Time Series Classification Based on Adaptive Feature Adjustment and Multi-scale AGRes2Net

Time series classification is an essential area of research in time series. To target the problem of unsatisfactory multi-scale feature extraction capability and the loss of features in deep learning for time series classification, an inter-module adaptive feature adjustment mechanism (IAM) multi-scale AGRes2Net full convolutional network model (IMAGRes2Net-FCN) is proposed. The time series is processed to add dimensional information to the dataset. A network model of FCN-AGRes2Net with fused IAM is constructed. Feature extraction is performed using FCN. Then, correlations between different AGRes2Net residual blocks are learned by the IAM, and the global features are acquired. The local features obtained by AGRes2Net multi-scale feature extraction are stitched with the global features obtained by IAM. Finally, the features are fed into the classification layer, and the classification results are obtained. The experimental results show that the accuracy of the proposed model is improved, and the PCE is reduced. Compared to the MRes-FCN, AGRes2Net, LSTM-FCN and MACNN on 14 datasets, including Coffee, ItalyPowerDemand, with others, accuracy is improved by 1.13-11.30% on average, the PCE is decreased by 0.14-5.04% on average.

Automated summary

In this study, a new multi-scale AGRes2Net full convolutional network model (IMAGRes2Net-FCN) is proposed to address the issue of unsatisfactory feature extraction capability and feature loss in deep learning for time series classification. The proposed model processes the time series data to add dimensional information and uses a neural network model for feature extraction. Correlations between AGRes2Net residual blocks are learned by an inter-module adaptive feature adjustment mechanism (IAM), and the local features obtained by AGRes2Net multi-scale feature extraction are combined with the global features acquired by IAM. Experimental results demonstrate that the proposed model achieves improved accuracy and reduced PCE compared to other existing models.