Abstract: ‍ ‍In fast time-varying multipath channel environments， channel estimation for Orthogonal Frequency-Division Multiplexing （OFDM） systems remains a challenging problem. Traditional Oversampling Complex Exponential Basis Expansion Models （OCE-BEM）， despite their ability to mitigate Inter-Carrier Interference （ICI） arising from Doppler shifts to some extent， are plagued by inadequate noise resistance and inherent modeling inaccuracies， severely limiting their applicability in complex communication environments. To overcome these hurdles， this paper introduces a novel channel estimation method that leverages a residual neural network. This method recasts the data fitting problem as an image denoising task， wherein the estimated channel impulse response is regarded as the original image and the true value as the denoised image. By constructing a data mapping relationship between the two， the discrepancy between the estimated and true values is effectively reduced. The proposed method deploys a residual neural network to eliminate noise errors introduced by the least squares estimation results of OCE-BEM. It fully exploits the data fitting capabilities of neural networks to mitigate the inherent modeling errors of BEM by establishing a robust mapping between the estimated and true channel values. Moreover， the method integrates the real and imaginary parts of the input and output data， thereby preserving the channel’s amplitude and phase information to the fullest extent. In the design of the network architecture， the employed residual structure not only enhances the network’s feature extraction capabilities but also effectively prevents gradient vanishing and gradient explosion during the training process， thus improving network stability and accelerating convergence. Simulation results demonstrate that the proposed method consistently exhibits superior estimation performance compared to conventional BEM estimation methods across a range of signal-to-noise ratios and Doppler conditions.