Abstract Accurate estimation of the Remaining Useful Life (RUL) is essential for optimizing cloud-centric fleet management services. However, existing deep learning approaches often lack the adaptability to process complex data streams from diverse assets in a multi-tenant cloud environment. To address these challenges, this study proposes PaTaNet, a Position Aware Temporal Adaptation Network designed for robust and efficient cloud-based RUL prediction. The architecture incorporates a semi-implicit positional encoding mechanism within the local temporal modeling path and an explicit sinusoidal encoding strategy within the global reasoning path. A dynamic fusion gate adaptively weights these branches for each sample, enabling the network to automatically handle the high variance typical of multi-tenant data. Extensive experiments on the CMAPSS dataset demonstrate that PaTaNet achieves competitive or superior results, verifying its efficiency and suitability for high-throughput cloud inference. To further assess its generalization capability for large-scale data lakes, we additionally evaluate the PaTaNet on the N-CMAPSS dataset, confirming that PaTaNet maintains stable predictive accuracy across different turbofan simulation platforms. These results validate that adaptive temporal representation and dynamic fusion significantly enhance the robustness and generalization of RUL prognostics under diverse industrial scenarios typical of cloud-based monitoring.
Wang et al. (Thu,) studied this question.