A Fault Diagnosis Method for Industrial Gas Turbines Using Bayesian Data Analysis

This paper presents an offline fault diagnosis method for industrial gas turbines in a steady-state. Fault diagnosis plays an important role in the efforts for gas turbine owners to shift from preventive maintenance to predictive maintenance, and consequently to reduce the maintenance cost. Ever since its birth, numerous techniques have been researched in this field, yet none of them is completely better than the others and perfectly solves the problem. Fault diagnosis is a challenging problem because there are numerous fault situations that can possibly happen to a gas turbine, and multiple faults may occur in multiple components of the gas turbine simultaneously. An algorithm tailored to one fault situation may not perform well in other fault situations. A general algorithm that performs well in overall fault situations tends to compromise its accuracy in the individual fault situation. In addition to the issue of generality versus accuracy, another challenging aspect of fault diagnosis is that, data used in diagnosis contain errors. The data is comprised of measurements obtained from gas turbines. Measurements contain random errors and often systematic errors like sensor biases as well. In this paper, to maintain the generality and the accuracy together, multiple Bayesian models tailored to various fault situations are implemented in one hierarchical model. The fault situations include single faults occurring in a component, and multiple faults occurring in more than one component. In addition to faults occurring in the components of a gas turbine, sensor biases are explicitly included in the multiple models so that the magnitude of a bias, if any, can be estimated as well. Results from these multiple Bayesian models are averaged according to how much each model is supported by data. Gibbs sampling is used for the calculation of the Bayesian models. The presented method is applied to fault diagnosis of a gas turbine that is equipped with a faulty compressor and a biased fuel flow sensor. The presented method successfully diagnoses the magnitudes of the compressor fault and the fuel flow sensor bias with limited amount of data. It is also shown that averaging multiple models gives rise to more accurate and less uncertain results than using a single general model. By averaging multiple models, based on various fault situations, fault diagnosis can be general yet accurate.