A Soft Sensing Technique Based on Weighted Variational Auto-Encoder and its Application in Industrial Processes

The steady development and popularity of data acquisition systems in the modern industry have led to a trend toward the application of soft sensors. Feature representation converting raw data into a more representative and easier-to-process feature matrix is an important step in building soft sensors. The powerful feature extraction capability of deep learning makes it a popular choice for modeling. However, the unsupervised training process of some typical deep learning methods such as Variational Autoencoder (VAE) makes the extracted features contain information that is not useful for modeling predictions. In order to learn how the extracted features are related to the target variables, this article presents supervised soft sensing based on the weighted Maximal Information Coefficient (MIC) for VAE (wMIC- VAE). wMIC-V AE computes the correlation between the advanced feature extracted by VAE and the output variable using weighted MIC. Thus, the extracted feature is supervised and weighted, making the model training process more purposeful and rational. The effectiveness and superiority of the proposed method were demonstrated by applying it to the Pure terephthalic acid (PTA) dataset.