SENSOR FUSION BASED UPON SELF ORGANIZING MAPS AND LOCAL KALMAN FILTERS FOR LOCAL FAULT DIAGNOSIS

Abstract

A challenging strategy for fault diagnosis is related to data fusion and feature extraction such as having local models of the plant integrated through patterns. To pursuit this goal a group of Kalman Filters is proposed, one per sensor in order to estimate a state of the plant per sensor. Thereafter, a self-organized map processes the information in order to determine the response areas within the map. Moreover, Kalman filter auto tuning procedure, becomes an interesting issue due to correlated noise amongst them as well as local state dependency at this point this may call a pre-processing stage. Furthermore, once global state vector is produced, the classification stage takes place as long as this vector exists as post processing stage. The resulting maps determine areas of fault free scenarios or nonlinear and no healthy responses classified as faults. This last point is obtained through a high accuracy of the system.