Intelligent Livestock Health Monitoring System (Embedded)

A leading global livestock company (hereinafter referred to as "the client" for confidentiality) is developing a next-generation intelligent health monitoring system aimed at minimizing farming losses and increasing operational efficiency. Their main objective is to deploy large-scale AI models directly on embedded devices for real-time analysis. By processing data locally, server workload is reduced while ensuring accurate alerts and consistent system reliability.

Project Challenges

Given the client's large-scale operations, the system must deliver highly responsive monitoring. Relying exclusively on cloud-based inference is susceptible to network instability and high costs. Meanwhile, embedded devices are limited in computing and memory resources, yet are expected to run multiple algorithms simultaneously. Acoustic environments within livestock houses are complicated and noisy; without effective microphone array signal processing, atypical sounds like coughing can easily be masked by background noise. The system should also support OTA (Over-the-Air Update) capabilities and be simple to maintain for sustained, stable operation on site.

Our Solution

We concentrated our efforts on the architecture and implementation for embedded devices. To make AI models suitable for local execution, we leveraged model quantization and pruning techniques, enabling them to run smoothly on embedded chips. To keep overall costs down, all inference is performed locally; only audio data relevant for algorithm improvement is uploaded, minimizing bandwidth usage and reducing the server's workload at its origin.

To address the noise challenge, we integrated several microphone array processing methods onto a single chip: beamforming, noise reduction, and direction of arrival (DOA) detection. By combining microphone array signal processing with AI inference, the system can pinpoint specific cough sources among groups and effectively suppress background noise, substantially improving both the accuracy and responsiveness of event detection.

For maintenance, we assisted the client with building a secure OTA update system, supporting phased deployment of algorithm and firmware updates. This streamlines maintenance, reduces offline service costs, and ensures capacity for future expansion and upgrades.

Project Outcomes

In continuous field trials at multiple pilot farms, the optimized models delivered edge processing latency within a few milliseconds, while alert accuracy remained over 90% of its original baseline. Running multiple algorithms on a single chip lowered hardware costs and reduced power consumption. The solution validated its reliability and scalability. Most significantly, the new system cut overall costs to roughly one-third compared to the previous generation and replaced a multi-chip setup with a single integrated chip, resulting in greatly enhanced system integration and reliability.

Thanks to this collaboration, the client found an optimal balance between cost, performance, and reliability. The real-world application of edge AI in complex livestock environments proved its substantial value.