Extraction and classification of non-stationary acoustic signals via dynamic subspace filtering

This paper presents a dynamic, subspace based approach for extraction and classification of non-stationary acoustic signals under noisy conditions. The stationary subspace methods commonly used for noise removal take the whole signal into consideration while determining the signal subspace. This, for a non-stationary signal implies that spectral variations that occur through time are not taken into account. Thus, the overall signal subspace formulated will correspond to noise subspaces at certain points of time due to the highly non-stationary nature of the signal, resulting in noise leakage. The method proposed in this paper performs a subspace analysis dynamically at various stages of the observation period, enabling the signal subspace to evolve according to the non-stationary nature of the signal to be extracted. For signal classification problems, this enables the classifier to latch onto details about the non-stationarity in the signal which in turn provides valuable information with regard to the uniqueness of each signal class. Finally, an analysis is conducted on the resultant signals to prove the viability of the proposed methodology for signal extraction and classification problems under heavy noise conditions where the signals in concern are highly non-stationary.