An Unsupervised Anomaly Detection Approach Based on Industrial Big Data

The development of condition monitoring system provides large amounts of operational data. These data present typical characteristics of multiple sources, polymorphism, diversity and mass, and can reflect the service quality and operating state of the equipment. How to use these data becomes one of the main problems in data analysis. To address the issue, this paper presents a generalized model for complex mechanical system anomaly detection based on the data collected from distributed control system (DCS). The Stacked Auto-encoder network is used to achieve the automatic extraction of the hidden features in multidimensional polymorphic data. The isolation forest (IF) method is used to achieve the anomaly detection, which only needs to use the DCS monitoring data during normal operation of the unit for network training and model fitting without fault data. And the method can realize the abnormal detection during the operation of the unit without the traditional signal processing method for feature extraction. The proposed method has been used for real compressor. The results show that the proposed approach can detect the anomalies without the need for fault data. And the proposed method is more effective than traditional methods.