Copyright © IFAC System Identification, Kitakyushu, fukuoka, Japan, 1997 Failure Detection and Identification using a Nonlinear Interactive Multiple Model (IMM) Filtering Approach with Aerospace Applications * Raman K. Mehra Constantino Rago Sanjeev Seereeram Scientific Systems Company, Inc. (SSCI) 500 West Cummings Park, Suite 3000 Woburn, MA, 01801 (USA) e-mail: rkm@ssci.com Abstract: In this paper, we propose a novel approach for Failure Detection and Iden- tification (FDI) in non linear systems based on the Interactive Multiple Model Extended Kalman Filter (EKF) approach. In the non linear-system FDI application, the main idea consists of representing each failure mode by a model and combining the outputs of EKF's based on different models in a near-optimal way. This IMM- FDI filter provides not only failure detection and identification but also a near-optimal estimate of the system state (even during a failure). The approach has been applied successfully to a problem of spacecraft autonomy for the detection and identification of sensor (gyro, star tracker) and actuator failures. The results of this application show that IMM-EKF detects and identifies failures much more rapidly and reliably than the multi-hypothesis EKF. Furthermore, it handles satisfactorily both permanent and transient failures. Keywords: Failure detection and isolation, Spacecraft autonomy, Extended Kalman filter, Fault tolerant systems 1 INTRODUCTION Fast and precise Failure Detection and Identifi- cation (FDI) is becoming a crucial task as sys- tems grow in complexity and more emphasis is put on autonomous performance, see (Mehra and Peschon, 1971; Willsky, 1976; Basseville, 1988; Brown, 1993; Basseville and Nikiforov, 1993; Rauch, 1994; Isermann, 1994). In some appli- cations, like deep space exploration, autonomous FDI is not only desirable but also a necessity as the delay between the spacecraft and an earth sta- tion becomes too large to make human-decision as- sisted operation viable (Mehra et al. 1994, 1995; Pell et al., 1996). The spacecraft FDI problem is one of a vast class of problems where the model representing the dy- namics of the system can switch (unpredictably) from one model to another. Failures of compo- nents (gradual or sudden), statistical changes in noise parameters, variations in physical structure (damage to the spacecraft), etc. can all render standard control techniques unreliable. The use 'This research was supported through contract # NAS7-1383 from NASA Jet Propulsion Laboratory (JPL) and the support of Drs. Bayard and Hadaegh from JPL is gratefully acknowledged. 407 of a single model (a "compromise model") in these situations usually gives very poor estimates, and these estimates are incomplete in the sense that they lack information relevant to the model in ef- fect (this is particularly important in the failure detection case). In the IMM approach (see Blom and Bar-Shalom, 1988; Bar-Shalom and Li, 1993, 1995), this problem is solved by having M filters running in parallel (see Figure 1) and incorporat- ing a Hidden Markov Model switching mechanism among the models. For the FDI problem, each failure hypothesis corresponds to a different model within the IMM, it Le. each model is constructed to map a failure hypothesis into the dynamics of the system. A related approach based on Multiple Hypoth- esis EKF, or MH-EKF was presented in Mehra et ai, 1995, where several EKF's are run in paral- lel, and a hard-switching decision is made at each sampling time based on the innovation sequences and likelihood functions of each filter. A common problem encountered with this approach is the de- lay in detection due to build up of likelihood func- tion for active hypothesis and divergence of in- dividual EKF states for the inactive hypotheses. The IMM approach proposed here overcomes this Copyright © IFAC System Identification, Kitakyushu, fukuoka, Japan, 1997 Failure Detection and Identification using a Nonlinear Interactive Multiple Model (IMM) Filtering Approach with Aerospace Applications * Raman K. Mehra Constantino Rago Sanjeev Seereeram Scientific Systems Company, Inc. (SSCI) 500 West Cummings Park, Suite 3000 Woburn, MA, 01801 (USA) e-mail: rkm@ssci.com Abstract: In this paper, we propose a novel approach for Failure Detection and Iden- tification (FDI) in non linear systems based on the Interactive Multiple Model Extended Kalman Filter (EKF) approach. In the non linear-system FDI application, the main idea consists of representing each failure mode by a model and combining the outputs of EKF's based on different models in a near-optimal way. This IMM- FDI filter provides not only failure detection and identification but also a near-optimal estimate of the system state (even during a failure). The approach has been applied successfully to a problem of spacecraft autonomy for the detection and identification of sensor (gyro, star tracker) and actuator failures. The results of this application show that IMM-EKF detects and identifies failures much more rapidly and reliably than the multi-hypothesis EKF. Furthermore, it handles satisfactorily both permanent and transient failures. Keywords: Failure detection and isolation, Spacecraft autonomy, Extended Kalman filter, Fault tolerant systems 1 INTRODUCTION Fast and precise Failure Detection and Identifi- cation (FDI) is becoming a crucial task as sys- tems grow in complexity and more emphasis is put on autonomous performance, see (Mehra and Peschon, 1971; Willsky, 1976; Basseville, 1988; Brown, 1993; Basseville and Nikiforov, 1993; Rauch, 1994; Isermann, 1994). In some appli- cations, like deep space exploration, autonomous FDI is not only desirable but also a necessity as the delay between the spacecraft and an earth sta- tion becomes too large to make human-decision as- sisted operation viable (Mehra et al. 1994, 1995; Pell et al., 1996). The spacecraft FDI problem is one of a vast class of problems where the model representing the dy- namics of the system can switch (unpredictably) from one model to another. Failures of compo- nents (gradual or sudden), statistical changes in noise parameters, variations in physical structure (damage to the spacecraft), etc. can all render standard control techniques unreliable. The use 'This research was supported through contract # NAS7-1383 from NASA Jet Propulsion Laboratory (JPL) and the support of Drs. Bayard and Hadaegh from JPL is gratefully acknowledged. 407 of a single model (a "compromise model") in these situations usually gives very poor estimates, and these estimates are incomplete in the sense that they lack information relevant to the model in ef- fect (this is particularly important in the failure detection case). In the IMM approach (see Blom and Bar-Shalom, 1988; Bar-Shalom and Li, 1993, 1995), this problem is solved by having M filters running in parallel (see Figure 1) and incorporat- ing a Hidden Markov Model switching mechanism among the models. For the FDI problem, each failure hypothesis corresponds to a different model within the IMM, it Le. each model is constructed to map a failure hypothesis into the dynamics of the system. A related approach based on Multiple Hypoth- esis EKF, or MH-EKF was presented in Mehra et ai, 1995, where several EKF's are run in paral- lel, and a hard-switching decision is made at each sampling time based on the innovation sequences and likelihood functions of each filter. A common problem encountered with this approach is the de- lay in detection due to build up of likelihood func- tion for active hypothesis and divergence of in- dividual EKF states for the inactive hypotheses. The IMM approach proposed here overcomes this