Key Variable Based Detection of Sensor Faults in a Power Plant Case Riku-Pekka Nikula*, Ville Laukkanen**, Esko Juuso*, Kauko Leiviskä* *Control Engineering Laboratory, P.O.Box 4300, FIN-90014 University of Oulu, Finland (e-mail: riku-pekka.nikula@oulu.fi) **Indmeas, Tietäjäntie 12, FIN-02130 Espoo, Finland Abstract: In this study, an approach for detection of sensor faults is presented. It is based on an identification principle, which takes into account linear relationships between process variables and a key variable. Simple linear regression models are created for variables with a strong correlation. Relative limits are defined for the variables using actual values and response values from the linear models. The limits are used to scale the variables and to monitor exceptionally high and low values. The proposed approach is tested with real data from a coal-fired power plant. To demonstrate the behaviour of the approach on a fault situation, simulated sensor faults are monitored. The results imply that the approach is applicable to cases with strong linear relationships. The approach assists in revealing inherent linear relationships in a process to support the development of sensor validation approaches and to create limits for monitoring. Keywords: fault detection, outlier detection, process monitoring, sensor validation, variable identification 1. INTRODUCTION Good quality of sensor data collected from an industrial process is an essential factor to reliable operation of the process. Sensor faults are almost inevitable even with the most advanced design of instruments especially in harsh industrial environments. A fault is to be understood as a non- permitted deviation of a characteristic property which leads to inability to fulfill the intended purpose (Isermann, 1984). Roughly speaking, sensor faults can be categorized as “hard faults” with abrupt changes and as “soft faults”, which are slowly developing failures (Goebel and Yan, 2008), (Frank, 1990). Simulation of four types of sensor faults, including bias, precision degradation, drifting and complete failure are presented in (Qin and Li, 1999). In real environments, sensor noise, deterioration, system dynamics, and changing conditions bring challenges to detection of sensor faults. It is important to distinguish sensor faults from process changes, because process changes can interfere with sensor fault detection. Process changes can be categorized into: (i) unmeasured, normal process changes; (ii) slow process degradation; and (iii) abnormal process changes (Qin and Li, 1999). The process is considered to be stable but it has changes in the operational state in this study. 1 The detection of sensor values that do not conform to the assumed behaviour is related to outlier detection. Outliers can be referred as statistical anomalies. The term outlier describes a type of data point that is not representative of the sample being considered. Typically, outlier detection methods have the assumption of identically and independently distributed (i.i.d.) data. In that case, the sample mean and variance give 1 The research was funded by the Finnish Funding Agency for Technology and Innovation (TEKES) through the Measurement, Monitoring and Environmental Assessment (MMEA) programme. good estimation for data location and spread. The popular “ rule” is based on the idea of detecting the observations lying further than three standard deviations from the mean (Oakland and Followell, 1990). Hampel Identifier replaces the outlier-sensitive mean with the median and the standard deviation with the median absolute deviation from the median (Pearson, 2002). Chiang et al. (2003) compare many outlier detection methods on data from the Tennessee Eastman Process simulator in their study. σ 3 A physical system that involves several sensors monitoring the operating state has usually certain relationships between the sensor values. The expected value of one sensor might be obtained from the remaining sensor values involved in the same relationship. The verification of sensor values with other information is called sensor validation, which is often based on redundancy of several sensors. Physical redundancy involves redundant sensors measuring the same parameter of the system (Goebel and Yan, 2008). Analytical redundancy, on the other hand, utilizes a functional relationship between the sensors that are of different types or positioned at different locations (Walker and Wyatt-Mair, 1995). In addition, categorization into spatial redundancy, temporal redundancy and knowledge-based redundancy is presented in literature (Frank, 1990), (Lee, 1994). Some sensor validation methods produce sensor health information from a single sensor (Ma et al., 1999), (Näsi et al., 2005). An industrial plant has probably thousands of sensors monitoring numerous targets. When new methods for monitoring the plant are put into operation, it is time consuming to find and define parameters for every object individually. In that case, an approach which takes several process variables and uses the inherent redundancies in the process to automatically identify certain characteristics and parameters is useful. The identified parameters can be