Exploring the Potential of Machine Learning Algorithms Associated with the Use of Inertial Sensors for Goat Kidding...

Area of Science:

• Animal Science
• Biotechnology
• Machine Learning

Background:

• Autonomous animal birth identification aids timely human intervention, protecting mother and offspring health.
• Wearable inertial sensors offer a low-cost, less invasive method for animal monitoring.
• Existing commercial solutions focus on cattle, leaving small ruminants like goats without automated birth detection.

Purpose of the Study:

• To develop an automatic birth monitor for goats using inertial sensing and machine learning.
• To implement the system on a network edge device for real-time alarm triggering.
• To evaluate the accuracy and computational cost of different detection algorithms.

Main Methods:

• Development of two concept drift detection techniques and seven kidding detection mechanisms.
• Utilized data classification models for detecting birth events.
• Tested and compared algorithm performance based on accuracy and computational expenses.

Main Results:

• Concept drift algorithms were ineffective for kidding detection.
• Classification-algorithm-based static learning models successfully detected kidding, even with imbalanced and small datasets.
• The developed algorithm showed behavioral changes four hours prior to kidding, identifying the kidding hour with 61% accuracy.

Conclusions:

• Machine learning models show promise for developing cost-effective, edge-deployable automatic birth monitoring systems for goats.
• The system's suitability for edge devices is high due to its computational efficiency.
• Further improvement of the learning process is anticipated with larger datasets.