Predictive maintenance (PdM) plays a key role in improving the reliability and efficiency of industrial systems by anticipating abnormal operating conditions through data-driven approaches. In this context, many PdM solutions rely on temporal windowing strategies applied to multivariate time series; however, the choice of observation and prediction horizons is often fixed a priori and rarely analyzed in a systematic way. This paper investigates the impact of temporal window design by formulating alarm prediction as a window-based classification problem, explicitly distinguishing between the observation horizon and the prediction horizon. A comprehensive experimental study is conducted on three real-world industrial datasets characterized by different process dynamics. Multiple machine learning and deep learning models are evaluated across a wide range of combinations of observation and prediction horizons. The results show that predictive performance is strongly influenced by the joint configuration of observation and prediction horizons, rather than by their individual selection. While increasing the prediction horizon generally reduces accuracy, the magnitude of this effect depends on system dynamics. Similarly, longer observation horizons substantially improve performance in processes with slow dynamics, while providing limited benefits in faster-evolving systems. Overall, the study highlights the importance of explicitly modeling temporal window parameters in predictive maintenance design, providing both methodological insights and practical guidelines for selecting appropriate temporal configurations in industrial alarm prediction tasks.
Caterino et al. (Mon,) studied this question.