Analyse schaltender und driftender Dynamik mit neuronalen Netzen

In this work, a method for unsupervised segmentation and identification of time series is presented. It can be used to analyze dynamical systems with nonstationary switching or drifting behavior { a phenomenon observable in many natural, real-world domains. As examples we examine the dynamics of speech and physiological systems, but also many other complex systems, e.g. the financial markets, are expected to show such type of behavior. The ansatz is based on hard or soft competitive learning and a competitive advantage for temporally adjacent data points in a time series. The competition is performed by a set of (neural network) predictors that compete for the prediction of the data points. Each predictor is capable of predicting only stationary dynamics. Therefore, the predictors specialize during a training phase on those segments of the data where the dynamics is stationary to a large extent. This is achieved by means of a low-pass filter on the prediction errors, which can be motivated by the assumption of a low switching rate between different dynamical modes. A continuously alternating dynamics can then be modeled by a drift segmentation algorithm, which is also presented in this work. An application to physiological wake/sleep data (EEG, EOG, respiration) shows that the proposed method detects dynamical changes with a high resolution. Moreover, it autonomously yields a segmentation into wake and sleep stages which is in good agreement with the manual segmentation of a medical expert