Time-Frequency Collaborative Learning for Imbalanced Ship Motion Data With Missing Labels in Sea State Estimation | JoVE Visualize

Area of Science:

Marine engineering and oceanography
Artificial intelligence and machine learning
Data science

Background:

Semi-supervised learning (SSL) is crucial for sea state estimation (SSE) but struggles with imbalanced and incomplete ship motion data.
Existing pseudo-labeling methods in SSE are limited by challenges like high class imbalance and missing data.
The need for robust SSL methods that handle data imperfections in ship motion analysis is significant.

Purpose of the Study:

To introduce BalanceSSE, a novel semi-supervised learning approach for sea state estimation tailored for class-imbalanced ship motion data.
To address the limitations of existing SSL methods in handling missing data and class imbalance in ship motion datasets.
To enhance the accuracy and applicability of deep learning models in sea state estimation.

Main Methods:

Dynamic Imputation (DIT): Dynamically imputes incomplete ship motion data by weighting different data dimensions.
Imbalance Temporal-Frequency Learning (ITFL): Utilizes time-frequency collaborative learning for pseudo-label generation and an adaptive confidence strategy for pseudo-label selection.
ClusterProx Classifier (CL): Enhances pseudo-labeling and estimation accuracy through clustering and proximity-based classification.

Main Results:

BalanceSSE demonstrates superior performance compared to state-of-the-art methods on both UCR and ship motion datasets.
Ablation studies confirm the significant contribution of each module (DIT, ITFL, CL) to the overall effectiveness of BalanceSSE.
The proposed method successfully handles class-imbalanced and incomplete data, improving sea state estimation.

Conclusions:

BalanceSSE offers a robust and effective solution for semi-supervised sea state estimation, particularly in the presence of data challenges.
The integration of dynamic imputation, imbalance learning, and advanced classification significantly boosts estimation performance.
This work advances the field of sea state estimation by providing a more applicable and accurate SSL framework for real-world ship motion data.