Plug-and-Play Fault Detection and Isolation for Large-Scale Nonlinear Systems With Stochastic Uncertainties

Abstract

This paper proposes a novel scalable model-based Fault Detection and Isolation approach for the monitoring of nonlinear Large-Scale Systems, consisting of a network of interconnected subsystems. The fault diagnosis architecture is designed to automatically manage the possible plug-in of novel subsystems and unplugging of existing ones. The reconfiguration procedure involves only local operations and communication with neighboring subsystems, thus yielding a distributed and scalable architecture. In particular, the proposed fault diagnosis methodology allows the unplugging of faulty subsystems in order to possibly avoid the propagation of faults in the interconnected Large-Scale System. Measurement and process uncertainties are characterized in a probabilistic way leading to the computation, at each time-step, of stochastic time-varying detection thresholds with guaranteed false-alarms probability levels. To achieve this goal, we develop a distributed state estimation scheme, using a consensus-like approach for the estimation of variables shared among more than one subsystem; the time-varying consensus weights are designed to allow plug-in and unplugging operations and to minimize the variance of the uncertainty of the fault diagnosis thresholds. Convergence results of the distributed estimation scheme are provided. A novel fault isolation method is then proposed, based on a Generalized Observer Scheme and providing guaranteed error probabilities of the fault exclusion task. Detectability and isolability conditions are provided. Simulation results on a power network model comprising 15 generation areas show the effectiveness of the proposed methodology.

Detection and Classification of Recessive Weakness in Superbuck Converter Based on WPD-PCA and Probabilistic Neural Network

“…Besides, the process of generating control command signals produce modeling uncertainties and other external disturbances [34]. There also present uncertainties in measurement and processes which may lead to false-alarm situations [35]. The parameter identification errors influence the data-driven fault detection design as well [36].…”

Distributed Fault Detection for Interconnected Large-Scale Systems: A Scalable Plug & Play Approach

“…On the other hand, most of the mentioned methods consider deterministic bounds for noises and uncertainties in order to suitably determine detection thresholds. Instead, with the aim of achieving a less conservative detection performance, a stochastic characterization of the noises and the definition of time-varying bounds is considered (see also the preliminary works [3], [41]). Moreover, instead of using a sensor network to monitor a system characterized by stochastic uncertainties, where each sensor takes noisy measurements of the entire state [41], in this paper only a part of the state is considered by each local estimation and detection unit thus significantly broadening the applicability of the proposed approach.…”

“…In this respect, an important feature of the proposed methodology is the possibility of unplugging a faulty subsystems in order to avoid or reduce the propagation of faults in the interconnected system, and the possible plug-in of the disconnected subsystem (once the issue has been solved), without the need of a global redesign of the estimators but only resorting to local operations. Compared with [3], in this paper the knowledge of the mean and the variance of the coupling uncertainty is not assumed to be known and the computation of a bound for the influence on the uncertainty of the neighboring estimates is presented. Furthermore, the assumption used in [3] that the state is fully measurable is here removed.…”

Model-Based Fault Detection and Estimation for Linear Time Invariant and Piecewise Affine Systems by Using Quadratic Boundedness

“…Therefore, fault diagnosis is a research field that has been in the front end of the technological evolution for a few decades and has attracted the attention from the research and industrial community, as testified by many important survey papers and books (see [1]- [3] just as examples). Recent effort has been directed at investigating solutions for monitoring distributed, large-scale and interconnected systems [4]- [11]. When dealing with model-based approaches [2], due to the presence of uncertainties, one of the main issues is the definition of thresholds for some residual signals defined to be sensitive to the presence of faults.…”

“…When dealing with model-based approaches [2], due to the presence of uncertainties, one of the main issues is the definition of thresholds for some residual signals defined to be sensitive to the presence of faults. Different solutions have been proposed, either considering deterministic bounds on the uncertainties so to guarantee the absence of false alarms [12], or a stochastic characterization of noises and disturbances in order to set bounds on the allowed falsealarms rate [11]. A major challenge is represented by the fact that, in order to be robust to the noises, thresholds are often conservative, thus leading to scenarios where the uncertainties may hide the presence of faults.…”