Deep Learning Based Normal Behaviour Modelling for a Cereal Double Screw Extruder

ABSTRACT Normal Behavior Modeling (NBM) is an artificial intelligence technique that is used to determine baseline-normalcy of an equipment or process in relation to predictive maintenance and anomaly detection. The Cereal Double Screw Extruder (CDSE) stands notably as a critical and sophisticated Equipment for a Cereal Processing line. The ability to accurately understand the machine performance at any given time is key to sustainability of the equipment and the overall process. For a Deep Learning (DL) NBM using SCADA time series data, RNN and their variants were used. The three architectures of interest were RNN, RNN-LSTM and RNN-LSTM-DAE. The lowest overall reconstruction loss was achieved with RNN-LSTM-DAE with overall reconstruction loss of 1.0007 x 10-4. The same model also had the lowest reconstruction losses of Critical Response Variables (CRVs). The Model, with given thresholds of CRVs of 0.04, 0.02, 0.02 and 0.1 for Main Motor Amperage, Gearbox Temperature, Barrel Section_2 Temperature and Barrel Section_3 Temperature CRVs respectively, was able to pick all known anomalies. Key Words: Cereal Double Screw Extruder, Critical Response Variables, Normal Behavior Modeling, Anomaly detection, data reconstruction, reconstruction loss, deep learning.