Thesis Project - Low Power Machine Learning Based Activity Monitor for Livestock: Summer 2020 - Present

Summary

I have led a total of seven students in developing a low power wireless sensor node (WSN) to monitor the health and activity of cattle. The current system consists of a Microchip WLR089U0 MCU, Bosch BMA400 accelerometer, and U-Blox ZOE-M8Q GPS. The prototype takes measurements from its various sensors, transmits them using the LoRaWAN protocol, and stores them in an AWS database. In order to further extend its battery life, we are using the BQ25570 power management IC (PMIC) and a solar panel.

The LoRa radio consumes the most power in our prototype. We have employed different approaches to reduce the amount of transmitted data and minimize power consumption. In one such approach, we are using machine learning in situ for classifying cattle behavior based on high frequency accelerometer data. This approach significantly reduces the data transmitted by processing it within the WSN itself.

There was no pre-existing infrastructure for collecting training data from the cattle since this is a sui generis project pioneered by me in my research group. As a part of my thesis project, I developed a novel method of collecting accelerometer data and combined it with video data so we could accurately label the accelerometer data with the cattle’s behavior. My method allows us to monitor cattle in a pasture for a 24-hour period with minimal human interactions. We are currently using this method to create a dataset that well encapsulates the natural behaviors of cattle living in a pasture.