Model-based Fault Diagnosis of Nonlinear System using Intermediate ANN-Hammerstein Approach

This paper outlines an approach for developing a fault diagnostic system for a nonlinear dynamic process using Hammerstein model. The nonlinearity is assumed to be static in nature and dynamics is assumed to be linear. Sequential functional approximation is done using a nonlinear block of feed forward neural network at input side cascaded with a state space model. The topology and the parameters of ANN structure are determined from near steady state data extracted from dynamic test data using wavelets while for the linear dynamic part a subspace approach is used. The FD system is based on the principle of analytical redundancy and generates a fault detection index with suitable thresholds for discrimination of fault-free and faulty behavior. The faults are isolated using fault isolation indices for the respective faults. The nonlinear model can be directly used for detection and isolation of the faults in sensors. For isolation of faults in actuators, a delta model linearized around the fault free input is proposed. The limitations of such model based system are discussed. This approach has been applied for modeling a CSTR plant operated over a wide range of operating conditions in open loop but could be extended to closed loop operation.