Thesis Project - Low Power Machine Learning Based Activity Monitor for Livestock

Summary

I led a team of seven students to develop a power-efficient wireless sensor node (WSN) designed to monitor cattle health and activity. The system utilizes a low-power MCU, 3-axis accelerometer to collect motion data, and a LoRa radio transmitter.

Edge Intelligence and Power Optimization

A primary challenge in livestock monitoring is the high power consumption of LoRa radios during long-range data transmission. In this work, I applied edge intelligence, implementing a decision tree classifier in situ to process raw accelerometer data directly on the WSN. By performing feature extraction and classification locally, we significantly reduced the amount of wirelessly transmitted data. Our measurements demonstrate that this approach:

• Drastically Reduces Power: Lowers average power dissipation by a factor of 50.
• Optimizes Link Budget: Enables the WSN to operate efficiently even at low data rates. These lower data rates offer improved link budget, which can be used to increase transmission range or relax antenna requirements.

Classification Performance and Methodology

Because no pre-existing infrastructure existed for this research, I developed a novel data collection method that synchronized accelerometer readings with video data for accurate behavioral labeling. Using our dataset, we trained a decision tree model to classify five distinct cattle behaviors: grazing, ruminating, lying, standing, and walking.

Key performance metrics for the model include:

• High Accuracy: Achieved an overall classification accuracy of 94% using tenfold cross-validation.