A hybrid modeling approach integrating first-principles with data for nonlinear process monitoring

Fault monitoring plays a crucial role in modern industrial systems. Timely and accurate monitoring is essential to mitigate economic losses and protect life safety. Traditional data-driven methods, such as auto-encoder (AE) and variational autoencoder (VAE), are widely utilized in nonlinear process monitoring due to their simplicity and efficiency. However, the interpretability and generalization capabilities of data-driven models are constrained by the absence of physical knowledge. To address this limitation, we have developed a physics-informed variational auto-encoder (PI-VAE) method in this study. This method incorporates physical mechanism information into the model’s loss function through equality constraints. The results obtained from applying the PI-VAE method to a 1000MW ultra-supercritical coal-fired power plant demonstrate its exceptional capability in fault monitoring. By integrating reliable physical data with the VAE model, the model ensures enhanced stability and improved generalization.