Multiple structured latent double dictionary pair learning for cross-domain industrial process monitoring

Abstract

Process data collected from real-world industrial operating environments have different distributions and lack real-time labeled samples, which causes the performance of process monitoring to decline. In this paper, a multiple structured latent double dictionary pair learning (MSLdDPL) method for cross-domain industrial process monitoring is proposed to address the issue of inconsistent distribution between acquired offline training set and online testing data. And it monitors the target domain with insufficient label information of data by learning the rich data knowledge in the source domain. The latent projective double reconstruction strategy is designed to integrate transfer learning and latent space mining into a unified framework, maximally bridging the distribution divergence between both different domains. In MSLdDPL, a multiple structured regularization is proposed to ensure the discrimination of coding coefficients and significant features. Meanwhile, an adaptive locality-preserved function is developed to jointly obtain the block-diagonal locality weights, so as to enhance the locality of coding coefficients and significant features in latent space. Then, the robust dictionary pair and a projection matrix are obtained for monitoring, further identifying abnormal sources with a reconstruction-based anomaly isolation technology. Extensive experiments verify that the proposed method performs better in the property and application of cross-domain process monitoring.