Application of Gaussian Process Classification to the Integrity Management of Energy Pipelines

Abstract Leveraging algorithms that autonomously evolve with data, machine learning has become integral and been extensively applied to the pipeline integrity management practice to create accurate predictive models, whose outputs can either be numerical or categorical. The latter corresponds to the classification analysis, whose models typically yield deterministic outputs, categorizing data points as of either the positive or negative class. However, practical engineering scenarios may demand a more nuanced approach. Additionally, the conventional “black-box” nature of many machine learning models often hinders their interpretability. This study introduces Gaussian Process Classification (GPC) — a probabilistic classification algorithm that has been found to be underutilized — to the realm of pipeline integrity management, where challenges in the classification accuracy and efficiency of specific problems have become noteworthy. Unlike some classifiers, GPC provides the probability that a data point belongs to the positive class, allowing for a more flexible decision-making process. Notably, users can self-determine the threshold probability. Moreover, the predictive process of GPC is analytically transparent, enhancing the overall interpretability of the machine learning model. The algorithm is applied to three distinct binary classification problems within pipeline integrity management: separating failure modes for cracked pipelines, differentiating failure modes for corroded pipelines, and predicting ignition upon rupture of onshore natural gas transmission pipelines, while each problem involves unique input features. The machine learning models for these problems undergo thorough training via hyperparameter tuning using the maximum likelihood method. This approach improves model robustness, mitigates overfitting, and optimizes performance. The developed GPC models are benchmarked against existing mechanics-based and log-logistic models for comparative evaluation, whose results demonstrate the accuracy and superiority of GPC in these specific scenarios. The present study demonstrates the value of probabilistic machine learning for improving the pipeline integrity management practice.