On the Identification of Decision Boundaries for Anomaly Detection in CPPS

In this work, a new data-driven approach to anomaly detection in Cyber-Physical-Production-Systems (CPPS) is developed, which uses the geometric structure non-convex hull to build a decision boundary for the classification of new observations. A novel algorithm based on Mixture-of-Experts model is presented to estimate n-dimensional non-convex hulls. Furthermore, a new method is proposed to solve the point in non-convex hull problem. With these two methods in hand, a novel algorithm for anomaly detection is developed. Compared with convex hulls based anomaly detection methods, our approach can handle data sets with arbitrary shapes. Since the presented geometric approach does not make any assumption about either the probability density or the structure of given data, it can be combined with different machine learning algorithms. The effectiveness of this approach is evaluated with real world data collected from wind turbines.