An algorithm for classification and outlier detection of time series data

This algorithm treats a time series as an image and segments the image into clusters of interest, such as "nominal data" and "failure mode" clusters. The algorithm uses density clustering techniques to identify sequences of coincident clusters in both the time domain and delay space, where the delay-space representation of the time series consists of ordered pairs of consecutive data points taken from the time series. "Optimal" clusters that contain either mostly nominal or mostly failure-mode data are identified in both the time domain and delay space.

A best cluster is selected in delay space and used to construct a "feature" in the time domain from a subset of the optimal time-domain clusters. Segments of the time series and each datum in the time series are classified using decision trees. Depending on the classification of the time series, a final quality score (or quality index) for each data point is calculated by combining a number of individual indicators.