Research on frequency attention-enhanced ResNet with Mel-spectrogram for poultry state recognition

Poultry condition recognition technology based on acoustic signals is becoming a key component in improving farming efficiency in the field of smart agriculture and livestock condition monitoring. Conventional methods mainly focus on manual empirical judgment or simple time-frequency feature extraction, which are difficult to capture the complex emotional expression patterns in poultry vocalizations. To address this challenge, this study proposes an improved residual network architecture (FAM-ResNet), which is integrated with the frequency attention module (FAM) to achieve high-precision classification of four core poultry emotional states (pleasure, relaxation, anxiety, and hunger) by utilizing the joint time-frequency representation capability of Mel spectrogram. Experimental results demonstrate that the model achieves a test accuracy of 99.30% on a self-constructed dataset, with F1 scores of more than 1.00 for extreme states such as anxiety and hunger. Innovatively, the FAM module enhances the network's ability to adaptively focus on key frequency components, providing a novel deep-learning-based approach for interpreting and monitoring the state of poultry.