Pipelines are mainly used to transport crude and refined petroleum, such as natural gas, worldwide. Monitoring pipeline health condition at offshore locations is challenging. Despite several attempts to develop leak detection systems, few can simultaneously detect the leak location and size. It is extremely difficult to obtain abnormal data such as actual leaks from a long-distance subsea pipeline. Dynamic modeling can be a good alternative to overcome this limitation. In this study, based on the dynamic model matched with the field, we conducted various flow simulations and selected the most sensitive variables. By changing these variables within an appropriate range, a machine-learning data set was generated. We used deep neural network methods to train the data and derived the optimal learning model. To improve the model accuracy, we adjusted the pipeline model section size not to exceed 20 m from the initial 50 m and designed models with a more detailed pipeline structure. The mean absolute error for each leak size was separately calculated to assess its effect on learning itself. Overall, the model showed excellent accuracy. However, for leak sizes of 0.5 cm, the accuracy appeared too low because the leak effect on mass flow, pressure, and the temperature was minimal. These parameters have been reported to have a great impact on the accuracy of machine-learning models. Therefore, the leak size detected was rearranged to perform data learning again. As a result, the model accuracy was improved by 80% compared to the initial learning model. Based on our study results, we proposed a flowchart for leak detection in the gas pipeline. The proposed procedure can be applied to various pipelines and support more efficient operation by detecting leaks in real-time.

管道主要用于在世界范围内运输原油和精炼石油，例如天然气。监测海上位置的管道健康状况具有挑战性。尽管曾多次尝试开发泄漏检测系统，但很少有人能够同时检测泄漏位置和大小。从长距离海底管道获取实际泄漏等异常数据极其困难。动态建模可能是克服这一限制的一个很好的选择。在本研究中，基于与现场匹配的动态模型，我们进行了各种流动模拟并选择了最敏感的变量。通过在适当范围内更改这些变量，生成了机器学习数据集。我们使用深度神经网络方法训练数据并推导出最优学习模型。为了提高模型精度，我们将管道模型截面尺寸从最初的 50 m 调整为不超过 20 m，并设计了更详细的管道结构模型。分别计算每个泄漏尺寸的平均绝对误差以评估其对学习本身的影响。总体而言，该模型显示出出色的准确性。然而，对于 0.5 cm 的泄漏尺寸，由于泄漏对质量流量、压力和温度的影响很小，因此精度显得过低。据报道，这些参数对机器学习模型的准确性有很大影响。因此，重新排列检测到的泄漏尺寸以再次进行数据学习。结果，与初始学习模型相比，模型精度提高了 80%。根据我们的研究结果，我们提出了燃气管道泄漏检测的流程图。