Consistent-Contrastive Network With Temporality-Awareness for Robust-to-Anomaly Industrial Soft Sensor

The semisupervised soft sensor has gradually become a more practical solution due to the difficulty in collecting labels for the industrial soft sensors. Currently, the commonly used manifold regularization assumes similar inputs result in similar outputs, but such similarity assumption fails when temporality exists. Moreover, the performance of the soft sensor is vulnerable to the anomaly, so anomaly is generally discarded before modeling. The latent information in anomaly has seldom been explored for modeling yet. To overcome the limitation of manifold regularization and explore the usability of anomaly, a novel semisupervised soft sensor, named consistent-contrastive network (CC-Net), is proposed to build a temporality-aware and robust-to-anomaly soft sensor. Specifically, CC-Net consists of two branches, namely label branch and feature branch, to map the input into label and feature, respectively. In the label branch, labeled samples are used in a supervised manner. For unlabeled samples, two regularization terms, consistency and temporal consistency, are designed to constrain the pseudo-labels consistent against noise and along time, respectively, which can adapt to temporality. In the feature branch, the idea of contrastive learning is applied to separate normal samples and anomalies. Such a design properly utilizes the distribution information in anomaly as a regularizer. With the designed regularization terms and the utilization of anomalies, CC-Net implements a temporality-aware and robust-to-anomaly soft sensor, which is demonstrated by real denitrification and desulfurization cases.