Classification Prediction of Natural Gas Pipeline Leakage Faults Based on Deep Learning: Employing a Lightweight CNN with Attention Mechanisms

The integrity of natural gas pipelines will decrease with an increase in operating time, thus causing pipeline leaks and accidents. However, it is challenging to improve the precision and automation of existing sensors to raise leak prediction and classification precision. Therefore, based on deep learning, a 1D convolutional neural network (CNN) incorporating the channel attention mechanism is proposed for recognizing and classifying the type of natural gas pipeline leakage. Firstly, the data reconstruction of the leaked acoustic signals, which have been classified by energy modes, is performed by feature augmentation and Bessel filtering. Subsequently, a lightweight CNN is proposed, and an attention mechanism is introduced to optimize the model performance. The results show that the training performance of the network with the attention mechanism is superior to that of the original network and the network with batch normalization. The attention mechanism network is then used to train the leakage signals with different features of engineering parameters. Finally, the test accuracy achieves 97.81%, validating the effectiveness of the proposed method for identifying and classifying natural gas leaks. It presents new ideas for the implementation of deep learning in the natural gas and chemical industries.