Pipeline block and pipeline leak may lead to serious accidents and cause huge economic losses, which have been urgent problems for gas transportation. In this work, active acoustic pulse-compression technology is first introduced to detect and locate these two anomalous events. The matched filtered signals are then normalized and input into one-dimensional convolutional neural network to achieve classification of not only pipeline block and pipeline leak but also normal event such as pipeline elbow which causes acoustic wave reflection as well. Neural network parameter optimization has also been carried out as well as the comparison with long- and short-term memory network. Experimental results demonstrate that compared with long- and short-term memory network, one-dimensional convolutional neural network has an improvement in efficiency due to the great reduction of running time. For non-aliasing pipeline events, both of the models can reach 100% classification accuracy. For aliasing pipeline events, despite the shorter time series and fewer features, the classification accuracy of one-dimensional convolutional neural network still reaches 100.00%, but that of long- and short-term memory network is only 93.89%. Furthermore, the smoothing and slight fluctuation of receiver operating characteristic curve and the high value of area under curve also verify the stability and good classification performance of the proposed trained model. Therefore, the one-dimensional convolutional neural network shows significant performance for pipeline events classification and has considerable potential and application prospect in gas pipeline safety monitoring.

管道阻塞和管道泄漏可能导致严重事故并造成巨大的经济损失，这已成为天然气运输的迫切问题。在这项工作中，首先引入了主动声脉冲压缩技术来检测和定位这两个异常事件。然后将匹配的滤波信号进行归一化，并输入到一维卷积神经网络中，以实现不仅对管道阻塞和管道泄漏进行分类，而且对正常事件（例如也会引起声波反射的管道弯头）进行分类。还进行了神经网络参数优化以及与长期和短期存储网络的比较。实验结果表明，与长期和短期记忆网络相比，一维卷积神经网络的运行时间大大减少，从而提高了效率。对于非混淆管道事件，两个模型都可以达到100％的分类精度。对于混叠管道事件，尽管时间序列较短，特征较少，但一维卷积神经网络的分类精度仍达到100.00％，而长期和短期存储网络的分类精度仅为93.89％。此外，接收机工作特性曲线的平滑和轻微波动以及曲线下面积的高值也证明了所提出训练模型的稳定性和良好的分类性能。所以，