Research And Application Of Intelligentlung Sound Detection System

Lung sound is a physiological sound that occurs between the human respiratory system and the outside world during respiration.It is one of the main physiological sound signals of the human body.Abnormal respiratory system,the earliest and most direct response is abnormal lung sounds.The study of lung sound signals has great significance in diagnosis and scientific research.Diagnosing diseases with stethoscopes plays an important role in lung auscultation.At present,the stethoscope used in clinical practice is still mainly mechanical.The stethoscope is prone to errors in the process of auscultation due to easy interference.In addition,due to the subjective reasons of the auscultation doctor,it is impossible to get objective diagnosis results.Nowadaysthe research focus of lung sound detection is to combine electronic signal analysis and machine learning algorithms to improve the performance of the stethoscope and exclude the subjective influence of doctors.It can be used in telemedicine,home medical and other public medical.In this thesis,we aiming at the existing problem of mechanical stethoscope,developthe intelligent lung sound detection system.The system mainly includes the lung sound signal acquisition module,preprocessing module,main control module,analysis module,feature extraction moduleand classification recognition module.Based on the above modules,a portable electronic stethoscope with intelligent diagnostic function is proposed.Based on the characteristics of lung sound signal,this system studies the characteristics of a variety of sensors in the acquisition of signals.According to the research and experimental results,a sensor based on piezoelectric principle,high sensitivity,stable performance and low noise is adopted.The preprocessing module is mainly used to amplify the lung sound signal and filter the high frequency noise.This module is a hardware circuit based on operational amplifier.For non-stationary characteristics of lung,the coincidence problem of heart and lung sounds in frequency domain.We studythe Fourier transform,short-time Fourier transform and wavelet transform,according to thede-noising requirements of the lung sound signal,the improved mixed de-noising technology based on wavelet transform is used to perform the pre-processed lung sound signal.It can effectively remove highly overlappingheart sound signalsin frequency domain,and achieved a good de-noising effect.Base on basic research of general signal feature extraction,this thesis makes an in-depth study on the Mel-scale Frequency Cepstrum Coefficient,and find that the method has good integrity and redundancy for the feature data extracted from the lung sound signal.In the classification and recognition part,from the backpropagation artificial neural network and the K-nearest neighbor classification algorithm,to the deep learning are studied in depth.On this basis,backpropagation neural network,K-nearest neighbor algorithm and deep learning network were used to classify and recognizenormal breath sounds and abnormal breath sound signal.The results show that a pre-training network with feature extraction function Google Net be used to extract the features effectively represent the original signal,furthermore Bi-directional Long Short-Term Memory neural network has anaccuracy rate of more than 90% in the classification and recognition of lung sound signals,and has good stability.Compared with back propagation neural network and K-nearest neighbor classification algorithm,deep learning has better classification effect.