Research on the inversion method of ice loads on ships in broken ice channels using deep neural networks

• Deep Neural Network (DNN) is used to identify the ice loads acting on ship structures. • The ice load dataset is obtained using the CFD-DEM coupled method. • Effective measures to mitigate noise interference are discussed. • The model test reveals the applicability of the DNN model in practical applications. Accurate ice load prediction is essential for ensuring the navigational safety of polar vessels, with full-scale ship measurements being the most reliable method for obtaining ice load data. However, due to harsh environmental conditions and the complexity of ship-ice interactions, directly measuring ice loads with force sensors is impractical. Instead, ice loads are identified through the structural response of the hull or adjacent regions. This study proposes an ice load inversion method using deep neural networks (DNN) to overcome the limitations of traditional load inversion methods. High-fidelity datasets are generated using a coupled computational fluid dynamics-discrete element method-finite element method (CFD-DEM-FEM). A deep feedforward neural network model is constructed, and the DNN is trained and tested using the generated high-fidelity dataset. The accuracy of the DNN-based ice load inversion method is evaluated, and a scaled model test is conducted to explore its potential for engineering applications. The results indicate that the proposed DNN model exhibits exceptionally high inversion accuracy under noise-free conditions, with a mean squared error of only 0.98 % for the test samples. More importantly, when non-Gaussian noise is introduced into the input data, the model’s performance is notably affected, with the average error increasing to 5.73 %. However, by optimizing the loss function to the Huber loss, the model’s robustness was significantly enhanced, successfully reducing the average error to 1.59 %. Finally, in the scaled hull grillage model tests, the correlation coefficient between the inverted and measured loads exceeded 0.99. This indicates that the proposed method is not only theoretically feasible but also has significant potential for engineering applications.