The work reports the systematic approach to the study of artificial intelligence (AI) in addressing the complexity of inline inspection (ILI) data management to forecast the risk in natural gas pipelines. A recent conventional standard may not be sufficient to address the variation data of corrosion defects and inherent human subjectivity. Such methodology undermines the accuracy assessment confidence and is ineffective in reducing inspection costs. In this work, a combination of unsupervised and supervised machine learning and deep learning has profoundly accelerated the probability of failure (PoF) assessment and analysis. K-means clustering and Gaussian mixture models show direct relevance between the corrosion depth and corrosion rate, while the overlapping PoF value is scattered in three clusters. Logistic regression, support vector machine, k-nearest neighbors, and ensemble classifiers of AdaBoost, random forest, and gradient boosting are constructed using particular features, labels, and hyperparameters. The algorithm correctly predicted the score of PoF from 4790 instances and confirmed the 25% metal loss at a location of 13.399 m. The artificial neural network (ANN) is designed with various layers (input, hidden, and output) architecture. It is optimized using an activation function to predict that 74% of the pipeline's anomalies that classified at low-medium and mediumhigh risk. Furthermore, it provides a quick and precise prediction about the external defects at 13.1m and requires the personnel to conduct wrapping composite. This work can be used as a standard guideline for risk assessment based on ILI and applies to industry and academia.
|ASCE-ASME Journal of Risk and Uncertainty in Engineering Systems, Part B: Mechanical Engineering
|Published - Mar 2023
- artificial intelligent
- deep learning
- inline inspection
- machine learning
- unsupervised machine learning