Sub-band anomaly detection and spatial localization

The modulating source assumption allows us to specify the problem of anomaly detection as a problem in robust statistics. The demodulation property of the Teager energy operator is used to jointly expose amplitude and frequency features of audio sub-bands. Spectral decomposition is achieved using a scaled wavelet function acting as a band-pass filter which halves its bandwidth for each decomposition level. Through successive transformation of variable, Teager energy of Laplace distributed sub-bands are redistributed into a Gaussian distribution for which parameters are estimated using the FAST-MCD algorithm. The robust Mahalanobis distances are used in the identification of minor anomalies. Tracked over a brief period, if an uncharacteristic change is detected in the context vector containing the robust sub-band means in units of the robust standard deviations, then the minor anomalies are relabeled as major anomalies and are used in a cross-correlation estimate from both channels of a two sensor array. The median of the resulting delay estimates in each sub-band provide the localization for the major anomaly.