Amrita Das

Date of Award

January 2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Civil Engineering

First Advisor

Sattar Dorafshan

Abstract

Metallic corrosion is an electrochemical process that can occur due to the formation of aqueous adlayers on the metal surface. It is often described as reverse metallurgy, as corrosion drives refined metals back to their lowest energy state, namely an oxidized, ore-like form. Environmental factors that promote corrosion include precipitation, high humidity, and chemical condensation, which can result from the hygroscopic nature of pollutants deposited on the surface. There is a significant impact of metallic corrosion on the U. S. economy including infrastructure, transportation, utilities, production, and manufacturing. The total direct cost of corrosion in the United States has been estimated at $279 billion annually, representing approximately 3.2% of the nation's Gross Domestic Product (GDP). However, timely monitoring can reduce these costs by 15–35%. Applying protective coatings is typically the first line of defense against corrosion, as these coatings act as inhibitors by forming a barrier between the steel surface and the surrounding environment. Current non-destructive inspection methods require the involvement of human inspectors. One of the significant limitations of this approach is the potential inaccessibility of defect-prone areas, which may result in undetected corrosion. Developing an autonomous methodology for corrosion detection became essential to mitigate this issue. This study focused on developing different non-contact and autonomous corrosion identification methods. In this study, a comprehensive cycle of developing a noncontact method has been demonstrated and divided into several steps. At first, a conventional image processing method was used where corrosion features were leveraged in YCbCr color space including other preprocessing steps such as contrast adjustment, histogram equalization, and adaptive histogram equalization. The model segmented 70% of corroded pixels correctly after improving the brightness to an optimum level. However, this conventional image processing method needs user input; hence, it has limited applications, e.g. one set of data from one type of structure. To elevate this method towards an autonomous corrosion detection methodology a deep learning (U-shaped encoder-decoder) network has been trained and tested to detect the corroded pixel from the steel structure’s images. Four types of models such as original UNet and changing the backbone with DenseNet121, EfficientNetB7, and ResNet34 were tested to classify the corroded pixels. Out of all these U-shaped models, the network with ResNet34 as the backbone outperformed the other models by predicting 93.4% corroded pixels precisely with a 90.77% intersection over union (IOU) value. In addition to this, an image classification model, AlexNet was trained and tested in real-time with the help of a customized payload integrated with an Uncrewed Aerial System (UAS). A human-machine interface was introduced to take the inspector’s input in sequential training of this model allowing the model to learn from human expertise. The model was then retrained on the inspector’s output and used in the next inspection. Before retraining 84.78% of images with corrosion were correctly predicted by the model. The result showed that the adapted deep learning model performance improved successfully with more inspection than expected. In particular, the number of reported false calls made by the model has reduced. The accuracy of image classification and semantic segmentation models depends on the expertise in labeling datasets. Moreover, detecting corrosion can be considered the first step in corrosion monitoring. The type of corrosion and corrosion without visual manifestation cannot be recognized with visual sensors. Therefore, the feasibility of hyperspectral imagery (HSI) in detecting early corrosion was investigated. For the bare steel, two types of corrosion tests were performed. The first one is progressive early corrosion without visual manifestation of corrosion, i.e., 2 hours, 4 hours, 6 hours, and 8 hours. The second one is intermittent corrosion, steel specimens have been corroded for different exposure periods such as 6 hours, 16 hours, and 24 hours. For the invisible or early phase of corrosion, the maximum change for the progressive corrosion of 8 hours found in the Near Infrared (NIR) range is 36.82% but for the Visual Near Infrared (VNIR) range, it is 89.09%. Similarly, for the visible corrosion at 24 hours, a 95.4 % change in reflectance was reported in the VNIR range, which is 54.54 % for NIR. After collecting the hyperspectral signature, these samples were coated with primer and topcoat. The performance of the hyperspectral sensor was validated with a Benford model, an empirical model to derive the reflectance value of multilayer coated material. A maximum deviation of 14% from the Benford model calculation was observed as the thickness increased to the range of 220–240 µm. The feasibility of the hyperspectral sensor to identify the corrosion underneath the coating for different exposure times was also investigated. The reflectance value decreased as the coating thickness increased for all samples. The spectral signatures for sound and corroded samples with the same coating thickness were not the same, which showed the feasibility of implementing a hyperspectral sensor in this regard. The first derivative of the spectrum was calculated to locate the maximum change, which was found in the 1350-1400 nm range. Notably, an absorption band for the corroded sample was identified in the 1340-1440nm range. However, this study has certain limitations. The dataset lacked diversity and balance for image-based, non-contact methodologies—an aspect that should be addressed in future research. Additionally, the performance of hyperspectral imaging (HSI) should be further validated using a larger set of corroded samples generated under controlled corrosion chamber conditions.

Download

Share

COinS