Research On Feature Extraction Of Pulse Signal And Clustering Of Pulse Image

Wrist pulse diagnosis is one of the most popular diagnostic method in the Chinese medicine.It has a long history and lasting heritage.Traditional pulse diagnosis methods uses tactile perception to obtain pulse information and to determine the corresponding TCM symptoms based on clinical experience.The method has diagnostic results with a strong subjectivity.Therefore,the lack of objective quantitative standard in the modern medical system has held back its development in the further.It is a powerful tool to inherit and develop pulse diagnosis by using the machine learning recognition method to study the pul se characteristics.The pulse image,which is defined according to the shape,rhythm and pressure of the pulse signal,is the definition standard of different pulse in pu lse diagnosis.The pulse feature extraction and clustering method is devoted to combining pulse diagnostic theory of TCM with machine learning,using the latest machine learning method to extract the pulse features,and mining the types of the pulse image by the clustering method.At present,there are four main drawbacks in pulse pattern recognition,for example,the expression of pulse image is weak,the convergence of multiple processes is difficult to be optimized,the method of identifying multi-period pulse is lacking,and the definition of pulse pattern is indefinite.This topic mainly aims to improve the four shortcomings of pulse pattern recognition method.In the aspect of single-cycle pulse feature extraction,the curvature segmentation method reduces the cycle segmentation error rate.The wavelet method eliminates the influence of the wake fluctuation in the single-cycle pulse train.In the part of feature extraction,the pulse signal is decomposed into periodical features a nd aperiodic features.In the experimental comparison of single-period features,the accuracy of the classification of the fused features after the period decomposition is improved by 2.7% compared with the one-dimensional feature combination.In the multi-cycle vein pulse feature extraction,a 14-layer convolutional neural network with one-dimensional input is constructed in this paper.The advantages of end-to-end learning avoids the problem of error transfer caused by multi-stage flow optimization.And multiple convolutional layers and multi-layer features fusion can enhance the expression of features greatly.The experimental results show that multicycle convolutional neural network characteristics compared gets an improved accuracy by 1.3% with the network without multi-layer features fusion.And it gets an improved accuracy by 4.7% compared with the VGG16 network.Finally,the pulse recognition accuracy is improved by 5.8% compared with the single-cycle twodimensional fusion feature.In the aspect of pulse image clustering and disease analysis,the paper constructed a clustering integration method based on mutual information value to cluster the pulse images and make some preliminary disease analysis based on the pulse images.The results show that some of the clustering pulse images are consistent with TCM pulse patterns and they also have found more detailed pulse images.In addition,by analyzing the distribution of the cardiovascular disease in different maps,the results showed that health,heart disease,hypertension and dyslipidemia can be classified by some pulse images.