Abstract: To investigate the vocalizations of broilers under different health conditions in poultry houses, and achieve non-contact, automated identification and monitoring of broiler health status, thereby enhancing the intelligence of poultry farming and improving animal welfare. Addressing the issue of acoustic feature distortion caused by respiratory diseases, a lightweight broiler health sound recognition model（MobileNetV3-ECA-GAM）that integrated efficient channel attention (ECA) and global attention mechanism (GAM) to realize “channel-spatial cascade enhancement” was proposed. This design strengthened the model’s capacity to capture pathological acoustic features. In the experiment, 100 one-day-old mixed-sex broilers were divided into two groups. The primary respiratory disease of interest was infectious bronchitis, a common condition in broilers. Controlled-environment audio data were collected at scheduled intervals from both healthy and diseased (treated) groups. Preprocessing steps included audio segmentation, spectral subtraction for noise reduction, and conversion to Mel-spectrograms. Results demonstrated that the MobileNetV3-ECA-GAM model achieved an outstanding 96.75% accuracy in broiler abnormal sound recognition tasks, validating the effectiveness of ECA and GAM in enhancing model performance, efficiency, and generalization. These findings indicated that the proposed model was well-suited for non-contact abnormal sound monitoring in broilers, providing both theoretical and technical support for intelligent poultry farming.

Key words: broiler vocalization, non-contact monitoring, Mel-spectrogram, MobileNetV3, ECA, GAM