Self-Supervised Time-Series Preprocessing Framework for Maritime Applications

This study proposes a novel self-supervised data-preprocessing framework for time-series forecasting in complex ship systems. The framework integrates an improved Learnable Wavelet Packet Transform (L-WPT) for adaptive denoising and a correlation-based Uniform Manifold Approximation and Projection (UMAP) approach for dimensionality reduction. The enhanced L-WPT incorporates Reversible Instance Normalization to improve training efficiency while preserving denoising performance, especially for low-frequency sporadic noise. The UMAP dimensionality reduction, combined with a modified K-means clustering using correlation coefficients, enhances the computational efficiency and interpretability of the reduced data. Experimental results validate that state-of-the-art time-series models can effectively forecast the data processed by this framework, achieving promising MSE and MAE metrics.