Identifying Mode Shapes of Turbo-Machinery Blades Using Principal Component Analysis and Support Vector Machines

Abstract

Manually identifying mode shapes generated from finite element solvers images is an expensive task. This paper proposes an automated process to identify mode shapes from gray-scale images of compressor blades within a jet-engine. This work introduces mode shape identification using principal component analysis (PCA), similar to approaches in facial and other recognition tasks in computer vision. This technique calculates the projected values of potentially linearly correlated values onto P-linearly orthogonal axes, where P is the number of principal axes that define a subset space. Classification was done using support vector machines (SVM). Using the PCA and SVM algorithm, approximately 5300 training images representative of 16 different modes were used to create a classifier. The classifier achieved on average 98% accuracy when tested using a test set of approximately 2000 images given P = 70. The results suggest that using digital images to perform mode shape identification can be achieved with high accuracy. Potential generalization of this method could be applied to other engineering design and analysis applications.