Layer-Wise Decomposition-Based Dual Attention Regression for Industrial Soft Sensing

With the rapid development of intelligent industrial systems, large volumes of process data are continuously collected from interconnected sensors. Effectively utilizing this data for accurate and interpretable modeling is crucial, especially given the dynamic behaviors and nonlinear couplings common in industrial processes, such as long-term trends, periodic fluctuations, and irregular disturbances. To address these challenges, this paper proposes a novel Layer-wise Decomposition-based Dual Attention Regression (LDDAR) framework for dynamic soft sensing. The model employs a stacked autoencoder to extract hierarchical representations and applies seasonal-trend-dispersion (STD) decomposition at each layer to disentangle distinct temporal patterns. A dual attention module is then introduced, where pattern attention captures correlations between raw inputs and decomposed patterns, and temporal attention models time-dependent dynamics. This design enables adaptive feature fusion across both variable and temporal dimensions. Experiments on real-world industrial datasets demonstrate that LDDAR outperforms baseline and ablation models, providing an effective solution for soft sensing in complex, nonlinear industrial environments.