The challenge
Control valves manipulate material flow throughout a refinery. Unplanned valve failures cause costly production downtime, while unnecessary preventive maintenance wastes resources. The refinery had no data-driven early fault detection capability to distinguish normal from faulty valve behavior.
At a glance
Method4-layer GRU recurrent neural network, reconstruction-error anomaly detection
DataPI / DCS historian time series
ValidationMaintenance engineer review of labeled periods
What we did
SWICON built a GRU-based recurrent neural network anomaly-detection model on control-valve time series from the plant historian: operational parameter, control mode, set point and process variable per valve, across flow, level and pressure control. Shutdown periods were filtered out and good/bad periods were labeled and validated with the refinery's maintenance engineers.
The result
The proof of concept validated that control-valve fault detection is achievable with machine learning from the available sensor data. Detectable abnormal behavior patterns were confirmed across six valve types, and maintenance engineers confirmed irregular operational-parameter trends prior to shutdown, meaning the signal precedes the failure. A conceptual plan for a group-wide rollout was delivered.
