Predictive Maintenance Under Absence of Sensor Data

In industrial settings, component breakdowns can cause production delays, until repaired or replaced, and incur high costs. To address this issue, many industries have adopted predictive maintenance, which is an approach that combines machine learning (ML) and condition monitoring sensors to estimate when the equipment is likely to fail. This allows for early repairs and efficient maintenance scheduling, reducing maintenance costs and downtime. However, installing the necessary sensors can be a significant undertaking for a large company with many industrial machines. To reduce the installation costs and labor required, we investigated an intermediate solution for estimating the remaining useful life (RUL) based only on construction data and their running lifetime. This paper examines how treating RUL estimation as a classification task (i.e. calculating the likelihood of breaking within a period of time instead of its lifespan), increases the volume of available data and allows the employment of ML techniques, which have demonstrated satisfying performance in classification and regression tasks. This method also allows us to integrate additional construction information for each individual component, leading to an increase in the prediction accuracy. Our approach is applied on real-world data from a large production company, forecasting how many smelting pots will malfunction in the near future, resulting in a two-fold increase in accuracy over the company's previous statistical life usage model.