Dynamic Pattern Classification Method Research Based On Deterministic Learning

Time series analysis is an indispensable part of big data analysis,and Time Series Classification(TSC)is an important research direction in time series analysis.The research of TSC is crucial for fields such as healthcare,finance,traffic control,and industry,as it has broad application value.In recent years,researchers have proposed numerous time series classification models,which can effectively understand the patterns and trends of time series data and provide more scientific basis for decisionmaking.This article innovatively applies a dynamic pattern recognition method based on Deterministic Learning Method(DLM)to solve the classification problem of dynamic patterns(time series with class period or regression properties can also be called dynamic patterns),and proposes a dynamic pattern classification method.This method uses dynamic distance metrics as the classification basis and achieves better classification performance than traditional distance metrics in large-scale dynamic datasets generated by dynamic systems.Furthermore,by combining the dynamic distance metrics of this method with traditional time-domain distance metrics,a dynamic pattern classification method based on multi distance metrics feature fusion is proposed,which has achieved good results in both public datasets(UCR datasets)and practical clinical problems(auxiliary diagnosis of adenoid hypertrophy in children).The main content of this study includes the following three aspects:Firstly,in response to the research question of dynamic pattern classification methods,this article extends the dynamic pattern recognition method to a dynamic pattern classification method for the first time,and uses deterministic chaotic exploration technology to construct a large-scale time series dataset to verify the method.On this large-scale dataset,we designed multiple classification scenarios to validate this method and compared it with various mainstream time series classification methods.The experimental results show that the dynamic pattern classification method still has good performance on large-scale datasets,and this classification method can solve the early classification problem of time series.Secondly,in response to the question of improving the performance of dynamic pattern classification methods,this paper proposes a dynamic pattern classification method based on multi distance metric feature fusion.Firstly,we use distance based feature extraction methods to improve the dynamic pattern classification scheme based on nearest neighbor classification by combining dynamic distance features with support vector machine(SVM)classification strategy.Secondly,in order to measure the similarity between time series data more fully and comprehensively,we not only consider the dynamic distance features obtained above,but also consider the distance features in the time domain,and propose a dynamic pattern classification method based on multi distance metrics feature fusion.In order to eliminate the magnitude difference between different distances,we propose a negative exponential transformation strategy.The effectiveness of the proposed method was verified on the UCR public dataset,and its classification performance is superior to existing distance feature based classification methods.Finally,in response to the practical clinical problems of auxiliary diagnosis of adenoid hypertrophy in children,the above dynamic pattern classification method was used to carry out research on auxiliary diagnosis of adenoid hypertrophy diseases based on nasal airflow.The nasal airflow signal can essentially be regarded as a nonlinear signal generated by the respiratory dynamics system,and the obstruction of the respiratory tract caused by adenoid hypertrophy can cause changes in nasal airflow.We used 74 samples collected from the Otolaryngology Department of Shandong Provincial Hospital from 2021 to 2022,as well as 21 samples we collected of healthy volunteer samples,to construct a dynamic dataset containing adenoid hypertrophy and health.The proposed method was validated on this dataset,and experimental results showed that it can provide high accuracy and can be used as an auxiliary diagnostic tool for adenoid hypertrophy in children.