Characterization of Cast Iron Microstructure Through Fluctuation and Fractal Analyses of Ultrasonic Backscattered Signals Combined with Classification Techniques

This work aims at evaluating the performance of pattern recognition methods in the identification of different microstructures presented by cast iron, namely, lamellar, vermicular and nodular microstructures, through the statistical fluctuation and fractal analyses of backscattered ultrasonic signals. The signals were obtained with a broad band ultrasonic probe with a central frequency of 5 MHz. The statistical fluctuations of the ultrasonic signals were analyzed by means of Hurst (RSA) and detrended-fluctuation analyses (DFA), and the fractal analyses were carried out by applying the minimal cover and box-counting techniques to the signals. The curves obtained from the statistical fluctuations and fractal analyses, as functions of the time window, were processed by using four pattern classification techniques, namely, principal-component analysis (PCA), Karhunen-LoSve transformation (KLT), neural networks and Gaussian classifier. The best results were obtained by Karhunen-LoSve expansion and neural networks, where an approximately 100% success rate has been reached for the classification of the different microstructures as well as for the training and the testing sets of events. The results presented correspond to an average taken over 100 randomly chosen sets of events. These results indicate that, within the techniques used, the Karhunen-LoSve transformation and neural network associated with the statistical fluctuation analyses (RSA and DFA) are the best tools for the recognition of the different cast iron microstructures. It is worthwhile pointing out that the microstructure classification was made by using backscattering signals acquired during pulse echo ultrasonic nondestructive testing only. Therefore, that approach is a promising method for material characterization.