Ship trajectory anomaly detection method based on encoder-decoder architecture composed of Transformer_LSTM modules

Objective In order to improve the accuracy and efficiency of ship trajectory anomaly detection, and solve the problems of traditional anomaly detection methods such as limited feature characterization ability, insufficient compensation accuracy, gradient disappearance and overfitting, an unsupervised ship trajectory anomaly detection method based on the Transformer_LSTM codec module is proposed.

Method Based on the encoder decoder architecture, the Transformer_LSTM module replaces the traditional neural network to achieve track feature extraction and track reconstruction. By embedding the transformer into the recursive mechanism of LSTM, combined with the cyclic unit and attention mechanism, self-attention and cross-attention can be used to calculate the state vector of the cyclic unit and effectively construct the long sequence model. By minimizing the difference between the reconstructed output and original input, the model learns the characteristics and motion mode of the general trajectory, and trajectories with a reconstruction error greater than the abnormal threshold are judged as abnormal trajectories.

Results AIS data collected in January 2021 is adopted. The results show that the accuracy, precesion and recall rate of the model are significantly improved compared with those of LOF, DBSCAN, VAE, LSTM, etc. The F₁ score is improved by 8.11% compared with that of the VAE_LSTM model.

Conclusion The anomaly detection performance of the proposed method is significantly superior to the traditional algorithm in various indexes, and the model can be effectively and reliably applied to the trajectory anomaly detection of ships at sea.