Abstract
The Five-hundred-meter Aperture Spherical radio Telescope (FAST) is the world's largest single-dish radio telescope. Its large reflecting surface achieves unprecedented sensitivity but is prone to damage, such as dents and holes, caused by naturally-occurring falling objects. Hence, the timely and accurate detection of surface defects is crucial for FAST's stable operation. Conventional manual inspection involves human inspectors climbing up and examining the large surface visually, a time-consuming and potentially unreliable process. To accelerate the inspection process and increase its accuracy, this work makes the first step towards automating the inspection of FAST by integrating deep-learning techniques with drone technology. First, a drone flies over the surface along a predetermined route. Since surface defects significantly vary in scale and show high inter-class similarity, directly applying existing deep detectors to detect defects on the drone imagery is highly prone to missing and misidentifying defects. As a remedy, we introduce cross-fusion, a dedicated plug-in operation for deep detectors that enables the adaptive fusion of multi-level features in a point-wise selective fashion, depending on local defect patterns. Consequently, strong semantics and fine-grained details are dynamically fused at different positions to support the accurate detection of defects of various scales and types. Our AI-powered drone-based automated inspection is time-efficient, reliable, and has good accessibility, which guarantees the long-term and stable operation of FAST.
Introduction: The Scale of the Challenge
The Five-hundred-meter Aperture Spherical radio Telescope (FAST), located in Guizhou, China, is an engineering marvel. Its reflecting surface covers an area equivalent to 30 soccer fields, composed of 4,450 individual aluminum panels. This vast surface is essential for its high sensitivity but is constantly exposed to environmental hazards like falling rocks, hailstorms, and high winds.
Maintaining the integrity of these panels is critical; even minor defects like small dents or holes can scatter radio waves, resulting in significant data loss. Historically, inspection has been a "nightmare" task, requiring human inspectors to physically climb the structure. This manual process is not only dangerous and slow but also prone to human error, especially when identifying small defects against the complex background of perforated aluminum.
AI-Powered Drone Inspection
To modernize this process, researchers developed an automated pipeline using a DJI Mavic 2 Enterprise Advanced drone. Equipped with a Real-Time Kinematic (RTK) module, the drone achieves centimeter-level positioning accuracy. This allows every captured frame to be precisely georeferenced, enabling maintenance teams to pinpoint the exact location of any detected damage.
However, standard computer vision models struggle with FAST's imagery due to two primary factors:
- Large-scale variations: Defects can range from less than half an inch to over a foot in diameter.
- High inter-class similarity: On the perforated aluminum surface, a shallow dent and a small hole appear very similar to standard detection algorithms.
The Cross-Fusion Mechanism
The core innovation of this research is the Cross-Fusion block, a plug-in module designed to enhance deep learning backbones. While traditional Feature Pyramid Networks (FPNs) typically combine layers through simple addition or concatenation, Cross-Fusion introduces a point-wise cross-layer attention mechanism.
As shown in the architecture above, the model dynamically decides how to weight information from different layers for every pixel. For a tiny hole, the system prioritizes fine-grained details from high-resolution layers. For a large dent, it leans on the broad semantic context provided by deeper layers. To bridge the "receptive field gap" between these layers without excessive computation, the authors utilized dilated max pooling. Furthermore, the block is initialized as an identity map, ensuring training stability by allowing the model to gradually learn the optimal fusion strategy from a pre-trained state.
Experimental Results and Generalization
The system was rigorously tested on a newly created FAST Surface Defect Dataset, containing over 34,000 labeled defects. The results were impressive:
- Performance: The model achieved a mean Average Precision (mAP) of 41.3%, outperforming established models like YOLOv5 (37.9%) and YOLOv7 (38.9%).
- Sensitivity: It successfully detected holes as small as 0.2 inches in diameter from drone altitudes—defects that are often invisible to the naked eye during manual climbs.
- Generalization: Beyond telescope inspection, the Cross-Fusion block improved performance on standard benchmarks like MS COCO and VisDrone, as well as specialized tasks like pavement crack detection, with negligible computational overhead (~0.1% increase in FLOPs).
Conclusion and Future Directions
This automated system represents a significant leap forward in the maintenance of large-scale scientific infrastructure. By combining high-precision drone hardware with an adaptive attention-based detection model, the researchers have created a reliable, time-efficient solution for FAST.
Future work aims to transition from offline processing to real-time, onboard edge computing. Optimizing the algorithm for embedded hardware would allow the drone to perform "active sensing"—deviating from its flight path to investigate suspected defects autonomously—and potentially paving the way for a fully autonomous "detect-and-repair" cycle.
Summary updated