This paper proposes a time-frequency domain signal reconstruction method for non-Gaussian interference in underwater acoustic communication pulse signals. The method is based on an improved deep residual network. The gradient disappearance problem in deep network training is solved by constructing adaptive residual blocks. The non-stationary characteristics of pulse signals are adapted by optimizing the residual connection mechanism. A multi-level residual structure is used to enhance the network's capabilities. The network uses a complex number field joint mean square error loss function. This loss function combines spectral consistency constraints. The loss function also combines a phase-sensitive mechanism. This improves the fidelity of the reconstructed signal in the time-frequency domain. The experimental results show that under simulated conditions, when the signal-to-noise ratio and signal-to-interference ratio are both low, this method is effective. It can restore pulse signals that are drowned out by noise and interference. It significantly improves the signal distortion ratio. It significantly improves the signal-to-interference ratio gain. Therefore, this method provides a reliable solution for signal processing in complex underwater acoustic environments.