Pipeline anomaly detection model using machine learning approach in oil and gas industry

Abstract

Anomaly detection plays a crucial role in the oil and gas industry by identifying irregular patterns within pipeline systems, signaling the occurrence of pipeline leaks and potential infrastructure failures. Given the extensive reach of these pipelines, effective leak detection ensures operational safety while minimizing environmental hazards and financial loss. Traditional leak detection methods, such as manual inspection and pressure monitoring, are prone to errors and delays, often resulting in high false positives and delayed leak detection. A review of existing research reveals key limitations: high false positive rates, failure to adequately address class imbalance, and limited exploitation of temporal and long-term dependencies inherent in pipeline data. This study presents a comparative evaluation of three Machine Learning (ML) models: Support Vector Machines (SVM), Random Forest (RF), and  Long Short-Term Memory (LSTM)  for pipeline leak detection. The research focuses on developing models that effectively minimize false positives while maintaining high sensitivity to actual leak occurrences. The Synthetic Minority Over-sampling Technique (SMOTE) was applied to mitigate class imbalance and enhance the detection of the minority-class. The results indicate that integrating SMOTE with optimization techniques, hyperparameter tuning, and threshold adjustment enhances the models’ performance by reducing false positives and improving the minority (leak) class detection. The LSTM achieved superior performance with 98% Accuracy and 97% F1-score, attributed to its ability to effectively capture temporal and long-term dependencies in pipeline data. These findings contribute to enhanced and reliable leak detection models, tailored to the operational safety and management of pipeline system.