A generalized version of the information matrix is introduced consisti
ng of the autocorrelation and cross correlation matrices of the shifte
d input and output data. Based on the concept of data correlation, a n
ew system realization algorithm is developed to identify a model direc
tly from input and output data. The algorithm starts with computing th
e information matrix to derive a special correlation matrix that in tu
rn produces the system observability matrix and the state-vector corre
lation. A system model can then be identified from the observability m
atrix in conjunction with other algebraic manipulations. The algorithm
leads to several different methods for computing system matrices to r
epresent the system model. An experimental example is given to illustr
ate the validity and usefulness of these methods with some comparison.