Residual generation for fault diagnosis thru recurrent nets

Isolability in model-based fault diagnostic methods is usually enhanced through structuring the residual space. For this, raw residuals are transformed to structured residuals by a set of z-domain polynomials. However, these polynomials have been known to have difficulty in presence of modeling bias and noise. In this paper, we propose the use of time-delay recurrent networks (TDRNs) as an alternative to these polynomials. For this, the TDRN is tuned based on a set of residual fault data. The effectiveness of TDRN for residual transformation is evaluated in simulation, by comparing the structured residuals obtained from the TDRN with those from polynomials. Simulation results show that the use of TDRN improves robustness in presence of modeling bias and noise.