Lightweight ConvTimeNet: a simplified fully convolutional model for multivariate time series classification

In recent years, to integrate the local feature extraction capability of CNNs with the long-range dependency modeling advantages of Transformers, researchers have proposed various hybrid architectures. Among them, the ConvTimeNet model has significantly improved the performance of time series classification by introducing deformable patch embedding and deep hierarchical full convolutional blocks. However, there are still questions about the specific contributions of the deep hierarchical full convolutional blocks in the model. To clarify their actual role, this study employed a comparative experimental approach and constructed a lightweight ConvTimeNet (LwConvTimeNet) model, which retains the deformable patch embedding but simplifies the convolutional structure. The experimental results show that the simplified network architecture achieved performance comparable to ConvTimeNet on multiple datasets.