Fault detection and distributed estimation with sensor networks

Abstract

A sensor network is a distributed system, consisting of plenty of embedded sensors that can be deployed over a large-scale physical environment. One of the major applications of sensor networks is to monitor the state of systems that are evolving in the sensing field. Thanks to the emergence of advancements in high-performance processors, nodes in a sensor network can not only collect measurements but coordinate to estimate the state of systems as well. This thesis proposes a monitoring architecture, where distributed state estimation and fault detection algorithms are implemented by every node in the sensor networks to track the system’s state while simultaneously detecting the faults occurred in either the monitored systems or the sensor networks. Several approaches for different monitoring tasks are presented in this thesis and classified into two main parts: distributed state estimation and fault detection algorithms in the monitoring architecture. In the first part, we present two distributed state estimation algorithms in the sensor networks for the monitoring of a system, which can be described by a centralized, decentralized, or distributed dynamic model. The first one is implemented over a sensor network, where the local estimator in each node consists of a filtering step – which uses a weighted combination of neighboring sensors information – and a model-based state predictor. The filtering weights and prediction parameters jointly minimize both the mean and variance of the prediction error in a Pareto optimization framework at each time step. Since each predictor uses the model of the whole system monitored by the sensor network, the algorithm over a sensor network can only monitor a centralized system or each subsystem of a decentralized system. For a distributed system, where several subsystems interact with each other, the second algorithm implemented over several sensor networks is introduced so that each sensor network can coordinate with neighboring networks to monitor the corresponding subsystem of the distributed system. The second part of the thesis is devoted to fault detection algorithms for process faults in monitored systems and sensor faults in sensor networks. The aforementioned estimation algorithm over a sensor network is applied to design process fault detection algorithm for a centralized or decentralized system. A residual is defined, and suitable stochastic thresholds are designed, allowing to set the parameters so to guarantee an upper bound of false alarms probability. For detecting sensor faults in the sensor networks, the centralized, decentralized, and distributed sensor fault detection schemes are proposed in a discrete-time framework. And the detection performance is compared by an industrial benchmark simulation in a continuous stirred tank heater (CSTH) pilot plant. Then a rigorous fault detectability and detection time interval analysis of the centralized sensor fault detection scheme is presented. The performance of proposed distributed estimation methods and effectiveness of presented fault detection methods are evaluated by extensive numerical and industrial benchmark simulations.