Fault detection and diagnosis for unknown nonlinear systems: a generalized framework via neural networks

The paper presents a generalized framework for the fault detection and diagnosis for unknown nonlinear systems via neural networks. The systems are assumed to be expressed by a general nonlinear model including a set of unknown parameters whose unexpected changes are defined as faults in the system. At first, it is assumed that the healthy (initial) values of these parameters are known during an initial operation period of the system. By incorporating the healthy parameter set into the structure of a neural network, a nonlinear model can be trained during this initial operation period. Using the trained neural network model, the diagnosis of the fault is achieved by directly estimating the changes of parameters. Both small and large faults are considered. The former leads to a linearized approach where least squares estimation is applied to estimate the size of the fault, whilst the latter results in a gradient based diagnosis algorithm. A simulated example is included to demonstrate the use of the proposed method