Stacked dynamic target regularization enhanced autoencoder for soft sensor in industrial processes

Stacked autoencoders (SAEs) have great potential in developing soft sensors due to their excellent feature extraction capabilities. However, the pre-training stage of SAE is unsupervised and some important information related to target variables may be discarded. Meanwhile, as the depth of the network increases, reconstruction errors continue to accumulate, resulting in incomplete feature representations of the original input. In addition, the dynamic nature of the data affects the predictive results of the model. To address these issues, the stacked dynamic target regularization enhanced autoencoder (SDTR-EAE) method is proposed, which adds the DTR and the original input information layer by layer to enhance the feature extraction. To adapt to the dynamic changes in data and extract target-related features, entropy weight grey relational analysis (EW-GRA) is used as the DTR term to constrain the weight matrix and suppress irrelevant features. To reduce the accumulation of information loss during the reconstruction, an information enhancement layer is introduced, where the original inputs and the information of the hidden layers of previous DTR-EAE units are added to the follow-up DTR-EAE unit. Finally, in the regression process, the DTR term is used again to fully utilize depth features for quality prediction and prevent overfitting. Experimental verifications using the debutanizer column and thermal power plant are conducted to validate the effectiveness of the proposed modelling method.