Gas Turbine Gas-Path Fault Diagnosis Based on Decision Fusion of Model-Based and Data-Driven Methods

Abstract Model-based methods (MBMs) and data-driven methods (DDMs) are currently the primary approaches for gas turbine (GT) gas-path fault diagnosis. However, the accuracy of model-based methods can be decreased due to the mismatch between the real engine and model, while data-driven methods struggle to handle untrained fault types. To solve this problem, this study proposes a fusion framework based on the Bayesian network (BN) for decision fusion of model-based and data-driven methods, aiming to enhance the reliability of the gas-path fault diagnosis system. Firstly, a health parameter estimation system is constructed using the component-level model (CLM) of the GT and the unscented Kalman filter (UKF). This system is then combined with a fault identification system based on the fuzzy inference approach to form a model-based diagnostic system. Subsequently, a data-driven diagnostic system is developed using historical operational data and the artificial neural networks. Next, the BN is adopted for decision fusion of diagnostic results. A new BN parameter calculation methodology is proposed based on the characteristics of MBMs and DDMs. The resultant Bayesian network can effectively fuse two different diagnostic methods. Finally, the fusion diagnostic results are compared with those obtained from the individual diagnostic systems. The results indicate that decision fusion method effectively overcomes the limitations of individual diagnostic systems and enhances the scope of health state identification, leading to improved overall performance in GT gas-path fault diagnosis. This research has significant importance for the performance monitoring and maintenance of GTs.