A Transformer-based fault detection method built on real-time data from microreactors

Abstract

Microreactor technology has garnered significant attention for its efficiency and precision in chemical production. However, research on data analysis within microreactors remains limited. Fault detection and diagnosis is crucial for ensuring safety in the chemical industry. Although many fault detection algorithms based on reconstruction deep learning methods have been proposed and tested using simulated data, these simulations often fail to account for disturbances that may occur in real chemical production processes. To address this gap, this paper presents a microreactor system capable of real-time data monitoring and proposes a Transformer-based hybrid model that incorporates cross-time and cross-variable attention mechanisms. The performance of this model is evaluated using both normal and abnormal data from water test and an oxidation process in the microreactor. Compared to traditional reconstruction-based methods, our model demonstrates a higher fault detection rate when applied to real-world data containing disturbances, highlighting its significant potential for improving process safety.