Unsupervised learning technique for surface electromyogram denoising from power line interference and baseline wander

We present a novel approach to single-channel power line interference (PLI) and baseline wander (BW) removal from surface electromyograms (EMG). It is based on non-negative matrix factorization (NMF) using a priori knowledge about the interferences. It performs a linear decomposition of the input signal spectrogram into non-negative components, which represent the PLI, BW and EMG spectrogram estimates. They all exhibit very different time-frequency patterns: PLI and BW are both sparse whereas EMG is noise-like. Initialization of the classical NMF algorithm with accurately designed PLI, BW and EMG structures and a carefully adjusted matrix decomposition rank increases the separation performance. The comparative study suggests that the proposed method outperforms two state-of-the-art reference methods.