Publications

Abstract : Scattering transform is proposed using machine learning to extract translational, rotational and deformation invariant information for the first time from vibration signals obtained from rolling element bearings (REBs). The core idea of scattering transform lies in the construction of a scattering network which is formed from a stack of signal processing layers of increasing width. Each layer is formed from the association of a linear filter bank with a non-linear operator. It uses a cascade of wavelet filter bank, modulus rectifiers and averaging operators to build a deep convolution network and computes multi-scale co-occurrence coefficients which are invariant to translation in time, rotation and deformation. The scattering transform coefficients are extracted as features from seven stages of a vibration signal prognosis data repository which are then input to a support vector machine (SVM) classifier. Vibration signals from the intelligent management system (IMS) bearing data centre are used to validate the proposed algorithm. Test results analysis and solution show that scattering transform can be used to obtain distinguishing features from seven bearing health stages with an average accuracy of 99 %. The results were compared with other feature extraction strategies on the same data and were found to be superior.