Application of statistical based data mining techniques to operational data for ground vehicle diagnostics

More and more systems, both military and commercial, have the capability to record and archive operational and performance data. Many of these systems record and archive data from existing sensors and signals used to control platform operation or provide feedback to the operators. The challenge is to make the best use of these available signals to provide information on system health and capability, even though most of the signals and sensors were not originally selected based on their ability to provide information about system health. Furthermore, the recorded data is usually sampled at low sample rates and is often averaged or aggregated before archival. For a demonstration program, data was collected from several U.S. Army vehicles and this data was used for this analysis. The focus of the research described in this paper was to quantify the normal range for each sensor output as a function of engine mode. This paper describes the application of three data mining techniques to evaluate the operational range and detect anomalies in the sensor data, which was collected over a six month time period that the vehicles were operated. The first approach, ANalysis Of Means (ANOM) focuses on characterizing normal variances in sensor signals within a single platform. Time series data from sensors are analyzed on a vehicle-by-vehicle basis to determine if any sensors within a particular vehicle are responding in an anomalous manner relative to the same or similar sensors across the group of vehicles. The second approach, ANalysis Of Variances (ANOVA) compares variances in sensor signals across a group of vehicles to detect anomalous sensor behaviors. The third approach uses clustering techniques from pattern recognition to identify vehicles with similar behavior and can identify vehicles whose behavior is anomalous compared to the other clusters. Each data mining technique is described along with preliminary results from the resulting data analysis.