Fault Detection and Analysis via Latent Space Differences Between the Plant and the Model Representing Normal Operation

Abnormal plant operations (faults) occur when equipment or an instrument no longer functions well. Detection of such faults is made more difficult by process control applications, which attempt to ameliorate the impact of faults by keeping the production on target. Closed loop controls cause the plant to operate in a narrow region, which in turn limits the validity of data-driven fault detection methods to such a limited scope. This work introduces a fault detection architecture that employs data generated from steady-state simulation models to build a latent space model (e.g., Principal Component Analysis, PCA) of the normal operation, thereby overcoming limitations inherent in the plant's historical data. In real-time, variables from the model and the plant are processed by identical copies of PCA. The pattern of the differences indicates the fault occurrence and gives insight into possible causes.