A method for generating ship collision risk test scenarios based on conditional variational autoencoders

To test and evaluate the decision-making capabilities of intelligent ships during navigation under varying collision risk conditions and increase the likelihood of encountering scenarios with controllable risk levels, this paper proposes a method for generating ship collision risk test scenarios based on conditional variational autoencoder (CVAE).

Using ship data from the automatic identification system (AIS), the relative motion parameters of ships are calculated. Combined with the "International Regulations for Preventing Collision at Sea （COLREGs）" and a fuzzy rule base, ship encounter situations and ship collision risk levels are classified to construct a ship encounter-risk dataset. The CVAE model is then trained on this dataset, with the relative motion parameters serving as input features and the encounter situation and the collision risk level as conditional variables. After training, the model generates corresponding test scenarios by inputting the conditional variables and the initial state of the test vessel. Additionally, during navigation, the distance to closest point of approach（DCPA）, time to closest point of approach (TCPA), and collision risk index (CRI) are dynamically displayed with a 10-second time step. Comparative experiments with a variational autoencoder (VAE) and random sampling methods are conducted to verify the accuracy and similarity of scenario generation across different models.

The results indicate that, under the specified encounter situations and collision risk levels, the CVAE model generates test scenarios that simultaneously satisfy both preset conditions with an accuracy of 93.54%. In comparison, the VAE model achieves only 0.08%, and the random sampling method 0.83%, representing improvements of 93.46% and 92.71%, respectively.

The proposed method improves the effectiveness, diversity, and realism of the generated encounter situations and ship collision risk scenarios, demonstrating superior accuracy in scenario generation for intelligent ship navigation.