Ice condition assessment using onboard accelerometers and statistical change detection

The presence of sea ice is the predominant risk for ship operations in the Arctic, and monitoring of ice condition around a vessel is crucial during all times of operation. This paper presents a system for online onboard assessment of ice condition. It is demonstrated that ice-induced accelerations in the bow section of the hull follow a bivariate t-distribution and that parameters of the distribution have a one to one relation to ice condition. The paper suggests a methodology to monitor the ice condition in real time through estimation of parameters that characterise the distribution of hull accelerations. It is shown how a Kullback-Leibler divergence measure can classify ice condition among a set of pre-trained conditions. An absolute measure of ice load is suggested as an alternative for situations when pre-training data are not available. The alternative algorithm quantifies the condition through the entropy of measured accelerations. The article presents a computationally easy methodology and tests against data collected during Arctic transit of an icebreaker. Further, the classification results are compared with the results from two standard methods from machine learning, decision tree and a support vector machine approaches. The results show that the statistical methods provide robust assessment of the prevailing ice conditions, independent of visual and weather conditions. The also comparison shows that the statistical classification methods, designed by process knowledge, provide steadier and more reliable results.