Deep learning reduces sensor requirements for gust rejection on a small uncrewed aerial vehicle morphing wing

TL;DR

End-to-end deep reinforcement learning forgoing state inference is shown forgoing state inference to efficiently alleviate gusts on a smart material camber-morphing wing and overcome challenges of computationally expensive modeling and expansive distributed sensing networks.

Abstract

There is a growing need for uncrewed aerial vehicles (UAVs) to operate in cities. However, the uneven urban landscape and complex street systems cause large-scale wind gusts that challenge the safe and effective operation of UAVs. Current gust alleviation methods rely on traditional control surfaces and computationally expensive modeling to select a control action, leading to a slower response. Here, we used deep reinforcement learning to create an autonomous gust alleviation controller for a camber-morphing wing. This method reduced gust impact by 84%, directly from real-time, on-board pressure signals. Notably, we found that gust alleviation using signals from only three pressure taps was statistically indistinguishable from using six signals. This reduced-sensor fly-by-feel control opens the door to UAV missions in previously inoperable locations.

Figures (7)

• Figure 1: Natural flyers use wing shape morphing to reject gusts. (A) Inspired by how birds change the shape of their wings to adjust for environmental changes, we implemented a trailing edge camber morphing mechanism. (B) The morphing wing consisted of 3 active sections driven by macro fiber composites (MFC). A rigid wing acting as a gust generator was mounted 30 cm upstream of the morphing wing with three active camber morphing sections within the University of Michigan 1’ x 1’ (30 cm x 30 cm) wind tunnel. (C) The morphing wing was designed with six pressure taps to sense gusts. (D) The gust generator deflected upwards (yellow) and downwards (green) at varying degrees (depicted by opacity) to create a variety of velocity wakes, (E) the magnitude of which was quantified with particle image velocimetry.
• Figure 2: Gust Rejection Percentage (GRP) provides a metric for controller performance and consistency. With Proximal Policy Optimization (PPO), we trained 10 controllers using six pressure taps for gust alleviation in this high lift environment. (A) We quantified controller performance by comparing the change in lift ($\Delta L$) of the actively controlled wing with that of the inactive baseline, where the magnitude of the arrows indicates GRP. (B) On average, the learned controllers rejected more than 84% of the $\Delta L$ produced by the tested gusts. Additionally, we measured consistency between tests, gust conditions, and trained controllers with the standard deviation between (C) ten (10) tests for one trained controller at one gust condition, (D) average gust responses for a single controller at each gust condition (6), (E) and the average responses at a single gust condition for each trained controller (10).
• Figure 3: The number of pressure taps significantly affected gust rejection performance. (A) We used the settled gust rejection percentage (GRP) to measure controller effectiveness for each pressure tap configuration. (B,C,D) Controllers relying on a single pressure tap rejected a significantly smaller portion of the gust than controllers using all six taps for all flight conditions (high lift: P < 0.006, medium lift: P < 0.001, low lift: P < 0.001) as represented by open circles. However, the difference between using three pressure taps and six pressure taps was not significant for each flight condition (high lift: P = 0.40, medium lift: P = 0.32, low lift: P = 0.67). (E) We quantified controller speed using rise time, the time needed to reduce the gust from 10% to 90% of the settled GRP. Due to the highly skewed nature of these results, we used the median to illustrate the central tendency for the speed metric. (E,F,G) Varying the number of pressure taps did not significantly affect the rise time.
• Figure 4: The third pressure tap lost sensitivity during downward gusts for the high lift flight condition. (A) Although the first three pressure taps produce sensitive pressure signals for the upward (positive) gust deflections, the third pressure tap is much less sensitive to downward gusts (16.7 %). (B) At the mild gust condition, the trained gust alleviation controllers using six pressure taps overshot zero lift error. (C) Particle image velocimetry (PIV) showed the environmental change in the incoming streamwise velocity experienced by the wing during different gusts. This change is measured by directly comparing the streamwise velocity at each position during a gust to that experienced during the neutral airflow. Blue represents a decrease in velocity at the specific position due to the gust generator, and red is an increase in velocity. The change in velocity is stronger over the front three pressure taps in the upward gust than in the downward gusts. The reduced change in velocity is most noticeable at the third pressure tap location.
• Figure 5: Data flow structure of our gusting wind tunnel experiment for controller training and testing. Training and testing were orchestrated using a Jupiter Notebook written in Python on a PC. The Python script informed the motor controller to rotate the turn table to deflect the gust generator to a desired magnitude and direction. The change in airflow in the wake of the gust generator was detected by the six pressure taps on the MFC morphing wing. The pressures were measured nd compared to a static pressure measured in front of the experimental setup using six differential pressure transducers. Signals from these pressure transducers were acquired by the NI-DAQ, and provided to the Python script. The Python script used this information for action selection. The selected action was provided to the NI-DAQ and transformed into an MFC Voltage signal which was then amplified to power the MFC camber morphing trailing edge of the wing. The lift produced by the change in camber was measured by the load cell, and provided to the Python script for reward calculation during controller training or performance measurement during controller testing.