Frequent pattern extraction based on data and prior knowledge fusion in gas turbine combustion system

Gas turbine combustion system (GTCS) works in the highly adverse environmental conditions of high temperature and high pressure. Because GTCS initial temperature is too high to be directly measured, exhaust gas temperature (EGT) is an alternative to detect performance of GTCS. However, various interferences, such as working and ambient conditions as well as compressor and combustion efficiencies and so on, have a comprehensive effect on EGT, which causes low detection accuracy to GTCS. So, it's necessary to build a model that can eliminate the interferences on EGT so as to monitor performance of GTCS effectively. Feature extraction and model structure are keys to modeling. In general, it's based on data-driven or mechanism analysis approaches to build the model. However, the former containing a low information density leads to too many input variables and the overcomplicated model structure. In addition, although the latter has a high information density, modeling of a gas turbine is a difficult task in practice, which still limits the application of model-based approaches. For deficiencies of the existing methods, this paper proposes a method based on data and prior knowledge fusion to build a model called Frequent Pattern model for detecting performance of GTCS. Firstly, it's based on prior knowledge to tentatively select the model feature. And it's found that Average Exhaust Gas Temperature (AET) contains all the influences of various factors on EGT. Secondly, it's determined by data analysis that there is a linear relationship between AET and EGT. Therefore, we conclude that a simple linear model can eliminate the interferences on EGT and characterize the structural differences of different flame tubes, by which performance of the GTCS can be monitored accurately and effectively. Thirdly, a contrast test is done to verify the conclusion. A linear model is built based on the proposed method and a nonlinear model is built only based on data-driven method. By comparison, it's concluded that the proposed method can not only simplify the model but also improve the model accuracy.