Fault detection of nonlinear dynamic systems with emperical input/output models

Abstract

Predictive maintenance has become a familiar concept in industrial fault detection regime. The ability to detect early warning signals in systems in the form of small changes in dynamic behavior is essential to anticipate failures. This implies that the fault detection method must be well tuned to a particular process. In general, accurate system models are an essential part in residual based fault detection. However, in complex nonlinear systems, the development of accurate models can be very difficult, thus usually other approaches are often selected. As an alternative to the nonlinear analytical models, neural networks have shown significant potential in accurately representing nonlinear systems. This paper utilizes a non-typical model structure which includes internal states, so that the identification method is similar to a nonlinear extension to (linear) subspace identification. This model is then used with a moving horizon observer to generate residuals. The experimental section shows the observer implemented on a thermo fluid system, with several introduced faults.