Reliability assessment of remaining useful life predictions of roller bearings using deep learning models

This study focuses on an accurate and reliable prediction of the remaining useful life (RUL) of the bearings. In recent times, machine learning based RUL predictions have gained traction to ensure safety, efficiency, and cost-effectiveness of industrial machinery. However, raw vibration signals acquired from run-to-failure experiments are typically noisy, non-stationary, and high-dimensional, making direct learning of degradation patterns challenging. This often undermines the reliability of the prediction due to inconsistency in the RUL predictions at a given time. This study proposes a reliable multi-stage prognostic framework that integrates robust signal pre-processing, data-driven health-state classification, and deep learning-based RUL estimation. The proposed pipeline employs a piecewise aggregate approximation for dimensionality reduction, a singular spectrum analysis for noise suppression, and a central-moment-based change point detection to automatically identify the onset of degradation. Subsequently, a health index is constructed to characterize the degradation process and useful service life. An extreme gradient boosting classifier is adopted to distinguish between healthy and degraded operating states of the bearing using vibration response. Later, a hybrid convolution with gated recurrent (CGR) neural network is proposed to predict the RUL. The method is validated using multiple run-to-failure bearing datasets with different operating conditions. The corresponding results show that the classifier achieves average F 1 score of approx. 0.993 with reliability indices > 2.0 , indicating a clear identification of health-state. The proposed CGR network exhibits attractive performance, with high regression accuracy in RUL prediction. The efficacy of the proposed network is also evaluated against state-of-the-art methods, including bidirectional long-short-term memory recurrent neural networks, gated recurrent units, causal dilated convolution-based residual densenet with channel attention and temporal convolutional network with transformer. Overall, the results confirm that the proposed framework enables accurate, computationally efficient, and reliability-informed RUL predictions, making it suitable for real-time condition monitoring and predictive maintenance applications in rotating machinery.