Data-driven remaining useful life prediction of rolling bearings via scattering transforms and long short-term memory networks

Remaining useful life prediction is an important metric in machinery prognostics where information on how much more life there is to a machine is predicted. The stages of a prognostic analysis are data collection, denoising, feature extraction, degrading zone classification and estimating the time to failure. Variational mode decomposition (VMD) is used to extract relevant modes from the data and remove unwanted frequencies or noise. The features are extracted through a non-linear unitary transform, which provides translationally and rotationally invariant, stable feature which delocalizes signal information into scattering decomposition paths. Support vector machine classifier is used to define the degradation zones in the data. A failure threshold is defined and the relative root mean square, a dimensionless measurement, assists in determining whether to begin making predictions. Long short-term memory networks are utilised to predict the remaining useful life and reconstruction error percentages are calculated for validation. Using PRONOSTIA bearing datasets, the suggested approach is verified, and its effectiveness is authenticated with that of alternative methods. The cumulative relative accuracy (CRA) scores show that the proposed method is an effective prediction methodology with an accuracy score of 89.00% and an average reconstruction error percentage of 28.00%. The findings demonstrate that the suggested approach is a reliable prediction methodology within the allowed error range and that it can be used in a variety of operational settings. • Modified VMD to suit bearing fault diagnosis in the current application. • New features capture bearing degradation better than classical statistics. • Utilising the scattering Transform algorithm for bearing fault detection. • Convert prognostic task to classification, then to RUL for improved prediction accuracy.