Image based real-time ice load prediction tool for ship and offshore platform in managed ice field

Akter, Shamima (2023) Image based real-time ice load prediction tool for ship and offshore platform in managed ice field. Masters thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (2MB)

Abstract

The increased activities in arctic water warrant modelling of ice properties and ice-structure interaction forces to ensure safe operations of ships and offshore platforms. Several established analytical and numerical ice force estimation models can be found in the literature. Recently, researchers have been working on Machine Learning (ML) based, data-driven force predictors trained on experimental data and field measurement. Application of both traditional and ML-based image processing for extracting information from ice floe images has also been reported in recent literature; because extraction of ice features from real-time videos and images can significantly improve ice force prediction. However, there exists room for improvement in those studies. For example, accurate extraction of ice floe information is still challenging because of their complex and varied shapes, colour similarities and reflection of light on them. Besides, real ice floes are often found in groups with overlapped and/or connected boundaries, making detecting even more challenging due to weaker edges in such situations. The development of an efficient coupled model, which will extract information from the ice floe images and train a force predictor based on the extracted dataset, is still an open problem. This research presents two Hybrid force prediction models. Instead of using analytical or numerical approaches, the Hybrid models directly extract floe characteristics from the images and later train ML-based force predictors using those extracted floe parameters. The first model extracted ice features from images using traditional image processing techniques and then used SVM and FFNN to develop two separate force predictors. The improved ice image processing technique used here can extract useful ice properties from a closely connected, unevenly illuminated floe field with various floe sizes and shapes. The second model extracted ice features from images using RCNN and then trained two separate force predictors using SVM and FFNN, similar to the first model. The dataset for training SVM and FFNN force predictors involved variables extracted from the image (floe number, density, sizes, etc.) and variables taken from the experimental analysis results (ship speed, floe thickness, force etc.). The performance of both Hybrid models in terms of image segmentation and force prediction, are analyzed and compared to establish their validity and applicability. Nevertheless, there exists room for further development of the proposed Hybrid models. For example, extend the current models to include more data and investigate other machine learning and deep learning-based network architectures to predict the ice force directly from the image as an input.

Actions (login required)

View Item