Towards Intelligent Fault-Tolerant Attitude Flight Control of Fixed-Wing Aircraft

Can a learned controller fly through a fault?

Classical fault-tolerant control relies on gain scheduling or hardware redundancy and on prior knowledge of the faults. We train self-learning attitude controllers with deep reinforcement learning that track references smoothly in nominal flight and adapt to unexpected failures.

Schematic. The controller tracks a roll reference; after an actuator fault the response deviates briefly and recovers. Illustrative trajectory, not a recorded result.

high-fidelity Cessna Citation 500 model: 6-DoF
fault scenarios beyond nominal flight: 6+
learning approaches: CEM-TD3 and PPO: 2

Simulation study on the TU Delft PH-LAB aircraft model, trimmed at 2,000 m and 90 m/s.

01 / The problem

Faults that were not anticipated.

Faults appear as sensor errors, structural or system failures, and unexpected phenomena. Passive and active fault-tolerant control strategies typically depend on knowing the fault in advance, which limits them to known fault types. Deep RL can adapt, but noisy control commands make hardware implementation difficult.

02 / Formulation

Track the reference, keep the actuators smooth.

Attitude control is posed as minimizing tracking error together with a penalty on changes between consecutive control inputs, subject to actuator magnitude and rate limits.

CEM-TD3

A neuro-evolutionary method: the Cross-Entropy Method refines a population of TD3 actors.

PPO

Clipped surrogate policy optimization with a learned value function.

Trade-off

The two approaches differ in how they balance tracking accuracy against control smoothness.

$$u^{*} = \arg\min_{u} \int_{0}^{t_f} \Big( L_s(s, u, r) + L_u(u) \Big)\, dt \quad \text{s.t.}\ \ |u| \le u_{\max},\ \ \Big|\tfrac{\Delta u}{\Delta t}\Big| \le \dot u_{\max}$$

L_s penalizes deviation from the reference; L_u penalizes abrupt changes in control input.

03 / Evaluation

Nominal flight and a battery of faults.

Evaluated conditions
Condition	Description
Nominal	Trim at 2,000 m and 90 m/s
Iced wings	Reduced maximum angle of attack, increased drag
Aft-shifted CG	Center of gravity moved aft by 0.25 m
Saturated aileron	Deflection clipped at ±1°
Saturated elevator	Deflection clipped at ±2.5°
Partial loss of elevator	Effectiveness scaled by 0.3
Jammed rudder	Rudder stuck at 15°
High dynamic pressure	Trim at 2,000 m and 150 m/s
Wind and sensor noise	Disturbances on dynamics and measurements

Condition definitions follow the paper.

04 / Findings

Stable, smooth, and robust in simulation.

The paper reports that the trained agents exceed the state of the art on nominal flight and on more than six fault scenarios, with each algorithm showing a distinct trade-off between tracking and control smoothness, and a stability analysis supports the stability of the neural-network controller.

05 / Scope

Simulation-based evaluation.

All results are from a high-fidelity simulation model; flight testing is outside the scope of the paper.

Reference

Cite this work

Zongo, A.B., Li, Q. (2025). "Towards Intelligent Fault-Tolerant Attitude Control of Fixed-Wing Aircraft." In: Yan, L., Duan, H., Deng, Y. (eds).Advances in Guidance, Navigation and Control. ICGNC 2024. Lecture Notes in Electrical Engineering, vol 1353. Springer, Singapore.

@inproceedings{zongo2025towards,
  title     = {Towards Intelligent Fault-Tolerant Attitude Control of Fixed-Wing Aircraft},
  author    = {Zongo, Alex B. and Qing, Li},
  booktitle = {Advances in Guidance, Navigation and Control (ICGNC 2024)},
  series    = {Lecture Notes in Electrical Engineering},
  volume    = {1353},
  year      = {2025},
  publisher = {Springer, Singapore},
  doi       = {10.1007/978-981-96-2264-1_15}
}