Diagnostic monitoring of high-dimensional networked systems via a LASSO-BN formulation

Quality control of multivariate processes has been extensively studied in the past decades; however, fundamental challenges still remain due to the complexity and the decision-making challenges that require not only sensitive fault detection but also identification of the truly out-of-control variables. In existing approaches, fault detection and diagnosis are considered as two separate tasks. Recent developments have revealed that selective monitoring of the potentially out-of-control variables, identified by a variable selection procedure combined with the process monitoring method, could lead to promising performances. Following this line, we propose the diagnostic monitoring that takes an additional step on from the selective monitoring idea and directs the monitoring effort on the potentially out-of-control variables. The identification of the truly out-of-control variables can be achieved by integrating the process monitoring formulation with process cascade knowledge represented by a Bayesian Network. Computationally efficient algorithms are developed for solving the optimization formulation with connection to the Least Absolute Shrinkage and Selection Operator (LASSO) problem being identified. Both theoretical analysis and extensive experiments on a simulated data set and real-world applications are conducted that show the superior performance.