Location of material flaws using wavelet analysis and neural network

Ultrasonic techniques combined with digital signal processing schemes have been successfully used to detect flaws in material for a number of years. Recently, wavelet analysis and pattern recognition have shown some advantages over the traditional methods, such as Fourier analysis to perform a time-frequency analysis of ultrasonic signals coming from material under testing. The aim of this work is to detect and locate material defects using ultrasonic signals generated by a 3.5 MHz transducer and pattern recognition. The detected echoes are pre-processed applying the Hilbert transform to produce the signal envelopes and then normalise them to properly feed the network. Signals are wavelet-transformed (Daubechies coefficients) to obtain the time-frequency data to train the network (ART2: Adaptive Resonance Theory class two). Consequently, a number of signals are chosen to train the network so patterns can be formed. Thus, the trained network is able to generate new patterns independently of the wavelet analysis. Aluminium material with artificially produced defects was put under test. A Krautkramer ultrasonic transducer immersed in water was used to generate and detect echoes coming from the testing material. The transducer was moved along a straight line at equally space distances and signals obtained for each position. Once the signals are acquired, a. few of them were chosen to train the network and the rest of them were used to experimentally validate our system. This approach can generate a pattern matrix containing the defects and their localisations. Visualisation of this matrix allows us to identify and localise the faults. Also, it is able to discriminate between flaws and echoes coming from corners or second-time around echoes.