Calibration of low-cost sound level meters using machine learning techniques

Abstract

The assessment of environmental noise pollution from complex sources that are not represented in traditional noise maps requires a long term monitoring network to manage citizen complaints, a network of low-cost sound level meters is a practical option to analyze real-world cases, however the accuracy of the measurements is a concern when this information is used to evaluate regulation accomplishment. Previous work on low-cost monitoring devices has been carried out in microphone comparison, processing board selection and recently in MEMS microphone use. In the present work the use of reinforcement learning techniques is explored to calibrate the sound pressure levels generated by a low-cost monitoring device using a class one sound level meter as reference in a continuous measurement setup. Machine learning based models are known to take into account strong non-linearities which are effective to get an alternative calibration method for low-cost sensors.