Neural Network-Based Method for Jitter Suppression of Roll Angular Rate in Dual-Spin Projectiles

To tackle interference and noise in highly dynamic and nonlinear roll rate signals of dual-spin projectiles, this paper proposes a lightweight temporal de-jittering neural network with a hybrid attention mechanism. Empirical analysis of roll angular velocity data reveals time-domain attenuation and concentrated spectral energy in the frequency domain. A hybrid feature extraction scheme is designed to address irregular sampling intervals, where hybrid attention adaptively modulates feature weighting. A novel approach incorporates prior noise modeling into roll rate de-jittering, training the network to predict ground-truth values at window endpoints for effective denoising. To satisfy real-time demands, the network architecture is streamlined for computational efficiency. Experiments show that the proposed method outperforms wavelet and median filtering in both accuracy and speed. It achieves an 80.5% error reduction and a 96.1% improvement in signal smoothness. Furthermore, its inference speed meets real-time constraints, enabling precise and stable roll rate outputs for dual-spin projectile attitude control systems.