Concrete structure damage classification and detection using convolutional neural network
MetadataShow full item record
A resilient infrastructure system remains a top priority for Canada as it is hinged to strong economies. Corrosion, ageing, aggressive environments, material defects, and unforeseen mechanical loads can compromise the serviceability and safety of existing infrastructures. Introducing Artificial Intelligence (AI) in smart structural health monitoring (SHM) can assist in building an automated and efficient infrastructure condition monitoring method to facilitate an effortless inspection and accurate evaluation of deteriorated infrastructure. Over the last decades, vision-based AI has proven successful in pattern recognition applications and motivated this current research to assemble a data-driven damage detection technique. To further explore the possibilities of deep-learning (DL) applications, this dissertation research aims to develop an autonomous damage assessment process using DL techniques to classify and detect two types of defects- crack and spalling on concrete structures. This research started with reviewing existing application of various DL-based technologies for damage detection of concrete structures and identifying the challenges and limitations. One major challenge in DL-based SHM technique is the lack of adequate real image database obtained from field inspection. To address these challenges, this research has created a diverse dataset with concrete crack (4087) and spalling (1100) images and used it for damage detection and classification by applying convolutional neural network (CNN) algorithms. For defects classification VGG19, ResNet50, InceptionV3, MobileNetV2, and Xception CNN models were used, whereas semantic segmentation process adopted Encoder-Decoder Models- U-Net and PSPnet. For both cases a detailed sensitivity analysis of hyper-parameters (i.e., batch size, optimizers, learning rate) was performed to compare their performances and identify the best-performed model. After assessing all the criteria, the best performance for defects classification was achieved by InceptionV3. On the other hand, for crack and spalling segmentation U-net outperformed the other models. Overall, the developed algorithms achieved an excellent performance in damage classification and localization and proved to be successful enough to offer an automated inspection platform for ageing infrastructures. The outcomes of this study signify that the data-driven CNN methods could be a promising solution for the condition assessment of deteriorating concrete structures. The research outcomes can be implemented by practitioners for condition assessment of existing infrastructure and recommending proper rehabilitation measures.