Optimal filtering and Bayesian detection for friction-based diagnostics inmachines

Non-model-based diagnostic methods typically rely on measured signals that must be empirically related to process behavior or incipient faults. The di fficulty in interpreting a signal that is indirectly related to the fundame ntal process behavior is significant. This paper presents an integrated non -model and model-based approach to detecting when process behavior varies f rom a proposed model. The method, which is based on nonlinear filtering com bined with maximum likelihood hypothesis testing, is applicable to dynamic systems whose constitutive model is well known, and whose process inputs ar e poorly known. Here, the method is applied to friction estimation and diag nosis during motion control in a rotating machine. A nonlinear observer est imates friction torque in a machine from shaft angular position measurement s and the known input voltage to the motor. The resulting friction torque e stimate can be analyzed directly for statistical abnormalities, or it can b e directly compared to friction torque outputs of an applicable friction pr ocess model in order to diagnose faults or model variations. Nonlinear esti mation of friction torque provides a variable on which to apply diagnostic methods that is directly related to model variations or faults. The method is evaluated experimentally by its ability to detect normal load variations in a closed-loop controlled motor driven inertia with bearing friction and an artificially-induced external line contact. Results show an ability to detect statistically significant changes in friction characteristics induce d by normal load variations over a wide range of underlying friction behavi ors. (C) 2001 Elsevier Science Ltd. All rights reserved.