Smart audio sensor on anomaly respiration detection using FLAC features

In this work, we present a system to conduct anomaly detection on respiration sounds. The approach can adaptively model the mainstream of respirations as well as detecting the irregular patterns. Firstly, we employ the acoustic feature we proposed in previous: local auto-correlations on complex Fourier values (FLAC). The FLAC features take advantage of both magnitude and phase information and extract temporal dynamics in time and frequency domains, which are favorable for representing non-stationary respiration sounds. We adopt the online learning scheme to adaptively model the pattern in respiration sounds using candid covariance-free incremental principle component analysis (CCIPCA) method. Then, the mainstream acoustic patterns can be encoded by a subspace in online manner. Consequently, the anomaly acoustic respirations can be distinguished by their deviation distances to the updating acoustic pattern subspace. The experimental results demonstrate the effectiveness of the proposed framework with FLAC features.