Fault Detection and Isolation Applied to a Ship Propulsion Benchmark Youmin Zhang ∗ N. Eva Wu ∗∗ Bin Jiang ∗∗∗ ∗ Dept. of Mechanical and Industrial Engineering, Concordia University, Montreal, Quebec H3G 1M8, Canada (email: ymzhang@encs.concordia.ca) ∗∗ Dept. of Electrical Engineering, State Univ. of New York at Binghamton, Binghamton, NY 13902, USA (email: evawu@binghamton.edu) ∗∗∗ College of Automation Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China (email: binjiang@nuaa.edu.cn) Abstract: Abstract: This paper describes a fault detection and isolation (FDI) scheme performed on the benchmark problem of a ship propulsion system. The model used for the ship propulsion system is nonlinear, for which two types of additive sensor faults, an additive incipient fault, and a multiplicative parametric fault are simulated. The estimation of the fault severity is accomplished by using an adaptive two-stage extended Kalman filter. A set of statistical detection variables is formed from the residuals of the bias and measurement estimates of the filter. These variables are then used in a threshold based hypothesis test to declare the occurrence of a fault and through a binary logic filter to identify the fault type. The simulation results showed that the developed fault detection and isolation scheme fulfilled some of the benchmark requirements reasonably well in the face of some prescribed perturbations in the model and disturbances of external signals. Keywords: Keywords: Fault detection and isolation, nonlinear adaptive two-stage Kalman filter, ship propulsion benchmark. 1. INTRODUCTION The ship propulsion benchmark (Izadi-Zamanabadi and Blanke, 1999) is a complicated and realistic test bed for fault diagnosis and fault-tolerant control. It presents a realistic simulation environment of a ship propulsion sys- tem under faults, disturbances and random noises. In the benchmark, not only additive abrupt faults are present, a multiplicative fault and an incipient fault are considered as well. Since the publication of the benchmark, several approaches have been developed to solve the benchmark problem from fault diagnosis and/or fault-tolerant control aspects. Blanke et al. (1998) designed an adaptive, non- linear observer for fault estimation of engine related faults in shaft speed sensor and engine gain. Research results from several groups were summarized in a book chapter (Izadi-Zamanabadi et al., 2000). Edwards and Spurgeon (2000) extended their results by using a dedicated sliding mode observer for fault detection on the benchmark. A sensor fault masking scheme of the benchmark is proposed in (Wu et al., 2006). A fault-tolerant control scheme has been developed in (Bonivento et al., 2003). In this paper, all faults entering the system are manipu- lated into additive random biases to the nonlinear ship propulsion system model. This enables us to approach fault diagnosis as a model based bias estimation problem. In this regard the utility of an earlier solution to a linear problem obtained by the authors (Wu et al., 2000) is expanded to solve the nonlinear benchmark problem. This estimator is further developed into a two-stage adaptive extended Kalman filter (EKF). Beyond monitoring the operation of a system, on-line diagnosis also provides basis for decisions on fault accommodation. Therefore, it is important that diagnostic outcomes are tested for their statistical significance before a drastic action is taken, such as the reconfiguration of a control law. These tests are part of an FDI (fault detection and isolation) process. In this work, a set of statistically significant detection variables is constructed out of a set of selected residuals of the two-stage adaptive EKF. Some of the residuals represent the estimated fault magnitudes. Others are measurement residuals from which the interested sensor faults are di- rectly observed. A fault occurrence is reported whenever a detection variable exceeds a set of threshold levels. The selection of the thresholds for these detection variables is dictated by the attempt to achieve not only a low probability of missed detection and a low probability of false alarm, but also a low probability of false isolation as well. Immediately following the declaration of a fault occurrence is a least squares linear regression analysis that confirms whether the detected fault is of an incipient or of an abrupt type, which is then followed by a fault isolation logic. The isolation logic depends on a highly compressed knowledge base obtained via extensive off-line analysis. The paper is organized as follows. In Section 2, the ship propulsion benchmark model and fault scenario are briefly described. An adaptive nonlinear two-stage Kalman filter Proceedings of the 17th World Congress The International Federation of Automatic Control Seoul, Korea, July 6-11, 2008 978-1-1234-7890-2/08/$20.00 © 2008 IFAC 1908 10.3182/20080706-5-KR-1001.4016