Design of sensor networks for simultaneous faults resolution and detection

RODRIGUEZ, LEANDRO P.; CEDEÑO, MARCO V.; SANCHEZ, MABEL C.

Atlanta

Congreso; 2014 AIChE Annual Meeting; 2014

Eventhough a great number of works in the literature devoted to the design of sensor networks (SNs) have focused on monitoring the normal process operation, the optimal location of instruments to effectively diagnose plant faults has central importance for safety, environmental protection and process economy. During the last decade, different sensor network design (SND) approaches for fault diagnosis were presented. Many of them introduced requirements on fault observability (O) and resolution (R). In this sense, Bagajewicz et al. (2004) developed a MILP formulation to obtain minimum-cost SNs subject to fault O and simple/multiple fault R, which was stated using matrix algebra concepts. Later on, Bhushan et al. (2008) addressed the robustness of the network to uncertainties/errors in the underlying failure cause-effect model and probability data by modifying a previous design formulation (Bhushan and Rengaswamy, 2002), which was based on quantitative information about the occurrence of faults and sensor failure probabilities. More distributed SNs were preferred to incorporate robustness to modeling errors. Regarding the treatment of inaccurate data, the main idea was to ensure that the constraints involving uncertain data were far from being active at the optimal solution. Recently, Rodriguez et al. (2013) proposed a new strategy to design an optimal SN able to resolve a set of key faults even under the presence of failed sensors or outliers. The procedure deals with key failures isolation when sensor malfunctions occur from a structural point of view. With this purpose key faults Resolution Degree constraints are incorporated into the minimum-cost design problem, which involves a linear objective function and a set of linear inequality restrictions. Other contributions analyzed the design of SNs devoted to detect and identify a set of process faults by means of particular monitoring strategies. These works considered the effect of measurement uncertainties, but took no account on the capability of the sensor structure to observe and resolve faults under the presence of sensor failures. In this sense, Musulin et al. (2004) and Gerkens and Heyen (2008) proposed methodologies to design SNs for processes monitored using Principal Component Analysis and model-equation residuals based techniques, respectively. The resulting optimization problems were solved using Genetic Algorithms. If the strategy used for process monitoring has been selected, it is advisable to design a SN which is able to simultaneously resolve process faults and detect them when that strategy is running on line. With this purpose, a new SND problem is formulated in this work that minimizes the instrumentation cost subject to the resolution of all faults and the detection capability of the monitoring technique. Fault detection based on Principal Component Analysis is selected, and the detection capability of this method is considered using the Fault Size Penalization model introduced by Musulin et al. (2004) which applies the concept of the Minimum Critical Fault Magnitudes (Wang et al., 2002). Furthermore key faults resolution and detectability should be attained even under the presence of sensor malfunctions or outliers. To tackle this issue, the optimization problem is appropriately reformulated. In this work new formulations of the sensor network design problem for fault diagnosis are presented, and their advantages over other existing procedures are discussed. Application results for a chemical plant with recycle are provided.