Estimation of shell characteristics using time-frequency patterns and neural network

Surface acoustic waves circumnavigating around an elastic shell carry out very valuable information on its mechanical and geometrical properties and can be used for ultrasonic inspection and non destructive evaluation as well as sonar classification. This information can be decomposed into several elementary components associated to various physical phenomena. Time-frequency representations (and, in particular, the Smoothed Pseudo Wigner-Ville Distribution) have been shown to be very relevant tools for describing these phenomena. Several experiments and simulations have shown, in the time-frequency plane, the importance of a frequency range associated to the so-called “coincidence frequency” (interaction between a Lamb wave and a Stoneley type wave). This pattern has been extracted from the time-frequency image for a great variety of shells in order to describe the direct problem with a reduced set of characteristics parameters. The inverse problem consists in estimating the mechanical and geometrical properties of the shell from these parameters. An innovative neural network approach has been developed for estimating these properties and has been applied to both simulated and experimental data. The method shows a very good accuracy (error less than a few percents on the estimation of the shell characteristics: thickness, density and shear wave velocity)