MLOps & Deployment
A model is useful only when its data, runtime, monitoring, and human workflow remain connected after training.
From research model to farm signal
Prepare the data
Organise recordings into healthy, unhealthy, and noise segments, then extract consistent acoustic features for training and evaluation.
Train and compare
Evaluate candidate CNN and LSTM approaches and use a multi-strategy Teacher ensemble to capture complementary predictive patterns.
Distil and package
Transfer useful Teacher knowledge into the approximately 970 KB Student LSTM, then package the inference path for edge hardware.
Deploy and observe
Run inference near the flock, expose clear results to the user, and collect approved feedback for later evaluation and improvement.
Deployment responsibilities
Signal quality
Account for microphone placement, sampling conditions, environmental noise, and missing or corrupted recordings.
Model health
Track confidence, class balance, false alarms, and performance drift as farms and recording environments change.
Human review
Keep alerts connected to farmer observation and veterinary review so predictions are not treated as standalone diagnoses.
Version control
Record model, feature-extraction, and data versions so results can be reproduced and updates can be rolled back when necessary.
Why the Student LSTM is the deployment choice
The presentation identifies the Student LSTM as lightweight and robust enough for the intended use case. Its compact size supports offline or low-connectivity operation, while the distilled training approach preserves knowledge from stronger models without requiring the full Teacher ensemble at runtime.
Deployment should still be accompanied by field testing, clear uncertainty communication, and a route to professional follow-up. Model compression improves feasibility; it does not remove the need for monitoring and validation.