Analysis of operational and mechanical anomalies in scheduled commercial flights using a logarithmic multivariate Gaussian model

This paper presents a machine learning approach to evaluate the performance of aircrafts using on-board sensor information on commercially scheduled flights with the aim to further improve system health monitoring strategies in air transportation. Logarithmic multivariate Gaussian models are trained to evaluate the performance of aircrafts at different flight phases (takeoff, ascent, cruise, etc.) separately. By including a forward synchronization, feature selection, and mini-batch training process, this model overcomes challenges introduced by the large size and high dimensionality of flight datasets. This framework also addresses the re-sampling issue in existing literature causing difficulties in handling time-series signals with different lengths. For demonstration and validation, the developed model is applied to analyze performance anomalies associated with the mechanical system and pilot operation in a historical flight dataset. Compared with existing literature focusing on similar datasets, this evaluation methodology shows promise in detecting performance anomalies especially at approach and takeoff phases. Therefore, the developed model is expected to be an effective addition to the current anomaly analysis and monitoring technologies for scheduled commercial flights. Applications include assisting transportation management systems by handling large amounts of historical flight datasets to analyze mechanical and operational anomalies, which may potentially improve future aeronautical system design and pilot training.