Abstract
Filter design is the essential problem in active noise control. In the case of FIR filter design, the whole problem can be formulated as a convex optimization. In practice, limited by computation capability, IIR filters are preferred but more challenging to design. On the one hand, the existence of zeros could cause instablity; on the other hand, the problem can not be formulated as a convex optimization problem. This work introduces a new IIR filter design method, which stablizes the system by modeling the ANC controller as casade parametric biquadratic IIRs and optimizes their cofficients by using machine learning algorithms. A neural network with zero input but learnable bias and network weights is used to generate the filter coefficients; and the loss function consists of a reward term for reducing noise and penalty terms for breaking the constraints. Training the neural network with gradient based machine learning algorithm directly leads to a group of coefficients that achieves good ANC performance and is subject to given constraints. It turned out that no big and deep neurual network is required for this method to perform well and therefore it is computationally efficient. Experiments showed that the proposed method was able to consistently yield good ANC filters for a variety of given acoustical environments.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have