Tri-sensor synergistic perception for aerial inspection of transmission assets via stacked attention and Pareto-aware co-optimization

Drone-assisted patrol of overhead lines has become standard utility practice, yet reliable recognition, robust under diverse weather conditions, of grid components from aerial imagery remains challenging. To narrow this gap, we introduce TriSense-Net, a perception framework that jointly exploits RGB photographs, long-wave thermal frames, and sparse LiDAR returns within a single end-to-end trainable pipeline. At its core, a heterogeneous encoder shares early convolutional stages across the three streams and then branches into modality-specific paths, while an affinity-driven alignment block re-balances inter-sensor contributions whenever any single channel becomes unreliable. Two complementary attention pathways — operating respectively along spatial coordinates and channel indices — are stacked in series, jointly amplifying scale-aware salient regions and discriminative filters while dampening sensor-specific noise. Training is further cast as a four-objective Pareto search coupling mean Average Precision, the number of false positives per image, on-board latency, and the train–test gap, in which an NSGA-III variant equipped with a cosine-scheduled inertia term steers the population toward solutions that are simultaneously accurate, computationally cheap, and transferable. For evaluation, we curated PowerTriplet-18 K, an in-house corpus of 17,847 spatially co-registered triplets gathered across twelve weather and illumination regimes. Under our protocol, TriSense-Net reaches 99.3% mAP@0.5 and 96.7% mAP@0.5:0.95, outperforming YOLOv8-X, Faster R-CNN with FPN, and MetaFusion by clear margins. A 26-week field trial along an 87.5-km, 220-kV right-of-way processed 127,853 triplets across 36 sorties at 42 ms per frame; of the 187 candidate anomalies surfaced by the system, line-crew inspection later confirmed 184 as actionable defects, corresponding to a field precision of 98.4% and a recall of 98.9%.