Towards Neural Network Classification of Terahertz Measurements for Determining the Number of Coating Layers

To determine layer thicknesses with terahertz time-domain spectroscopy, the number of layers must usually be known. However, in some applications the number of layers varies along the surface, so that the number of layers at a specific measuring location can be unknown. Our approach is to use an artificial deep neural network for estimating the number of layers at a preliminary stage for common terahertz algorithms. This work describes the selection and evaluation of a feedforward neural network. This neural network allows a good estimation of the number of layers confirming the usefulness of the proposed approach.

I. Introduction

LAYER thickness measurements of multi-layer systems using terahertz time-domain spectroscopy have been frequently investigated in the past. Usually, the number of layers (NOL) on an examined surface must be known when using terahertz layer thickness algorithms. In some applications, however, the number of coating layers varies along the surface, e.g. on certain wind turbine rotor blades. Therefore, the NOL at a measuring location can be unknown, making measurements of layer thicknesses more demanding. To determine NOL with terahertz time-domain spectroscopy (THz-TDS), simple pulse counting is usually unsatisfactory due to multiple reflections within the layers. Even when using alternative sensor systems such as computer vision, the NOL can remain unknown: Additional top layers might be detected visually, but different puttying under the surface is harder to detect.