Data-driven flight path monitoring technique using recurrent neural network for the safety management of commercial aircraft

Abstract

Aviation spins, particularly at low altitudes, significantly contribute to fatalities due to limited recovery time. Standard recovery procedures typically only become eligible after a spin is fully developed, by which time multiple turns may have already resulted in substantial altitude loss. The primary challenge in upset prevention is heavy reliance on the pilot's situational awareness, which is only effective before the spin has been fully developed. To address this issue, this study proposes an early detection capability to significantly enhance immediate response actions, potentially mitigating altitude loss and enabling pilots to recognize the initial signs of upset conditions. This research introduces a real-time predictive tool based on a novel recurrent neural network (RNN) model that utilizes data from the NASA Generic Transport Model (GTM)-a research platform designed for experimental flight case studies-to predict nonlinear flight responses during the critical initial seconds of a spin. Rigorous validation against ground truth data demonstrates the RNN model's superior predictive capabilities in detecting incipient spin phase, offering an essential tool for proactive spin management and reducing the risk of ground collisions. This early detection capability empowers pilots to identify the initial signs of upset conditions and make informed operational decisions, ultimately improving aviation safety. This advancement underscores the potential of advanced machine learning technologies to transform safety protocols by enabling earlier and more effective intervention strategies, thereby preempting catastrophic events.

Keywords: Air traffic management; Flight safety; Incipient spin detection; Machine learning-based flight monitoring.

Figure 1. Schematic of the proposed model framework for spin trajectory prediction and aviation safety.

Figure 2. Schematic of the RNN cell; where x_t is the model input variable, h_t is the history variable of the RNN at time t, and A is the RNN activation function.

Figure 3. Schematic of the proposed RNN-based model framework; input variables of model are the equivalent airspeed, angle of attack, sideslip, angular velocities, and the output variable is the simulation trajectory.

Figure 4. Schematic of the GTM framework, which has subsystems labeled thrust, engine, gravity, wind tunnel that contain computation of the aerodynamics with six-degree-of-freedom state variables as input for the full state of the aircraft at time increment ∆t.

Figure 5. Typical occurrence of spin with three main phases, which are incipient, fully developed, and recovery phases.

Figure 6. Loss function with training and validation data for the RNN-based model.

Figure 7. Sample comparison between the ML model and GTM predictions for (A) altitude during incipient and fully developed spin phases, (B) expected range of altitude absolute error.

Figure 8. Comparison of altitude distributions: (A) histogram and (B) boxplot of ground truth and ML model for altitude prediction.

Figure 9. (A) Mean incipient spin altitudes of the ML model and ground truth, with 95% confidence intervals; (B) parity plot of predicted vs ground truth incipient spin altitudes for all 100 new test cases.

Figure 10. Spin upset scenario phases prediction using GTM.