This project is concerned with the development of a machine learning-based DOA estimation of underwater acoustic signals with the view of improving the efficiency and reliability of the signal processing in complex underwater environments. The feature analysis and classification of the underwater acoustic signals are done with the help of convolutional neural networks (CNNs). Through multi-step purification of the signal data and lowering the sampling frequency of the pre-processed signal data, a high-performance arrival angle assessment framework is constructed. The results indicate that the algorithm achieves high accuracy, recall and F1 in different noise intensity levels; particularly, low noise environment with an accuracy of 95%. 2%. The application verification of practical application also confirms that this technology has strong applicability and stability in complex seabed conditions.