Diagnosis of abrupt faults using variable-structure neural network

In this work, a variable-structure neural network (VSNN) is proposed for fa ult diagnosis. It is a hybrid between the feedforward network (FFN) and the recurrent network (RecN). Similar to the Kalman filter approach, the filte r gain is adjusted according to the ratio of noise and error covariance. Wh en some of the states are not measurable, the VSNN naturally leads to a Rec N-like architecture. This is exactly the problem formulation for fault diag nosis. A chemical reactor example is used to demonstrate the effectiveness of the fault diagnosis scheme. Results show that the variable-structure neu ral network can detect and isolate incipient faults in an effective manner.