Generalized Cross-Domain Industrial Process Monitoring via Adaptive Discriminative Transfer Dictionary Pair Learning With Attribute Embedding

Abstract

Real industrial process data from various domains often exhibit divergent distributions, may occupy distinct feature spaces, and are occasionally unlabeled, which limits the effectiveness of conventional process monitoring methods. To address these challenges, we propose an adaptive discriminative transfer dictionary pair learning (ADTDPL) method with attribute embedding for generalized cross-domain industrial process monitoring. Specifically, this method aligns the feature spaces of source and target domains by the aligned transfer reconstruction, enabling the transfer of knowledge through a common synthetical dictionary. Concurrently, semantic attributes relevant to process knowledge are seamlessly fused into data information via attribute embedding, enhancing the transferability and interpretability of dictionary pairs. Considering the relative significance of marginal and conditional distributions, an adaptive distribution consistency function is designed to better reduce the distributional discrepancies. And the discriminative structure regularization is developed to ensure the discrimination of the dictionary pairs and their corresponding coding coefficients. Furthermore, in the absence of target domain labels, a novel selective pseudo-labeling strategy is advanced to adaptively update pseudo-labels. The superior performance of our method for cross-domain process monitoring is verified on the Tennessee Eastman platform and in practical aluminum electrolysis processes (AEPs).