The amplitude and phase of the signal transmitted to a channel is distorted by channel dispersion. To compensate the received signal, a kind of equalization technique should be used. In order to equalize the channel distortion, it is necessary to training the equalizer before normal data exchange. Most of the telephone line modems require relatively long training periods, during which timing and carrier synchronization and equalizer training should be completed. However, in the case of burst mode transmission, such as polling systems, the amount of information to be transmitted at a time is small. Therefore, a short training period is required to increase the transmission efficiency. To adapt the equalizer to the channel during the short period, a fast start-up equalization algorithm should be used.
In this study, we select two fast algorithms, RLS(Kalman) and fast Kalman algorithms, for the fast start-up equalizer and analyze their convergence behaviors in the CCITT V.22bis modem. As a result, the convergence speed of the two fast algorithms is very fast compared to that of the LMS algorithm. Using these algorithms, the training period can be reduced to 35\% that specified by the CCITT. Moreover, the convergence speed is not affected by the initial settings of tap coefficients of the equalizer even if the channel noise varies. On the contrary, convergence behavior of the LMS algorithm strongly depends on channel noise variation. So selection of the step size parameter is very cautious.
We have implemented real time V.22bis modem using a TMS320c30 DSP processor employing the two fast start-up equalizer algorithms. We confirm their superior performance to that of the LMS algorithm.