Arc Fault Detection through Model Reference Estimation

In most arc fault circuit breakers, arc detection is accomplished through the signature analysis of remotely sensed branch currents, with high frequency spectral components being indicative of arcing. This paper presents an alternative approach based upon system identification. A model is assumed for the load on the distribution bus. This model is updated continuously by comparing measured voltages and/or currents to the values predicted by the model. The resulting prediction errors are used to adjust the model in real time. In aviation loads, even in nonlinear loads that are rapidly and repetitively engaged and disengaged, we find that the model successfully adapts to give a good description of the load. However, when an arcing fault is present, its chaotic nature prevents a successful model identification and the prediction errors remain large. The large, continuous prediction errors provide a means of fault identification in a short time and with a high level of confidence.