At present, most of the analog circuit fault detection is aimed at component-level fault detection. In this paper, we propose a module-level soft fault detection method for the core module of a typical underwater acoustic sensing system. We use Simulink to implement function-level modeling of typical underwater acoustic sensing systems. The simulation data are obtained by establishing multiple measurement points, and the Short-Time Fourier Transform(STFT) is used to represent different fault modes so that the convolutional neural network can learn effectively. We design a deep learning model based on convolutional neural network, introduce an expansion bottleneck layer and an attention mechanism to improve the learning ability of the network. The model is then pre-trained by simulation data. Collecting experimental data for circuits that have been really applied to engineering practice, using less experimental data to fine-tune the model. Finally, the experimental results show that the proposed method can realize the soft fault detection of the core module of typical underwater acoustic sensing system under the premise of a small amount of real circuit data. Display omitted •Aiming at the core module of the underwater acoustic perception system, we propose a fault detection method at the module level.•A deep learning network model is proposed to realize the soft fault detection of the core module of the underwater acoustic perception system.•We use simulated data to pre-train the deep learning model to reduce the dependence on real data.•The experimental data is come from underwater acoustic perception systems that have been used in practical engineering.