Fault detection mechanism of a predictive maintenance system based on autoregressive integrated moving average models

Abstract

The industrial world is amid a revolution, titled Industry 4.0, which entails the use of IoT technologies to enable the exchange of information between sensors, industrial machines and end users. A major issue in many industrial sectors is production inefficiency, with process downtime representing a loss for companies. Predictive maintenance, whereby maintenance is performed only when needed and before a failure occurs, has the potential to substantially reduce costs. This paper describes the fault detection mechanism of a predictive maintenance system developed for the metallurgic industry. Considering no previous information about faults is available, learning happens in an unsupervised manner. Imminent faults are predicted by estimating autoregressive integrated moving average models using real-world sensor data obtained from monitoring different machine components and parameters. The models’ outputs are fused to assess the significance of an anomaly (or anomalies) along the time domain and determine how likely a fault is to occur, with alarms being issued when the prospect of a fault is high enough.