The Autoencoder（AE） in the deep neural network is globally optimized through two neural network modules at the transmitter and receiver，and uses end-to-end training to improve the reliability of the communication system. However，the existing research on the AE does not have a special design for the channel，especially for the multipath effect of the time-varying underwater acoustic channel，and thus it is difficult to make flexible adjustments，which reduces the practicability of the method. This paper proposes an Attention-Autoencoder network model to improve the adaptability of the underwater acoustic communication system channel environment. Based on the Attention network's characteristic that it can efficiently filter out key information from a large amount of information，an Attention mechanism for the underwater acoustic channel is designed. The mechanism can increase the ability of the network to extract the characteristics of the underwater acoustic channel and greatly improve the adaptability of the system. Simulation verification and lake test results show that the communication system based on a comparision of. Attention-Autoencoder network model with the AE model in the literature and the underwater acoustic communication system without the introduction of neural networks，has a higher channel environment adaptability.