過去的研究中，判斷船舶是否處於航行、錨泊或捕魚狀態等，大多依據船對地面的速度(Speed over ground，SOG)當作參考指標辨識船舶狀態。國際法規的規定和自願裝設下，越來越多船舶裝設船舶自動辨識系統(Automatic Identification System，AIS)，AIS提供巨量的船舶軌跡和眾多的船舶特徵，可以使用AIS提供的特徵來提高偵測船舶行為的機率。 本論文使用台灣沿岸AIS動態船位資料，預先選擇與船舶狀態行為有相關的屬性，採用深度學習方法，建構多層雙向長短期記憶網路模型，分別對拖網、延繩釣和曳繩釣三種漁船作業行為以及貨輪的錨泊行為進行預測。 In past researches, the identification of the ship’s state was mainly based on the ship’s speed over ground (SOG) to determine whether the ship is in a state of navigation or mooring or fishing. Under the provisions of international regulations and voluntary installations, more and more ships are equipped with an automatic identification system (AIS). AIS provides a huge number of ship trajectories and numerous ship features. Features provided by AIS can be used to increase the probability of detecting ship behavior. This paper uses AIS dynamic data along the coast of Taiwan and selects attributes related to ships' behavior. Deep learning is used to build multi-layer bi-directional long short-term memory networks. Finally, the fishing activities of three types as well as the anchoring activity of cargo ships are predicted.