Soft Sensor Development Based on Deep Extended Variational Autoencoder with Just-in-time Learning

Soft sensor based on deep learning have shown excellent performance in predicting key variables of industrial processes. However, most methods lack an update mechanism, leading to a gradual decline in the predictive performance of the models over time. To address this issue, this paper proposes a soft sensor development method based on a deep extended variational autoencoder with just-in-Time Learning (JITL-DE-VAE), which includes two stages: offline training and online updating. In the offline stage, there is a significant accumulation of reconstruction errors in the multi-layer Variational Autoencoder (VAE), and a large amount of invalid information in the features extracted during the model’s pretraining phase. To overcome these challenges, we introduce the Deep Extended Variational Autoencoder (DE-VAE). In DE-VAE, the hidden features from the previous layer and the input are used as the input to the next layer during the pretraining phase, enhancing the utilization of features at each layer of the network. Additionally, quality variables are incorporated during the feature extraction process to filter out redundant information that is irrelevant to the quality variables. In the online updating phase, the impact of time-varying processes on model performance is addressed by introducing the concept of just-in-time learning (JITL). The model updates itself by searching for similar samples from the historical database based on the input samples. Finally, the effectiveness and superiority of the proposed JITL-DE-VAE are verified using industrial debutanizer column.