A method is discussed that enables detection of burstlike deviances in normal stochastical signals using autoregressive modeling combined with the sequential probability ratio test. Corrections are made to better estimate the beginning of the event, and Student distribution is used taking into account the actual amplitude distribution.