Due to the long-term effects of wind and wave corrosion and hydraulic erosion on underwater structures, various degrees of damage such as cracks, holes, and corrosion may occur. When underwater attachments cause lift off disturbances, various interference signals are introduced, increasing the difficulty of defect detection. In order to maintain the safety and stability of reinforced concrete structures in underwater tunnels in a timely manner, a defect detection method for reinforced concrete structures in underwater tunnels based on ultrasonic echo signals and the CNN is proposed. The collection method and formation mechanism of ultrasonic echo signals for reinforced concrete structures in underwater tunnels are analyzed first. After obtaining the ultrasonic echo signals, noise is removed from the signals through empirical mode decomposition and the sparse table algorithm. An ultrasonic defect detection model for concrete structures based on the multi-attention FasterRCNN structure is constructed, the denoised ultrasonic echo signal is input into the model, the multi-attention mechanism is applied to extract the characteristics of the ultrasonic echo signal, and a balanced feature pyramid network is used to achieve feature fusion. The fused features are generated into defect candidate boxes through regional generation networks, and defect position information and category information are output at the fully connected layer of the model to complete defect detection of reinforced concrete structures in underwater tunnels. The experimental results show that this method can accurately remove noise from the echo signal of reinforced concrete structures and shows high denoising performance. When conducting defect detection, it can quickly detect the defect category and position, and provide the defect depth and diameter. The detection results are accurate and do not show significant errors.