A Cattle Behavior Recognition Method Based on Graph Neural Network Compression on the Edge.

Area of Science:

• Agricultural Technology
• Machine Learning
• Animal Science

Background:

• Cattle behavior monitoring is crucial for health and management.
• Current server-based recognition leads to high power consumption and latency.
• Edge computing offers a solution for real-time, low-power livestock management.

Purpose of the Study:

• To develop an edge-based cattle behavior recognition method.
• To reduce power consumption and computational latency in livestock monitoring.
• To enable precise and scientific livestock management through intelligent devices.

Main Methods:

• A wearable device integrating data acquisition and edge inference using a high-performance embedded microcontroller.
• A sequential residual model utilizing Inertial Measurement Unit (IMU) and displacement data for feature extraction.
• Graph Neural Network (GNN) compression with an Actor-Critic model for optimal pruning under Floating-Point Operations (FLOPs) constraints.

Main Results:

• The proposed method effectively classifies cattle behavior in real-time on edge devices.
• Significant reduction in computational latency and power consumption was achieved.
• The system demonstrates effectiveness for low-power, long-term cattle behavior monitoring.

Conclusions:

• The edge-based GNN compression method enables efficient and accurate cattle behavior recognition.
• Real-time edge inference is advantageous for reducing latency and power usage in livestock management.
• The developed system supports precise and scientific livestock management through intelligent, low-power devices.