Abstract: In order to solve the problem that it is difficult to accurately estimate the fast time-varying channel state information in orthogonal frequency division multiplexing (OFDM) communication system of high-speed railway in high-speed mobile scene, an OFDM channel estimation method based on multi-scale residual attention network was proposed. Firstly, we design the multi-scale channel feature extraction structure. For the low-quality channel matrix, we apply the multi-scale convolution kernel to extracting the shallow multi-dimensional feature information, which can improve the extraction performance of channel feature information with different scales. Then, a multi-scale residual attention cascade depth network is constructed for channel feature reconstruction and mapping. The local residual feedback is combined with CBAM (convolutional block attention module) attention mechanism to promote the fusion and utilization of deep features and improve the reconstruction and mapping ability of OFDM channel matrix. Finally, the sub-pixel convolution reconstruction is used to generate a high-resolution channel matrix to complete the channel estimation. The analysis in both frequency domain and time domain show that the proposed channel estimation method is better than other methods in terms of accuracy and complexity of channel estimation and can satisfy the needs of OFDM channel estimation.