Application of Computer Vision Technique For Crack Monitoring Of Concrete Cube Subjected To Compression

Abstract

To ensure safety and serviceability of concrete structure, monitoring its strength and deterioration due to environmental exposure is necessary. Physical inspection of structure is carried out at regular interval by trained professionals for monitoring condition of struc- ture and evaluating capacity of structures to bear expected load. Based on evaluation of structure, strategy for strengthening is determined. Physical inspection of large scale structures is time consuming, costly and risky. Visual inspection has limited exposure area and is unable to keep real time images for future reference & evaluation. There is also a chance for human error during observation and reporting condition of existing structure. Recent advances in imaging devices, computational resources and artificial intelligence- based methodologies have enabled automatic and accurate damage detection. Computer vision technology integrates image acquisition using cameras, image processing as well as artificial intelligence techniques for damage identification in structures. It reduced human effort and made it easy to develop strategy for condition monitoring of structures. One of the primary defects in concrete structure is crack. Concrete cracking occurs due to many reasons like shrinkage, heaving, premature drying, excessive loading etc. and it leads to reduction in its strength. Therefore crack identification and its monitoring in concrete structure is essential for ensuring safe functioning of important infrastructure facilities like building, bridges, dam, pipelines, road pavement etc. The present major project work is an initiative to explore computer vision technique in crack identification in concrete struc- tures. In present project work concrete cube under compressive loading is considered to demonstrate concept of computer vision for crack identification and monitoring. The main objective of the work is to understand various aspects of computer vision methodology for crack detection on surface of concrete cube subjected to compressive loading. Camera and lights are used to capture real time images with proper illumination when concrete cubes are subjected to loading. For identification of crack on concrete surface, machine learning technique is adopted. A convolution neural network (CNN) based approach is implemented for accurate crack identification. The approach comprises of dataset generation , transfer learning- based re-training of a pre-trained convolution neural network model and crack identification. Crack and non-crack contour images ex- tracted from images of the standard concrete cubes are used for training CNN. During

compressive loading of concrete cubes load is measured simultaneously using load cell. Data obtained from load cell and camera are communicated to computational devices for further processing. Using CNN, cracks on concrete surface are identified and further processing is carried out to evaluate length, area of each cracks. Using synchronized ob- servation of load and crack identification, various analysis are performed such as load v/s crack parameters like length & area, individual crack propogation and location. Different grade of cubes are cast and tested under compression load. Various features of cracks like numbers, location, length, area etc. are extracted for concrete cubes of different grades and load v/s crack parameters are presented graphically. Location and detailing of every crack from its initiation to termination is presented in the report. The total area of cracks as well as length of cracks at failure are observed in addition to load at first crack initiation. Crack-load analysis obtained from computer vision technique in present study can be used in identifying various characteristics of concrete. From crack-load analysis carried out during compressive testing of cubes, it is observed that for low strength concrete after 75% of total capacity, crack initiation is started. For the concrete with higher compressive strength crack is initiated after 25-30% of ultimate load. Maximum 6% area of total concrete surface is cracked at time of failure and average crack area is 2.61%. Maximum 4% cumulative crack length from total area of concrete surface is cracked at the time of failure and average value is 2.58%. Most of the cracks are terminated before failure and the end stage of cracks is due to spalling of certain portion of concrete. A very few cracks remain active till the failure. Simultaneously multiple cracks occur after about 70% loading of total capacity. So occurrence of multiple cracks indicates that cube is reached 70% of its strength. Average accuracy of 98% is observed during the performance evaluation of algorithm for crack identification. It is found that the model provided high recall and precision. The results justified the use of convolution neural network-based approach for accurate crack identification. Similar approach can be easily adopted by practicing professional and engineers to carry out damage assessment of real life structures.