USING MODEL UPDATING TECHNIQUE TO TRAIN NEURAL NETWORK FOR FAULT DETECTION

摘要

Vibration monitoring and fault detection of componentsin manufacturing plants involve a detailed analysis of acollection of vibration data in order to establish a correlationamong changes of the measured data and the correspondingfault. This work presents an alternative proposal which intentis to exploit the capability of model updating techniquesassociated to neural networks to reduce the amount ofmeasured data. The updating procedure supplies a reliablemodel that permits to simulate any damage condition, whichallows to establish a direct correlation between the deviation ofthe response and the corresponding fault. The learning of thenet is performed using a compressed spectrum signal createdfor each specific type of fault. Different fault conditions for aframe structure are evaluated using simulated data and finally,the capability of the proposal is demonstrated usingexperimental data.