Mechanism–data–text multimodal fusion for remaining useful life prediction of selective catalytic reduction catalysts in utility boilers

Abstract Selective catalytic reduction (SCR) catalysts play a pivotal role in controlling NO x emissions from coal‐fired utility boilers. However, their performance inevitably degrades under complex operating conditions, making accurate prediction of remaining useful life (RUL) essential for effective maintenance and emission compliance. Conventional life prediction approaches, such as empirical models, often fail to capture the coupled effects of multiple degradation factors and overlook valuable maintenance‐related textual information. This study proposes a mechanism‐data‐text multimodal fusion framework based on an improved gating‐and‐attention iTransformer (GA‐iTransformer) model for catalyst health monitoring. First, a cross‐modal data preprocessing strategy integrates an enhanced spectral residual algorithm, dynamic time warping, and BERT‐based text embedding to align high‐quality time‐series data with textual information. Second, a mechanism‐driven kinetic model quantifies catalyst activity through a novel state of health (SOH) index. Finally, GA‐iTransformer employs dynamic gating and cross‐attention mechanisms for hierarchical fusion of time‐series and text features, enabling accurate multi‐step SOH forecasting and RUL estimation. Validation on datasets from four SCR reactors across two 600 MW utility boilers demonstrates that the proposed method outperforms five baseline models, achieving average MAE, RMSE, and MAPE of 1.789%, 2.545%, and 2.306%, respectively, with RUL prediction errors within approximately 2 days. The proposed approach offers a robust tool for predictive maintenance of SCR systems, reducing replacement costs and ensuring emission compliance.