Defining, Analyzing, and Clustering Drive Cycles for Engine Applications Through Feature Engineering

Abstract

An ordered sequence of measurements, or time series, is the raw representation from which many analytics and data-driven decisions are made in the automotive industry. Volvo Penta is a company determined to become more data-driven to stay ahead of the competition. In this endeavor, signals from various manufactured engines in the field are logged as time series data. Through analysis of this data lies the information from which to obtain the repeated behavioral pattern of an engine in use, otherwise known as drive cycles. This thesis aims to segment time series data from an engine placed in a log stacker into drive cycles and engineer, extract, and select relevant features descriptive of engine usage from these. From the cycles with accompanying features, clustering techniques will be applied to group the drive cycles into behaviorally distinct categories. Drive cycles were defined by segmenting consecutive engine signals whenever the gap between recorded signal values exceeded ten minutes. Filtering was then applied to remove uninformative and outlier cycles. From the remaining cycles, features were derived, which, after selection, yielded eight temporal and statistical features that formed the basis for clustering. Using these features, the clustering algorithms k-means, agglomerative, and meanshift were employed, resulting in clusters that depicted two main engine behaviors. These were high- and low-load behaviors. The exception was the resulting clusters produced by mean-shift, which exhibited a single dominant behavior and a few deviating cycles. A consensus partition integrated all three methods and agreed with the solution provided by k-means. This thesis demonstrates that feature-based, unsupervised clustering can group drive cycles segmented from engine field test data into behaviorally distinct categories. These findings further the goals of Volvo Penta’s mission to become more data-driven and demonstrate the potential of clustering as a first step in areas such as engine simulation and predictive maintenance.