Neural network based fault detection using different signal processing techniques as pre-processor

The purpose of vibration monitoring is to detect faults occurring in machinery, in order to maintain safety and minimize the breakdown cost. The authors of this paper monitored the condition of two meshing spur gears with the ratio of 1:2, where intentionally a gear fault (a welded blip) was introduced on the loaded driven gear. The signals obtained from the faulty gear and the good or reference gear were preprocessed by using three spectral analysis techniques: Fourier transform, Power Cepstrum, and Wavelet transform. For each type of preprocessing a separate artificial neural network was trained and tested to distinguish the faulty gear from the good gear. Although similar work has been done before, the authors of this paper has expanded the work on to the transient signals by using Wavelet on the whole transformation rather than the amplitude of the meshing frequency. In order to achieve this the whole transformation was discretized for the artificial neural networks (ANNs) inputs. This is different from the commonly practiced method which selects the meshing frequency band.