Predicting the remaining life of oil pipeline circumferential welds based on hybrid machine learning-based methods

Circumferential welds are often considered critical junctions in oil pipelines. Considering that the failure of circumferential welds in pipelines can lead to economic losses and environmental pollution, timely maintenance of these welds is crucial, which requires accurately estimating the remaining life of welds. This paper proposes a comprehensive framework with hybrid machine learning-based methods for circumferential welds remaining life prediction. A backpropagation (BP) neural network is developed to identify circumferential welds with abnormal detection levels related to cracking defects. Then, another BP neural network and support vector regression are utilized to establish a time-series-based model for predicting the remaining life of circumferential welds. The model is then optimized for accuracy using a stacking method. The proposed methods are applied to real data from a pipeline, and the results indicate that the optimal model for abnormal circumferential weld detection achieves a training set accuracy of 99.44 %, a test set accuracy of 99.71 %, a recall rate of 0.97, and an F1 score of 0.98. The optimal prediction model for the remaining life of circumferential welds has root mean square errors of 1.36, 3.28, and 0.67. The research results demonstrate that the models have high accuracy and good performance.