Source separation using independent component analysis techniques for machine fault detection in the presence of background noise

Early and accurate fault detection on a machine is crucial in preventive maintenance in order to prevent accidents that may be catastrophe to the user. However, direct measurement using vibrometer is common usage which may impractical in an industry that uses many rotating machines. In this paper, we evaluate the independent component analysis techniques which is time-, frequency-domain and multistage ICA for remote condition monitoring by analyzing sound emitted from the machines. We used electrical water pumps with normal, and intentionally introduced faults to the pumps with unbalanced, misaligned and bearing defect. These machines worked simultaneously and then recorded in an anechoic chamber to obtain the baseline data and then at an open area to simulate the real plant situation. We assumed that the sounds mixed convolutively at which required microphone array as sensor as an interface before separation. The results suggest that the proposed technique performed accurately in mean-square-error (MSE) sense. This implies that the proposed technique may be suitable for further implementation in real plant setting with adverse environment to replace current technique using direct measurement.