Assessing Likelihood of Failure for Unpigged Pipelines Using Supervised Machine Learning

Abstract System-wide risk assessment is an integral process for quantitatively evaluating the risk levels of pipeline segments. This involves utilizing the best available information, including inline inspection (ILI) data, survey data, pipeline attributes, historical mitigation information and environmental properties pertinent to the integrity threats. The assessed risk provides the basis for planning mitigation activities, and ensuring that the risk level is maintained at an acceptable level. Estimation of likelihood of failure (LOF) is a crucial component of the risk assessment, which is subsequently combined with the consequence of that failure to assess risk. In the absence of ILI data, the LOF for unpigged pipelines were often estimated using conventional statistical models. These models delineate the relationship between historical failure rates and the factors influencing the occurrence or prevention of integrity threats. While such traditional regression models offer the advantage of interpretability, they are constrained in capturing the intricate nonlinear relationship between the LOF and pertinent variables. Additionally, they lack the flexibility to characterize the complex interactions among multiple predictor variables. To improve the accuracy of LOF assessment for unpigged pipeline segments exposed to external corrosion, this paper introduces a multi-layer model that was developed using supervised machine learning techniques and structural reliability method. The model consists of three layers. In the initial layer, a series of machine learning models are employed to predict key statistics that characterize the corrosion severity of the pipeline segments. These machine learning models were trained and validated using pigged pipeline segments, with ILI data considered as the ground truth for the corrosion severity. The predictor variables utilized by the model include pipeline attributes, factors contributing to corrosion, factors resisting corrosion, and survey data (e.g., DCVG). In the second layer, the previously predicted corrosion states serve as the input for probabilistic corrosion assessment models, facilitating the evaluation of the probability of leak (POL) and probability of burst failure (POF) through Monte Carlo simulation. In the final layer, the assessed POL and POF are aggregated into the LOF. This paper describes the framework of the multi-layer model, outlines the data requirement for model development, and illustrates the process of model calibration and validation. Furthermore, practical considerations in the model development are discussed, ensuring the general alignment between model predictions and engineering judgements. The enhanced LOF accuracy will enable the operators to make more cost-effective mitigation decisions for unpigged pipelines based on the results of risk assessments.