Evaluating the performance of artificial neural networks for estimating the nonmetallic coating thicknesses with time-resolved thermography

Abstract

Coatings are commonly applied to products both for protection and to enhance the appearance. In the manufacturing process, it is essential for coatings to be applied uniformly and with appropriate thickness. However, the coating thickness measurement of nonmetallic thin film remains a challenging task. This paper reports the results of research to explore the potential of regression and artificial neural network (ANN) models for estimating the thickness of nonmetallic coatings. The developed ANN models yielded a lower-error rate than a corresponding regression model. Three parameters were used as inputs to the ANN model to train 1880 network models with six different training algorithms. The input parameters were: temperature increments collected from 90 samples at 1 Hz for 59 s; rising temperature rates and slope of scaled log-reconstructed Laplace-transformed temperature along the real axis; and dEx(T) values derived from the temperature increments. The relative interval radius (RR) goes below 1% in ANN models with 99.5% confidence when dEx(T) is used as an input parameter. In addition, the RR value varies directly as the standard deviation of the modeling sample size varies, but the RR values stay at the same order of magnitude. Further testing found that the ANN models will not guarantee an acceptable prediction under untrained conditions.