Data-driven soft sensors for multi-scale feature extraction in quality prediction of industrial processes

To address the challenges of strong variable coupling, multi-scale feature extraction, and poor robustness to missing data in industrial process quality prediction, this paper proposes a Dilated Multi-scale Convolutional Gated Network (DMCGN) for data-driven soft sensors. The DMCGN integrates dilated convolution layers (dilation rates r = 1, 2, 3) to expand the receptive field, multi-scale convolutional kernels (3 × 3, 5 × 5, 7 × 7) to capture local spatio-temporal features, and a gated aggregation module to adaptively suppress redundant information. Validations on two real industrial datasets (sulfur recovery unit, SRU; debutanizer column process, DCP) show that the proposed method outperforms state-of-the-art models: for SRU, the RMSE is 0.0095 and R² reaches 0.9731; for DCP, the RMSE is 0.0081 and R² is 0.9984. Additionally, the DMCGN maintains high prediction accuracy under 10%–30% data missing rates, providing a reliable solution for real-time quality monitoring in complex industrial processes.