Ice Water Navigation
I have decided not to finish my Master's research after one semester. Below is a description of the work I did.
The objective of my master’s research is to develop machine learning models to forecast ice loading on Arctic marine vessels and offshore structures, enabling proactive control systems.
The research involves two primary steps:
Computer vision to detect ice (Mask-RCNN).
Regression model to forecast ice loading predictions.
Detecting Ice Floe:
To detect ice floe, I used the Mask-RCNN machine learning model. This approach offers the advantage of defining precise detection edges rather than bounding boxes, enabling more accurate identification of ice floe.
Below are some results demonstrating the detection of large ice floes:
Dividing up ice detections: To divide up ice data, a polar coordinate system is overlayed on the ship's center. According to literature, ice beyond 2 ship lengths fore, and 0.5 ship lengths aft have no effect on a ship's ice loading, filtering the data fed into the NN.
Overlayed coordinate system on the model ship and detected discrete ice floe centers:
Predicting Ice Loads:
To forecast ice load predictions, a Long-Short Term Memory (LSTM) model is used to forecast ice loading predictions.
Some key parameters fed into NN:
Vessel speed & heading relative to individual ice floe.
Ice concentration as a percentage of each polar cell (as shown above).
Floe perimeter as a percentage of each polar cell (as shown above).
Actual ice collision loads measured in testing.
This approach would enable a similar offshore supply vessel to assess the surrounding ice floes and navigate through them in a way that minimizes the risk of adverse impacts on the vessel. For example, the system could determine an optimal speed based on the ice conditions, considering factors such as collision forces, or select the most fuel-efficient route while avoiding high-risk areas.
