Remote Sensing Using Hyperspectral Imaging for Paint Thickness Evaluation and Degradation Assessment

Abstract

Protective coatings are essential in many industries because they protect structures by isolating the underlying materials, or substrate, from harmful environmental factors. When coatings begin to fail or degrade, the exposed substrate deteriorates. Substrate deterioration and corrosion demand maintenance with high costs and may pose potential safety risks. Typically, paint condition assessment techniques rely heavily on visual inspection, which is limited to detecting mechanical or advanced stages of degradation. Other techniques like eddy current gauges are used to measure paint thicknesses, as thickness reduction is a sign of paint degradation, but this contact method is not practical for assessing large surface areas. Many industries tend to carry out scheduled repainting based on operating hours without considering the actual paint condition. Spectroscopy methods, like Fourier-transform infrared spectroscopy (FTIR), are an effective and reliable method for paint assessment, but impractical in the field setup due to scalability issues. On the other hand, hyperspectral imaging (HSI) which is classified under spectroscopy techniques has emerged as a new method with a high potential for paint condition assessment. In our study, we have developed a comprehensive framework that combines HSI and machine learning (ML) to predict both paint thickness and the condition of the paint under accelerated ageing tests. Aluminium plates painted with aliphatic polyurethane-based paint have been prepared with different paint thicknesses and various ageing states. The HYSPEX SWIR 384 camera for data collection within a spectral range of 930-2505 nm and a QUV chamber was used to age the samples according to ISO 11507 Method A standard with a maximum period of 1600 hours. Various machine learning (ML) models were explored to predict paint thickness. Traditional methods like Nu-SVR and SGD performed well, but when their performance compared to DNN’s, with four hidden layers, DNN stood out. It reached an RMSE of 21.6 µm and R2 of 0.97, predicting thicknesses from 43 to 499 µm. For the degradation assessment, the 1D CNN model delivered the highest performance with an R2 of 0.94 and RMSE of 125 hours in predicting ageing hours from HSI data for 500- and 1000-hour aged paint samples. Extending the model to 250 and 800 aging hours, R2 achieved 0.90 and an RMSE of 157 hours. The obtained results demonstrated that our model captures well the relationship between paint thickness/degradation and the reflectance values. These results demonstrate the effectiveness of HSI for paint condition assessment. This solution can lead to more reliable inspections and better maintenance strategies. Future work will include field experiments on a large-scale structure.