Bidirectional Self-Iterative Domain Adaptation for Industrial Virtual Sensor Modeling

Data-driven virtual sensors have been widely applied to estimating important yet difficult-to-measure quality-relevant variables in industrial processes. However, the time series collected in complex industrial processes exhibits great nonlinear and dynamic features due to varied operating conditions. Additionally, the distribution discrepancy between the training and testing data poses a great challenge for traditional virtual sensor approaches in making an accurate prediction. Existing virtual sensor approaches pay little attention to the combined challenge of distribution discrepancy and dynamic feature transfer, referred to as the dynamic domain adaptation challenge. In this work, we propose a novel Bidirectional Self-Iterative Domain Adaptation (BiSDA) method to maintain the underlying dynamic feature. To realize an efficient feature transfer and guarantee a stable prediction, the proposed BiSDA narrows the feature disputation gap via transfer component analysis (TCA). We demonstrate the superiority of the proposed method via a real-world industrial case.