Paint stripping is a frequently required process to renew paint coatings on an aircraft. Typically the paint stripping processes used today remove the whole coating or require time-consuming experiments for layer specific removal of the paint. Otherwise it could lead to a defect of deeper coat layers or the substrate. Typically for each paint system process parameters have to be determined for each surface. Here we developed a process for laser-based paint removal that is based on using a neural network with hyperspectral imaging.

Other neural networks that can be found for classification of hyperspectral data usually use fully connected layers. The neural network used does not use a fully connected layer so that it can fully exploit the advantages of parallelization. In addition, spatial correlations can be used.

For the training and validation sample sheets were produced on which a two- or three-layer paint structure with primer and topcoat or primer, basecoat and clearcoat was applied. The paint was removed strip by strip in four cycles. The aim of the classification was to specifically determine the top coat or, in case a clearcoat was applied, to indicate the basecoat of the paint structures. It was also investigated whether the number of laser cycles can be determined in order to establish a guide value for the residual coating thickness.

The architecture of the neural network was able to increase the speed of classification by a factor of 14 against a comparable architecture with fully connected layers while maintaining the same accuracy. The paints can be reliably identified using the method presented. The basecoat and topcoat are particularly relevant for determining the laser parameters. These can be determined with a certainty of 99%.