Machine Learning for Predicting Production Disruptions in the Wood-Based Panels Industry: A Demonstration Case

Abstract

In this paper, we study the application of Machine Learning (ML) in detecting and predicting Ahead-of-Time (AoT) production disruptions in a Portuguese Wood-Based Panels Industry. Assuming an Industry 4.0 concept, the analyzed ML classification task presents several challenges, such as a high number of Internet of Things (IoT) sensors, high-velocity data generation and extremely imbalanced data. To solve these issues, we adapt and compare five state-of-the-art ML algorithms for anomaly detection. Moreover, we pre-process the big data and employ a Selective Sampling (SS) technique to train and test computationally efficient ML models. Overall, high-quality results were obtained by an eXtreme Gradient Boosting (XGBoost) model, both in terms of detection and AoT prediction of production stoppages. Finally, we applied an eXplainable AI (XAI) technique based on sensitivity analysis to the XGBoost model, enabling the understanding of the impact of the sensor inputs on the disruption condition.