Improved PCA-based Fault Isolation using Sparse Group Lasso

In industrial process control, data-driven fault detection and isolation methods have developed rapidly due to the easy availability of large amount of data. In fault isolation, principal component analysis(PCA) based contribution plot is a standard tool. The problem of PCA based contribution plot is that they are affected by the so called smearing effect. In fact, industrial process variables can be classified into groups according to their correlation or process structure, hence it is straightforward to consider the group-wise fault isolation problem. This paper introduces the sparse group Lasso as a regularization method to improve the fault isolation ability of PCA based contribution plot. The sparse group Lasso term considers both group-wise sparsity and within-group sparsity. Hence more accurate diagnosis can be obtained. In order to solve the optimization problem of sparse group Lasso, an efficient algorithm based on ADMM(Alternating Direction of Method of Multipliers) is proposed. Application study to the Tennessee Eastman(TE) process shows that the proposed method can better isolate faulty variables than competitive methods.