AN ARTIFICIAL NEURAL NETWORK AS A TOOL FOR THE INVERSION OF ULTRASONIC DISPERSION DATA FOR MATERIAL CHARACTERIZATION

摘要

The use of ANN and Fuzzy Logic algorithms in the identification of material parameters from ultrasonic dispersion data appears to be a promising area worthy of further research. In this study the feasibility of inversion of ultrasonic data for determining stiffness of composites has been demonstrated. The above ANN was trained using a very broad scope of material data set. The assumption that so little is known about the material is unrealistic. In a realistic situation, the range for possible elastic constants for a given material would be much less than those defined in this experiment, so the search space could be limited to encompass only those possible values, significantly increasing the precision of the network outputs. These smaller ranges would also require identification of more subtle changes in the curves by the ANN, which would probably call for an increase in the resolution of the input vector. This work has shown the feasibility of applying artificial neural networks to the problem of inverse modeling of ultrasonic dispersion data for unidirectional composites laminate plates. More work is needed to increase the network precision, possibly based on an apriori knowledge of the experimental materials or an increase in the size of the data training sets. Also, increase in the resolution and confining the analysis to specific regions of the data sets with most sensitivity will optimize the inversion and improve the performance. Such analysis has been performed earlier and the regions of sensitivity are well known. Experimental verification of this inversion has not been conducted in this paper and future work must include experiments.