Fault Detection via Autoencoder Latent Space Differences Between Reference Model and the Plant Operation ⁎ ⁎ We are deeply grateful for the financial support provided by the Ontario Research Fund and the McMaster Advanced Control Consortium (MACC).

Abnormal plant operations are caused by disturbances, process measurement faults, or malfunctioning equipment. Steady-state or dynamic models of the process units are widely available. Since continuous process plants operate under closed-loop control and available plant data often covers a narrow operating window, the process model can generate normal operating data over a wider window to train an autoencoder to represent that data. For deployment in real-time, the plant model accepts process inputs from the plant and calculates outputs; one instance of the autoencoder accepts data from the plant, and the other accepts data from the model. The occurrence of a process fault leads to Differences in the latent space variables of the two instances of the autoencoder, which enables fault detection. Compared to a traditional PCA-based fault detection framework, an autoencoder-based framework can model nonlinear processes, which is not possible by using PCA or dynamic PCA.