Combining model-based and learning-based anomaly detection schemes for increased performance and safety of aircraft braking controllers

In aircraft, the braking system is a safety-critical and heavily used component of the landing gear, prone to significant wear. Anomalies arising in the wear dynamics can degrade the performance of the braking system and compromise the safety of ground handling maneuvers. In this work, we tackle the problem of detecting incipient anomalies in aircraft brakes in a tightly coupled implementation with the Brake Control Unit (BCU). Two complementary approaches are presented. The first one is an observer-based architecture designed on the longitudinal aircraft dynamics that returns physically interpretable outputs connected to the wear process and allows us to improve braking performance online. The second one is an end-to-end convolutional autoencoder-based architecture that returns an anomaly score computed on data collected by the BCU with inherent robustness to modeling uncertainty, which the model-based one does not. A combined architecture that allows one to exploit the features of both model-based and learning-based approaches is proposed, which shows its capability of optimally blending the two. The approaches are evaluated in a MATLAB/Simulink multibody simulation environment that is able to replicate the braking actuator wear dynamics, demonstrating remarkable performances in anomaly detection, anti-skid control performance, and safety improvement.