Consecutive Threshold Learning for Interpretable RUL Forecasting in Industrial Time Series

Industrial remaining useful life (RUL) prediction is often affected by sudden spikes and irregular signal changes that make degradation patterns unstable. To solve this problem, this study presents a consecutive threshold learning (CTL) method that can find and remove short-term spikes while keeping the main trend of equipment wear. The approach uses a simple time-based prediction model built on TC-WaveNet and adjusts the spike level with dynamic thresholds that follow signal energy and time order. Tests on the C-MAPSS and other industrial datasets showed that CTL cut false spike alarms by 27% and reduced mean absolute error by 15% compared with normal deep models without thresholds. In addition, the model provides gradient-based maps that show how faults develop over time, making the results easier to understand for engineers. Overall, CTL improves the stability and clarity of RUL prediction and can be applied to real-time condition monitoring and maintenance planning in industrial systems.