A Novel Algorithm for Identifying Patterns from Multisensor Time Series

Pattern identification from multisensor time series is an important problem in many measurement, detection, and monitoring related applications. This paper introduces a generic approach to detect varying-length patterns and identify their class using predefined templates. In reality, the measured phenomena representing the same class can occur in slightly different ways which makes the resulting patterns vary in "shape" and time. A template-based approach calls for a sound processing of the manifold sensor signals in order to perform the needed comparisons reasonably. We use a template-specific quantization to normalize the signals and to manage their fluctuations. Dynamic time warping (DTW) handles their temporal variations. Our algorithm detects multiple patterns from a single DTW matrix in a computationally efficient way, without windowing. The background of this paper is in aeronautical fatigue research, and we evaluate the proposed algorithm with flight maneuver identification from flight monitoring data.