Abstract:

Buried pipelines are prone to external corrosion perforation, and leakage during service. Accurate prediction of external corrosion rate is of great significance for formulating reasonable pipeline maintenance strategies. By sorting out the factors affecting the external corrosion rate of pipelines, soil properties, stray currents, cathodic protection, and physicochemical properties were identified as the main influencing factors, and 60 sets of relevant external corrosion data were collected along a gathering pipeline in a certain region of China. Subsequently, corrosion information was extracted through multi-physics and data-driven approaches, and a Physics Guided Neural Network (PGNN) model based on physical mechanism guidance was constructed. On the basis of the conventional loss function, this model introduces physical mechanism constraints as penalty terms, adjusts corrosion factors and retrains the model to ensure that the training direction conforms to the corrosion mechanism. Genetic Algorithm (GA) is used to optimize hyperparameters, and the GA-PGNN corrosion rate prediction model is formed. Finally, Shapley Additive Explanations (SHAP) analysis quantifies to measure the influence of corrosion characteristics on corrosion rate from both global and local perspectives. The results demonstrate that compared to conventional Back Propagation (BP) and GA-BP models, the GA-PGNN model achieves superior performance with Mean Absolute Percentage Error (MAPE), Mean Squared Error (MSE), and correlation coefficient (R²) of 3.71, 1.51×10⁻⁵, and 0.9935, respectively. The GA-PGNN model exhibits smaller and more balanced mean absolute SHAP values, indicating consistent dependency on diverse corrosion factors and effective utilization of information from each factor. Conversely, both conventional BP and GA-BP models fail to capture accurate corrosion mechanisms, occasionally yield conclusions contradictory to physical principles, and predictions show significant randomness and instability. The GA-PGNN framework provides actionable insights for enhancing the integrity management of buried pipelines.

Key words: buried pipeline, external corrosion rate, physics-guided neural network, shapley additive explanations, loss function

摘要： 埋地管道在服役过程中易出现外腐蚀穿孔泄漏问题，准确预测外腐蚀速率对于合理制定管道维护策略具有重要意义。通过梳理管道外腐蚀速率影响因素，明确了土壤性质、杂散电流、阴极保护及理化特性为主要影响因素，并在我国某地区集输管道沿线收集了60组相关外腐蚀数据。随后，从物理和数据两个维度挖掘腐蚀信息，构建了基于物理机制引导的神经网络模型(Physics-Guided Neural Network，PGNN)。该模型在常规损失函数的基础上，引入物理机制约束作为惩罚项，通过调整腐蚀因素并重新训练模型，确保训练方向符合腐蚀机理，并利用遗传算法(Genetic Algorithm，GA)对超参数进行寻优，形成GA-PGNN腐蚀速率预测模型，最后通过事后可解释算法(Shapley Additive Explanations，SHAP)从全局和局部两个角度衡量腐蚀特征对腐蚀速率的影响程度。结果表明，与常规神经网络(Back Propagation Neural Network，BP)模型和GA-BP模型相比，GA-PGNN模型的平均绝对百分比误差、均方误差和相关系数分别为3.71、1.51×10^-5、0.9935，在3种模型中表现最佳；GA-PGNN模型的平均绝对SHAP值较小，且较均衡，说明模型对不同腐蚀因素的依赖程度基本一致，可以从每个腐蚀因素中获取有用信息；常规BP模型和GA-BP模型均无法捕捉到正确的腐蚀规律，甚至得到与腐蚀机理相悖的结论，预测结果存在较大的偶然性和随机性。研究结果对于提高埋地管道完整性管理水平具有借鉴意义。

关键词: 埋地管道, 外腐蚀速率, 物理机制引导神经网络模型, 事后可解释算法, 损失函数

CLC Number: 

• TE988.2

• X937