Keyhole arc welding is a cost-effective process to weld thick materials. To successfully apply this process, the state of the keyhole must be accurately detected and used in feedback control. In this paper, the dynamic behavior of the plasma reflection is described using the reflection arc angle (RAA). It is found that the RAA series can be considered an autoregressive moving-average (ARMA) process. The parameters of the ARMA are recursively estimated using the extended least squares algorithm. It is found that the recursive estimates of the model parameters change as the state of the keyhole changes. A discriminator has been proposed to determine the state of the keyhole based on the recursive estimates of the model parameters.