Statistical and machine learning for the analysis and prediction of faults and abnormalities in high dimensional sensor data on vessels

The research work conducted within this project relates to developing and applying modern statistical and machine learning analytics on top of vast amount of time series data collected from high dimensional marine propulsion systems. This is done in order to optimize different aspects of marine vessels operation and maintenance. Innovative, distributed and intelligent sensor technology is now available on most vessels. We have an unique access to such sensor data, which are for the first time streamed from various vessels in operation to onshore control centers. The purpose of the thesis is to investigate how these data can be used to generate deep knowledge about processes and dynamics, which lead to abnormal operations or even faults on vessels. The research conducted currently in this projects concentrates on deriving the condition of onboard rotating machinery. This particular class of equipment, or instance propulsion motors, are critical both with respect to the safety of and with respect to the operation of the ship. Being able to detect emerging failures provides the capability to plan and schedule maintenance, and thus effectively reduce downtime and improve the safety of critical operations.

-

This project contributes statistics and machine learning methodology for exploiting new high-dimensional and high-frequency sensor data on board vessels at sea, to improve marine safety, efficiency and availability. Shipping is and will be one of the leading Norwegian industries. This project develops innovative solutions for networks of sensors, programmed to measure many variables which are known to be important in describing the status of various types of machineries and components of a vessel. ABB has unique access to such sensor data, which are for the first time streamed from vessels in operation to ABB?s control centre in Norway. The purpose of the thesis is to investigate how these data can be used to generate deep knowledge about processes and dynamics, which lead to abnormal operations or even faults on vessels. This thesis focuses on two challeges: (i) Data management of huge data time series of different scale, collected over longer periods and on many vessels: how to organise, match and annotate such data in order to be able to investigate patterns and perform analysis? (ii) High dimensional, scalable statistical methods are needed, together with efficient machine learning approaches, in order to detect abnormalities or faults appearing in some components and in some scales, as rapidly as possible. The time series are high dimensional, with hundreds of sensors on each vessel and the signal of abnormality can reside in any of the stochastic multivariate structures of the time series. In addition, multivariate dependence can be at various lags. The series need to be analysed for surprising and abnormal behaviour. The sample frequency of the sensors varies very much. We will study how to handle these variable sample frequencies without resorting on massive missing values, both in storing and in analysing. Real time sensor data lead to huge data sets, which challenge storage and algorithms. All our methods and algorithms must scale computationally.