EAAI Journal 2025 Journal Article
FA-SconvAE-LSTM: Feature-Aligned Stacked Convolutional Autoencoder with Long Short-Term Memory Network for Soft Sensor Modeling
- Ping Wu
- Zengdi Miao
- Ke Wang
- Jinfeng Gao
- Xujie Zhang
- Siwei Lou
- Chunjie Yang
The advancement of soft sensor technology has enabled the real-time estimation of critical parameters in complex industrial processes, where direct measurement through hardware sensors is often infeasible. Industrial process data typically exhibit both spatial correlations and temporal dependencies, necessitating sophisticated modeling approaches to capture these characteristics effectively. In this study, a spatio-temporal model, termed the feature-aligned stacked convolutional autoencoder with long short-term memory, is proposed to develop soft sensors for nonlinear dynamic industrial processes. The proposed model begins with the systematic training of a stacked convolutional autoencoder using a layer-by-layer pre-training technique. This approach facilitates the extraction of high-level spatial feature representations from the process variables. To address the issue of feature misalignment in the spatial features extracted by the stacked convolutional autoencoder, a feature alignment strategy is implemented, ensuring that the extracted spatial features are properly aligned. Subsequently, the aligned spatial features are fed into a long short-term memory network to capture temporal dependencies, with quality variables serving as the output for soft sensor development. The effectiveness and superiority of the proposed method are demonstrated through experiments conducted on two industrial processes: the sulfur recovery unit and the multiphase flow process. Comparative analyses with other state-of-the-art methods reveal that the proposed model achieves the highest performance, with R 2 values of 0. 86222 for the sulfur recovery unit and 0. 94307 for the multiphase flow process, outperforming all compared methods.