Shipping is a major driving force of the global economy, as seen by the 90% of the volume of the yearly trade transported by ships. As a result, shipping has a significant financial, environmental impact. Maritime maintenance can be used to safeguard shipping's impact by improving safety and avoiding accidents. This is especially true, when considering that nearly 22% of all the accidents between 2011 and 2017 were attributed to improper maintenance. Consequently, maritime maintenance can be used as a hazard mitigation tool, improve ship safety by reducing accidents. Modern maritime maintenance is best applied through predictive maintenance schemes, which take advantage of the developments of Shipping 4.0. Under this scope, the goal of this thesis is the development of a compound novel data-driven and reliability-based predictive maintenance framework for ship machinery system. The novel framework tackles the areas of maritime predictive maintenance holistically by addressing the topics of critical equipment selection, data preparation, fault detection and diagnostics. Each of the framework's topics are developed in individual methodologies and assessed in unique case studies demonstrating their effectiveness in the respective tasks. Initially, the methodology for the critical equipment selection includes the novel combination of Fault Tree Analysis with data clustering for the identification of critical equipment, as applied in the case of an LNG Carrier. As a result, the most critical components are identified by taking into account reliability indices and repair costs for the considered components. Identifying critical components improves safety, as it focuses the maintenance efforts in items whose failures can have economic consequences and safety implications. Next, the methodology for the data preparation is developed, which includes the novel integration of the kNN and MICE algorithms for the imputation of missing data. Combining these two algorithms al lows for the novel integration a data-driven approach with domain knowledge in a single imputation model. The imputation methodology is applied in the case of a Chemical Tanker, showcasing the effectiveness of the novel method against a pure MICE and pure kNN approach. The treatment of missing values can improve ship safety, as it safeguards information contained within datasets and leads to more accurate condition assessing models. Following that, a novel Fault Detection methodology is established based on Expected Behaviour models, using Machine Learning, and Exponentially Weighted Moving Average control charts. This methodology aims at detecting developing faults in their early stages while avoiding the shortcomings of black-box approaches and having reasonable data requirements for training. Lastly, the diagnostics methodology is formed, which includes the novel integration of pre-processing and Machine Learning-based Fault detection with a diagnostic network using Bayesian Networks. The resulting methodology can identify the root cause of a detected fault, without using black-box Neural Network approaches, nor complicated and time-consuming physics-based models. Even though the Fault Detection and diagnostics methodologies are developed individually, they are both evaluated in the same case of a Bulk Carrier. The use of the same case study was dictated by restrictions in collecting additional data and by the use of the output of the Fault Detection methodology in the diagnostics. The detection of developing faults and the identification of their root-cause has a profound effect on ship safety, while also allowing for targeted maintenance actions.
|Date of Award||9 Oct 2020|
- University Of Strathclyde
|Sponsors||University of Strathclyde|
|Supervisor||Iraklis Lazakis (Supervisor) & Gerasimos Theotokatos (Supervisor)|