Interpretable modeling ship steering dynamics via Kolmogorov-Arnold Network with self-attention mechanism: an experimental study

This paper explores the cross-engineering field extension of the Kolmogorov-Arnold Network (KAN) to ship steering dynamics modeling. KAN is a novel network architecture with a mathematical foundation distinct from the classical Multi-Layer Perceptron (MLP). To further enhance modeling accuracy, a self-attention mechanism is integrated with KAN, forming an Attention-KAN framework. To evaluate the potential of KAN as a robust alternative to MLPs in modeling ship dynamics and to assess the effectiveness of the Attention-KAN mechanism, a series of lake-based experiments utilizing the Qiuxin-VII trial ship (operated at East Lake Laboratory, Hubei Province, China) are conducted to collect and construct the ship steering motion datasets. Comparative tests are performed among MLP, KAN, Attention-MLP, Long Short-Term Memory (LSTM) and Attention-KAN to analyze and evaluate the performances of the trained ship steering models. The results demonstrate that Attention-KAN achieves higher accuracy in building a lightweight ship steering dynamical model, and the integration of the self-attention mechanism yields further improvements over KAN. Moreover, KAN offers better interpretability in modeling the ship steering dynamics, enabling a more transparent understanding of the model's internal characteristics. • Kolmogorov-Arnold Network (KAN) is utilized to explore the cross-engineering field extension of ship steering dynamics modeling. • Self-attention mechanism is integrated with KAN to analyze the internal characteristics of input data and capture the local information. • A series of lake-based experiments utilizing the Qiuxin-VII trial ship are conducted to construct the dataset and evaluate the effectiveness of the proposed method. • KAN offers better interpretability in modeling the ship steering dynamics, enabling a more transparent understanding of the model's internal characteristics.