Applying Machine Learning to a 1-D CMOS-MEMS Anemometer for Angle Detection and Range Extension

Abstract

Through machine learning (ML), the measurement range of a 1-D CMOS-MEMS anemometer has been extended by a factor of 8.3, while enabling 360$^{\circ }$ directional measurement. Without the use of ML, the angle of attack of the flow was inseparable from the wind speed using the sensor output due to its dependence on both parameters simultaneously. Random forest and gradient boosting ML algorithms have been evaluated for their performance. The random forest regression performed best in all tests, extending the sensor's measurement range from 1.2 to ${\mathbf{10}}$ m/s for all directions, with a ${\mathbf{3.9}}$% full-scale error for speed and ${\mathbf{5}}$% for direction. Gradient boosting performed slightly worse (${\mathbf{4.3}}$% and ${\mathbf{6.6}}$%) but did have much smaller model sizes (<${\mathbf{1}}$%). A Shapley additive explanation analysis was performed to determine the impact of different sensor outputs on the ML prediction, giving key insights into ways to improve sensor designs. Despite the implicit symmetry of a 1-D sensor's output for positive and negative wind angles, the ML models can extract small (hidden) features from the data, which contain information on the direction. The 1-D configuration in combination with ML allows for a state-of-the-art accuracy in both speed and direction, with a significantly smaller sensor footprint (0.245 mm$^{2}$).