Reliable anomaly detection with low false alarm & troubleshooting for a critical air-blower that feeds combustion air to major parts in the refinery.
A semi-supervised anomaly detection solution based on expert-annotations provides alerts in real-time for deviations from normal behavior. A cause for the deviation is also indicated.
The refinery has an air-blower asset, which has had 4 critical stoppages in last 7 years due to maintenance needs or a failure. Each of these stoppages would cost the refinery between € 1-10 M and hence it was regarded as one of the fore runners for advanced condition monitoring & anomaly detection. Given the complex surrounding of the blower, its multiple operating modes & other diverse modes, building a data-driven solution with low false alarms & reliable detection performance was non-trivial. Manual data monitoring & analysis proved to be time consuming. Traditional unsupervised anomaly detection methods & simple threshold-based approaches failed due to the complex nature of the process.
MultiViz was used by the experts to identify the multiple normal modes and failure modes. When unidentified patterns or anomalies are detected, a unique sensor ranking algorithm provided a list of causes for the abnormality. A prototype for anomaly detection with low false alarm was realized to monitor the asset continuously.
Anomaly detection with low false alarm for their most critical asset. Reduced data preparation from weeks to hours.