Marine pipelines are crucial for fluid transportation in offshore oil and gas extraction. However, they are prone to damage due to internal corrosion and external environmental factors, which can cause mechanical failure and pose significant threats to industrial safety. Therefore, it is essential to develop accurate damage identification methods. Traditional methods based on modal analysis or conventional machine learning often have limitations in accuracy and damage quantification. In this study, a pipeline vibration experimental setup was constructed to simulate the vibration response of pipelines under random excitation using a shaker. Vibration response data from multiple locations were collected to build a neural network dataset. A new deep learning model, Multi-Source Information Deeply Integrated Dendritic Network (MIDI), was proposed by integrating dendritic networks with multi-head attention and weighted feature fusion mechanisms. Comparative experiments with other neural networks, such as wavelet neural networks and recurrent neural networks, demonstrated that MIDI has excellent performance in identifying and quantifying crack damage in marine pipelines under various scenarios. This research provides a more accurate and reliable solution for marine pipeline damage identification.