Adapting an MSE controller for active noise control to nonstationary noise statistics

In feedforward active noise control, a primary noise signal is used to generate control signals via a set of loudspeakers. In problems where many loudspeakers are used, it becomes difficult to adjust the controller filters to fast changing spectral properties of broadband primary noise signals. Many parameters then need to be readjusted simultaneously, which limits the tracking performance of e.g. filtered-x LMS adaptation algorithms. Here we propose and evaluate two feedforward control methods based on linear quadratic gaussian control that adapt to the often time-varying statistical properties of the feedforward noise signal, in a partly indirect adaptive design. For both proposed methods, the time is divided in batches and the control laws are updated for each time batch based on repeated estimates of the noise statistics. In the first method, the estimates of the noise statistics are incorporated into the controller, which is updated for each batch, whereas in the second method the controller is kept constant and a predictor for the noise is updated. The first method seems promising, and shows a gain in attenuation of about 5 dB over a controller that disregards the feedforward noise statistics. The second method, however, shows no such advantage.