Flight record data-driven pitot tube fault feature extraction and fault prediction

Pitot tube failures pose a considerable risk to flight safety; however, current air data computers exhibit limited effectiveness in detecting such failures and lack predictive fault diagnosis capabilities. This study addresses this limitation by utilizing flight record data to explore methodologies for fault feature extraction and prediction specific to pitot tubes. Initially, ten fault-related parameters are constructed, and the random forest importance scoring algorithm is employed to select the most relevant parameters. Based on these, four time-domain features, including average value, are extracted. Subsequently, a fault prediction model is developed using a fully connected neural network trained on the extracted features. Case validation conducted on a fault dataset from ten aircraft confirms the effectiveness of the proposed fault prediction model.