Chicken disease detection and localization using multi-noise separation and acoustic recognition.

Area of Science:

• Poultry Health
• Bioacoustics
• Machine Learning

Background:

• Early detection of chicken diseases is vital for poultry industry sustainability.
• Environmental noise and complex acoustic features challenge disease identification.
• Current methods require improvement for timely and accurate disease diagnosis.

Purpose of the Study:

• To develop an acoustic-based method for early detection and localization of chicken diseases.
• To overcome limitations posed by environmental noise and acoustic feature extraction.
• To establish a novel early warning system for poultry health management.

Main Methods:

• Established a standardized acoustic dataset for Avian Infectious Laryngotracheitis (AILT), Mycoplasma, Newcastle Disease, and healthy chickens, verified by PCR.
• Developed a fusion model (MSA-BiFPN-EfficientNetV2) integrating multi-spectral channel attention (MSA) and bidirectional feature pyramid networks (BiFPN).
• Employed log-Mel spectrograms for end-to-end sound classification and TDOA-GCC-PHAT for sound source localization.

Main Results:

• The fusion model achieved 96.17% accuracy and 96.28% F1-score on the test set.
• Specific diseases showed high performance: Mycoplasma (97.34% F1-score) and AILT (97.09% F1-score).
• The localization strategy achieved an average error of 0.085 meters, accurately pinpointing diseased chickens.

Conclusions:

• The proposed acoustic recognition method offers a promising solution for early chicken disease detection and localization.
• The MSA-BiFPN-EfficientNetV2 model effectively handles acoustic challenges in noisy environments.
• This technology can be implemented as an effective early warning system to enhance poultry biosecurity and sustainability.