딥 언롤링을 이용한 오실레이터 기반 적응 노치 필터의 데이터 기반 파라미터 학습

초록

This paper presents a data-driven parameter learning approach for oscillator-based adaptive notch filters (OSC-ANF) using deep unrolling. Conventional OSC-ANF fixes the bandwidth parameter , causing a trade-off between convergence speed and steady-state accuracy: small yields fast convergence but poorer steady-state performance, while large improves accuracy but slows tracking under abrupt frequency changes. We reformulate the OSC-ANF iterations as an unrolled network and learn a time-varying via backpropagation. The loss combines frequency estimation error and an oscillator residual to preserve model consistency, and a Reset mechanism is added for rapid re-convergence after sudden frequency shifts. Simulations at SNRs of 0, 5, and 10 dB show improved tracking robustness across noise conditions with low computational complexity suitable for embedded implementations.

키워드