MGSD-Net: A multi-geometric signature diagnostic network for rotor fault diagnosis under irregularly sampled inertial streams

This article addresses propeller fault diagnosis from irregularly timestamped on-board unmanned aerial vehicle (UAV) sensor streams. Conventional baselines force the asynchronous streams onto a uniform grid via zero-order hold (ZOH), linear, cubic, binning, or backward-fill interpolation, which either dampens the high-frequency vibration spectrum (sinc-squared attenuation) or fabricates Runge-type overshoots that corrupt downstream attention statistics. We argue that the appropriate variable for irregular avionics data is the rough path itself rather than the resampled tensor: its truncated path signature is invariant to time reparametrisation and encodes inter-channel multi-scale interactions in a single algebraic descriptor. We propose MGSD-Net, a multi-geometric signature diagnostic network that consumes raw timestamped streams, computes signature features under four parallel geometries (global-backward, global-forward, local-tight, local-wide), and feeds the concatenated features into a bidirectional Transformer encoder for end-to-end damage classification. On the UAV-Realistic propeller-damage benchmark, a focused five-seed clean audit gives MGSD-Net a macro-F1 of 82.15 ± 2.57% directly on raw streams, compared with 71.13± 3.06% for the audited deployable temporal convolutional network (TCN) × ZOH baseline, for a mean gap of +11.02 percentage points. The worst MGSD-Net seed is 77.81%, which remains above the best TCN seed of 73.44%, and evaluating TCN with the same small batch size does not close the gap. In a focused five-checkpoint burst-loss audit against TCN × ZOH, MGSD-Net loses 6.85 pp at 70% burst drop, whereas TCN loses 16.35 pp, and MGSD-Net remains ahead by +20.40 pp at that endpoint.