Research On Classification Methods Of Pulse Signal Based On Unbalanced Samples

Pulse diagnosis is one of the four clinics of traditional Chinese medicine,which is a common way for Chinese medicine to diagnose diseases.Pulse diagnosis contains a wealth of pathological information and is a painless and non-invasive diagnostic method."The pulse is the key to the doctor,if the doctor does not observe the pulse,there is no evidence,and if the certificate is not different,there is no solution"(Xu Chunzhen "Ancient and Modern Medicine"),from which we can see the importance of pulse diagnosis in the diagnosis and treatment of the disease in traditional Chinese medicine.With the rise of artificial intelligence in recent years,the combination of machine learning and medical diagnosis has also promoted the development of pulse diagnosis in TCM.In the actual scene,there is a serious data imbalance problem in the pulse data?Training the classifier directly based on unbalanced pulse data can easily lead to poor classification effect of the classifier on a small number of pulse signal samples.However,in the real world,people tend to pay attention to the corresponding diseases in a few samples of pulse signals,for example,in both health and cancer samples,cancer is more important.If the classifier does not identify a small number of disease samples well,misclassifying them into majority samples will result in significant losses.Based on the above problems,this paper analyzes the characteristics of the pulse signal,and studies from the aspects of feature selection,unbalanced processing of the pulse signal,feature fusion,and multi-classification of the pulse signal.In terms of feature selection of pulse signals,PCA algorithm is used to reduce the dimension of the pulse signal,and then classification separability criterion is used to select the features of the pulse signal,so as to obtain the optimal feature set.The processing of unbalanced pulse signal is mainly solved from data and algorithm.At the data level,this paper improves the downsampling method based on genetic algorithm.This paper improves the weak classifier for measuring chromosome fitness value in genetic algorithm into a strong classifier suitable for pulse signal.In order to select the information-rich majority of pulse signal samples,improved algorithm tries to maximize the performance of the original classifier,and minimizes the loss between the original majority of the pulse signal samples and the majority of the pulsed signal samples after the downsampling.At the algorithm level,this paper improves the gravitation method based on fixed neighbors.The main idea of the improved algorithm is to treat the pulse signal samples in the training set as quality entities with gravitational effects between the entities.By using the pulse signal to test the gravitational sum of sample K's neighbor points to determine the label of the test sample.The experimental results show that the two improved algorithms can improve the classification performance of unbalanced pulse samples to some extent.The multi-set canonical correlation analysis method is used to fuse the Gabor feature,STFT feature,pulse two-dimensional matrix feature and wavelet feature of the pulse signal,and adopt one-to-one multi-classification method,multi-classification method based on undirected graph and error correction based on the coded multi-classification method completes the multi-classification experiment of the pulse signal samples.