Wind estimation by multirotor drone state using machine learning with data rotation and reduction

Abstract

The problem of wind estimation by multirotor drone is suitable for machine learning (ML), because of the unknown drag coefficient changing with orientation. In previous work, we studied this problem experimentally to train an ML model that estimated wind by drone state alone. The primary drawback of this work was a decrease in performance between randomly selected and new flight data. This work applies data rotation and reduction to overcome this. Rotation allows the model to generalize to arbitrary coordinates, and reduction addresses data imbalance present in the original data. Two models are trained on experimental flight data: a gated-recurrent-unit (GRU) and a long-short-term-memory (LSTM) neural-network. The better performing GRU achieved 0.48 m/s root-mean-square-error on new flight data, although the LSTM performs similarly. These models approach the experimental limits of performance, which is determined by the spatial variation of wind as measured by the error between two separate anemometer readings.