Unsupervised Anomaly Detection in Aviation Piston Engines Across Multiple Flight Conditions Using Stacked Deep Auto-encoders

In recent years, anomaly detection has gained significant importance in general aviation airworthiness research. As the primary propulsion system for general aircraft, the aviation piston engine (APE) plays a critical role in ensuring flight safety. However, current research predominantly focuses on anomaly detection in large aviation engines, such as turbojet, turboshaft, and turbofan engines, with limited attention given to APEs. Moreover, the diverse operating conditions of APEs present additional challenges that must be addressed in anomaly detection process. This paper proposes a stacked Deep Auto-Encoder approach for Multi-flight condition anomaly detection (MDAE) specifically designed for APEs. The proposed methodology comprises three main stages: Initially, feature processing is performed on APE operational data collected through the Full Authority Digital Engine Control (FADEC) system. Subsequently, a clustering-guided approach is implemented for APE working condition identification, where APE features are subjected to K-means clustering, and aircraft flight phase information is integrated to refine the clustering results. Finally, stacked deep auto-encoders are employed to analyze APE data within each identified working condition. Experimental evaluations are conducted using four representative fault datasets collected during routine flight operations. Comparative analysis between the MDAE approach and traditional methods demonstrates superior detection performance, while some problems are also addressed for future improvement and optimization.