Neural network-assisted sliding mode observer for fault diagnosis of flying wing UAV actuators

Actuator degradations in flying-wing UAVs are difficult to diagnose because strong cross-axis aerodynamic coupling and limited control authority can mask weak loss-of-effectiveness (LOE) and incipient jamming under modeling mismatch, wind gusts, and sensor noise. To address this issue, this paper develops a disturbance-fault-decoupled neural network-assisted sliding-mode observer (NN-SMO) for real-time actuator fault detection and magnitude estimation. A nonlinear 6-DOF flying-wing model is used as a physically consistent basis, and a decoupling formulation is introduced to construct a fault-dominant subsystem, improving the structural observability of weak actuator faults. On this subsystem, an radial basis function network approximates the fault term online while a sliding-mode injection suppresses bounded uncertainties; a Lyapunov analysis provides verifiable convergence conditions and gain-selection guidance. The method is evaluated on a closed-loop Simulink platform under a fixed fault-injection time (t = 120 s) and is benchmarked against an EKF-residual detector and lightweight deep temporal models (TCN/LSTM/Transformer) under an identical protocol. For a 30% LOE fault and a small-angle jamming fault, NN-SMO detects the onset within 1.07 and 0.63 s, respectively, achieving area under the curves of 0.971 and 0.986 and root-mean-square errors of 0.047 and 0.031. Performance degradation remains below 5% under +10% sensor noise and ±8 m s⁻¹ gust disturbances, indicating robustness suitable for safety-critical onboard monitoring.