Domain Adaptation Soft Sensing with Parameter Transferring

Soft sensor modeling often encounters a distribution discrepancy problem, when working conditions or environmental factors change. Such problem leads to an insufficient number of training data samples for an accuracy regression model construction. In addition, a soft sensor constructed for a specific mode is unlikely to obtain reliable prediction results for other modes. This paper presents a new transfer learning-based soft sensor model to handle the domain adaptation issue with a transferring parameter, which is suitable for multi-mode processes with limited target training samples. The difference between the source and target domains is considered as a parameterized maximum mean discrepancy regularization term in the objective function, based on which a trade-off between minimizing the two domains' difference and maximizing the prediction performance on the target domain's testing samples can be realized. Furthermore, an alternating optimization algorithm is formulated to optimize the transferring parameter along with the output weights. The proposed method is expected to fully leverages the limited target samples and the related source ones simultaneously to construct an adaptive target soft sensor. Comparative studies with several popular soft sensing approaches are conducted to demonstrate the effectiveness and advantages of our approach.