Non-Negative Novelty Extraction: A New Non-Negativity Constraint for NMF
Yohei Kawaguchi, Takashi Endo, Kenji Ichige, Koichi Hamada · 2018
To develop a sound-monitoring system for checking machine health, we propose an algorithm for extracting novel sound (i.e. abnormal sound) from the monaural signal including the learned noise. Conventional approaches use non-negative matrix factorization (NMF) or semi-supervised NMF (SSNMF) for separating the novel sound and the learned noise. However, the conventional approaches assume the low-rankness of the novel sound, which is not necessarily satisfied. In addition, SSNMF has a hyper parameter that cannot be tuned i.e. the number of atoms corresponding to the novel sound. To solve the problems of NMF and SSNMF, our algorithm applies a constraint for the novel sound, i.e. only the non-negativity without low-rankness. Our algorithm can be interpreted as a simultaneous optimization version revised from the state-of-the-art non-negative matrix underapproximation (NMU). Experimental results indicate that the extraction performance of the novel sound is higher than conventional approaches.