Dynamic Time Warping for Phase Recognition in Tribological Sensor Data

This paper analyzes the potential of dynamic time warping (DTW) for recognizing phases of tribological sensor data. The three classes in these time series—run-in, constant wear, and divergent wear—are distinguished by their long-term trend and curvature. A set of reference data for each class is needed for the classification. Each time series in the reference set represents a typical shape of this class. The classification is done by computing the DTW between a given time series and each reference time series, and assigning it to the class with the minimum distance. In experiments on simulated and real-world time series, we show that DTW is capable of correctly classifying whole time series representing one class. Additional experiments are done to analyze the capability of DTW to classify a time series that is only a part of the entire time series representing one class. During these experiments, limitations arose that demonstrated the importance of the choice of good reference data.