Abstract: Aiming at the problem that the track information of AIS（Automatic Identification System） of ships in the narrow-mouth section of waterway is prone to error and missing, this paper conducts a statistical analysis of the actual waters and the basic situation of ships in the narrow-mouth section of Ningbo Strip Broom Gate and proposes a method to reconstruct AIS ship trajectories. The proposed method constructs a pre-input model by normalizing the anomalous data and uses a multi-layer feedforward network restoration method based on Dempster-Shafer evidence theory with bi-directional long and short-term memory. The method is compared with other machine learning methods such as BP（Back Propagation） and LSTM（Long Short Term Memory） through experimental demonstration, and the results show that it outperforms other methods in all four dimensions of trajectory restoration. Moreover, the average loss rate of machine learning is 0.048 3% within 20 consecutive lost points, which is lower than other methods. The repaired ship trajectory and data are more complete and consistent with the ship motion pattern.