Structure-Guided Graphical Lasso for Process Monitoring

To improve the monitoring performance of Gaussian Graphical Lasso-based methods. This paper proposes a novel fault detection and isolation approach using the structure-guided graphical lasso (SGGL). Unlike the conventional graphical lasso techniques, the proposed method employs both log-likelihood terms to decompose the graph model as the skeleton part and the perturbation part. To obtain sequential accurate graphs, the graph learning process is guided by the reference graph using two regularization terms: the ℓ2,1-norm and the Frobenius norm. By imposing the ℓ2,1-norm on the difference between graph skeletons, SGGL is able to capture different groups of abnormal variables. Whilst sporadic fault information can be retained in the perturbation part by the Frobenius norm. Once the structure-guided graph is inferred by our developed ADMM algorithm, the fault detection and isolation can be conducted by quantifying the amount of skeleton change and perturbation degree. The better performance of SGGL is then illustrated by a practical glass melter process.