Trend-fluctuation correlated attention unit for remaining useful life prediction

Health indicators (HIs) are measures that consolidate complex machinery characteristics into a single dimension and have been extensively studied for predicting the remaining useful life (RUL). Recent research has proposed HI-prediction models based on recursive cell structures that demonstrate promising performance. However, recent cell structures have limitations, such as saturation in hidden states and difficulty learning interdependent patterns among various components. Accurate RUL predictions require minimizing error accumulation in the recursive HI-prediction process, which can be achieved using a cell structure capable of capturing interdependent patterns. This study proposes a new cell structure termed the trend-fluctuation correlated attention unit (TFCAU) that employs a gate mechanism to capture interactions between divided trends and fluctuations in HI and an attention mechanism for robust long-term predictions. The superiority of the TFCAU was validated using a bearing dataset from the PRONOSTIA platform and a real-world dataset from the Korean weapons system. Comparative experiments using six sub-datasets demonstrated that TFCAU outperformed the state-of-the-art model—the gated transient fluctuation dual-attention unit—achieving a 169.37% improvement in the score metric and a reduction of 49.93 units in the mean-absolute error. An extensive ablation study experimentally confirmed that the TFCAU effectively captured the interdependent characteristics between trends and fluctuations. TFCAU delivered excellent performance for both rotating and non-rotating systems, suggesting its potential applicability across multiple operational scenarios.