Research · Audio & sensing · 2022 – 2024

ChickenSense

A low-cost deep learning system that estimates broiler feed intake by listening to pecks with piezoelectric sensors.

Diagram of the experimental pen: a feeder on a scale fitted with piezoelectric sensors, a preamp and PC on a frame above, and an RGB-D camera, next to a photo of the real installation.

92%pecking-event accuracy
91%F1 score
8 ± 7%error on daily feed intake
19 daysof 24/7 recordings

Knowing how much each flock eats is one of the most useful signals on a broiler farm, but feeder scales are expensive and hard to deploy at scale. ChickenSense estimates feed intake from sound instead: piezoelectric sensors attached to the feeder pick up the vibration of every peck.

Method

Signals were recorded around the clock for 19 consecutive days. A labeled subset split events into two classes: feed-pecking and non-pecking (vocalizations, anomalies and silence).

Each signal is cleaned with a noise-removal algorithm and a band-pass filter.
Two views are extracted: a spectrogram and the signal envelope.
A dual-branch, VGG-16-based CNN learns 2-D features from the spectrogram and 1-D features from the envelope, followed by a binary classification head.
The trained model is run over the full recording, and detected pecks are converted into estimated feed consumption.

Results

The classifier separated pecking from non-pecking events with 92% accuracy and a 91% F1 score. Compared with scale measurements, daily feed intake estimates had an 8 ± 7% mean percent error. Hourly estimates reached an R² of 0.71 and a Pearson correlation of 0.85.

Time series of feed consumption over about 20 days, comparing scale weight with the piezo-based estimate, with camera snapshots at several peaks. — Hourly feed consumption measured by the scale versus estimated from the piezoelectric signal.

Because the hardware is inexpensive and works per feeder, the approach has the potential to be implemented in commercial farms.

Citation

@article{amirivojdan2024chickensense,
  title     = {ChickenSense: A Low-Cost Deep Learning-Based Solution for Poultry Feed Consumption Monitoring Using Sound Technology},
  author    = {Amirivojdan, Ahmad and Nasiri, Amin and Zhou, Shengyu and Zhao, Yang and Gan, Hao},
  journal   = {AgriEngineering},
  volume    = {6},
  number    = {3},
  pages     = {2115--2129},
  year      = {2024},
  publisher = {MDPI}
}