Research And Application Of Cardiopulmonary Sound Separation Based On Non-negative Matrix Factorization And Neural Network

Accurately diagnosing the condition of heart and lung system and adopting effective treatment methods are the guarantee for the early recovery of patients with cardiopulmonary diseases,and the accurate diagnosis of heart and lung system diseases depends on the accurate of physiological signals.In clinical practice,methods for examining the physiological characteristics of the cardiopulmonary system include electrocardiogram,chest X-ray,and auscultation of cardiopulmonary sound,etc.Since cardiopulmonary sound signals contain more comprehensive cardiovascular and cardiopulmonary pathological information,and are more convenient than other diagnostic tools,diagnostic efficiency is also relatively high,so auscultation of cardiopulmonary sound is the main diagnostic method for cardiovascular disease and respiratory disease in clinical applications.In the course of clinical auscultation,the heart sound and lung sound signal collected by the stethoscope is usually a mixed signal with both background noises.When the doctor auscultates the heart sound signal,it is disturbed by the lung sound signal,and the lung sound signal is processed,the heart sound signal will interfere with the lung sound signal in turn.In addition,indoor and outdoor murmurs are also obstacles for doctors to accurately diagnose heart and lung diseases.In order to overcome the above difficulties,the most critical technology is the separation of mixed signals.Therefore,this paper uses the separation scheme designed by the non-negative matrix factorization theory and the long-short-term memory network combined with Hidden Markov Models separates the signals of the heart and lung sounds,the main content as follows:1)Study how to use non-negative matrix factorization to separate single-channel cardiopulmonary mixed signals.Based on the analysis of the existing non-negative matrix factorization algorithm,we propose a new non-negative matrix factorization algorithm combining non-negative matrix factorization and autoregressive models,and name this method as autoregressive regularized nonnegative matrix.Then we used this method to design a separation scheme that separates the heart sound and the lung sound from a single-channel heart-lung sound mixed signal.2)Using non-negative matrix factorization method based on autoregressive regularization to separate the cardiopulmonary sound from the simulation experiment,combined with the necessary simulation code and the experimental real data,it fully reflects the entire experimental process.Moreover,the experimental data were analyzed effectively,and the superiority of the proposed method was proved by the experimental results and comparative analysis.3)Study use separation scheme combining Long Short Term Memory networks and duration-dependent hidden Markov model to separate single-channel mixed cardiopulmonary sound signals and obtain medical signals of interest.In the background of the rapid development and widespread application of neural networks and deep learning,relevant people have used various types of neural networks to conduct applied research in various fields.There are many people who use them to separate heart sounds and lung sound signals from mixed heart and lung sound signals,and background noise.On the other hand,the HMM process can be statistically modeled based on the observation output of the timing signal,and can effectively segment the timing signal.Moreover,with the advent of duration-dependent hidden Markov model,the segmentation effect can be greatly improved according to the background characteristics.We combine the physiological characteristics of heart-lung sound signals and effectively to propose a mixed signal segmentation and separation scheme.We use the Long Short Term Memory networks to perform preliminary separation of the mixed signals(or to understand them as denoising),and then divide the preliminary separated interest results through the duration-dependent hidden Markov model to obtain different sub-state sequences of the signals of interest.Then we use Long Short Term Memory networks to segment and details or local denoising for the sub-state sequence for each state.After denoising,each sub-state is spliced into a complete first processed post-signal according to the characteristics of physiological signals.Next loop iteration complete signal separation--interest signal segmentation--sub-state local detail denoising--splicing this process,output the signal of interest when a certain termination condition is satisfied,then we separate the interest from the mixed heart and lung sound signals signal.4)Using the proposed Long Short Term Memory networks and duration-dependent hidden Markov model segmentation separation scheme to simulate the heart and lung sound signal separation.Simulation experiments show that the separation process of Long Short Term Memory networks and the segmentation process of duration-dependent hidden Markov models promote each other and influence each other,making the separation and segmentation effects all increase.Then the result of the whole simulation experiment was analyzed,which proved the feasibility and superiority of the p roposed method.