Unplanned shutdowns due to failures in chemical production facilities cause large financial losses. Additionally, 'soft failures' caused by deviation of flow rates from intended setpoints can cause degradation in product yield or even total loss of production output, not to mention potentially posing safety issues to personnel and equipment.

 

Machine learning has proved to be effective at predicting early warning of valve failures up to one month ahead of time, even by using only normal operation data for training to perform anomaly detection - in this case NO labelled examples of valve faults is required. Flow rate monitoring can be implemented in a number of ways requiring some supervised signal such as yield or actual flow rate. (See paper from Shell to right for examples of this application).