K. Wang, O. Q. H. Tristan, X. Zhang, X. Fu and Z. Qin, "Data-centric AI practice in maritime: securing trusted data quality via a computer vision-based framework," 2024 IEEE Conference on Artificial Intelligence (CAI), Singapore, 2024, pp. 414-417, doi: 10.1109/CAI59869.2024.00082.
Abstract:
The advancement of data-driven Artificial Intelligence (AI) applications becomes increasingly important in optimizing maritime operations. Establishing trustworthy decision-making processes hinges upon precise diagnosis of data quality—a fundamental prerequisite. However, prevalent statistical-based methodologies encounter inherent challenges, such as requiring precise threshold settings, overlooking contemporary insights from recent data, etc. To address these challenges, this study proposes a vision-inspired framework for the classification of Automatic Identification System (AIS) data quality issues, particularly within highly imbalanced datasets. Four typical data quality issues are included in this study, namely Zig-zag Value, Identity Theft, Temporal Missing, and Abnormal Constant Value. The overall framework includes a graphical transformation process to represent the spatial information of the trajectories, a data augmentation process to mitigate the class imbalance issue, and a deep learning model for image classification. Extensive experiments show that 1)The proposed method could achieve an impressive 99.29% accuracy and 99.27% F1 score in AIS data quality issue classification; 2)The ConvNeXtV2 model, an enhanced convolutional neural network, demonstrated its superiority in this application, overtaking other state-of-the-art models by 2.14% in accuracy, 2.35% in F1 score, and 3.40% in MCC;3) The MixUp-based data augmentation method outperformed other imbalance learning strategies such as CutOut, Focal Loss, WeightedLoss, etc. As one of the first few practices on data-centric AI in the maritime sector, this study promises to notably reinforce maritime data reliability, fostering enhanced decision-making processes industrywide
License type:
Publisher Copyright
Funding Info:
This research / project is supported by the Singapore Maritime Institute (SMI) - Programme of Maritime AI Research in Singapore
Grant Reference no. : SMI-2022-MTP-06