Abstract: Roll eccentricity in rolling mills has an important influence on the delivery gauge of rolled strips. An online back propagation algorithm with memory range was used to identify roll eccentricity and compared with the simple online back propagation algorithm on identification. The results show that the roll eccentricity identification method with memory range has a faster convergence and a better anti-noise performance. It can make the thickness fluctuation decrease about 95 96 when the eccentricity's magnitude, phase and frequency change, consequently compensate thickness error induced by roll eccentricity and improve the delivery gauge of rolled strips effectively.