Abstract
This study presents a forensic diagnostic framework aimed at enhancing the early detection, fault classification and remaining useful life (RUL) prediction of rolling bearing failures. The proposed network integrates a novel three-stage machine learning formulation - (1) identification of health state using voting ensemble, (2) prognostic analysis via a hybrid convolutional neural network and gated recurrent unit (CNN-GRU), and (3) fault type identification through the segment anything model (SAM) based on time-frequency representations. The ensemble and CNN-GRU models are trained on both time- and frequency-domain features from vibration signals, while SAM leverages this data in visual sense through iterative masking for zero-shot spatial-temporal fault segmentation. Pre-processing techniques, including piecewise aggregate approximation and singular spectrum analysis, are used to denoise and compress the vibration response without impacting key statistical traits. The proposed e-CNN-GRU-SAM network demonstrates better accuracy in diagnosing fault types, predicting RUL and identifying root causes under different operational conditions. This is established using diverse operating benchmark datasets that simulate induced and real-world degradation scenarios for generalization. Thus, the proposed framework offers a comprehensive forensic analysis toolkit for diagnosis and prognosis of bearings.