To efficiently extract fishing status from automatic identification system （AIS） data and provide critical technical support for combating illegal， unreported， and unregulated （IUU） fishing， this study proposes a hybrid deep learning model that integrates a multi-scale convolutional neural network （CNN） and a long short-term memory （LSTM） network， enhanced with a multi-head self-attention （MHSA） mechanism. The model takes AIS trajectory segments from four typical vessel types （longliners， purse seiners， trawlers， and squid jiggers） in the Northwest Pacific Ocean as input， comprehensively extracting features such as vessel speed， heading， spatiotemporal location， and their dynamic changes. To address the issue of class imbalance， a class weighting mechanism was incorporated during model training. The experimental results， based on 2020 AIS data， show that the model achieves recognition accuracies of 85.4% for non-fishing and 84.8% for fishing states， demonstrating good stability and discrimination ability. The model exhibits good recognition performance for trawlers， purse seiners， and squid jiggers， with accuracies exceeding 88%. However， the recall rate for longliners is relatively low （73.12%）， indicating that their complex operational behavior poses a challenge for the model. The analysis also reveals that different vessel types exhibit significant variations in fishing speed distribution， diurnal patterns， and seasonal spatial distribution. The developed model effectively identifies the fishing status of multiple vessel types， compensating for the limitations of traditional rule-based or clustering methods in complex behavior recognition. This research can provide technical support for improving dynamic monitoring and regulatory efficiency， enabling IUU fishing behavior detection， and supporting fishery resource assessment.