Oil and gas anomaly detection using feature engineering techniques and machine learning

Anomalies in oil and gas systems are defined as abnormal operational changes of the system such as sudden changes in pressure, changes in temperature, instabilities in the flow of the oil or gas or malfunctions in the valves that disrupt the normal conditions of operation. These anomalies are critical to identify early because they may cause severe outcomes, such as equipment damage, down-time in the production process, safety risk, and expensive environmental accidents. With the advent of modern oil fields that feed voluminous multivariate sensor data using the IIoT infrastructure, manual monitoring becomes impractical and intelligent automated detection methods must be implemented. This thesis introduces a comprehensive anomaly detection framework for oil and gas operation based on advanced feature engineering techniques and a combination of ML and statistical learning models. Using the publicly available 3W Dataset 2.0.0, which is a multivariate time-series of 42 oil wells in normal and faulty operation condition, the thesis deals with major industrial issues including sensor noise, nonlinear operating regimes, class imbalance, and reduced interpretability of models. Six machine learning algorithms (XGBoost, Random Forest, SVC, KNN, Decision Tree and Logistic Regression) and three statistical learning models (Lasso, Ridge, and ElasticNet Logistic Regression) were tested after rigorous preprocessing (data cleaning, temporal segmentation and statistical and temporal descriptors extraction: mean, standard deviation, RMS, energy, IQR, skewness, kurtosis, and zero-crossings). The model robustness was estimated in three experiment conditions, clean data, noisy data, and noisy data with hyperparameter optimization via Optuna. The results indicate that the ensemble approaches, especially XGBoost and Random Forest, demonstrated better accuracy, recall, and MCC in all conditions, with XGBoost tuned to reach 99.93% accuracy and 99.87% recall. Noise injection showed which simpler linear models were very vulnerable, and Optuna tuning significantly enhanced model stability.