MODWT, PCA and Decision Tree based Fault diagnosis of Gear

A gearbox is an important power transmission equipment. Its maintenance is a top requirement because it is prone to a variety of failures. For gearbox fault diagnosis, techniques such as vibration monitoring have been widely used. Also, when it comes to machine Condition monitoring and fault diagnostics, feature extraction is the crucial step. For a classifier to perform accurately, it must have the appropriate discriminative information or features. Hence, this paper proposes a signal processing methodology based on Maximal overlap discrete wavelet transform (MODWT) and a dimensionality reduction technique i.eprincipal component analysis (PCA) to reduce the dimensionality of the feature space and obtain an ideal subspace for machine fault classification. Firstly, the raw vibration signature is denoised with the help of a state-of-the-art MODWT signal processing technique to identify the hidden fault signatures. Then various traditional statistical features are extracted from this denoised signal. These multi-dimensional features are then processed with PCA and further, the Decision Tree is used for fault classification. Performance comparison of the proposed method with traditional raw analysis and without application of PCA is presented and the proposed method outperforms at every level.