Research On Feature Extration For Ultrasonic Echo Signals Based On Time-Frequency Analysis Methods

The thesis, supported by the National Natural Science Funds ’Research on quantitative recognition for bonding flaw of composite plates based on ultrasonic testing’, has proposed multiple adaptive methods in time-frequency domain to analyze ultrasonic echo signals, resulted accurate experiment results indicate some new direction for automatically identification of composite plates.Based on a fact the data collected with the echo acquisition circuit are typical nonlinear and non-stationary signals, the time and frequency domain analysis methods processing non-stationary signals are widely used in the paper.With the basis of feature extraction using the traditional EMD method, an improved EMD approach based on mirror extension is developed to reduce the end effect and optimize the decomposition process. Simulation results verified that the EMD method based on mirror extension decreases the end effect obviously.The mode mixing produced by traditional EMD method will lead to inaccurate intrinsic mode functions(IMF) which could not represent real vibration modes of original signals. In the paper, another improved EMD method, namely EEMD, is proposed to eliminate mode mixing. EEMD decomposes the signal mixed by original signal’s and Gaussian White Noise essentially. The Hilbert spectrum of IMFs obtained by the decomposition of original signals illustrates the EEMD method cripples the mode mixing effectively. In addition, with the cognition on the sensitivity between IMFs and original signals, a sensitivity assessing algorithm is used to choose appropriate IMF components which could represent the real vibration modes of original signals.The LMD method introduced in chapter5is an adaptive decomposition method, which is similar to the EMD method, while the LMD method outperforms the EMD approach in terms of reduction of the energy leakage, iterations and lack of the signal scale. Therefore, the LMD method is also treated as one of the main approaches for extracting features of echo signals.This paper has introduced various time-frequency approaches including EMD(including the improved algorithm), EEMD and LMD for processing echos signals to obtain feature vectors representing de-bonding levels, which are classified with the support vector machine approach to realize quantitative identification of bonding flaws. Resulting identification consequence with high-precision verified that aforementioned methods reveal the time-frequency regularity of non-stationary signals effectively and recognize bonding flaws of composite plates properly.