Research Of Recognition Algorithms Of Lung Sounds Based On Genetic BP Neural Network

Lung sounds occurs while breathing，as the important human physiological signalsplays important roles in clinical because it is one of the early signals of lung disease aswell as abnormal lung sounds often precedes lung organ disease. In recent years the studyof lung sound has become a hot research project in the world because of its scientific andclinical value. However，as its un-stationary and randomness，the complicacy of itsgeneration mechanism and conduction principle，and the disunity and diversity of itsacquisition device end analysis method，the results of the study of lung sounds are differentin the world. Lung auscultation is an important tool for diagnosis of respiratory illness.With the successive development of computer and electronic technology，the lungdiagnosis in future will realize the electronic collection，computer intelligent analysisfunction. The study of lung sounds diagnosis will get great achievements in identifyingpositive and abnormal lung sounds and inferring the patients disease categories. It will playa supporting role in disease diagnosis.The lung sounds in this paper are obtained by electronic stethoscope including clinicalnormal and abnormal lung sounds. After processed lung sounds by pretreatment based onthe filter and segmentation，then get respectively statistical feature parameter extraction byWelch power spectrum analysis and wavelet analysis. We finally used the neural networkand genetic neural network to recognize and classify lung sounds signal and compare itsrecognition rate and choose the best classifier. This paper constructs a relational model oflung sounds and lung diseases，through identifying wheeze，crepitus and crackles toforecast each kind of lung sounds corresponding to the respiratory disease category.Firstly，we finished the lung sound pretreatment. We obtained four types of lung soundsignals(normal，wheeze，crepitus and crackles) from clinical. After wave filtering and cyclesegmentation we get lung sound samples for analysis. We improves the wavelet filteringalgorithm of filtering the PCG and get heart sound interference components through thewavelet filtering，then obtain purified lung sounds by using the original lung sound minus the heart sound interference. Based on the Viola integral waveform technique，we get thecharacteristic waveform lung sounds and then select cycle segmentation of lung sound.Secondly，we got the lung sound feature. Using improved the Welch power spectrumstatistical characteristic value and the wavelet coefficients statistical characteristic value oflung sound as the signal feature. Statistical characteristics of the Welch power spectrumvalues of lung sounds is combined feature including mean，median，geometric mean，harmonic mean，trimmed mean，standard deviation，four points range，mean absolutedeviation. Lung sound of the wavelet coefficients statistical characteristic value is alsocombined feature including the mean of the absolute values，each with the waveletcoefficient energy，standard deviation of wavelet coefficients between adjacent sub bands，average absolute value of the wavelet decomposition of the second layer to the seventhlayer detail coefficients. Above two respectively characteristics are used as thecharacteristics of the new combination features of lung sound. Compared with thetraditional Welch power spectrum and wavelet coefficient，these two types of new features’dimension is reduced in order to improve the efficiency of classification.Finally，we finished the classification of lung sounds. We study the artificial neuralnetwork(BP) classification principle and genetic algorithm. We use genetic algorithm tooptimize network weights and threshold of the BP neural network and finally getgenetic-BP neural network(GABPNN) named optimization of network. In this paper weused the Welch power spectrum statistical characteristic value and the wavelet coefficientsstatistical features respectively as the input of BP neural network for lung soundrecognition，the recognition of Welch power spectral characteristics is83%，its higher thanthe rate in the wavelet coefficient features rate(of81.1%). As a contrast，we use combinedwith genetic algorithm and BP neural network (GABPNN) classifier with Welch powerspectrum as the characteristics to identify lung sounds. The identification results show thatthe recognition rate of GABPNN is higher，it get the rat of89%，as the BP neural network’srecognition is83.0%. The identification error of GABPNN is smaller than BP neuralnetwork’s and its network is much more stable.This paper’s algorithm is implemented in the MATLAB. The final design of the userinterface can realize loading the lung sound，waveform display and signal analysis，andfunctional disease prediction.