Deep learning-assisted active noise control in a time-varying environment

The success of active noise control (ANC) is largely deTermined by the fidelity of the estimated secondary path, which encapsulates the “room acoustics” between the secondary sound source and the error sensor. In a time-invariant system the secondary path is usually measured and hard-coded in the controller prior to the ANC operation. When ANC is to be performed in a time-varying environment, however, the estimated secondary path should be updated accordingly, a task that poses many challenges in Terms of efficacy, cost, and user comfort. In this paper we present a deep learning-assisted secondary path update technique, in which deep neural networks are trained to estimate the secondary path in real time according to changing boundary conditions. The feasibility of the technique is tested in an airborne duct, where the error sensor is allowed to move along the duct to simulate changes in boundary conditions. Results have shown that even in the face of a dramatic change in boundary conditions, the ANC system equipped with the present update scheme is capable of reducing broadband noise by up to 10 dB.