On using the sequential regression (SER) algorithm for long-term signal processing

The use of the sequential regression (SER) algorithm [1, 2] for long-term processing applications is limited by two problems which can occur when an SER predictor has more weights than required to predict the input signal. First, computational difficulties related to updating the autocorrelation matrix inverse could arise, since no unique least-squares solution exists. Second, the predictor strives to remove very low-level components in the input, and hence could implement a gain function that is essentially zero over the entire pass-band. The predictor would then tend to become a "no-pass" filter which is undesirable in certain applications -- e.g., intrusion detection [6]. Modifications to the SER algorithm that overcome the above problems are presented, which enable its use for long-term signal processing applications.