A Novel Frequency-Based Global-Scope Adaptive Mixer for Sea State Estimation

Real-time sea state estimation (SSE) is essential for issuing timely warnings about hazardous sea conditions and optimizing ship route selection. In this study, we investigate SSE using the wave buoy analogy, which leverages floating ships as large wave buoys by utilizing ship motion sensor data to estimate the current sea state. Through the analysis of SSE data, we identified three key challenges: frequency distribution differences, multiscale characteristics, and class imbalance. To address these challenges, we propose the frequency-based global-scope mixed prototype network (FGMPN) for SSE. The core of FGMPN is a stackable multiscale feature extractor that integrates a 1-D convolutional neural network and FGA-mixer modules to extract local and global patterns, respectively. The FGA-mixer module is specifically designed to capture frequency difference features, with a group multilayer perceptron (MLP) employed to adaptively learn weights for different frequency components. To address class imbalance, we incorporate a prototype classifier—a distance-based approach that constructs class prototypes based on representative features. The effectiveness of FGMPN is demonstrated through experiments on two simulated SSE data sets and a real-world SSE data set. We also evaluate the generalization capability of FGMPN using the public UEA archive. Furthermore, the importance of each module is validated through ablation analysis. Most importantly, FGMPN shows robust performance on real-time ship motion data, underscoring its practical applicability in real-world scenarios.