Anomaly Detection of a Reciprocating Compressor using Autoencoders

Abstract

This study introduces a novel approach for early fault detection using an autoencoder under time-varying conditions of a reciprocating compressor. The main strategy of this unprecedented method functions by combining a thermodynamic equation of compressor's discharge temperature with sensors' data to increase the prediction accuracy. This equation enables the model to identify the relationships between variables including the temperature, pressure and molecular weight of gas, thus alleviating the problem of poor data quality. Energy spectrum of vibration signals in the frequency domain was also used as additional features. The model was trained to recognize normal operations with 5-year data sampled every one minute. Two months before a machine shutdown was considered as abnormal period, of which the model wanted to identify it. The result suggested that the model can differentiate between normal and abnormal operations by a substantial margin.