A Synthetic Approach to Validate the Usability of Anomaly Detection Model

Abstract

In recent years, air compressors for industrial use have played a vital role in a vast number of industrial applications. The ability to pressurize air has been a boon to productivity at factories. Use of the air compressor under unexpected situations for a long time will cause serious damage to the air compressor and disrupt the manufacturing process. Therefore, anomaly detection is critical in air compressor maintenance, one of the challenges for many researchers using machine learning to prognostic and health maintenance air compressors. In this paper, we propose excellent machine learning for anomaly detection in an air compressor. We utilize a one-class support vector machine(OCSVM) to deal with anomaly detection in time series data and unsupervised learning. To design the experiment, we employed two datasets: one from a Taiwanese company with no abnormal data and no domain expert for labeling, and another from the FEMTO-ST (PHM IEEE 2012) Institute with domain experts for labeling. To demonstrate the promising performance, we used the PHM dataset to design experiments on both synthetic and measured data.