Implementation and characterization of a machine learning algorithm for bat detection running on a low-power microcontroller

Details

Supervisors

Bol, David

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil électricien, à finalité spécialisée

Abstract

Biodiversity has long been a concern for environmental authorities, and bat species have not been spared from the global biodiversity loss. Acoustic surveys are the main method for censuses of bat faunas, but their use has remained limited, partly due to their cost and the difficulties of call detection. Several algorithms using machine learning to detect bat calls have been developed and have shown great performance. However, most of them are designed for computers and cannot be implemented on low-power embedded devices for real-time operation. In this work, an algorithm for bat detection inspired by state-of-the-art algorithms is designed and implemented for a platform based on a Cortex-M4. Multiple design choices are made to speed up computation, including the use of mel-scaled filterbanks and quantization of the model. The dataset to train the algorithm is collected in Louvain-la-Neuve with the device, and a teacher model is used to generate annotations. This dataset is used to train the CNN in the algorithm and to optimize various parameters. The resulting system requires 11.47 ms to process 9.8 ms of audio, while consuming 7.45 mA in execution and 115 μA in standby. Taking the sensing duty cycle into account, the proposed system is estimated to have an average precision between 60.29 % and 69.55 % on the call detection task, and has been validated in the field to be able to detect bat calls. Performance improvements based on hardware modifications and possible applications of the proposed solution are discussed.