成果速递：IOM实验室提出一种在工业过程中整合深度学习与可解释性的软测量建模方法

Title: Integrating Deep Learning With Interpretability for Soft Sensors in Dynamic Industrial Process Monitoring

Authors: Xiaoyu Yao; Renzhi Gao; Xiangwei Kong; Hefu Liu;

Journal: IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

URL: https://doi.org/10.1109/TSMC.2026.3740240

abstract: 

Intelligent system monitoring that enables real-time estimation of industrial process states is essential for ensuring reliable operation and decision-making in complex industrial systems. However, although deep learning has substantially improved the modeling of nonlinear and dynamic industrial processes in developing soft sensors for key quality estimation, its closed-box nature limits model interpretability and system transparency. To address this issue, this study proposes a deep-interpretable hybrid spatiotemporal fusion network (DIH-SFN) for intelligent monitoring of dynamic industrial systems. DIH-SFN integrates interpretable variable–response modeling with spatially guided temporal learning in a unified framework. In addition to spatiotemporal feature extraction and fusion for accurate prediction, DIH-SFN provides a novel interpretability solution through response curves of the predicted system states. These response curves reveal the marginal effects of process variables on system states, thereby supporting operators in identifying influential factors and meaningful operating regions. DIH-SFN is evaluated on three representative industrial system cases. Experimental results demonstrate that DIH-SFN outperforms existing state-of-the-art methods in prediction performance and provides more faithful explanations. Exploratory analyses show how DIH-SFN provides intuitive and interpretable insights into system operation to support operational optimization and enhance system reliability.