The Abnormal Detection Method of Ship Trajectory with Adaptive Transformer Model Based on Migration Learning

Abstract

The traditional modeling method of ship trajectory data is unable to deal with multiple features or ignore the temporal relationship. Moreover, due to the problems of data imbalance and lack of data labels, it is a great challenge to build ship behavior analysis and anomaly detection models based on trajectory data. Based on transfer learning and Transformer architecture, this paper proposes an anomaly detection method for adaptive Transformer model fitting trajectory timing distribution characteristics. Firstly, the trajectory data are preprocessed to simulate the characterization relationship of time-series data. Multiple different subsequences are divided by quantization of temporal distribution similarity and matching of temporal distribution. From the perspective of transfer learning, each subsequence is regarded as multiple source domains of independent distribution. The distribution correlation weight evaluation method is used to learn and fit the distribution characteristics of subsequences, to effectively describe the similarity and correlation effects between the distributions of each subsequence. At the same time, the attention mechanism is used to build the dynamic training Transformer network structure, mining the dependencies between data points in the sequence, and realizing a domain adaptive model that can effectively fit the multivariate characteristics of trajectory data. Finally, the anomaly detection task is completed by calculating the anomaly error. The example verification shows that, compared with the existing model, the model in this paper can more accurately realize the modeling of the normal trajectory motion characteristics, lay a foundation for the trajectory anomaly detection, and show a certain practical and engineering application value.

Kexin Li (1998-), master student.