Multivariate Time Series Classification Based on MCNN-LSTMs Network

Abstract

Multivariate time series data has high latitude, variable length, coupling, multi-scale and other characteristics. The existing multivariate time series classification methods often extract a single type of feature through complex artificial feature engineering or deep neural network, and do not fully exploit the multi-class features of multivariate time series. Therefore, this paper proposes an end-to-end multi-scale neural network model MCNN-LSTMs for multivariate time series classification. Firstly, using multi-scale entropy and Inceptions structure ideas, the subsequences of each channel are convolved in time dimension by using one-dimensional convolution kernels of different sizes to extract high-level multi-scale spatial abstract features. Secondly, the extracted multi-scale spatial features are input into the FC-LSTM network to further extract their temporal features, and then the temporal and spatial features of the captured features are fused. Finally, the fused features are input into the fully connected layer for classification. The model is tested and evaluated on multiple data sets. The experimental results show that the network model has better classification effect than the existing multiple representative time series classification methods.