Classification of materials using artificial intelligence techniques based on modal properties generated by vibration

Abstract

The motivation of this project is to integrate the technology of signal processing and materials characterization into developing a system of non-destructive test of material identification. The purpose of this research is to obtain the modal parameters of stainless steel SS 304, aluminium 1100 and glass, validating the parameters by finding and comparing the elastic constants of the materials to theoretical and conventional testing values, and classifying the parameters using linear discriminant analysis (LDA), k-nearest neighbor (k-NN), and artificial neural network (ANN) according to their respective material types. The modal parameters were obtained by modal analysis method, a vibration technique that employs impulse excitation of the materials by using impact hammer and analysis of the frequency response function (FRF) resulted from the excitation through peak-picking on the stabilization diagram. The values for modal parameters were validated by LMS Modal Synthesis, a program in the modal analysis software, normality test by MiniTab 17, a statistical software, and by comparing elastic constants of materials between the ones obtained from exploiting the modal parameters, conventional testing and theoretical values. LMS Modal Synthesis compares the percentages of the correlation and error of two FRF signals; one being the signal from the exact experimental values and another from the synthesize signal generated by the software itself. Normality test analyses on how closely the modal parameters will follow the normal distribution based upon the Anderson-Darling test. Elastic constant determination shows how credible and precise the values of modal parameters based upon the correlation of the elastic constants from the exploitation of modal parameters with the experimental and theoretical values. The validated modal parameters are the used as the features for classification by three different classifiers. LDA gave the best performance for this research. The architectures are then used for classification of modal parameters with the addition of noise to further test the reliability of the classification system. All the results and analysis are presented and discussed thoroughly in the thesis.