Vibration Monitoring of Gas Turbine Engines: Machine-Learning Approaches and Their Challenges

摘要: In this paper, condition monitoring strategies are examined for gas turbine engines using vibration data. The focus is on data-driven approaches, reason a novelty detection framework considered the development of reliable models that can describe underlying relationships processes taking place during an engine's operation. From data analysis perspective, high dimensionality features extracted and complexity two problems need to be dealt with throughout analyses type. latter refers fact healthy engine state nonstationary. To address this, implementation wavelet transform get set from signals nonstationary parts. problem high-dimensionality addressed by "compressing" them kernel principal component so more meaningful, lower-dimensional used train pattern recognition algorithms. For feature discrimination, scheme based one-class support vector machine algorithm chosen investigation. main advantage, when compared other algorithms, learning being cast as quadratic program. developed strategy applied detecting excessive levels lead failure. Here, we demonstrate its performance experimental operating different conditions. Engine designated belonging engine’s “normal” correspond fuels air-to-fuel ratio combinations, in which experienced low vibration. Results such schemes achieve satisfactory validation accuracy through appropriate selection parameters machine, width γ optimization penalty parameter ν. This was made searching along fixed grid space values choosing combination provided highest cross-validation accuracy. Nevertheless, there exist challenges discussed suggestions future work enhance similar schemes.