Novel control-aware fault detection approach for non-stationary processes via deep learning-based dynamic surrogate modeling

Abstract

The use of surrogate models for forecasting dynamic behaviors of processes is a promising approach for optimizing process operation and control. This study aims to utilize the powerful prediction capabilities of deep learning-based dynamic surrogate models (DSMs) for fault detection in non-stationary processes, taking into account the impact of control actions on faults. A novel control-aware fault detection approach, utilizing data-driven dynamic surrogate modeling of closed-loop controlled processes, is proposed to detect potential faults under various control actions and operating conditions. DSMs for both one-step and multi-step ahead prediction are developed and combined to detect various types of faults, based on predictions and residuals while considering control effects. High-fidelity dynamic simulations are used to build first principles-based closed-loop process models and generate data under various operating conditions to train the DSMs. Deep learning methods such as long short-term memory and gated recurrent units are employed for dynamic surrogate modeling, and their performance is compared to select the optimal method. The proposed approach is demonstrated through two case studies using a continuous stirred tank reactor and a distillation column with an advanced control structure. The results consistently demonstrate that the approach can achieve both accurate fault detection and reliable false alarm avoidance under different dynamic operating conditions and faulty situations with varying severities and fault types.