Sputum sounds are a consistent characteristic of every breath in pneumonia patients. Based on this, this research have designed a novel portable continuous monitoring system that collects 30 seconds of respiratory signals, performs adaptive wavelet thresholding to denoise the signals, and uses a dual threshold method to extract all 1-3 second respiratory sub segments. Optimized Mel-frequency cepstral coefficients are then extracted from these sub-segments for classification and recognition. This research proposes an adaptive wavelet threshold design method based on Bayesian Occam's rule, providing dual threshold methods and related thresholds suitable for this study. This method improves the frequency domain distribution of Mel filter banks and optimizes the support vector machine classifier. This research proposes a feature transformation method based on sine mapping and an ensemble learning method to further improve the classification accuracy of the model. Compared to directly recognizing the 30-second signal, this approach reduces the data volume, avoids overlap of respiratory spectra, and integrates the recognition results of multiple respiratory segments, achieving a recognition accuracy of nearly 100% for the 30-second signal. These optimization methods can be extended to other machine learning models, providing valuable guidance for research in this field.

A Sputum Sound Collection and Sputum Deposition Degree Diagnosis System Based on Improved Support Vector Machine Method