| With the rapid development of the Internet and information technology,various industries face increasingly large and complex data.One of the main tasks in the application and development of advanced technology today is how to efficiently and intelligently model the characteristics of these data to understand and analyze the valuable information behind the data accurately.Among these complex types of data,multivariate time series data is a typical type of data.Cross-correlation,autocorrelation,and dynamicity are common characteristics of time series data.Among them,cross-correlation represents the correlation between variable dimensions;autocorrelation represents that each univariate time series data has its own temporal dependency relationship;dynamicity refers to the continuous changes and conceptual drifts in statistical characteristics of multivariate time series data over time.Therefore,modeling and analyzing multivariate time series is a challenging task.In recent years,deep learning technology has made rapid development.Due to its advantage of automatically learning nonlinear and complex relationships in data,it has been increasingly applied to multivariate time series data modeling and analysis.However,currently,time series modeling techniques based on deep learning mainly improve their expression ability by constructing complex models,lacking consideration for the noise and instability inherent in time series data.Existing methods often suffer from insufficient generalization ability and poor robustness when facing time series data with complex characteristics and noise.Considering that localized sensitivity can effectively measure the robustness of deep learning models to input disturbances,this thesis aims to achieve robust time series modeling and focus on better utilizing localized sensitivity to enhance the robustness and generalization ability of deep learning.The time series modeling methods are studied around the three complex characteristics of time series data:cross-correlation,autocorrelation,and dynamism.Specifically,research in this thesis is performed using multivariate time series classification,complex multivariate time series prediction,and multivariate time series streaming anomaly detection as typical tasks.Main contributions of this thesis are as follows:(1)Current deep learning methods,which utilize cross-correlation inside time series data to enhance the semantic classification information,may be prone to overfitting and affected by noise.Therefore,this thesis proposes a robust multivariate time series classification method based on frequency domain information and graph adaptive network.This method first uses a frequency representation learning module to achieve automatic suppression of noisy frequency components in each variate dimension,so as to assist in more accurate modeling of the cross-correlation among variates.Then,a graph adaptive network is proposed to adaptively mine the potential corss-correlation and achieve dynamic data augmentation.Finally,by minimizing the localized generalization error,the parameters of frequency representation learning module and the graph adaptive network are optimized.It improves the robustness of the time series classification model and breaks through the limitations of previous methods which are dominated by minimizing the training error to mine and utilize the cross-corrletaion characteristic.Experimental results of Parkinson’s disease detection task based on signals of plantar sensors show that compared with other comparison methods,the proposed method yields the best results in all three performance metrics.(2)To tackle the difficulties of modeling temporal autocorrelation in complex time series data and the complicated model problem,this thesis proposes a robust multivariate time series forecasting method based on adaptive reversible normalization and sensitivity-based multi-view ensemble.This method first makes the non-stationary statistical information of the original input data invisible to the time series forecasting model through adaptive reversible normalization,thereby enabling the forecasting model to model temporal autocorrelation more easily.Then,two different views are constructed.The two views are 1)the hierarchical local temporal information extracted by the Temporal Convolutional Neural Network and 2)the key context sequential information captured by the Bi-directional Long Short-Term Memory Neural Network with Temporal Attention.Meanwhile,the multi-view learning theory is utilized to improve the prediction accuracy of each single view.Finally,a dynamic ensemble method based on localized sensitivity is used to assign higher weights to the more robust views of each sample,improving the forecasting accuracy and robustness of the multi-view ensemble model.It can also avoid the difficulty of parameter tuning and training of a single complicated model.Experimental results of short-term solar irradiance forecasting tasks show that the performance of the proposed method is superior to commonly used time series prediction models.For example,the proposed method yields 1.34%-2.29% lower root mean square errors than those of the best comparison method on four different datasets.(3)Current streaming anomaly detection methods do not perform well in both concept drift and fine-grained detection.This thesis proposes a robust multivariate time series streaming anomaly detection method based on conditional normalizing flow and memory.This method estimates the probability density of each time series variate using its complementary variates as the condition and a conditional normalization flow to achieve fine-grained real-time anomaly detection at the variate level.At the same time,this method utilizes memory to dynamically update the anomaly detection threshold and maintain time-varying normal data.Moreover,a localized sensitivity-based sample weighting method is adopted to update the conditional normalizing flow model online to enable the model to assign higher weights to newly distributed samples while reducing the impact of potential abnormal samples during online model updating.Thus,the robustness and adaptability of streaming anomaly detection methods to concept drift are enhanced.This thesis demonstrates the effectiveness of this method through systematic comparative experiments on four public datasets.In summary,this thesis models three complex characteristics(i.e.,cross-correlation,autocorrelation,and dynamism)of time series data and conducts research on representative multivariate time series classification,complex multivariate time series forecasting,and multivariate streaming anomaly detection tasks.Based on the localized sensitivity,corresponding model training methods,dynamic ensemble methods,and sample weighting-based online updating methods are proposed,which can improve the robustness and generalization capability of deep learning and achieve robust modeling for characteristics of time series data. |
PDF Full Text Request