This paper presents an offline fault diagnostics method for highly degraded industrial gas turbines. The method recasts gas path analysis to an inference problem using Bayesian networks where the health condition of each component is quantified in comparison to an expected value. The health parameters are inferred from available gas path measurements, which are sometimes erroneous due to sensor faults or miscalibration. The sensor errors should be inferred as well as the health parameters. Thus, typically in gas path analysis the unknowns are more than the knowns. To address this issue, the present method uses multiple Bayesian network models each of which contains a subset of the unknowns. Their results are averaged according to how much each of the models is supported by the data. Although this method has been reported successful for the faults affecting a few unknowns, its results are still less accurate and confident when it is applied to highly degraded gas turbines. Such gas turbines are likely to have health parameters deviated from the new and clean condition as well as have component faults and sensor errors. Because of this, the present method must infer too many unknowns at the same time to result in a solution with high confidence. In addition, this method cannot differentiate normal or expected degradation from an actual fault. These issues are resolved by fusing extra information to the method. First of all, a sensor calibration report, if available, eliminates the sensor errors from the unknowns. Consequently, the number of possible subsets decreases, and so does the number of Bayesian models. Second, a degradation model provides meaningful prior guesses for the health parameters. It is equivalent to change the point of reference from a brand new gas turbine to a normally degraded one. It will be demonstrated that the method accompanying with the degradation model and the sensor calibration report shows significant improvement in accuracy and confidence.
展开▼
