Robust failure detection from the generalized likelihood ratio test

The generalized likelihood ratio (GLR) test employs two steps in identification of failures. The first involves the use of Kalman filter for the unfailed system to generate the innovation, or residual sequence. In the second step, by means of description of various hypothesized failure modes, the likelihood ratio is computed and maximized. Comparison of this maximum ratio to a threshold value yields the failure decision (failed or unfailed). The authors extend this strategy for dynamic stochastic systems with unknown inputs.