Adaptive maneuver load alleviation via recurrent neural networks

THE load alleviation technique has gained increasing prevalence in the design of modern aircraft over the past decades because it is able to reduce additional structural loads due to both wind gusts and aircraft maneuvers, as well as get decreased structure weight, extended structure life, and enhanced flight performance. Gust load alleviation aims at improving ride comfort, which is not a concern for high-performance fighters. Hence, only maneuver load alleviation (MLA) has been taken into consideration in the load alleviation of fighters. One of the main criteria of MLA is the reduction in the bending moment near the wing root, with the normal load factor remaining the same. The MLA problem has been studied since the 1970s [1-6]. For example, Anderson et al. synthesized the longitudinal flight control system for an F-4E model by using classic control to realize load alleviation [1].Woods-Vedler et al. demonstrated a rolling maneuver load alleviation system on the wind-tunnel model of an active flexible wing and presented a systematic approach for designing rolling MLAcontrol laws [3]. Gaulocher et al. used model predictive control to attenuate the structure response of sudden roll maneuvers [4]. Paletta et al. designed a maneuver load control system for longitudinal maneuvers of a high-altitude performance demonstrator, and they reduced the wing-root bending moment by approximately 20% while following the desired normal load factor law [5]. All the MLA techniques mentioned were not fully adaptive, and the control laws had to be redesigned under a wide range of flight conditions. Consequently, an improved adaptive controller should be proposed for high-fidelity models and variable flight conditions, such as the Mach number and dynamic pressure. The past two decades have witnessed the development of numerous adaptive controllers. Among them, the controllers based on recurrent neural networks (RNNs) have found successful applications to many fields [7-11], owing to their applicability and fast self-learning ability in real-time control. In this Note, an adaptive controller based on RNNs was synthesized to realize an MLA system based on the identification of the aeroelastic model of a fighter first. Then, the adaptive controller was demonstrated via case studies under different flight conditions. The study showed that the controller required only the inputs and outputs of the aeroservoelastic system to realize system identification and the desired maneuver flight with a reduced wing-root bending moment.