GEOSTATISTICAL ANALYSIS AND DEEP LEARNING BASED PREDICTION FOR CORROSION SURFACES OF STEEL PLATES

The assessment of corrosion damage is an essential part of the maintenance of steel structures. The numerical analysis of the properties of corrosion surfaces and the accurate prediction of corrosion surfaces are of great significance. In this study, four kinds of unpainted steel plates, SM400A, SM490A, SMA400AW, and SMA490AW, were used for corrosion experiments under the artificial seawater corrosive environment ISO 16539 Method B, and two atmospheric exposure environments in different regions. The corrosion depths of the steel plates were measured by a laser focus measurement system. Semi-variogram was used in the geostatistical analysis to investigate the spatial autocorrelation structure of the corrosion surfaces. By using this method and the ordinary kriging technique, a method was proposed to simulate the spatial characteristics of the corrosion surfaces. The simulation results indicated that the corrosion depth and surface morphology of the corrosion surface were in high agreement with the experimental results. In addition, a deep learning model based on generative adversarial network (GAN) was used to build a prediction model of the corrosion surface. The spatial properties of the prediction model were verified using the geostatistical analysis method proposed in this study, and the results showed that the predictions had the same spatial properties as the actual corrosion surface.