Online Time Series Changes Detection Based on Neuro-Fuzzy Approach

Abstract

The problem of fault detection of time series properties is attractive for many researchers in different areas for a long enough time. The results of the solution of this problem are used in many areas, such as monitoring of the manufacturing processes, control of moving object, bioinformatics, medical diagnostics tasks, and video stream processing. Nowadays, a fairly large number of approaches are proposed for solving this problem. Among popular approaches, there are methods, which are based on statistical analysis of time series, mathematical models of objects that generate these time series, pattern recognition, clustering, and artificial neural networks. The situation is more complicated if the information is fed for processing in online mode, and changes of signal properties can have both abrupt type (faults, outliers, and anomalies) and enough slow drift. At that, these time series can be represented in the vector or matrix sequences form and have not only stochastic character but also chaotic one. In this case, the approach based on computational intelligence methods, first of all, the neuro-fuzzy models with online learning algorithms, can have the most effectiveness. In cases where changes in monitored objects have a smooth slow nature, and as a result, it is impossible to establish a crisp boundary between segments of time series. In this situation, the use of fuzzy clustering methods is effective. At the same time, since algorithms of fuzzy clustering (both probabilistic and possibilistic) are intended to operate in batch mode, their online modifications are proposed, which essentially present the gradient procedures for minimizing conventional fuzzy goal functions. Thus, neuro-fuzzy algorithms are proposed for the fuzzy segmentation of multidimensional time series, which allow detecting in a real time both abrupt and smooth changes in the properties of stochastic and chaotic sequences.