In recent years, leakage detection models using artificial intelligence have been widely used by researchers. In this study, acoustic sound data were recorded at 10 observation points, focusing on ductile iron pipe and vinyl polyethylene. Leakage detection models were built using recurrence plots (RPs) and a convolutional neural network (CNN). Using the same number of RPs for training data and testing data, we analyzed the effect of data variations on model performance. The results showed that our proposed approach can improve accuracy at several points, although the amount of training data information was advantageous in previous work. There were cases in previous work that had poor accuracy, but when implementing our proposed approach, it improved the accuracy to over 80% when using the 8-point model. The increase in accuracy depends on which interval is used in the test data because each interval contains information about different properties. For multi-point models, the effect of increasing the number of RPs and data variations was not clarified in the previous work. However, our study confirmed that the increase in data variation contributed more than the RP number.