Remaining useful life prediction with uncertainty quantification for risk-aware maintenance decision-making

Abstract

Remaining useful life (RUL) prediction is a key task in predictive maintenance; however, most existing methods emphasize point estimation accuracy, limiting their ability to characterize predictive uncertainty and support maintenance decision-making. To address this issue, this paper proposes RQDNet, an integrated framework that unifies RUL prediction, uncertainty quantification, and maintenance decision support. Specifically, a two-stage TCN-BiGRU temporal encoding architecture is employed to extract degradation features, while a temporal attention mechanism is introduced to enhance the representation of critical time steps. Based on this architecture, interval-valued RUL estimation is achieved via multi-quantile outputs, and a risk-aware maintenance strategy is further developed to translate prediction results into maintenance actions. Experiments on the public C-MAPSS and XJTU-SY datasets show that RQDNet achieves the lowest RMSE, MAE, and NASA Score among the evaluated methods, although the margins over the strongest recent Transformer- and graph-based baselines are modest. The proposed framework also provides a favorable balance between prediction-interval coverage and width, maintains quantile ordering consistency, and yields lower normalized maintenance cost, failure rate, and unplanned outages under the simplified validation-oriented maintenance protocol.