Classify Multivariate Time Series by Deep Neural Network Image Classification

Inspired by the success of deep neural net-works(DNN) in image classification tasks, we proposed a multivariate time series classification(MTSC) method, which fulfills the indefinite-length and indefinite-dimensional time series classification by transforming time series into gray images and classifying images by image classification neural networks. We explored two frequently used image classification neural networks on two imaging methods and explored transfer learning performance when we pre-trained the networks on ImageNet dataset. Both networks in the experiments achieves competitive results compared to the previous time series classification networks. The experimental results show, by using fix-sized images, ResNet18 achieves 2.54% improvement in overall average accuracy compared with the other deep neural network models.