Asynchronous Dual‐Channel TCN for Real‐Time UAV Sensor Health Monitoring Under Resource Constraints

The perception, localization, and reliable fusion of unmanned aerial vehicles (UAVs) in industrial inspection are strongly influenced by sensor health. Existing studies mainly address class imbalance and limited fault data, but rarely consider how limited onboard resources affect real‐time fault diagnosis algorithms for UAVs. From a flight safety view, we propose a real‐time health monitoring method for asynchronous sensor systems, the Asynchronous Dual‐Channel TCN (ADC‐TCN). First, we design a dual‐channel architecture with an adaptive gating mechanism to ensure diagnostic speed and accuracy in both offline training and online deployment. Causal, dilated 2‐D temporal residual blocks are proposed to capture couplings between the time and feature axes across the fast and slow channels. Second, we introduce a cross‐attention module for efficient cross channel fusion. Finally, to address the mismatch between offline training and online deployment, we introduce a stochastic channel deactivation training strategy and design a combined loss function to ensure efficient training. We evaluate ADC‐TCN offline and on an NVIDIA Jetson Orin Nano against ten baselines. The method achieves detection and classification F1 scores above 93% with a per‐inference latency of 1.78 ms, and an energy cost of 12.6 mJ per inference, demonstrating reliable performance under asynchronous updates and on‐board resource constraints.