Adaptive process monitoring via online dictionary learning and its industrial application

For industrial processes, one common drawback of conventional process monitoring methods is that they would make an increasing number of false alarms in cases of various factors such as catalyst deactivation, seasonal fluctuation and so forth. To address this issue, the present work proposes an online dictionary learning method, which can fulfill the process monitoring and fault diagnosis task adaptively. The proposed method would incorporate currently available information to update the dictionary and control limit, instead of keeping a fixed monitoring model. The online dictionary learning method are more superior than conventional methods. Firstly, compared with some traditional offline methods based on small amounts of historical data, the proposed method can augment train data with online dictionary updating, thus it copes with time-varying processes well. Secondly, the proposed method enjoys a lower computational complexity than the offline learning method with mass data, which is appealing in the era of industrial big data. Thirdly, the proposed method performs more reliably than the existing recursive principal component analysis-based methods because it can resolve the anomaly of principal component or non-orthogonality of eigenvectors problem which was often confronted in the recursive principal component analysis-based methods. Finally, some experiments were designed and carried out to demonstrate the advantage of the online dictionary learning.