LSSW-SDAE: A Local Sparse Spatiotemporal Window-Driven Stacked Denoising Autoencoder for Industrial Process Monitoring

Industrial process monitoring via soft sensors remains challenging due to time delays, strong disturbances, and time-varying characteristics. Therefore, a local sparse spatiotemporal window-driven stacked denoising autoencoder (LSSW-SDAE) soft sensor model is proposed in this paper. The testing data processed by time difference are initially projected into the Gaussian space, and the subset is retrieved. The least absolute shrinkage and selection operator is simultaneously applied to construct the local sparse spatiotemporal window (LSSW), which reduces the impact of noise introduced by irrelevant information. Subsequently, the stacked denoising autoencoder (SDAE) is deployed to extract latent features from the LSSW for the prediction of key quality. With the Kullback–Leibler divergence-based update rule, the model can rapidly adapt to dynamic changes. Finally, the LSSW-SDAE is validated by two industrial process cases. In the Debutanizer column process, the LSSW-SDAE achieves root mean square error (RMSE) and coefficient of determination (R2) of 0.0153 and 0.9923 in monitoring butane concentration, respectively. In a simulated wastewater treatment plant, the LSSW-SDAE achieves RMSE and R2 of 0.0444 mg/L and 0.9970 in monitoring effluent five-day biochemical oxygen demand, respectively.