Creating an Anomaly Detection Model to Prevent Plant Shutdown

Situation

‍

Some anomalies are much more difficult to detect and predict. In certain situations, for example, an anomaly might occur even when a process appears to be functioning normally. For these more involved cases, data scientists can apply a machine learning technique known as an anomaly detection model. Anomaly detection models can identify rare items, events, or observations that deviate significantly from the rest of the time-series dataset and that are known to be outside normal operating behavior. They can be used to set up monitors and alerts when special deviations occur.

Problem

‍

At this chemical plant, fluid from one process occasionally was leaking into a compressor and damaging it. This led to a full-plant shutdown and significant loss of production. Engineers already had determined that vibrations were causing the leak, but they had no way to monitor for this problem to occur. Furthermore, operational experts could not intervene in time to prevent damage and avoid the costly shutdown.

‍

‍

Solution

‍

• Attach vibration sensors to the compressor
• Load time-series data into MLHub
• Train the TrendMiner Anomaly Model with different types of vibrations
• Create a new machine learning model tag based on the trained model
• Activate a monitor on the newly created tag to detect for irregular vibrations
• Set alerts to allow key stakeholders time to intervene before it was too late

‍

‍

Results

‍

• The soft sensors created through machine learning models detected an anomaly at the compressor that indicated there was a leak, even though other process parameters appeared to be normal
• When the anomaly was detected, process experts received an alert with enough time to make changes before the compressor was damaged
• Engineers were able to avoid a complete plant shutdown

‍