Fault Detection Model of Rotating Machinery Using Machine Learning: Case Study of Oil and Gas Company

This paper introduces a machine learning-based approach for fault detection in rotating machinery within the oil and gas industry, focusing on a Gas Turbine-Compressor (GTC) unit. This paper compares the performance of three supervised classifiers (Support Vector Machine, Random Forest, and k-Nearest Neighbors) using accuracy, precision, Fl-score, and the Matthews Correlation Coefficient (MCC) that particularly suited for data imbalance environments. The analysis leverages historical condition monitoring data comprising 20 operating parameters collected from historical condition monitoring data. The methodology involves a two-stage process: initial data processing through Self Organizing Maps (SOM) clustering, followed by supervised classification. SOM facilitates unsupervised clustering to identify normal, alert, and fault conditions. Data imbalance within clusters is addressed using Synthetic Minority Over-Sampling Technique (SMOTE) on the training set. Among the classifiers, Random Forest exhibits superior performance with 98.5% accuracy, while SVM and KNN achieve 98.2% and 97.8% accuracy respectively. The consistency of evaluation results between training and testing sets indicates robustness and rules out overfitting concerns. Model explainability is enhanced through Shapley Additive Explanations (SHAP) Treexplainer, revealing key parameters influencing model outcomes and asset conditions.