BCG Signal Feature Analysis And Recognition Based On High-order Statistics

In modern society,due to factors such as high work pressure,environmental pollution,and food safety,people are prone to emergent heart problems,and heart health issues are getting more and more attention.As a result,there is an increasing demand for monitoring of heart activities outside the hospital.The Ballistocardiogram is a non-invasive and non-sensory cardiac activity detection method that is suitable for daily cardiac monitoring in the family.When the heart contracts,blood ejects from the blood vessels and causes the surface of the skin to vibrate,the periodic weak vibrations are traced as the BCG signal.In this paper,a joint acquisition system based on piezoelectric thin-film sensor is designed,and the BCG signals before and after exercise are collected without feeling.The feature extraction,feature analysis and pattern recognition of BCG signals in different states are carried out,and the status model of BCG signal is established to provide assistance for further diagnosis and treatment.The main work is organized as follows:(1)Analysis and extraction of time domain features of BCG signals.After preprocessing the BCG signals,a threshold judgment method is proposed to mark H,I,K and L waves with smaller amplitude based on the localization of J waves by local maximum method.The results show that the proposed method can mark the waveform feature points of BCG accurately.The heart rate before and after exercise was calculated based on the BCG signal and the pulse signal respectively,and the result shows that the difference was small.The time intervals between the H-I,I-J,J-K,I-K,and H-L waves was analyzed and extracted,as well as the JK amplitude and the IJ amplitude ratio,and the differences in the time intervals between BCG signals pre-exercise and post-exercise is compared.In addition,from the principle that the BCG signal is a vibration signal,time domain features such as energy,root mean square,and peak coefficient are extracted.(2)Analysis and extraction of higher order statistic features of BCG signals.Compared with the traditional second-order statistics such as variance,covariance and power spectrum,higher order statistics not only preserves the signal amplitude information,but also preserves the phase information of non-stationary signals.Since higher order statistics have achieved good results in the analysis of EEG signals,ECG signals and other physiological signals,this paper proposes to apply higher order statistics to the analysis of BCG signals.It uses bispectrum to analyze the nonlinear coupling phenomenon for the BCG signal pre-exercise and post-exercise.At the same time,the three-dimensional graph and the contour line view of the bispectrum amplitude of the BCG signal,and the three-dimensional graph and the contour line view of the bispectrum phase were analyzed.Aiming at the shortcomings of large computational in bispectral analysis,a bispectrum slice with less computational complexity is proposed to analysis the amplitude and phase of BCG signal.The maximum of slice spectrum amplitude and the average phase of the slice spectrum is extracted.In addition,skewness coefficients,kurtosis and kurtosis factors are also extracted.The experimental shows that there are significant differences in the characteristics of higher order statistics with different state.(3)BCG signal feature analysis and extraction based on wavelet transform multi-resolution analysis.In order to study BCG signals in detail,a bispectrum analysis and feature extraction method based on small transforms is presented.The signals pre-exercise and post-exercise are decomposed by wavelet.The characteristics of BCG signals are analyzed and studied in different frequency domain subbands.In this paper the db5 wavelet is used to decompose the BCG signal in 6 layers.The subband wavelet coefficients are reconstructed to obtain the subband BCG signal.Spectrum of each subband signal is analyzed,then the subband BCG signal is analyzed using bispectrum and slice spectrum.Through comparison of experimental,we selected the subband energy and the maximum slice spectral amplitude of each subband as the features of BCG signal for pattern recognition.(4)Establish BCG signal classification model based on BP neural network.The parameters of BP network are optimized,and the recognition rate is 79.5% when 33 feature parameters are input.The feature parameter combination with the highest recognition rate is selected as the best input of the network,and the final network recognition rate reaches 83.332%.In the end,the classification effect of pulse signal and BCG signal is compared using the optimized BP neural network model.The results shows that the extracted features and the established classification model have good stability and applicability.