|
|
|
“A Signal Processing Approach to Robust Jet Engine Fault Detection and Diagnosis”
by Martin Gawecki
August 2014
As in any mechanical system, entropy is continually fighting our best efforts to preserve order. Engineers, mechanics, and pilots have all helped in the process of engine health management by perceiving and identifying faults in aircraft. The complexity of these systems has gradually increased, necessitating the evolution of novel methods to detect engine component problems. As airlines and manufacturers have begun to develop capabilities for the collection of ever more information in the age of "Big Data," an opportunity for such a method has presented itself to the signal processing community.
This work will address the development of reliable fault detection and diagnosis algorithms, built around the collection of various types of engine health data. Engine Health Management (EHM), has so far relied on rudimentary readings, the diligence of maintenance crews, and pilot familiarity with expected equipment behavior. While the majority of EHM advances are inexorably tied to the field of mechanical and aerospace engineering, signal processing approaches can make unique contributions in effectively handling the oncoming deluge of complicated data.
During the scope of this work, two broad approaches are taken to address the challenges of such an undertaking. First, the feasibility of vibration and acoustic sensors is examined in controlled experimental conditions to determine if such information is useful. This in turn will be used to develop modern detection/diagnosis algorithms and examine the importance of sampling frequency for EHM systems in this context. Here, ii this work offers several important contributions, chief among which are: excellent results for "stationary" phases of flight, a consistent fault detection rate for synthetic abrupt changes, fast responses to component failures in high-frequency data, and well-defined clustering for nominal samples in lower-frequency (1Hz) data.
Second, this work describes an improved Gas Path Analysis (GPA) approach that utilizes information from traditional sensors (pressures, temperatures, speeds, etc.) to produce relevant high-quality simulated data, develop a correspondence between simulated and real-world data, and demonstrate the feasibility of fault detection in these scenarios. Here, the chief contribution is the establishment of a close agreement between synthetically simulated faults and nominal data from real engines. Building on this, a reliable fault detection and diagnosis system for "stationary" and "transient" flight phases is developed, while adapting high quality simulated full flight data to low-frequency (1Hz) real world correspondences.