
Electric power lines are essential for transmitting electricity from generation plants to end-users. Vulnerable to damage in outdoor settings, strand breakage poses a significant threat to supply reliability and safety. Detecting such breaks is crucial to ensuring uninterrupted power flow and preventing incidents.
UAVs have gained popularity in power line inspections due to their maneuverability and image-capture capabilities. However, manual analysis or centralized cloud processing with large deep learning models is inefficient and burdensome. Real-time, on-site detection using UAV resources is preferable. Existing solutions fall short in accuracy and efficiency.
A two-stage defect detector
To address the challenges of power line segmentation in UAV-captured images, a two-stage defect detector is proposed. The first stage conducts power line segmentation to extract power lines and crop local image patches. In the second stage, the detector performs classification on these image patches, identifying defects and their positions.
Power Line Segmentation Network: Power line segmentation is challenging due to slender power lines and background noise. A lightweight segmentation network, a better version of the bi-phase advanced network (BA-NetV2), is introduced. It enhances the original BA-Net with a larger capacity, higher-resolution feature maps, and the ability to capture long-range semantic relations.
Postprocessing Method: After segmentation, a hierarchical clustering approach identifies line segments belonging to the same power line, providing position and width information. The cropping method utilizes a sliding window approach, ensuring comprehensive coverage of power lines and potential defects.
Patch Classification Network: The patch classification network recognizes broken strands in image patches. A light-weight network with a classification head and an auxiliary segmentation head is designed. Multi-task learning is employed to utilize limited training data effectively. A combined loss function unifies classification and segmentation losses, with hyperparameters controlling their contributions.
Overall, this architecture aims to achieve accurate and efficient power line strand breakage detection, enabling effective power line monitoring and maintenance.
Experiment and results
The study assesses the accuracy and efficiency of a proposed strand breakage detection method through a series of experiments. The research consists of three main parts: evaluating the BA-NetV2 performance, assessing the patch classification network, and comparing the overall strand breakage detector with state-of-the-art object detectors. These experiments aim to validate the detector's precision and applicability for UAV-based power line inspections.
Performance Evaluation of BA-NetV2: A detailed analysis is presented for the modifications made to upgrade BA-Net to BA-NetV2, which includes reducing parallel branches, increasing channel numbers, and implementing dilated convolution. Experimental results indicate that these modifications enhance both accuracy and inference speed. BA-NetV2 outperforms several segmentation networks, showcasing its effectiveness in power line feature extraction.
Performance Evaluation of the Patch Classification Network: The study compares the proposed method with commonly used general classification models and a defect recognition network. Results show that the proposed method achieves high precision, average precision, and recall, outperforming the compared models. This success is attributed to the multi-task learning strategy and the cropping of local image patches.
Overall Performance of the Proposed Strand Breakage Detector: The research evaluates the overall performance of the strand breakage detector and compares it with state-of-the-art object detectors. The proposed method demonstrates significantly higher performance in terms of precision, recall, and F1-score. While its inference speed is slightly lower than end-to-end detectors, it remains suitable for real-time processing in most cases.
These findings suggest that the proposed strand breakage detection method, with its multi-stage pipeline and local patch cropping strategy, is a promising solution for UAV-based power line defect detection, offering high accuracy and efficiency, even with limited defect samples.
Conclusion
In summary, researchers introduced a real-time broken-strand detection method tailored for UAV-based power line inspections. It employs a multi-stage pipeline encompassing power line segmentation, image patch cropping, and patch classification. This approach effectively utilizes normal power line images and detailed local feature data, addressing challenges posed by slender power line shapes and the scarcity of strand breakage samples.
Experimental results demonstrate that the proposed model performed with superior accuracy compared to object detection techniques and real-time processing on embedded edge devices. It also enhanced the adaptability of the power line segmentation network, BA-NetV2, to power line features. Finally, the proposed model shows high accuracy in patch classification due to the multi-task learning strategy. Future work will focus on improving inference efficiency, addressing complex backgrounds and tower scenarios, and practical deployment in power line inspection hardware systems.
Deep Learning Enhances Canola Weed Detection