Advanced imaging techniques for damage characterization of concrete

Abstract

Reinforced concrete is widely used for the construction of bridges, buildings, highways, dams, power plants and many other infrastructures. Monitoring the structural health of these infrastructures is essential for their uninterrupted functioning and assessing their deterioration due to loading and environmental factors. The monitoring also helps in the estimation of its load-carrying capacity, serviceability and need for repair. Cracking in concrete structures is one of the critical parameters representing the health of the structure. Concrete cracking occurs due to many reasons like shrinkage, heaving, premature drying, and excessive loading which leads to a reduction in the strength of structures. Trained personnel monitor the development of cracks and their progression at the critical locations of the structures through a physical vision at regular intervals of time. Structures like bridges, buildings, roads, tunnels, historical monuments and many others are inspected at regular intervals to estimate their deterioration and to prevent further accidents which may directly affect human life. Different methods like acoustic, ultrasonic and image processing-based inspection methods have been deployed to carry out an assessment of such concrete structures. Physical inspection of structures for health monitoring is time-consuming, costly and risky. Automatic detection of cracks in concrete structures of various shapes and scales holds worthy importance in the area of structural health monitoring. Advances in image acquisition, processing techniques, and computational resources have made vision systems, a cost-effective and accurate technique for structural health assessment. The present work is aimed to address the concrete crack detection problem by developing a novel system using machine vision and deep learning. The work covers concrete crack monitoring by identifying the location of the crack, the number of cracks, the length of the cracks, and the area of the cracks. The methodology is applied for crack monitoring of concrete cube of standard dimensions 150 mm × 150 mm × 150 mm subjected to compressive loading. An innovative system has been developed by integrating machine vision and convolutional neural networks to acquire real-time images of concrete surfaces, detect concrete cracks, and extract parameters related to cracks, such as the number of cracks, location, length, and area, in synchronization with the applied compressive loading. The present system is implemented to capture 1 sample of image and load data per second, for crack monitoring during compression testing of concrete cubes of size 150 mm × 150 mm × 150 mm. The crack x detection methodology presented in the work offers a better understanding of concrete cube behaviour with respect to applied compression load. An integrated experimental setup is developed to acquire consistent images of the concrete cube along with the applied load during the compression testing of the cube. The setup comprises of a state-of-the-art industrial machine vision system, load cell and load cell data acquisition system. Several concrete cubes are tested for compressive load and surface images of the cubes are acquired. Due to random unevenness on the surface of concrete blocks, designing an efficient crack detection algorithm becomes challenging. The problem of crack detection is first addressed using subtraction, line emphasis and iterative threshold-based method. Additional method based on percolation as well as machine learning are also applied for the crack detection. Each of these methods lacks in meeting expected performance in terms of accuracy and precision in crack detection. Convolutional neural network based methodology has been applied to overcome the challenges of variations in the concrete cube surface images, and similarity in the intensity level of dents, noises and spots present on the cube surface. With the use of the developed experimental setup, dataset collection and preparation are carried out for binary contour images and greyscale contour images of cracks and non-cracks. Initially, the neural network based method is used with a binary image dataset. The results obtained with the binary dataset are not satisfactory and hence greyscale data is used thereafter. But due to data imbalance between the crack and non-crack contours, the performance of the network with greyscale images do not have good recall and precision. Finally, the network is trained with a balanced and augmented greyscale image dataset to achieve satisfactory results in crack classification. Despite the challenges in accurate crack identification on uneven surface of concrete cube, the novel methodology shows improved performance in crack detection. The Inception v3 model is having lower error rate, faster convergence and reduced computation time for object identification. Therefore, Inception v3 architecture is trained to detect the cracks from the image of concrete cube surface images in the most accurate manner. The Inceptionv3 model is trained and validated using more than 80,000 crack and 80,000 non-crack images dataset prepared manually using the concrete cube surface images. Popular data augmentation techniques are used to generate the training dataset. An average of 97.49% accuracy and 7.38% cross-entropy are achieved in the training whereas 97.67% accuracy and 7.69% cross-entropy are achieved in the model validation. The training is carried out with a batch size of 100 and 5,000 epochs. An average accuracy of 99% has been achieved during the performance evaluation of crack detection on concrete cubes as presented in the results. The average values of precision, recall and F – score is obtained as 0.88, 0.98 and 0.93 respectively. The performance of the proposed methodology is compared with the other strategies reported in the literature based on image acquisition device, infrastructure and CNN model. In comparison, it is observed that the performance of the present methodology is 2-10% better in terms of crack detection accuracy. In addition to that, the proposed methodology is also compared with other methods having similar approach like machine learning based crack detection, crack classification using the binary dataset, imbalance greyscale dataset and balanced greyscale dataset. The comparison justified the use of data augmentation techniques and a greyscale contour dataset prepared for accurate crack detection. The outcome of the machine vision system in graphical form is presented for various parameters of cracks like the number, location, length, the area with respect to compressive load for different concrete cubes. Comprehensive crack-load monitoring and analysis is presented for multiple concrete cubes. A detailed crack-load analysis data is presented which comprises of analysis of individual crack-related data, crack-highlighted images of concrete cubes, load versus progression in the crack length, load versus progression in the crack area as well as cumulative crack length and crack area with respect to load for the concrete cube. The observations obtained from the crack load analysis can provide vital information and improve the understanding of concrete behaviour subjected to loading. The developed machine vision and deep learning based method is a step forward in the structural health monitoring of real-life concrete structures like buildings, bridges, and pavements. The machine vision system can be implemented on different concrete structures for acquiring real-time data on crack development and progression. The developed framework will be an effective tool for engineers working in the domain of structural health monitoring of concrete structures.