Edge intelligence for poultry welfare: Utilizing tiny machine learning neural network processors for vocalization...

Area of Science:

• Agricultural Technology
• Machine Learning
• Animal Behavior

Background:

• Poultry flock health is vital for sustainable agriculture.
• Machine learning and speech analysis offer potential for real-time flock monitoring.
• Limited research exists on using Tiny Machine Learning (Tiny ML) for continuous poultry vocalization monitoring.

Purpose of the Study:

• To develop and deploy Tiny ML models on low-power edge devices for monitoring chicken vocalizations.
• To address challenges in implementing Tiny ML in agricultural settings, including memory, processing, and battery constraints.
• To accurately identify and categorize chicken vocalizations linked to emotional states like discomfort, hunger, and satisfaction.

Main Methods:

• Creation of a diverse dataset of poultry vocalizations in collaboration with avian researchers.
• Utilization of Digital Signal Processing (DSP) blocks on the Edge Impulse platform for spectral feature generation.
• Development and application of a one-dimensional Convolutional Neural Network (CNN) model for vocalization classification.
• Implementation of noise-robust Tiny ML algorithms to enhance accuracy and reduce background noise.

Main Results:

• The Tiny ML models demonstrated high performance in classifying chicken vocalizations.
• Average accuracy and F1 scores were 91.6% and 0.92 before background noise removal.
• Accuracy improved to 96.6% and F1 scores to 0.95 after implementing noise-robust algorithms.

Conclusions:

• Tiny ML models can be effectively deployed on low-power edge devices for continuous poultry vocalization monitoring.
• The developed algorithms show significant potential for improving the assessment of flock health and welfare in real-time.
• Addressing challenges like background noise is crucial for practical and accurate implementation in sustainable farming.