Adaptive Transfer Learning of Cross-Spatiotemporal Canonical Correlation Analysis for Plant-Wide Process Monitoring

Abstract

Multivariate statistical methods have gained significant popularity in past decades. However, process dynamics and insufficient training data usually result in degradation or even failure of a trained model. To deal with these problems, this paper proposes a novel process monitoring method, called cross-spatiotemporal adaptive boosting transfer learning (CS-AdBoostTrLM). Different from the standard methods, CS-AdBoostTrLM has the following advantages: first, source domain (SD) data, which are discarded by the factory, can be re-enabled to alleviate the issue of insufficient training data. Second, cross-spatiotemporal canonical correlation analysis is proposed to achieve the domain adaptation between the SD data and target domain data, so as to overcome the negative transfer. Third, the particle swarm optimization algorithm is used to optimize the local detection model, in such a way that the integrated detection model can converge to the optimality globally. Finally, the data from the wastewater treatment plant and chemical plant are analyzed to demonstrate the effectiveness of the proposed method.