Research On Classification Of T-CNN Time Series Based On Gram Matrix

In recent years,with the development of information technology,time series has become more widely used,such as disaster monitoring,security analysis,finance and business,including massive data with time attributes.These data have the characteristics of large scale and many types,which contain great value.Therefore,how to accurately classify time series is the basis of flow data event analysis and data mining,and also the focus and difficulty in the field of data flow.Time series data is a real value data waveform obtained by monitoring a certain process according to a given sampling frequency.It is recorded continuously with time stamp change and is not affected by system environment and other factors.Time series data has not only the numerical elements of the conventional data,but also the time elements.Most of these methods do not fully retain their timing,loss characteristics during classification and low classification accuracy.The existing time series classification methods are mainly based on symbolic aggregation method,time domain distance method,Shapelet classification method and convolution neural network classification method.Therefore,this paper proposes a time series classification method based on Gram matrix,which transforms time series into time domain images without loss and proposes an improved convolution neural network T-CNN to classify time domain images.The specific contents are as follows:Firstly,the time series will inevitably carry a certain degree of normal background noise in the acquisition process,and the time domain image transformed by the Gram matrix shows the normal distribution.So the adaptive wavelet threshold denoising method is used to filter the normal background noise carried in the time series.After denoising,the time series is normalized and transformed into time domain images with Gram matrix,which preserves its timing and improves the classification accuracy.Secondly,the transformed time domain images are input into the convolution neural network for classification.The convolution layer computing of traditional convolution neural network needs to move continuously convolution kernels according to the given step length and iterate to obtain the convolution result.This paper uses the Toeplitz matrix to construct the convolution kernel,obtains the Toeplitz convolution kernel,uses the Toeplitz convolution kernel and the time domain image to calculate the matrix product,obtains the same result as the traditional convolution calculation,replacing the traditional convolution calculation and reducing the time complexity.Thirdly,in the full connection layer of convolution neural network,the idea of triplet network is introduced to compute the difference functions between the same class and different classes in time series,and the above two difference functions are added to the square loss function of CNN model to obtain the triplet loss function.And then propose the T-CNN classification model to improve the accuracy of the convolution neural network.Compared with the existing four time series classification methods,the experimental results show that the proposed method has better classification performance.Lastly,experimental validation is performed on multiple real datasets and compared with four existing time series classification methods.Experimental results show that the proposed method has better classification efficiency and higher classification accuracy.