Machine fault diagnosis through an effective exact wavelet analysis

Continuous wavelet transforms (CWTs) are widely recognized as effective tools for vibration-based machine fault diagnosis, as CWTs can detect both stationary and transitory signals. However, due to the problem of overlapping, a large amount of redundant information exists in the results that are generated by CWTs. The appearance of overlapping can smear the spectral features and make the results very difficult to interpret for machine operators. Misinterpretation of results may lead to false alarms or failures to detect anomalous signals. Moreover, as conventional CWTs only use a single mother wavelet to generate daughter wavelets, the distortion of the original signal in the resultant coefficients is inevitable. Obviously, this will significantly affect the accuracy in anomalous signal detection. To minimize the effect of overlapping and to enhance the accuracy of fault detection, a novel wavelet transform, which is named as exact wavelet analysis, has been designed for use in vibration-based machine fault diagnosis. The design of exact wavelet analysis is based on genetic algorithms. At each selected time frame, the algorithms will generate an adaptive daughter wavelet to match the inspected signal as exactly as possible. The optimization process of exact wavelet analysis is different from other adaptive wavelets as it considers both the optimization of wavelet coefficients and the satisfaction of the admissibility conditions of wavelets. The results obtained from simulated and practical experiments prove that exact wavelet analysis not only minimizes the undesirable effect of overlapping, but also helps operators to detect faults and distinguish the causes of faults. With the help from exact wavelet analysis, sudden shutdowns of production and services due to the fatal breakdown of machines could be avoided.