Machine learning framework for poultry disease detection using vocal pattern analysis

Current poultry health monitoring methods rely on visual inspection and manual assessment, which are laborious, unreliable, and frequently overlook early indicators of disease. Vocal Pattern Analysis is an innovative non-invasive method of automated health evaluation in poultry farming. This study analyzes vocal patterns to classify the health status of poultry, and builds and evaluates a machine learning system for a web-based real time classification system for this condition. A total of 346 audio samples were initially collected, comprising Healthy (n=139), Noise (n=86), and Unhealthy (n=121) recordings. After excluding 86 Noise samples as invalid data, 260 samples were used for binary classification (Healthy vs Unhealthy). Audio data extensive features extraction: 41 features composed from the Mel-frequency cepstral coefficients (MFCC), spectral features, zero crossing rate, chroma features, mel-spectrogram calculate statistics, RMS energy, and tempo. We compared two machine learning algorithms, Random Forest and Support Vector Machine (SVM). Both algorithms achieved 96.92% test accuracy, with macro-averaged metrics of 96.9% precision, 96.9% recall, and 96.9% F1-score. With regard to unhealthy detection, the Random Forest model achieved 96.67% sensitivity and SVM achieved 93.33% sensitivity, successfully differentiating healthy and unhealthy poultry vocalizations. We developed and deployed a web-based application that showcased the classification in real-time. It leads us to believe that ML based vocal pattern analysis can provide an effective, non-invasive method for poultry health monitoring. The system developed in this work could be applied for the automated health surveillance of commercial poultry.