The Separation of Vibration Components Based on Sparse Nonnegative Tensor Factorization

Abstract

The vibration signals collected for machinery fault diagnosis tend to contain multiple excitation sources. Independent Component Analysis (ICA), the common approach of Blind Source Separation (BSS) to separate mixed source signals, is challenged when the data distributions are not known in advance. To overcome this problem, the source separation method based on sparse nonnegative tensor factorization (SNTF) is proposed. Firstly, the multi-channel vibration signals are constructed into a third-order time-frequency tensor by means of Short Time Fourier Transform (STFT). Then, the spectrum tensor is decomposed into three factors by performing Sparse NTF, and the source signals are reconstructed separately utilizing the different subspace. Eventually, the practical experiment on a two-stage gearbox indicates the performance improvement crediting to the sparseness constraint and verifies the effectiveness of the separation method.