AI-enhanced predictive maintenance in hybrid roll-to-roll manufacturing integrating multi-sensor data and self-supervised learning

This paper proposes a new AI-assisted predictive maintenance framework for hybrid R2R manufacturing that combines multi-sensor data and self-supervised learning. It equips a multi-sensor Internet of Things (IoT) infrastructure to measure various conditions, such as temperature, vibration, and pressure, from the production line and employs Contrastive Predictive Coding (CPC) (a self-supervised learning model) for multi-sensor data representations without the scarce labelled data. This framework was tested on a lab-scale roll-to-roll chemical vapour deposition (R2R CVD) reactor system, and we show that the CPC model accurately predicts the underlying process dynamics and is capable of anomaly detection and failure prediction. Our system was able to acxshieve 96.2% accuracy in failure prediction with an AUC-ROC of 0.94 and an F1 score of 0.88. Our method was able to predict a process anomaly 12 min in advance, which could potentially be used to pre-empt a possible shutdown given enough time to intervene. Therefore, our framework not only predicts the failure before it occurs but also optimises the maintenance schedule, thereby improving robotic manufacturing efficiency and reducing downtime. The comparison between our method and the traditional predictive maintenance method shows that our method can predict failure earlier and give more advance notice when compared with the former in most cases. Finally, this study enriches the body of research on smart manufacturing by showing that integrating multi-sensor data with self-supervised learning can boost the effectiveness of predictive maintenance in hybrid R2R processes. It can be further expanded to achieve highly accurate predictions for other R2R manufacturing applications, which will help to improve production efficiency, reduce maintenance costs, and enhance product quality.