Robust soft sensor modelling with a contractive enhanced adaptive variational autoencoder

Abstract Reliable industrial soft sensing is challenged by nonlinear process behaviour and abnormal sensor measurements. To address these problems, this work proposes a contractive enhanced adaptive variational autoencoder (CEA‐VAE) for quality‐variable prediction. The model combines bounded statistical features, residual‐adaptive latent regularization, and local sensitivity control. The physics‐informed bounded boundary‐statistics enhancement (PI‐BBSE) path is designed to use sensor limits defined from plant information or expert knowledge. In the present experiments, the minimum and maximum values of each sensor in the clean training data were used as proxy bounds. A separate copy of each normalized sensor signal is clipped to these bounds to calculate minimum and maximum features, while the original normalized input is retained in the convolutional path. Residual‐adaptive Kullback–Leibler (KL) weighting adjusts latent regularization based on prediction residuals, and a Jacobian‐based contractive penalty reduces the sensitivity of the posterior mean to input changes. All components are trained jointly for input reconstruction and quality prediction. The model was evaluated in ten runs using data from a real industrial polyester process under clean conditions and 3%, 5%, and 8% point contamination. The average MSE of CEA‐VAE across the four conditions was 9.7% lower than that of the convolutional neural network VAE (CNN‐VAE). The ablation study showed that the PI‐BBSE path using proxy bounds provided the largest performance improvement.