Abstract: Underwater target recognition is a key technology in underwater detection. However, due to the application of modern ship noise reduction technologies, which result in reduced signal-to-noise ratios and blurred spectral features of target signals, traditional recognition methods based on artificial features exhibit limitations in their ability to characterize these features. To address this issue, this paper proposes a hybrid deep learning model combining Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN), aiming to leverage the advantages of both network structures in feature extraction and sequence modeling to achieve multi-level analysis of underwater acoustic target signals. A signal preprocessing workflow including pre-emphasis, frame-based windowing, and energy normalization was designed, and the LSTM-CNN hybrid model was constructed based on this workflow. The proposed hybrid model was experimentally validated, achieving at least a 1.2% improvement in accuracy, recall, F1 score, and G-mean compared to traditional fully connected network models. However, the experimental results also indicate that the model's generalization ability across different pulse width signals remains insufficient and requires further research and improvement.