Enhancing Anomaly Detection in Oil and Gas Rotating Machinery through a Physics-Informed Data Filtration Pipeline

The reliability of anomaly detection models in the oil and gas sector is frequently compromised by the poor quality of raw sensor data from rotating machinery. This paper introduces a multi-stage, physics-informed data filtration pipeline designed to systematically remove operational noise and confounding artifacts from vibration data prior to model training. By applying this pipeline, the accuracy of an Isolation Forest-based anomaly detection model was significantly improved from a baseline of 56% on unfiltered data to a robust 92%. This enhancement transforms the system from an unreliable alerting mechanism into a trustworthy decision-support tool for predictive maintenance (PdM). The proposed framework represents a practical shift towards proactive maintenance strategies, which are critical for reducing unplanned downtime, minimizing operational costs, and enhancing safety in high-stakes industrial environments.