Open Collections

Zimmerman, Steven

University of British Columbia

Multirotor drones are an ideal platform for a range of environmental studies, but wind measurement by anemometer payload suffers from the downwash of the propellers. This work presents a machine learning (ML) based disturbance observer, which implicitly estimates the wind based on the drone state without requiring a dedicated anemometer. Experimental data is collected by flying an instrumented quadrotor in proximity to two reference anemometers. Four ML models are developed: a long-short-term-memory (LSTM) neural network, an artificial-neural-network, a gated-recurrent-unit (GRU) neural network, and a Gaussian-process-regression. These models are trained with variations in their inputs, considering the addition of drag estimations by rigid-body equations of motion and control inputs, both of which improve performance. Two key processing methods are developed: Data augmentation by global coordinate frame rotation enforces rotational invariance and grid-based data reduction removes data imbalance. The best performing GRU achieves 0.48 m/s root-mean-square-error on unseen complete flight data, although the LSTM performs similarly. This approaches the experimental limits of performance, as reference anemometers cannot be exactly co-located with multirotors in flight. The difference between both spatially offset anemometers accounts for 73% of the GRU estimation error. These results are useful for emissions measurement, gas source localization and other applications.

Text

2022-05-26

Attribution-NoDerivatives 4.0 International

http://hdl.handle.net/2429/81638

Mechanical Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nd/4.0/