Weighted LDA Based Dimensionality Reduction for Ultrasonic Flaw Signal Classification

Abstract

Ultrasonic inspection is the most successful non-destructive testing technique for detection of flaws in engineering materials. Generally, discrete wavelet transform (DWT) is widely used to extract features for ultrasonic flaw signal. Directly taking the DWT coefficients as input vectors for training classifier, however, may result in computation complexity or even poor classification performanc. We propose a weighted linear discriminant analysis (WLDA) method to address the problem. In this study, DWT is first applied for the time-frequency analysis of ultrasonic flaw signals, and the wavelet coefficients are extracted as the initial features. After that, the proposed WLDA, which can effectively estimate the within-class and the between-class scatter through calculating similarity based weighting function, is used to reduce the dimension of original features. Finally, the features in new lower dimensional space are taken for flaw classification. Experimental results show that compared with the original and state of art linear discriminant analysis methods, WLDA is helpful for improving classification performance.