The Source Separation of Multi-channel Vibration Signal Based on Nonnegative Tensor Factorization

Abstract

the traditional theory method of blind source separation needs priori knowledge, while in the absence of priori knowledge, how to separate the source signal from the composite signals of the power device, which is great significance in the fault diagnosis of mechanical equipment. Aiming at this problem, the separation method of the multi-channel vibration source signals based on nonnegative tensor factorization (NTF) is proposed. Firstly, the data of multi-channel vibration signals are used to construct the tensor of time-frequency by Short Time Fourier Transform (STFT), generally in the process of STFT the window length is set based on experience, aiming at the above problem, the optimal window length of STFT based on mean entropy both in the direction of time domain and frequency domain is proposed; Secondly, it is proposed that the optimal number of source signals is estimated by utilizing the index of the convergence error , the iterative steps and the nuclear consistency generated in the process of NTF; and then NTF is performed again according to the optimal number of source signals for the time-frequency tensor data; furthermore, the time-frequency distribution of the source signals is reconstructed by utilizing the matrix obtained by NTF, and the source signals are obtained by the inverse of STFT; finally, the fault frequency characteristic can be presented clearly after the envelope demodulation for the reconstructed source signals. The experimental results of the plunger pump verify the effectiveness of the methods mentioned above.