The Fault Sources Identification Technology Based On Independent Component Analysis

The Independent Component Analysis (viz. ICA) can denoise and separate the useful messages out efficiently for fault diagnosis. But ICA limits the number of input signals equivalent to or more than the number of independent components. This paper firstly studied the ICA filtration character, revealed that SNR has no influence on ICA separation, compared two algorithms: FastICA and Infomax, and indicated that ICA cannot identify the phase. Secondly, the ICA application in the initial fault diagnosis of Gearbox Transmission System has been compared with the performances of traditional signal analysis methods. Finally, for the ICA limitation, this paper proposed the generalized definition of noise ICA and the additional virtual channels ICA denoise method. Under the premise of only a few sensor signals, this method replaces the other sensor signals by the reference signal, which have been saved when normal operation, to achieve the ICA separation. The experiments proved that the new separared signal, which not belong to the reference signals, just reflects the fault character. The method attained an acceptable effect on separating the turbine shaft vibration signals.