Signals find applications in the Non-Destructive Evaluation (NDE) of metals and other materials like composites as well as in the field of Medical Imaging. Back scattered echoes generated by the reflection of ultrasonic waves contain information pertaining to the location, size and characteristics of the defects present, along with the material and the path information about the object under evaluation. The accurate detection, location and sizing of the defects and other imaging applications are limited by the ability to precisely estimate the information contained in the ultrasonic echoes. Model based signal parameter estimation is a technique, wherein an underlying model for the signal is assumed and parameters of the model are estimated

Signal parameter estimation of complex Ultrasonic RF Signals in the presence of noise as well as multiple overlapping echoes using a hybrid search algorithm will be discussed. This hybrid algorithm comprises of an evolutionary algorithm (Genetic Algorithm) and a gradient based algorithm (Levenberg-Marquardt Algorithm). This technique was found to overcome problems of earlier methods that are marked by the presence of local minima. In this paper, the hybrid algorithm is shown to demonstrate higher robustness when compared to a purely gradient based approach and faster convergence when compared to a purely evolutionary approach, for both simulated and experimental signals.

The hybrid algorithm, when applied for signal parameter estimation problems in the ultrasonic NDE context combines the advantages of a gradient based approach and an evolutionary approach, providing higher robustness than the gradient based methods and higher speed when compared to the genetic algorithm. Robustness is needed especially in cases where an inspection of the estimated parameters does not reveal their optimality like in the case of overlapping echoes. In this paper the improvement in the range of the initial guess allowed was shown to be more than 500% when compared to the Gauss-Newton method. Similarly, the reduction in the computation time was observed to be about 90% when compared to the Genetic algorithm used alone.

    Radiographic Testing using X-rays is  commonly used NDT method for detecting internal welding flaws. It is based on the ability of X-rays  to pass through metal and other materials opaque to ordinary light, and produce photographic records by the transmitted radiant energy. Because different materials absorb either X-ray  to different extent, penetrated rays show variations in intensity on the receiving films. That provides a means to examine the internal structure of a weld. Traditionally, experienced workers are required to evaluate the weld quality based on radiographic images. Therefore, the results very much depend upon the capability and experience of the operator. Unfortunately, the manual interpretation process is time-consuming and the results could be very subjective, inconsistent, and sometimes biased. Therefore, it is desirable to develop a computer based automated defect recognition (ADR) system to increase the objectivity, consistency, accuracy, and efficiency of RT inspection.

Digital image processing techniques are employed to lessen the noise effects and to improve the contrast, so that the principal objects in the image can be more apparent than the background. Feature extraction is necessary to obtain a set of features that can describe the characteristics of welding defects. These features should be small in number and high in discriminatory power. Pattern classification methods are needed to analyze feature data and make a prediction of the defect type. Pattern classification algorithms might differ in efficiency and accuracy. Therefore, two renowned supervised algorithms: Artificial Neural Network (ANN) and Support Vector Machine (SVM) were investigated.

The approach is listed as follows:

    The multiple mode and dispersive nature of ultrasonic Lamb waves complicate their signal interpretation. Others have shown that it is possible to obtain a plate’s dispersion relationship by using the two dimensional Fourier Transform to operate on multiple, equally spaced waveforms. Unfortunately, the need for exact, spatially sampled data restricts the practicality of the 2D-FFT for inspection applications. In contrast, Time-Frequency Representations (TFR) requires only a single RF signal for analysis.

This work explores the use of Wavelet Transform combined with theoretical dispersion relationship to automatically identify modes (without source position known a-priori) by exploiting their dispersive nature, a widely considered unfavorable property,  of the ultrasonic Lamb waves. Four different methods to identify modes are discussed and an efficient algorithm is presented. The algorithm also identifies the possible source location of the modes.

It was found that the techniques using adaptive template generation prove effective for the interpretation of complex Lamb wave signals. The technique requires the material properties and geometry of the sample as inputs for analytically deriving the templates for individual modes using the dispersion curves. This technique was found to work well for cases where the multiple modes were separated in the dispersion curve. However, difficulties were encountered in regions where modes intersected.