Abstract

An online 2-D changepoint detection algorithm for sensor-based fault detection, is proposed. The methodology con- sists of a differential detector which looks for characteristics across datasets at a particular instant, and a standard detector which when combined can identify anomalies and meaningful change- points while maintaining low rates of false-alarm generation. A key aspect of changepoint detection methodologies is the setting of relevant thresholds which are typically based on empirical trial and error. Here, a statistical methodology is adopted which provides the engineer with a trade-off between correct detection and false-alarm rates, thereby informing decision making at the design stage. The efficacy of the techniques is demonstrated through application to two industry case studies of fault detection on Industrial Gas Turbines, and are shown to readily provide an early warning indicator of impending failures.