Research on time series classification based on convolutional neural network with attention mechanism

Time series classification(TSC) is an interesting and worthy research problem in the field of machine learning. Thus, many convolutional neural network(CNN) algorithms have been proposed to improve the classification accuracy. Among these algorithms, most models solve this task by designing different neural network architectures. In addition, we are inspired by the successful application of the attention mechanism in the computer vision field, which can extract critical information that is beneficial to the target task from the input. Therefore, in this article, we apply 5 attention mechanisms to 6 neural networks and construct 30 models to study classification of time series. Specifically, we choose the attention mechanism to focus on the effective information in the time series from the channel dimension or the spatial dimension. We evaluate the performance of our constructed models on the UCR archive [1], and the experimental results show that the model that processes time series from multiple scales obtains the better results.