Investigation of Effectiveness of Statistical Methods to Gas Turbine Vibration Diagnostics and Synthesis of SVM Based Approach

• Main Author: VIN, TAN AL
• Format: Final Year Project
• Language: English
• Published: IRC 2017
• Subjects: TJ Mechanical engineering and machinery
• Online Access: http://utpedia.utp.edu.my/17930/1/Dissertation%2818180%29.pdf; http://utpedia.utp.edu.my/17930/

Gas turbine is used for generating electricity since 1939. Gas turbine is a kind of internal combustion engine(IC) which it compress and mix the air and fuel for combustion. Thus, hot gases that produce after the combustion will spin the turbine to generate power. Gas turbine is one of the most widely used technology in power generation. This project aims to explore different kind of statistical method in vibration analysis this understand the theory behind each method which have potential to develop Support Vector Machine (SVM) based turbine vibration diagnostic model. The effectiveness of statistical methods in vibration analysis will be evaluated. Lastly, the statistical analysis results will be used as input to develop SVM based turbine vibration diagnostic model. In this paper, the evaluation of effectiveness on statistical vibration analysis methods will be done by observing the trend of the analysis result plots. Once the completed the evaluation, the results from the high effectiveness methods will be used as input to train the SVM model. The SVM model will be applied to estimate the instantaneous Remaining Useful Life percentage. As a result, two statistical vibration analysis methods: Feature Extraction and Empirical Mode Decomposition, are effective enough to show the failure trend clearly. The SVM based diagnostic model is able to estimate the Remaining Useful Life percentage with above 60% accuracy. Analysis of the final results show that this SVM based diagnostic model is comparable to other publication, which the SVM diagnostic model has the possibility to overestimate the result due to insufficient of available vibration data for model training.