Learning from Irregularity: Continuous-Time State Space Models for Asynchronous Multimodal UAV Sensor Fusion

Precise multirotor UAV rotor fault diagnosis relies on multimodal sensor fusion, yet onboard avionics suffer from inherently asynchronous and irregularly sampled data streams caused by bus arbitration delays, heterogeneous sampling rates, and EMI-induced packet loss. Existing methods depend on forced interpolation, which destroys high-frequency micro-vibration signatures of incipient rotor damage. This paper proposes the Irregularly Sampled Fault Diagnosis (ISFD) paradigm, leveraging continuous-time state space models for end-to-end diagnosis directly from raw irregularly-timestamped sensor streams without interpolation. The architecture integrates a time-aware continuous embedding, an Ebbinghaus forgetting curve-driven SSM backbone with spectrally bounded projections, and an event-driven ODE fusion layer for asynchronous multimodal alignment. Validated on the first UAV fault dataset preserving raw microsecond-level timestamps from a custom quadrotor testbed, ISFD achieves 89.06% macro F1-Score, surpassing the best interpolation-aligned baseline by 13.74 percentage points. Under 70% EMI burst packet loss, all baselines collapse by 10–21 pp while ISFD improves by +0.68 pp, validating the continuous-time decay matrix’s physical regularization.