Active noise control with on-line estimation of non-Gaussian noise characteristics

Active noise control (ANC) is a methodology for attenuating noise based on adaptive signal processing algorithms. ANC is well assessed for the attenuation of Gaussian noise, but the rejection of non-Gaussian impulsive noise signals represents a much more critical task that may even impair algorithm convergence. To overcome this problem the adaptive filter weight update process must be modified by discarding or discounting samples associated with impulsive noise. This can be done either by modeling the impulsive noise with a non-Gaussian distribution such as the Symmetric alpha-stable (SaS) distribution or by applying outlier detection method. With both approaches the accuracy in the noise description appears to be crucial for effective noise reduction. This paper proposes two novel approaches for the attenuation of impulsive noise both for invariant and time-varying noise distributions. The first one is based on the on-line estimation of an SaS model of the noise probabilistic description. The second relies on a simple on-line recursive procedure that reliably estimates amplitude thresholds for outlier detection. Both methods compare favorably with competitor approaches, while maintaining a sufficiently low algorithm complexity. Several examples are shown to demonstrate the algorithms' effectiveness.