FFT-deep learning denoising method for resonant integrated optical gyroscope

The noise of a resonant integrated optical gyroscope (RIOG) remains a critical bottleneck for ultra-high precision applications. This paper presents a cascaded "physics-prior and data-driven" denoising framework: fast Fourier transform (FFT) comb filtering was first employed to suppress discrete spectral lines at 0 Hz, 50 Hz, and harmonics; a lightweight convolutional neural network (CNN)-long short-term memory (LSTM) hybrid network then deeply extracted spatiotemporal noise characteristics. The network architecture is composed of: 64-point sliding windows (step=2), two 1D convolutional layers (16×8 and 32×4) for local features, a 32-unit LSTM for long-term dependencies, and a linear output layer for noise regression. Experiments on 1 kHz-sampled, 2000000 points field data show that the proposed method reduces angular random walk (ARW) to 0.28 ^∘/h and bias instability to 2.52 °/h. These results show an approximately 50% reduction in both ARW and bias instability compared to our previously reported method, which marks a significant advance in gyroscopic noise suppression. This study demonstrates the strong generalization of "frequency-domain physical constraints and spatiotemporal deep learning" for high-performance gyroscope signal processing.