Multivariate Time Series Classification (MTSC), one of the most essential tasks for time series data mining, has continuously been attracting great interest over the past decades. For MTSC, previous methods usually extracted temporal features directly from the entire time series, which neglects the inherent multi-periodic and correlated characteristics of multivariate time series. In this paper, we propose a novel network, called Temporal feature Flip Fusion Network(TCFNet), to extract abundant temporal and correlated features from the multi-scale subsequences of multivariate time series, which are decomposed by different periods. To obtain the appropriate multi-scale subsequences, we first transform the time series from the time domain to the frequency domain by utilizing the Fast Fourier Transform(FFT). The original time series is decomposed into subsequences of different scales based on the multiple periods according to top-k frequencies. Then, all the subsequences of each scale are mapped onto a two-dimensional space and we design a two-branch feature extraction and fusion block to extract the features. After handling k scales, all the features are concatenated to complete the MTSC task. To verify the performance of our model, we do experiments on 26 UEA benchmark multivariate time series datasets. The results show that our model gains better accuracy compared to the baselines.