Statistical Causality Analysis: Use Cases In Oil and Gas Applications

When a compressor trips or a process drifts, engineers usually work out the cause by scrolling through hundreds of sensor trends by hand. This paper tests several statistical and machine-learning ways to rank which sensors most likely drove an event, and wraps them in a simple in-house tool for operations staff. The main lesson is that the best method depends on the physics: lag-aware tests suit slow liquid systems, while clustering and linear-coefficient ranking suit fast-responding gas systems. No single method should be trusted on its own.

Key results

Granger causality worked best for liquid systems, where an upstream change shows up in other sensors only after a delay; clustering and linear-coefficient ranking worked best for gas systems, which respond almost instantly
In a separator gas-temperature trip, the tool linked the temperature rise to a recycle valve opening, which shortened troubleshooting
In a hot-oil high flash point case, clustering and PowerBI's linear key-influencer analysis both pointed to blanketing-gas temperature as the driver, while the best tree model (AdaBoost) ranked it only 11th. Switching from fuel-gas to nitrogen blanketing gave a lower and steadier flash point
Time-lagging improved the tree and linear models, but choosing the right lag remained hard, so the authors recommend combining several methods with input from domain experts

Approach

Built a dataset of past shutdown and optimization cases with root causes confirmed through interviews with the operations team, merged with historian sensor data sampled every 15 seconds or 1 minute
Compared Granger causality tests, hierarchical clustering, tree-model variable importance (via AutoML cross-validation) and linear-model coefficient ranking, back-testing each against the known causes
Feature engineering: missing-value filling, min-max scaling, inverted valve-opening signals and time-lagged features
Built a Python-backed UI that filters tags by system and equipment type, ranks Granger p-values against the first-out tag, and exports the dendrogram as a searchable indented tree in HTML

The problem

Root-cause analysis of unplanned shutdowns and process upsets on offshore platforms relies on slow, manual review of thousands of sensor tags and is not always data-driven.

Tools & methods

PythonGranger causality testhierarchical clustering (dendrogram)AutoML cross-validationAdaBoost regressorlinear regressionPower BI Key Influencersmin-max scaling

Summary written for this site from the published paper. For figures, data and the full method, read the paper on OnePetro.