Application Research Of Time Series Data Clustering Algorithm Based On Deep Neural Network

Time series is a common form of data,such as stock price trends,electricity consumption data,patient index data,etc.As an important part of data mining,time series clustering research has attracted much attention.Real-time time series data has high-dimensional,high-frequency noise and other characteristics.Therefore,unsupervised time series clustering research is extremely challenging.Existing time series clustering algorithms can be divided into two categories:original data-based methods and feature-based methods.The idea of the method based on the original data is to design a similarity measurement method based on priori knowledge in a specific field on the original data and for different data scenarios.However,the time series is high-dimensional data in most scenarios,and the data is much scattered in the high-dimensional space.The clustering method based on the original data has insufficient accuracy and poor scalability.Feature-based methods mainly include dimensionality reduction and feature extraction of the original time series to obtain low-dimensional features that can be easily classified,and then clustering methods are used to classify the low-dimensional features.The features extracted by such methods are often linear,but non-linearity is an important feature of time series that cannot be ignored.Deep learning can learn the mapping of high-dimensional data to low-dimensional features and extracts features of data.Therefore,the time series clustering research method based on deep learning has emerged.However,the existing time series clustering methods based on deep learning have not fully extracted the non-linear features of the time series and they are more sensitive to noisy data,and may occur overfitting.Moreover,the data at different time points in the time series have different weights for features.Therefore,we studies the time series clustering based on deep learning methods,and the main work of the thesis is as follows:(1)In view of the problem that the deep autoencoder used in the time series clustering method based on the deep autoencoder model is easily affected by noise,this paper proposes a time series clustering model based on the denoise encoder(Time Series Clustering Based On Denoise Encoder,DTSC),DTSC uses a denoise encoder to make the features obtained by the encoding more robust.The experimental results on the public time series data set UCR verify the validity of the DTSC model.(2)Further,to solve the problem of the problem of distraction in deep learning methods when extracting time series features,the attention mechanism is added to the encoder of the deep encoder,the time series clustering based on element attention mechanism and denoise encoder is proposed,The model(Time Series Based On Denoise And Cell Attention,DATC)also introduces the nonlinear function Gelu to extract the nonlinear features of the time series,which makes the extracted low-dimensional features easier to classify and improves the accuracy of clustering.The experimental results on the public time series data set UCR verify the validity of the DATC model.(3)Based on the above research,we apply the proposed DATC model to the cluster analysis of kidney disease patients.Through the DATC model,the patient time series data is clustered into different categories in an unsupervised manner,and then the difference characteristics among categories and the hidden features of all patients in the same category are analyzed for clinical medical reference.