NULL-SPACING PREPROCESSING FOR UAV CLASSIFICATION FROM RF IQ SIGNALS UNDER ISM-BAND INTERFERENCE
Main Article Content
Abstract
Unmanned aerial vehicle (UAV) classification from radio-frequency (RF) in-phase/quadrature (IQ) signals in industrial, scientific, and medical (ISM) bands is often degraded by background noise and co-channel interference. This paper investigates Null-Spacing as an IQ-domain preprocessing step that attenuates structured interference before short-time Fourier transform (STFT) or continuous wavelet transform (CWT) representation. The interference subspace is estimated from Wi-Fi/Bluetooth samples in the training set, and each IQ signal is projected onto its orthogonal complement. The method is evaluated on the CardRF dataset with five UAV classes under additive white Gaussian noise (AWGN), Wi-Fi, Bluetooth, mixed interference, and jointly negative signal-to-noise ratio (SNR) and signal-to-interference ratio (SIR) conditions. With the STFT + convolutional neural network (CNN) configuration at η = 0.99, Null-Spacing improves macro-F1 from 82.82% to 84.56% and accuracy from 82.95% to 84.61%. Under negative-SNR AWGN, macro-F1 increases by 9.165 percentage points. Control experiments further show that the gain does not come from arbitrary projection alone, but from an interference subspace derived from Wi-Fi/Bluetooth references. The results support Null-Spacing as a conditional preprocessing step: it improves robustness under several difficult noisy and interfered conditions, but is not uniformly beneficial across all signal conditions