Deep Learning Architecture for the Classification of Time Series Data of Different Domains Using Convolutional Neural Network

A time-series dataset (TSR) is a collection of linked, chronologically and temporally gathered data points. These dataset objects are experimented for classification or prediction in the field of machine learning. Time series forecasting calls for complex methods like DTW, LSTM, or ARIMA. To classify them, however, techniques such as bag-of-words, neural networks with convolution, interval-based (TSR forest), shapelet-based algorithms, and k-nn with dynamic temporal warping could be applied. Most of the works in the available literature applies time-series data directly to CNN. Here, the CNN library functions for image classification are employed, converting time-series data into the images before inputting. This article uses CNN to classify five univariate datasets from the UCR collection: Adiac, Beef, 50Words, Lightning7, and Olive oil. Deep learning architecture (CNN) are discovered to be superior to conventional artificial neural networks (ANN), since the latter are seen to be more time and space intensive and to have a lower accuracy rate.