Research on Abnormal Lung Sound Recognition Method Based on S-ECA-Transformer

Respiratory diseases and traumatic infections are generally accompanied by lung infections. Lung sounds can effectively determine the pulmonary infection level. At present, the lung sound judgment is completed by professional physicians through auscultation. Real-Time detection and assessment of lung sounds in infected individuals represent a crucial clinical challenge. To address this issue, an abnormal lung sound recognition method based on S-ECA-Transformer was studied. The primary research components are as follows: Firstly, a real-time lung sound acquisition system was designed to collect lung sound data of different infection levels. Under the professional physician guidance, the collected lung sounds were annotated to establish a feature dataset (primary care, secondary care, tertiary care, healthy states). Secondly, the lung sound denoising algorithm based on Butterworth filter and wavelet decomposition was studied, the Mel spectrum feature information of lung sound was extracted after denoising. Finally, the improved channel attention mechanism ECA-NET was introduced, and the optimized CNN model was combined with the Transform-Er model to propose the S-ECA-Transformer classification model. This model combined the strengths of CNN spatial features and Transformer sequence coding, providing global feature perception and local feature extraction capabilities. Experiments show that the algorithm achieves 94.5 %, 94.5 %, 94.5 % and 0.94 in recognition accuracy, precision, recall and F1 score respectively. Compared with other deep learning models, this algorithm can effectively identify abnormal lung sounds and determine the infection level.