On Change Detection in the Complexity of the Time Series with Multiscale Renyi Entropy Processing

The objective of the paper refers to change detection in the structure of time series data, with direct application in fault detection of industrial processes and their components. The detection criterion is based on a combination of multiscale analysis technique with Renyi entropy on each scale, followed by a cumulative sum evaluation to estimate the point of change. Two case studies are considered. One considers a synthetized mixed signal, by mixing a deterministic signal with one having random components. The other one uses real vibration signals generated by some faults in the bearings of rotating machines. The method generates different values of the entropy for different structures of the time series and allows change detection of the incipient faults in bearings, i.e. small amplitude of the vibration signals, and changes in the structure and complexity of these signals. From fault detection point of view, the results are at the same level of qualitative performance with other approaches based on multiscale analysis and other type of entropies, e.g. Multi Scale Entropy (MSE). From change detection in the structure of the signals, the proposed method is superior, being more sensitive to random components and smoother. The proposed method could be adapted to other categories of difficult signals or processes, as those from medical area.