Ultrasonic Flaw Signal Classification And Identification Study Based On Empirical Mode Decomposition And Rough Set Attribute Reduction

Ultrasonic inspection is one of the most successful nondestructive testing(NDT) techniques for quality assessment and defect detection in engineering materials. The echo signals often contain a wealth of information of defective materials such as the flaw’s location, magnitude and category and so on. Therefore the key of ultrasonic inspection is how to process the echo signals in order to obtain more useful information about the material itself. Although ultrasonic nondestructive testing have made great progress, it also requires further research work to make use of signal processing and pattern recognition technology in defects qualitative evaluation, automatic identification and intelligence aspect. Because of the above reasons, this article has done the following work.Firstly, this article introduces the basic principles of empirical mode decomposition, and analyzing advantages and disadvantages of the empirical mode decomposition and wavelet decomposition in non-stationary signals and nonlinear signal processing. Then the ultrasonic echo signals were decomposed by using the empirical mode decomposition, and analyzing the intrinsic mode functions in the time domain and frequency domain. Finally, it summarized the characteristic parameters of the echo signal of the time domain and frequency domain.Secondly, because the initial feature set contains much redundant information based on analysis of time domain and frequency domain, which needs dimension reduction for improving the classification speed and accuracy. Rough set attribute reduction is choosed to achieve it there. Spectral clustering method is used for feature value discretization. After that, rough set attribute reduction can be applied to implement feature selection so that it will help to achieve defect identification signal.Thirdly, the measured data of ultrasonic echo signals were constructed training and validation samples in the frequency domain characteristic parameters. Then the BP neural network is trained on MATLAB,and then an experiment is made to verify the validity of neural network for diagnosing the defect type of ultrasonic echo signal. The experimental results show that BP neural network has a better classification effect.Finally, Compared with the other feature extraction method-empirical mode decomposition, wavelet decomposition and principal component analysis, the method based on rough set attribute reduction and empirical mode decomposition has better classification results compared to other common methods.