Multilevel Approach to Sensor Streams Analysis

Monitoring of the systems, which are described with numerous time series, can be a complex task. Too much data is difficult to follow even by an expert human user. In presented work we focused on understanding dynamics of such complex systems and presenting the results in a humanlycomprehensible way. Instead of following the dynamics of a system through numerous time series, the result of our methodology is a directed state graph (see Figure 1) equipped with corresponding transitional probabilities. To achieve such a result we extract different features (markers) from the time series, aggregate them in a sliding window and pack them into state vectors. We perform clustering on top of a set of such vectors and calculate transitional probabilities between the clusters.