Intelligent Gas Turbine Engine Condition Monitoring Method Based On Performance And Vibration Data Reconstruction

Gas turbine engines are important industrial power equipment that play a significant role in aviation,petrochemical,marine,military and other fields.As gas turbines operate in extreme high-temperature and high-pressure environments for long periods of time,failures can occur.To avoid unnecessary losses due to these failures,researchers have designed health management systems for gas turbines to help them operate stably for long periods,reduce unnecessary maintenance,save costs and prevent failures.Although traditional methods based on mechanism analysis and physical modeling can solve most gas turbine monitoring problems,there are still no practical solutions in some complex industrial scenarios.For example,when data parameters are incomplete,the physical mechanism model cannot be used;the gas turbine’s variable operating conditions may cause drastic changes in parameters that make it difficult to effectively distinguish between working conditions and failures;industrial on-site data has large and variable noise,and traditional threshold values can cause high rates of missed alarms and false alarms,among other issues.Based on these problems,this paper proposes a robust self-supervised deep learning model based on parameter reconstruction that uses gas path performance and wideband vibration data,which are the most common and can reflect most faults in gas turbines.It also combines the reconstruction error-based anomaly monitoring indicators with trend analysis to build an intelligent monitoring method for industrial scenarios.The main research contents are as follows:(1)We reviewed existing monitoring and diagnostic methods based on gas turbine systems and proposed a system analysis framework for gas turbine monitoring and diagnosis.Using this framework,we designed a reconstruction model that combines simulation methods and unsupervised anomaly detection methods in traditional methods.(2)We analyzed the beta-VAE model based on variational autoencoders in the existing unsupervised learning field,understood its principle,and designed a progressive training method that works with it.We also learned data fusion methods in the field of deep learning.Combining these two technologies,we designed a complete data reconstruction model that can extract disentangled representations in data unsupervised or semi-supervised and has strong robustness.It can accurately reconstruct gas path performance or vibration spectrum data with high precision.(3)We analyzed the problems of existing fixed threshold methods and designed dynamic thresholds based on maximum likelihood estimation that vary with operating conditions.Based on dynamic thresholds and reconstruction residuals,we designed the PAM anomaly indicator for gas path performance.Based on the characteristics of the spectrum,we designed the VAM anomaly indicator for vibration spectrum.We analyzed the role of trends in monitoring and designed a variable scale trend extraction method.Finally,we combined anomaly indicators and trend analysis to design a monitoring method for gas path performance and vibration spectrum of gas turbines.(4)We analyzed the data acquisition in practical industrial scenarios and designed an adaptive normalization method based on the characteristics of industrial data.Finally,using the proposed preprocessing method,the above reconstruction model and monitoring system,we completed gas turbine fault monitoring and preliminary diagnosis based on gas path performance and wideband vibration data on a real dataset.