Comparison of Isolation Forest and One Class SVM in Anomaly Detection of Gas Pipeline Operation
Kelvyn Lukito, Aditya Firman Ihsan, Hasmawati Hasmawati · 2023
Detecting data anomalies in the operational process of oil and gas pipelines is very important to reduce the risk of disasters, which can adversely affect human safety, the environment and financial aspects. Failure to do so can lead to catastrophic results. The problem is also supported by several catastrophic events that have occurred in several areas of oil and gas production facilities in several regions. To solve this problem, it is necessary to implement a suitable monitoring system that aims to prevent potential losses caused by leaks or over-pressurization of natural gas pipelines. Among the many machine learning algorithms available for anomaly detection such as Feed Forward Neural Network, Linear Regression, KNN, Random Forest, and Support Vector Machine and unsupervised machine learning models such as Principal Component Analysis (PCA) and Hierarchical clustering, One-Class SVM and Isolation Forest are the most prominent. However, these algorithms have their own advantages and disadvantages regarding their performance. This study aims to compare the performance of machine learning algorithms in classifying and detecting data anomalies in offshore natural gas pipeline operational datasets. The assessment is based on ROC-AUC Curve, Confusion Matrix, Sensitivity, and Specificity. The findings indicate that the Isolation Forest model outperforms the One-Class SVM, with a ROC-AUC value of 90%, compared to the One-Class SVM's value of only 61%. Furthermore, the Isolation Forest exhibits a Sensitivity value of 98%, in contrast to the One-Class SVM's 41%, and a Specificity of 81%, compared to the One-Class SVM's 80%.