In order to monitor whether there is leakage in the water supply pipeline, aiming at the problems of low manual detection accuracy, difficult removal of background noise and low practicability of equipment, a water supply pipeline leakage sound recognition method based on lightweight convolutional neural network CNN is studied. Firstly, the real-time sound acquisition system in the water supply pipeline of the underwater robot is designed, and the sound information collected by the system is uploaded to the upper computer by using UDP/IP communication technology. Divide the sound in the water supply pipeline into leaky and non-leaky sound and make it into a data set, extract the Mel spectrum feature map information of leaky audio and non-leaky audio samples, and select lightweight convolutional neural network ShuffleNet V2 for training and recognition according to the real-time performance; Secondly, the attention mechanism of CBAM is introduced to improve the Unit1 of ShuffleNet V2 and We proposed Unit1_y. At last, the improved network is compared with the lightweight networks such as MobileNet V3 and Resnet18. The test results show that the improved network has the best effect on water supply pipeline leakage sound recognition, and the recognition rate of the test set reaches 92.14%, which verifies the effectiveness of the algorithm.