The Application Of Statistics To The Analysis Of Pulse Signals

Pulse feeling is an important part in the"Four Diagnoses"and the most special diagnostic method in traditional Chinese medicine. With the development of the modern computer technology, people hope to employ the modernized instrument and approach for pulse signal processing to study and probe into the pulse-feeling diagnosis objectively and provide some judgment basis for clinic medicine.The parameter model analysis of time series is always the most basic and effective method in engineering application field, which is holding a wide attention of the scholars from signal processing field in domestic and abroad. One of the essential characteristics of time series is the dependence between neighboring observed values. It is an undoubtedly good idea to describe this physical phenomenon with the time series data matched model, in order to get an accurate reflection and depiction of the law of their own developments and changes. Moreover, it is also widely applied in various fields to use the model coefficients to represent the characteristics of the original data. This paper extracts characteristics from the collected pulse signal data with this approach and provides lower dimensional data for later classification.BP Nerve Network (NN) is the most widely-used artificial NN so far and exerts great advantages especially in non-linear fields. The standard BP NN has some inherent shortcomings, which may cause a great waste of resources in applications. Therefore, this paper,based on the characteristics of the standard BP Algorithm, adopts an improved BP Algorithm, i.e. LM Algorithm. By comparison between these two algorithms, it is found that LM Algorithm is more suitable for the study in this paper. The experiments show that it has a high identification rate to use the eigenvectors extracted with the approach mentioned before as the samples for network training and testing, and then identify the druggers from normal people.Aimed at addressing the problem that the differences in pulse signal characteristics between druggers and normal people are too delicate, this paper employs two kinds of time series statistic to extract the characteristics, namely to extract the characteristic parameters from the collected data with self-correlation coefficient and AR model coefficient respectively. The most critical and important thing for modeling is to determine the order: too high or too low order can neither depict the precise essential characteristics of the pulse signals and will consequently impact on the identification work. This paper combines the two widely used approaches, correlation analysis and residual analysis, to determine the order of the pulse signals. It is proved by a number of experiments that when the order is 3, it will reflect the original signal characteristics best. This reduces the original 90 dimensional vectors to 3, facilitating the identification of the BP NN.Among the 30 subjects in the study, 20 are arranged for training, with the druggers and normal people fifty-fifty; the remaining 10 are for testing, with the druggers and normal people also fifty-fifty. When the LM algorithm is used in AR model coefficients for identification, only one drugger, numbered"b12", out of the 10 subjects is misidentified as normal person. Therefore, we can see it is feasible to identify the pulse signals in Chinese medicine based on statistic analysis parameters of time series and improved BP NN.