An enhanced CNN framework for accurate RUL prediction with uncertainty quantification

Remaining Useful Life (RUL) prediction is a critical component of Prognostics and Health Management (PHM). It aims to predict the reliable operational time of equipment in the future by monitoring and analyzing its current state. This capability enables enterprises to enhance operational efficiency, reduce downtime, optimize maintenance decisions, and lower costs while ensuring safety and reliability. As a key technological approach in modern industrial equipment management and intelligent maintenance, RUL prediction has garnered crucial attention in both academia and industry. Despite the substantial progress achieved in RUL prediction techniques, there remain momentous challenges and limitations in current research and practical applications. The most prominent issue lies in the lack of uncertainty quantification (UQ). Existing methods predominantly focus on point estimation, overlooking the inherent uncertainties in the prediction process. To address this issue, this paper proposes an enhanced convolutional neural network with UQ (ECUQNet). The model first generates a point estimation distribution of the prediction results, which is then passed through an UQ module to obtain the probability density distribution of the predicted points. This enables the evaluation of prediction reliability. Experimental results on the C-MAPSS dataset demonstrate that the proposed method achieves competitive RMSE performance, highlighting its effectiveness. Furthermore, the use of UQ validates the robustness and reliability of ECUQNet.