Time Series Classification Based On Gaussian Process

Chronologically recorded data can be known from all aspects of life time, miningresearches of such data (time series) have theoretical significance in the area ofscientific and engineering application. In this paper, the time series classification andtime series anomaly detection problems were researched and discussed. Byintroducing the approaches of semi-supervised learning and one class classification,which are combined with the machine learning method of Gaussian process, a timeseries of semi-supervised Gaussian process classification was constructed and oneclass Gaussian process for time series anomaly detection model was built, then it wasused in mining the time series data sets, and verified by experiment.First, on the basis of traditional Gaussian process classification model, aSemi-supervised Gaussian Process of Time Series Classification Algorithm (SGPTC)was proposed. This paper analyzed the characteristics of time series data, conductedthe measurement of the similarity features of a time-series, then semi-supervisedtraining data set was constructed by using the labeled samples and unlabeled samples,meanwhile the value of the confidence level of the training classification was judged,thus the semi-supervised classification can be constructed by using achieved trainingset, until the best SGPTC classification model are captured. Finally, the effectivenessof the algorithm can be proven through simulating the data sets of experiment andanalyzing UCI datasets.Secondly, to deal with the problem in the field of specific applications that theabnormal samples are difficult to obtain, an anomaly detection algorithm for timeseries based on One Class Classificaton of Gaussian Process (OCC_GP) is proposed,which is combined with the idea of one class classification. The algorithm combinesthe a priori and Gaussian process regression theory, selects RBF as the kernel function,and constructs a set of feature vectors, which can be used to guide the construction ofone class classifier, through analysis the characteristics of the target time series data.The effectiveness of OCC_GP in the field of the anomaly detection is verified byapplying the algorithm according to the simulated data sets and TE data sets achievingduring industrial process.Finally, combining the characteristics of the time series data and the Gaussianprocess, the semi supervised Gaussian classification based on time series propertieswas realized and a prototype system for the anomaly detection according to the time series was designed and implemented in this paper, which not only demonstrated thefeasibility of the SGPTC classification model and the OCC_GP anomaly detectionalgorithms, but also achieved a corresponding comparison algorithm. Thus it cansatisfy the different requirement of different users for each algorithm function.