Limited data based deep learning for steel member coating condition inspection method

Abstract

Coating is commonly used to protect steel structures from corrosion, abrasion, and cracking, among other things. The thickness and abnormal conditions of the coating must be monitored to ensure the structural stability of the steel structure. Unfortunately, depending on the type of coating and the external environment, using the taught machine learning model was unfeasible, and securing enough data for learning may be limited depending on the environment. As a result, this research created an algorithm that performs optimization learning for target structures with limited data, estimates a wide range of coating thicknesses, and diagnoses abnormal conditions. First, a wide range of dynamic thermal image data and vision data were acquired using thermal image cameras and vision cameras appropriate for steel structure coatings. By substituting the information from the relevant thermal data and the coating thickness obtained from less than 10 parts, the coating thickness of the entire surface was estimated. Furthermore, newly occurring abnormal conditions were classified using thermal image and visual convergence data without the need for additional learning. The suggested method has been tested on a variety of coating materials, and when fewer than ten data points were used, it demonstrated thickness estimation accuracy of more than 95%. Also, the classification accuracy of 86.4% for newly anomalous circumstances.