Bridging SCADA Noise and Real-Time Field Alerts: A Comparative Benchmarking of Machine Learning and Lightweight Time-Series Models for Oilfield Pipeline Anomaly Detection

Unplanned flow disruptions, severe scale accumulation, and pressure anomalies in offshore oil and gas production pipelines cost the energy sector millions of dollars in operational downtime and present severe environmental hazards. Traditional SCADA monitoring systems rely heavily on static single-variable thresholds that trigger frequent false alarms during routine transient flow shifts, creating alarm fatigue for control-room operators. Conversely, while complex deep learning architectures can capture non-linear multi-sensor relationships, their high computational overhead and black-box nature pose severe challenges for real-time edge deployment on remote production platforms. This study presents a comprehensive benchmarking framework evaluating classical machine learning models (Logistic Regression, Random Forest, XGBoost, LightGBM) against lightweight temporal sequence architectures (LSTM Autoencoders) using Petrobras's open-source 3W dataset. Sensor streams including Permanent Downhole Gauge pressure (P_PDG), Temperature and Pressure Transducer measurements (P_TPT, T_TPT), and choke downstream temperature (T_JUS-CK) were preprocessed using rolling statistical lag windows and class-weight balancing. Our empirical results demonstrate that gradient-boosted decision trees (XGBoost/LightGBM) paired with 15-second statistical lag features achieved a Macro F1-score of 0.914 and a PR-AUC of 0.932, outperforming the LSTM Autoencoder (Macro F1: 0.876) while reducing inference latency by 78% (12.4 ms vs 56.2 ms per sample). These findings show that lightweight ensemble learning offers a highly effective, low-latency strategy for real-time pipeline integrity monitoring.