Research On Intelligent Fault Detection And Diagnosis Of Gas Turbines Under The Circumstance Of Few Fault Samples

The gas turbine is a power machinery with wide applications.Fault detection and diagnosis of gas turbines has great significance to its safe and reliable operation.With the boom of artificial intelligence and big data technique,data-driven intelligent fault detection and diagnosis of gas turbines is becoming increasingly popular with researchers.However,fault samples are rare or even unavailable in actual applications.For newly-run gas turbines,they are healthy during the early stage of their operation and their operational data are almost all normal samples.When the newly-run gas turbines run for some time,a few fault samples can occur in the historical data of gas turbines.At this time,the number of normal samples are much larger than the number of fault samples and the data distribution is seriously class-imbalanced.Even for gas turbines that have run for a long time,the class-imbalanced problem is still quite serious.The circumstance of few fault samples brings many challenges to data-driven fault detection and diagnosis of gas turbines.Aiming at the problem of intelligent gas turbine fault detection and diagnosis under the circumstance of few fault samples,this paper carries out the following studies.Firstly,this paper analyzed their typical gas path fault characteristics and compared the effects of gas path faults on measurable parameters with the effects of various ambient conditions and operational conditions on measurable parameters through nonlinear component level models of single-shaft gas turbine and three-shaft marine gas turbine.The thought of suppressing the interference of various ambient conditions and operational conditions and emphasizing faults are proposed for gas turbine fault detection and diagnosis under the circumstance of few fault samples.Secondly,aiming at the situation where newly-run gas turbines only have normal samples during their early stage,this paper first solved this problem from purely data-driven prospective,and explored the fault detection performances of four one-class pattern recognition methods including one-class support vector machine,principle component analysis,isolation forest and local outlier factor.Experiments on Taurus 70 single-shaft gas turbine and a three-shaft marine gas turbine shows that the four one-class pattern recognition methods can realize fault detection to some extent.Among the four methods,one-class support vector machine with RBF(Radical Basis Function)kernel function obtained relatively good fault detection performance.Thirdly,aiming at the situation where newly-run gas turbines only have normal samples during their early stage,this paper introduced physical principles of gas turbines and proposed the normal pattern extraction method for fault detection.Through Brayton cycle and basic principles of gas turbines,the intrinsic mapping relationships among measurable parameters in healthy gas turbines was revealed and the inferences of ambient conditions and operational conditions was eliminated.Accurate normal pattern extraction as well as sensitive and robust fault detection were realized through nonlinear autoregressive with exogenous inputs(NARX)network.The proposed method obtains 98.67% detection accuracy for normal data of Taurus 70 gas turbine and 99.96% detection accuracy for fault data of Taurus 70 gas turbine.Comparison experiment with above four one-class pattern recognition methods shows that the proposed normal pattern extraction method significantly improves the fault detection accuracy of single-shaft gas turbines.Fourthly,aiming at the situation where newly-run gas turbines only have normal samples during their early stage,this paper proposed the normal pattern group composed of multiple normal pattern models for three-shaft marine gas turbines.The normal pattern group was identified by long-short term memory(LSTM)network,a neural network with strong temporal data mining ability.Though the collaborative decision of multiple normal pattern methods,sensitive and robust fault detection of three-shaft marine gas turbines was realized.Experiments in three-shaft marine gas turbines show that the normal pattern group obtains 96.97% detection accuracy for normal data of and more than 99.86% detection accuracy for all 13 gas path fault data.Comparison experiment with single normal pattern models and above four one-class pattern recognition methods shows that the proposed normal pattern group method significantly improves the fault detection accuracy of gas turbines.Finally,aiming at the seriously class-imbalanced problem in gas turbine fault diagnosis,the performances of random oversampling,SMOTE(Synthetic Minority Oversampling Technique),Borderline-SMOTE,random undersampling,Near Miss undersampling and weighted support vector machine in class-imbalanced gas turbine fault diagnosis were systematically compared.Experimental results show that oversampling method and sample weighting method can both improve the fault diagnosis accuracy in the case of class-imbalanced data.The accuracy improvement of oversampling method is more significant than that of sample weighting method.In conclusion,this paper realized sensitive and robust fault detection in the case of only normal data and improved fault diagnosis performance in the case of class-imbalanced data for both single-shaft gas turbine and three-shaft marine gas turbine,This paper can provide an effective guideline for fault detection and diagnosis of gas turbines in the circumstance of few fault samples.