Coupling Pattern Recognition With State Estimation Using Kalman Filter for Fault Diagnosis

This paper deals with a diagnosis tool based on a pattern recognition approach associated with Kalman interpolator/extrapolator. The first aim is to decrease the number of measurements to realize while increasing the learning database contents using a Kalman state estimator. The second one is to estimate, from the initial set of measured data, future states of the studied process. A 5.5-kW induction motor bench is used as an application to validate this approach. First, a signature is determined in order to monitor the different operating modes evolution. Diagnostic features are extracted only from current and voltage sensors. Then, a feature selection method is applied in order to select the most relevant features for diagnosis. Finally, a Kalman filter algorithm is developed in order to interpolate the known states and to predict evolution toward new ones. A new diagnosis tool is then designed handling continuous evolution (severity, load) inside the different operating modes (healthy, stator fault,…).