A Data Analytic Engine Towards Self-Management of Cyber-Physical Systems

With the increasing complexity of cyber-physical systems, it is essential to enhance their self-management capabilities (e.g., self-protection, self-optimization). This paper presents a data-oriented approach to achieving that goal, given that a large amount of measurements can be collected in current systems. We investigate typical data characteristics in physical systems, and identify that the collected data from those systems exhibit a wide range of diversities. Following those observations, a new analytic engine is proposed and developed to extract knowledge from measurement data streams in physical systems. The engine treats each attribute in measurements as a time series and contains an ensemble of models, each attempting to discover a specific data property accordingly, such as periodicity, pairwise dependency and so on. Therefore time series are profiled based on their properties captured by engine models. The extracted data profiles can be further used to facilitate several management tasks of system status monitoring and online anomaly detection. Our experimental results in a real power plant have demonstrated that our analytic engine can correctly profile heterogeneous time series in the system, and successfully detect a number of abnormal situations in the system operation including some system inspection events as well as component faults.