SWIFT-Mamba: A Lightweight State Space Model Toward In Situ Verification of Airborne Wavefront Sensing Signals

Unmanned Aerial Vehicle (UAV)-borne laser wavefront sensing technology holds significant application prospects for high-precision vibration detection in the field; however, the acquired signals are highly susceptible to complex, nonlinear environmental noise interference. Existing high-precision deep learning denoising models typically rely on massive computational resources, making them difficult to deploy on resource-constrained edge devices. Consequently, practical engineering exploration is often confined to an inefficient "blind sampling followed by offline processing" mode, incurring a high risk of data invalidation. To explore solutions for real-time quality control at the edge, this paper proposes a lightweight time-frequency state space model (SWIFT-Mamba), aiming to provide an efficient algorithmic foundation and an engineering proof-of-concept for portable devices moving toward in situ verification. Through rigorous evaluation on over 60,000 laboratory-measured and controlled synthetic wavefront vibration data samples, SWIFT-Mamba achieves an average Signal-to-Noise Ratio (SNR) gain of 19.64 dB and a Scale-Invariant Signal-to-Distortion Ratio (SI-SDR) gain of 16.10 dB, with an extremely low computational overhead requiring only 0.066 M parameters and 0.147 GFLOPs. Experimental results demonstrate that while significantly reducing computational costs, the proposed model can effectively extract the physical manifold of the signal and precisely preserve high-frequency phase features.