Design of Feedback Active Noise Cancellation Filter Using Nested Recurrent Neural Networks

We examine the problem of Feedback (FB) Active Noise Cancellation (ANC) and propose a novel DSP-inspired neural network for designing FB ANC filters. It reinterprets the ANC problem as a neural network optimization issue, thus allowing the use of modern machine learning tools. At the same time it also benefits from classical DSP techniques as it retains and extends the state of the art for efficient generation of anti-noise. The proposed approach leverages machine learning algorithms to learn stochastically optimal FB ANC filter coefficients under a variety of conditions and constraints that make the problem intractable for classical methods. One challenge with FB ANC filter design is to guarantee the stability of the IIR filter structure as well as the feedback loop around it. The proposed method meets this stability requirement while providing new avenues for improved system efficiency and adaptability. We demonstrate its effectiveness through simulations using real-world models.