Predictive Maintenance of Critical Equipment for Liquefied Natural Gas Liquefaction Process Meeting global energy demand is a massive challenge, especially with the quest of more affinity towards sustainable and cleaner energy. Natural gas is viewed as a bridge fuel to a renewable energy. LNG as a processed form of natural gas is the fastest growing and cleanest form of fossil fuel. Recently, the unprecedented increased in LNG demand, pushes its exploration and processing into offshore as Floating LNG (FLNG). The offshore topsides gas processes and liquefaction has been identified as one of the great challenges of FLNG. Maintaining topside liquefaction process asset such as gas turbine is critical to profitability and reliability, availability of the process facilities. With the setbacks of widely used reactive and preventive time-based maintenances approaches, to meet the optimal reliability and availability requirements of oil and gas operators, this thesis presents a framework driven by AI-based learning approaches for predictive maintenance. The framework is aimed at leveraging the value of condition-based maintenance to minimises the failures and downtimes of critical FLNG equipment (Aeroderivative gas turbine). In this study, gas turbine thermodynamics were introduced, as well as some factors affecting gas turbine modelling. Some important considerations whilst modelling gas turbine system such as modelling objectives, modelling methods, as well as approaches in modelling gas turbines were investigated. These give basis and mathematical background to develop a gas turbine simulated model. The behaviour of simple cycle HDGT was simulated using thermodynamic laws and operational data based on Rowen model. Simulink model is created using experimental data based on Rowen’s model, which is aimed at exploring transient behaviour of an industrial gas turbine. The results show the capability of Simulink model in capture nonlinear dynamics of the gas turbine system, although constraint to be applied for further condition monitoring studies, due to lack of some suitable relevant correlated features required by the model. AI-based models were found to perform well in predicting gas turbines failures. These capabilities were investigated by this thesis and validated using an experimental data obtained from gas turbine engine facility. The dynamic behaviours gas turbines changes when exposed to different varieties of fuel. A diagnostics-based AI models were developed to diagnose different gas turbine engine’s failures associated with exposure to various types of fuels. The capabilities of Principal Component Analysis (PCA) technique have been harnessed to reduce the dimensionality of the dataset and extract good features for the diagnostics model development. Signal processing-based (time-domain, frequency domain, time-frequency domain) techniques have also been used as feature extraction tools, and significantly added more correlations to the dataset and influences the prediction results obtained. Signal processing played a vital role in extracting good features for the diagnostic models when compared PCA. The overall results obtained from both PCA, and signal processing-based models demonstrated the capabilities of neural network-based models in predicting gas turbine’s failures. Further, deep learning-based LSTM model have been developed, which extract features from the time series dataset directly, and hence does not require any feature extraction tool. The LSTM model achieved the highest performance and prediction accuracy, compared to both PCA-based and signal processing-based the models. In summary, it is concluded from this thesis that despite some challenges related to gas turbines Simulink Model for not being integrated fully for gas turbine condition monitoring studies, yet data-driven models have proven strong potentials and excellent performances on gas turbine’s CBM diagnostics. The models developed in this thesis can be used for design and manufacturing purposes on gas turbines applied to FLNG, especially on condition monitoring and fault detection of gas turbines. The result obtained would provide valuable understanding and helpful guidance for researchers and practitioners to implement robust predictive maintenance models that will enhance the reliability and availability of FLNG critical equipment.

Document Type


Publication Date