SmartEars: An intelligent method for monitoring poultry respiratory illnesses based on audio signals
Description
Sound datasets from real farming environments are scarce, prompting us to release a portion of 5-second data segments with labels after multiple rounds of data cleaning. We have disclosed 2,000 segments for each of the three categories (Healthy, Sick, None - no chicken sound), totaling 6,000 five-second audio clips. We make this dataset publicly available, to contribute to the advancements in research related to the detection of respiratory diseases based on poultry vocalizations. In large-scale poultry farming, respiratory diseases affect the health of chickens, leading to a decline in the quality and yield of both meat and eggs. Effective monitoring of these diseases is crucial to reducing their impact and enhancing the quality and yield. Currently, most monitoring methods still rely on manually monitoring chicken vocalizations, which are time-consuming, labor-intensive, and require specialized personnel, making 24/7 monitoring unfeasible. Existing intelligent methods are often limited to laboratory environments where individual chickens are monitored separately. These approaches do not meet the industrial and commercial requirements of poultry farms, where a diverse set of complex auditory signals may be captured. These signals include not only chicken vocalizations but also complex noises from cages, chicken behaviors, human activities, mechanical ventilation systems, and other backgrounds noises. In this study, we design a deep learning-based intelligent recognition algorithm capable of accurately distinguishing abnormal chicken vocalizations among complex sound signals. Furthermore, we integrate this algorithm into a distributed health monitoring system - SmartEars, enabling continuous collection of various sound signals and performing real-time recognitions, thereby providing round-the-clock monitoring of the respiratory diseases of chickens in real production environments. We collected 11,686 audio slices from actual farming environments, which were labeled through multiple rounds of annotations by veterinary experts, resulting in a high-quality dataset for model training. Additionally, we used Logfbank to capture critical audio features to assist model learning. We also designed five data augmentation techniques to prevent overfitting and improve model performance. Finally, we compared multiple models on an independent test dataset and selected RegNet as the best model, which achieved the highest accuracy of 96.03%. To validate the effectiveness of our approach, we compared the annotation results of SmartEars with seven veterinarians over the same dataset. The results demonstrated that SmartEars with an accuracy of 93% outperformed human veterinary experts with accuracy from 85% to 93%. SmartEars has been deployed in 3 large poultry farms located in Hebei, China, and it has successfully identified a number of outbursts of chicken diseases, such as a confirmed event around March 19, 2024, demonstrating the effectiveness of SmartEars.