A Lightweight Frequency Multilayer Perceptron for Sea State Estimation Based on Ship Motion Data

Abstract

Accurate and real-time sea state estimation (SSE) is paramount for ensuring the safe and efficient operation of autonomous marine vessels. Current SSE methodologies fail to fully exploit the rich temporal and spectral information embedded within ship motion data, typically employing overly complex architectures that have not adequately explored the potential of frequency-domain multilayer perceptrons (F-MLPs) for capturing global patterns. This paper presents a novel hybrid time-frequency model that addresses these critical limitations. Our proposed framework innovatively integrates time-domain convolutional neural networks (CNNs) with frequency-domain MLPs, yielding a streamlined architecture that demonstrates superior efficiency in extracting complex features from vessel motion data. Theoretically, we establish the mathematical equivalence between frequency-domain MLPs and time-domain convolutional operations. Comprehensive experimental evaluations demonstrate state-of-the-art performance, achieving the highest accuracy on 8 out of 17 benchmark datasets - representing a 1.50% improvement over the leading baseline model FapFormer. Furthermore, on two specialized ship motion datasets, our model exhibits significant performance enhancements, with F1-scores surpassing other advanced baselines by 2.0% (World Wide dataset) and 4.0% (North Atlantic dataset) respectively. Remarkably, these achievements are accomplished with only 55K parameters, establishing new standards for parameter efficiency in SSE systems.These findings collectively highlight our model's exceptional balance between architectural efficiency and estimation performance, offering a promising solution for developing more accurate, computationally efficient, and readily deployable SSE systems for autonomous maritime operations.

Keywords: Autonomous ship, multilayer perceptron, hybrid time-frequency, sea state estimation, time series classification, Lightweight model