Anomaly Detection for Coal-fired Boiler via Adversarial VAE Based Method

As the most important part of thermal power plant, a sudden shut down of a coal-fired boiler may cause a huge loss of the operating revenue. Thus, the coal-fired boiler needs to be monitored timely and comprehensively to enhance its stability and reduce shut down times. Today, it has become the norm rather than the exception to use learning-based anomaly detection methods to identify faults in a real time. The various components of a coal-fired boiler are typically monitored and generating massive multivariate time series. However, due to the complex patterns and little useful labeled data, it is a great challenge to detect anomalies from these time series data. Conventional methods generally require a large number of labeled data, but cannot guarantee high accuracy because training sets contain sparse labeled anomaly samples in most cases. In this paper, we propose an anomaly detection framework using adversarial variational auto-encoder (VAE). We first use two discriminators to adversarially train an auto encoder to learn the normal pattern of multivariate time series, and then use the reconstruction error to detect anomalies. Furthermore, in order to assist operators better diagnose anomalies, we provide anomaly interpretation based on reconstruction error of univariate time series. We con-duct experiments on the dataset collected from a coal-fired boiler. The results show that our method significantly outperforms other state-of-the-art methods.