The problem of roll eccentricity compensation in steel-strip rolling mills is examined. A novel roll eccentricity sensor is proposed for the task of retrieving the desired reference signal from the available roll-force measurement. The sensor implements a modified comb filter structure which can be arranged in either tunable or adaptive form. Simulation results based on typical experimental data indicate that the proposed sensor is appropriate and offers distinct advantages over other methods. The proposed comb filter is new in the sense that it consists of a number of second-order notch filter models with specially constrained bandwidth characteristics. An approximate gradient-based adaptive algorithm is used for the frequency estimation. The important feature of such an algorithm is that it is robust and is simple to implement. The advantages of the sensor include fast convergence and accurate parameter estimates. No extra positional sensors are required since this information can be easily obtained from the phase estimates. Finally, the sensor is not based on approximate mill models; it treats the problem as a harmonic series buried in noise. The resulting solution is based on time domain measurements that characterize the mill roll eccentricity conditions during operation. >