Evaluation of Machine Learning-Based Regression Techniques for Prediction of Oil and Gas Pipelines Defect

Magnetic flux leakage (MFL) signals allow the scale of metal failure defects present on a pipeline to be observed, located, and measured. An advanced and reliable intelligent pipeline monitoring system is needed to protect the local ecosystem from disastrous effects due to malfunctioning pipelines. MFL plays a vital role in gas pipeline examination; various studies have explored smart MFL based defect prediction systems. Machine learning-based defect prediction systems allow predicting the characteristics of the oil and gas defects in real-time. As fault detection control is applied to low-performance embedded systems, prediction algorithms' reliability is of utmost importance. The intention was to extend the work of previous researchers and provide insight into the behavior of the selected classifiers with time as a robustness factor, an experimental design that constitutes the novelty of this study. For the said purpose, a prerecorded dataset comprised of MFL signals has been utilized. Prediction of defect length has been performed with several approaches, including linear regression (LR), linear regression with stochastic gradient descent (LRSGD), support vector machine (SVM), Gaussian process regression (GPR), boosting regression tree ensemble (BSTE), binary decision tree (BDT), stepwise (SW), and artificial neural network (ANN). The results indicated that ANN yields the best prediction results for MFL based defect predictions, followed by GRP regression analysis. The presented results can be utilized to design and implement a defect length prediction model to be deployed underwater, providing real-time prediction results.