A data mining approach to reveal patterns in aircraft engine and operational data Srini Sundaram, Iain G.D.Strachan and David A. Clifton Oxford BioSignals Ltd. 174, Milton Park, Abingdon, Oxfordshire, OX14 4SE +44 (0) 1235 433574 srini.sundaram@oxford-biosignals.com Steve King, John Palmer, Rolls-Royce plc, EHM Global Capability Group, PO Box 31, Derby, DE24 8BJ. Abstract Modern day aircraft engines are embedded with sensors acquiring vast amounts of data related to engine performance such as pressure, temperature, efficiency, and speed during takeoff, climb and cruise stages of flight. These performance data are stored in performance reports. In addition, the maintenance actions across the fleet, service history and observations during flight are logged into service reports and pilot logs respectively. A typical fleet-wide performance dataset can be very large. Efficient operation and maintenance of a fleet requires proper use and analysis of the available data, searching for potentially useful trends within large datasets. Such trends may allow a fleet specialist to establish engine behaviour profiles under different conditions, and may provide indication of abnormalities in engine condition, or of hazardous events. Currently, fleet specialists use various data analysis tools to identify and analyse abnormal behaviour. Data is typically obtained from multiple sources, making this is a complex challenge. To address this challenge, this paper introduces a data mining tool capable of assisting fleet specialists by searching for useful patterns in large datasets, generating reliable, timely alerts when “abnormal” patterns are identified. 1. Introduction Fleet management requires decision support tools to provide indication of potentially abnormal events during engine operation. The information is usually maintained in performance or service reports containing data from a fleet of engines. These datasets are typically multivariate and examples of events are rare compared to the quantity of available data. This paper describes the application of two techniques to the analysis of large fleet-wide datasets: Visualisation techniques are used to project such high-dimensional data into two dimensions for visual inspection providing knowledge about structure in the high- dimensional dataset, which can inform the process of constructing a model of normal system behaviour (1) as required for automated novelty detection.