Diagnosis of abrupt faults using variable-structure neural network

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