Abstract
The widespread popularity of unmanned aerial vehicles enables an immense amount of power lines inspection data to be collected. How to employ massive inspection data especially the visible images to maintain the reliability, safety, and sustainability of power transmission is a pressing issue. To date, substantial works have been conducted on the analysis of power lines inspection data. With the aim of providing a comprehensive overview for researchers who are interested in developing a deep-learning-based analysis system for power lines inspection data, this paper conducts a thorough review of the current literature and identifies the challenges for future research. Following the typical procedure of inspection data analysis, we categorize current works in this area into component detection and fault diagnosis. For each aspect, the techniques and methodologies adopted in the literature are summarized. Some valuable information is also included such as data description and method performance. Further, an in-depth discussion of existing deep-learning-related analysis methods in power lines inspection is proposed. Finally, we conclude the paper with several research trends for the future of this area, such as data quality problems, small object detection, embedded application, and evaluation baseline.
Introduction
Power lines are the lifeblood of modern society, but they are prone to damage from harsh weather, corrosion, and biological interference. Traditionally, inspection required hazardous manual ground surveys or expensive helicopter flights—methods that are high-cost, high-risk, and notoriously inefficient.
The advent of Unmanned Aerial Vehicles (UAVs) has solved the data collection problem by capturing massive amounts of high-resolution imagery. However, this "data explosion" has created a new bottleneck: manual analysis. This paper reviews the transition from traditional computer vision, which relied on rigid hand-crafted rules, to deep learning systems that can automatically identify components and diagnose subtle faults in complex, cluttered environments.
Analysis Methodology: Detection and Diagnosis
The literature on power line inspection is generally categorized into two primary stages of analysis:
1. Component Detection
This stage involves identifying the "where" and "what" (e.g., towers, insulators, conductors, and fittings). Traditional methods used color models and shape features like Canny edge detection. Modern deep learning employs object detection frameworks like Faster R-CNN, SSD, and YOLO. These models extract "deep features" that are far more resilient to changes in lighting and background clutter than traditional methods.
2. Fault Diagnosis
Identifying the "condition" of a component (e.g., missing insulator caps, broken conductor strands, or tower corrosion) is the next critical step. A popular strategy is the use of Cascaded Networks, where a detection network first crops a Region of Interest (RoI), and a second, specialized network (like a U-Net for segmentation) identifies the specific fault within that crop.
Key Insights and Trends
A pivotal observation in this review is the "2016 Inflection Point." While deep learning revolutionized general computer vision in 2012, its application to power lines didn't explode until 2016. Since then, the number of publications has grown exponentially, as researchers have adapted these powerful architectures to the specific challenges of the power grid.
Other critical insights include:
- Domain Knowledge Integration: Pure "black box" deep learning is often improved by incorporating physical constraints, such as the repetitive geometric structure of insulators.
- The "Small Object" Struggle: Standard architectures often fail on tiny components like pins or bolts. In a 6000x4000 pixel aerial image, a critical pin might only occupy 60x40 pixels, requiring specialized "zoom-in" strategies or Feature Pyramid Networks (FPN).
- Class Imbalance: In the real world, 99% of components are healthy. This makes training fault-detection models difficult. Researchers are increasingly turning to data augmentation and Generative Adversarial Networks (GANs) to create synthetic faults for training.
Performance and Future Directions
Deep learning has set new performance baselines in the field. SSD-based models have achieved ~94.7% mean Average Precision (mAP) for insulators, while cascaded frameworks have shown a recall of ~95.5% for identifying missing insulator caps.
Despite this progress, several "open research issues" remain:
- Embedded Deployment: Most current models require heavy GPUs. There is a pressing need for lightweight models that can run on-board UAV hardware like the NVIDIA Jetson TX2.
- Standardization: The lack of a massive, standardized evaluation baseline (an "ImageNet" for power lines) makes it difficult to objectively compare different researchers' algorithms.
- Multi-Modal Fusion: Future systems should better integrate visible imagery with infrared (for hotspots) and LiDAR (for vegetation encroachment) to provide a holistic view of grid health.
Summary updated