An intelligent underwater robot was developed for crack detection of underwater buildings. The robot has constant temperature control and low energy consumption drive, and has functions of data collection, navigation, and positioning independently in ultra-low temperature deepwater environment. Based on the image data collected by the robot, the original CNN (Convolutional Neural Network) model was improved on the basis of image preprocessing, deep convolution network theory and fracture feature data annotation. Hence, an FPECNN (Feature Pyramid Engagement Convolution Neural Network) model was proposed to extract different types of cracks. The FPECNN model was applied to the crack detection project of Lianhua Hydropower Station. The calculation results demonstrate that the FPECNN model stands at a high level in detection rate, recall rate and F value, reaching 97.26, 98.04 and 96.65, respectively, consuming only 3.12 s. It is also well universal, robust, and viable as it adapts to most crack data, and refrains from low detection rate and low efficiency of conventional CNN model in the detection of underwater buildings. With this intelligent robot, the inspection personnel can be relived from the harsh, heavy and dangerous field work in cold underwater, and the huge economic loss caused by the abandonment of water in traditional inspections can be avoided, and improve the detection efficiency and accuracy.