Creating a Multi-Model Artificial Intelligence Framework to Predict the Operational Availability of a Laboratory-Scale Ship Machinery Plant

Abstract

In interests of autonomous and unmanned operation of seagoing vessels by both commercial entities and the United States Government, significant research has been conducted for safe navigation and cybersecurity. This research has contributed to the reduction of required onboard personnel. However, research directed toward reducing required underway personnel ensuring reliable operation of shipboard machinery systems is limited. Machinery reliability has become a primary restriction for unmanned and autonomous operation. Due to complexities driven by plant machinery size and inter-connectivity of systems, traditional methods for improved reliability such as redundancy and component design for high reliability are insufficient to provide necessary reliability for achieving unmanned and autonomous operation of vessel machinery plants over desired duration of deployment. Given the inability to address component faults and failures, a need exists to focus research efforts on operational resilience, or the ability to continue operation in fault prone and present environments.

To improve operational resilience, this work proposes use of Artificial Intelligence (AI) to perform prognostics and diagnostics on plant machinery systems to understand state of health and predict vessel operational availability. With knowledge of system capabilities until failure, fault mitigation techniques may be employed. These techniques include modification to mission operations or more complex applications such as control based fault mitigation to maintain operational capabilities. Heretofore, research for ship machinery system prognostics and diagnostics have been focused at component and subsystems levels to acquire input data from hardware. Applications of prognostics and diagnostics at the system level are prevalent in literature in instances with input data obtained from software simulation models of hardware systems. Due to the lack of hardware based failure data, prognostics and diagnostics of ship machinery plants is largely unexplored. In this work, a laboratory scale ship machinery plant (MLSMP) is designed, constructed and leveraged to obtain lacking run to failure (RTF) data. The MLSMP consisted of a cooling system, fuel system, emulated diesel generator sets, energy storage system, electrical system, mission system, propulsion system, and real time control and data acquisition system. The MLSMP was used to obtain 100 RTF profiles for common faults and failures of machinery systems and illustrate three potential control mitigation strategies for the fault prone environment.

The constructed dataset served as input data to explore potential AI models, including the selected Long Short-term Memory (LSTM) Recurrent Neural Network (RNN) model. These models aimed to detect individual system failures and predict when a system would fail to support operational mission demands, which are utilized to create a multi-model prediction algorithm for the MLSMP. The developed plant-level algorithm is tested and evaluated using the 100 RTF profiles to demonstrate successes and predict accuracy concerning input parameter selection. The LSTM model performed well in the diagnostic and prognostic tasks for the cooling system. The models performed well for the more complex fuel system, although errors increased as system complexity increased.

Efforts under this PhD research provide a significant step towards the operation of unmanned and autonomous operation of ship machinery plants. These efforts include the construction of a laboratory based ship machinery plant, obtaining run to failure data for the laboratory based plant, constructing and evaluating an LSTM driven multi-model framework for prognostics and diagnostics of the MLSMP, and showcasing the potential for unconventional control methods to maintain operational availability in the presence of machinery system faults.