Robust Data-Driven Fault Detection: An Application to Aircraft Air Data Sensors

Fault detection (FD) is important for health monitoring and safe operation of dynamical systems. Previous studies use model-based approaches which are sensitive to system specifics, attenuating the robustness. Data-driven methods have claimed accurate performances scale well different cases, but algorithmic structures enclosed operations “black,” jeopardizing its To address these issues, exemplifying FD problem aircraft air data sensors, we explore develop a robust (accurate, scalable, explainable, interpretable) scheme using typical data-driven method, i.e., deep neural networks (DNN). guarantee scalability, inertial reference unit measurements adopted as equivalent inputs DNN, database associated with 6 aircraft/flight conditions constructed. Convolutional (CNN) long-short time memory (LSTM) blocks used in DNN performances. enhance robustness also two new concepts: “large structure” corresponds parameters that can be objectively optimized (e.g., CNN kernel size) via certain metrics accuracy) “small conveys subjective understanding humans class activation mapping CNN) within context object detection). We illustrate optimization process devising large structure, yields (90%) scalable (24 diverse cases) interpret small structure mapping, promising results solidifies DNN. Lessons experiences learned summarized paper, believe instructive addressing problems other similar fields.