Adversarial relationship graph learning soft sensor via negative information exclusion

Abstract

The development of soft sensors in process industries necessitates learning the dynamic variable relationships caused by physicochemical reactions, whilst avoiding noise interference that degrades prediction performance and explainability. To address this, an adversarial relationship graph learning soft sensor is proposed, comprising both relationship learning and prediction modules. Irrelevant and false variable relationships caused by noise are treated as negative information, quantified through mutual information loss. They are captured by alternately adversarial training the self-attention network and graph autoencoder. By excluding negative information, a suitable variable relationship graph is constructed. The graph convolutional network then mines information from the data and relationships for accurate prediction. Two practical cases verify the model’s physical consistency and demonstrate superior performance compared to several common models.