The identification method of pipeline leakage aperture based on the deep neural network has a high identification rate. However, its application in industrial environment and real-time processing is greatly limited due to the large number of parameters and large memory consumption due to its complex structure. To address this issue, an optimized convolution improved dense block lightweight neural network is proposed for the pipeline leak aperture identification. Firstly, a new multi-convolutional dense block is constructed by combining the deeply separable convolution with the heterogeneous convolution to extract the features of leakage signals. Then, the convolutional attention mechanism is used to classify the weight of features to realize the importance distinction of features. Finally, the results are obtained by classifier. Experimental results show that the recognition accuracy of the proposed method is 96. 59% , and the number of parameters is only 781 KB. While ensuring high recognition accuracy, the number of parameters and floating point numbers are greatly reduced, the training time is also reduced, and the real-time response ability is improved, which has guiding significance for practical industrial monitoring applications.