| dc.description.abstract | Machine learning has gone way beyond a ground-breaking technology a few decades ago to now taken for granted in many day-to-day activities. It is now providing new ways for manufacturing, assembling, operating, monitoring, and maintaining products. Typical application areas include performance optimization, quality improvements, fault detection and predictive maintenance. In this paper application of machine learning algorithms to forced response prediction during the design and analysis of a typical gas turbine compressor blade is reported. The forced response prediction process typically involves utilizing harmonic or time domain computational fluid dynamics (CFD) analyses to compute the forcing and the aero damping, to calculate reserve factors that represent the high cycle fatigue life of the blade. This time-consuming process is generally limited to the later phases of the design cycle and can lead to hundreds of calculations if one must address all the resonances in a typical twin shaft running range. A neural network trained using historical data is used to directly predict the reserve factor with high confidence without the need for costlier high fidelity CFD by using just the finite elements (FE) predicted parameters. This allows to perform high-fidelity aero-mechanical assessment at an early stage in the design process. Further, application of image recognition using a convoluted neural network to aid in the identification of FE predicted Modeshapes is also demonstrated, which can also improve the quality of the reserve factor predictions. | |
