With the continuous expansion and extension of pipeline networks, leakage incidents in pipeline systems have become increasingly frequent, severely impacting the surrounding environment and residents. Consequently, the safety monitoring of pipeline systems has become particularly important. Acoustic leak signal recognition technology plays a pivotal role in detecting and locating pipeline leaks. This technology is gradually transitioning from traditional methods to deep learning approaches, which has introduced new opportunities and challenges to the field. However, the accuracy and generalization capability of acoustic leak signal recognition still require further enhancement. After pre-processing, the acoustic signals are converted into Mel-Frequency Cepstral Coefficients (MFCC) feature maps as inputs, transforming the acoustic signal recognition problem into an image recognition problem. The CNN is then employed for learning and classification to identify the presence of leaks. The dataset includes acoustic signals collected under various environmental conditions. Experimental results indicate that the proposed improved CNN model performs exceptionally well in leak detection tasks, achieving a high accuracy rate. Compared to sound signal recognition methods based on Support Vector Machine (SVM) and Alex Net, the proposed method demonstrates a significant advantage. The CNN-based acoustic leak signal recognition method proposed in this paper not only improves the accuracy of leak detection n but also provides an effective technical means for the safety monitoring of pipeline systems. Future work will focus on further optimizing the model and validating its application in actual industrial environments.