Anomaly Detection on Compressor Application: Detailed Evaluation of Statistical and Machine Learning Techniques

Unmanned remote compressors on wellhead platforms trip often, and each unplanned shutdown costs production and needs a boat trip to restart. The team sorted five years of shutdown events, tested supervised, regression, rule-based and unsupervised anomaly-detection methods against them, and deployed a hybrid of unsupervised anomaly scores and expert rules. Back-testing suggests most avoidable shutdowns can be caught early or diagnosed in advance, so crews arrive with the right fix.

Key results

Back-testing indicates about 70% of unplanned controllable shutdowns could be mitigated, either by early warning or by pointing to the likely root cause
The earlier bidirectional LSTM classifier reached an 85.28% F1 score on training data but gave many false positives on live data and was abandoned
Supervised methods only worked for common failure types with many labelled examples, while unsupervised anomaly scores gave the most encouraging results
In an example case, the anomaly score and the top-error sensor (fuel gas pressure) flagged a bad fuel supply several times before the shutdown

Approach

Grouped thousands of shutdown events from five years of records into planned, unplanned-controllable and uncontrollable cases, and focused only on unplanned-controllable ones
Tested data resampling and signal smoothing, which strongly affected results, along with data size and processing time
Compared four approaches: supervised shutdown classification (including an earlier bidirectional LSTM), regression on run-time drivers, rules and z-scores from subject matter experts, and unsupervised anomaly detection
Ranked unsupervised models by synthetic AUC, back-tested anomaly scores against real shutdown histories, and set thresholds for high precision to avoid flooding operators with alarms
Deployed a hybrid of ML anomaly scores and rules, with time-series visualization plus an SQL-backed table of recommendations and their acknowledgements

The problem

Frequent unplanned shutdowns of unmanned remote compressors cause lost production, and existing prediction attempts either did not generalize or raised too many false alarms.

Tools & methods

unsupervised anomaly detectiondimensionality-reduction anomaly modelbidirectional LSTMz-score / rule-based logicsynthetic AUCback-testingSQL databasetime-series visualization

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