Deviation Prediction and Correction on Low-Cost Atmospheric Pressure Sensors using a Machine-Learning Algorithm

Atmospheric pressure sensors are important devices for several applications, including environment monitoring and indoor positioning tracking systems. This paper proposes a method to enhance the quality of data obtained from low-cost atmospheric pressure sensors using a machine learning algorithm to predict the error behaviour. By using the extremely Randomized Trees algorithm, a model was trained with a reference sensor data for temperature and humidity and with all low-cost sensor datasets that were co-located into an artificial climatic chamber that simulated different climatic situations. Fifteen low-cost environmental sensor units, composed by five different models, were considered. They measure - together - temperature, relative humidity and atmospheric pressure. In the evaluation, three categories of output metrics were considered: raw; trained by the independent sensor data; and trained by the low-cost sensor data. The model trained by the reference sensor was able to reduce the Mean Absolute Error (MAE) between atmospheric pressure sensor pairs by up to 67%, while the same ensemble trained with all low-cost data was able to reduce the MAE by up to 98%. These results suggest that low-cost environmental sensors can be a good asset if their data are properly processed.