Second Order Sliding Mode Control of Flying Wing Aircraft Based on Feedforward Neural Networks
Yuecheng Song, Zhenbao Liu, Junwei Han, jinbiao yuan, Wen Zhao, Qingqing Dang
AI summary
Problem
Flying-wing aircraft are highly nonlinear and difficult to stabilize with traditional linearized controllers, which degrade outside specific operating points and struggle with wind disturbances and actuator chattering.
Approach
The method combines an offline feedforward neural network for high-fidelity aircraft modeling with a second-order sliding mode controller whose parameters are dynamically optimized by DDPG reinforcement learning, alongside a DNN-based wind disturbance observer.
Key results
- FNN-based unified nonlinear model captures complex aerodynamics across the entire flight envelope
- DDPG auto-tuning eliminates manual gain scheduling for the MIMO second-order sliding mode controller
- DNN disturbance observer enables accurate real-time wind estimation and proactive compensation
- Validated robustness across Simulink, Software-in-the-Loop, and Hardware-in-the-Loop simulations
Why it matters
This framework provides a deployable, robust control solution for complex UAV platforms, bridging the gap between advanced nonlinear control theory and real-time embedded implementation.
Abstract
The flying-wing aircraft control problem is a major concern. In this paper, a new control strategy is introduced. First, a Feedforward neural network (FNN) modeling is introduced. Then, a second-order sliding mode control is applied, with the parameters generated from Deep Deterministic Policy Gradient (DDPG) reinforcement learning. To study the disturbance rejec- tion performance, wind disturbance is applied to the aircraft using a deep neural network as an disturbance observer for different types of winds. Finally, All three simulations: Simulink, Software In The Loop, and Hardware In the Loop are applied to show the effectiveness of the proposed strategy. The simula- tion results show that the proposed method demonstrates good robustness in various conditions. Note to Practitioners—This paper is motivated by the tra- ditional linearized flying-wing aircraft controller with the effectiveness of the FNN and reinforcement learning on UAV applications. The controller can be designed for various situa- tions without changing the parameters by modeling the aircraft through the FNNs. The theoretical framework proposed in this paper combines the multiple-input multiple-output (MIMO) sliding mode strategy with the reinforcement learning with higher accuracy modeling. This method reduces the settling time, overshoot and steady error. More realistic effects should be considered for deployment into real aircraft. The parameters of the neural networks should also be adjusted in real appli- cations.The FNN networks and DDPG have low computational footprints and are readily deployable on embedded systems like Received 18 June 2025; revised 29 August 2025; accepted 8 September 2025. Date of publication 23 September 2025; date of current version 7 October 2025. This article was recommended for publication by Associate Editor C. M. Abdissa and Editor C. Seatzu upon evaluation of the reviewers’ comments. This work was supported in part by the National Natural Science Foundation Fund under Grant 52072309 and Grant 62303379, in part by the Key Program of the National Natural Science Foundation of China under Grant U2433208, in part by the Key Research and Development Program of Shaanxi under Program 2019ZDLGY14-02-01, in part by Shenzhen Fun- damental Research Program under Grant JCYJ20190806152203506, in part by the Natural Science Foundation of Shaanxi Province under Grant 2023- JC-QN-0665, in part by Suzhou Municipal Science and Technology Bureau under Grant ZXL2023177, in part by the Aeronautical Science Foundation of China under Grant ASFC-2018ZC53026, in part by the Foundation of Yunnan Key Laboratory of Unmanned Autonomous Systems under Grant 202408ZD01, and in part by the National Key Laboratory Foundation of Helicoptor Aeromechanics under Grant 2024-ZSJ-LB-02-03. (Corresponding author: Zhenbao Liu.) Yuecheng Song and Junwei Han are with the School of Automa- tion, Northwestern Polytechnical University, Xi’an 710072, China (e-mail: illidan st227@mail.nwpu.edu.cn; jhan@nwpu.edu.cn). Zhenbao Liu, Wen Zhao, and Qingqing Dang are with the School of Civil Aviation, Northwestern Polytechnical University, Xi’an 710072, China (e-mail: liuzhenbao@nwpu.edu.cn; zhaowen@nwpu.edu.cn; dangqingqing@ nwpu.edu.cn). Jinbiao Yuan was with the School of Civil Aviation, Northwestern Poly- technical University, Xi’an 710072, China. He is now with Qing’an Group Company Ltd., Xi’an 710077, China (e-mail: yuanjinbiao0414@126.com). Digital Object Identifier 10.1109/TASE.2025.3613383 Pixhawk. The DNN observer may require model compression for the smallest processors. While the current framework requires per-aircraft training to achieve optimal performance, this process is conducted offline. The resulting controller gains are then fixed for reliable real-time operation, providing a clear pathway for implementation on specific UAV platforms.