Companies continuously must optimize their maintenance policies: to reduce equipment downtime, ensure the safety of the systems, and decrease maintenance costs, which can represent a significant portion of production expenses. Predictive maintenance is a policy able to address these needs thanks to the support of machine learning algorithms that could be adopted to predict anomalous conditions. Due to the large number of components that make up an industrial plant, these algorithms can sometimes become complex or, conversely, may fail to detect some complex relationships among operating parameters. This study presents preliminary results of a multi-stage approach in which two classification algorithms are put in series, aiming first to detect potential anomalies in components and, if none are found, to monitor the clogging levels of the main component in the technological process. The model was applied to a cyclonic bag filter equipped with a fan. The first model evaluates some anomalies in the fan, and subsequently, the second model monitors the clogging levels of the filter bags. Preliminary results showed performance close to 85%, while additional parameters will be introduced in the future to expand the number of monitored components and to estimate the remaining useful life of the filter bags.
Suppini et al. (Thu,) studied this question.