Fault diagnosis of gearboxes using feature extraction in time domain

In the vibration signals of a faulty gearbox, there exist not only the response excited by source of faults such as rotating frequencies of shafts, mesh frequencies of gear pairs, their higher harmonic components and modulated components, but also natural and random vibrations of system, components and parts, and noise interferences. Thus it is difficult to recognize and capture typical fault frequencies. For this purpose, an approach was proposed for fault diagnosis of gearboxes, by combining temporal feature extraction based on independent component analysis (ICA) and fault classification based on forward least squares approximation support vector machine (FLSA-SVM). Multiple ICA neural networks were batch trained by multi-classes vibration data separately, each corresponding to one of fault patterns of gearboxes. A feature extractor was structurally constructed by the trained ICA networks, followed by a post-processing unit of feature extraction based on one second-order measure, i.e. residual total correlation (RTC). Experimental results show that the proposed approach is efficient and effective for fault diagnosis of gearboxes, even if the used train sets are of small size. The integrated ICA-RTC feature extractor is able to extract low-dimensional representative features from time-domain data, and to overcome some difficulties in traditional frequency-domain feature extraction.