Long-Short-Term Memory Based on Adaptive Convolutional Network for Time Series Classification

Deep learning technology is effective to solve time series classification tasks. The existing deep learning algorithms with fixed step convolution cannot effectively extract and focus on multi-scale features. Based on the complexity and long-term dependence of time series data, an end-to-end model, called as Adaptive Convolutional Network Long-Short-Term Memory (ACN-LSTM), is proposed in this paper. This network is composed of two branches: long-short-term memory (LSTM) and adaptive convolution neural network (ACN). To control the transmission of sequence information, fully extract the correlation information of time series, and enhance the discriminative power of the network, LSTM uses memory cells and gate mechanism. ACN obtains local features of time series by stacking one-dimensional convolutional neural block (Conv1D) and then the multi-scale convolutional neural block is used to capture different scales of information. Meanwhile, to adaptively adjust the feature information between layers, an inter-layer adaptive channel feature adjustment mechanism (ACFM) is proposed. ACN-LSTM not only fully extracts long-term time correlation information, but also enables neurons to adaptively adjust their receptive field sizes, thus, it obtains more accurate classification results. The experiment results on 65 UCR standard datasets show that the proposed ACN-LSTM achieves highest arithmetic and geometric mean rank, and the lowest mean error, which are 2.492 and 2.108, and 0.127, respectively, compared with other models, which indicates that it is effective in univariate time series classification.