An online learning soft sensor method based on PLS kernel algorithm is presented for industrial processes with time-varying characteristics.By recursively learning representative samples

this method

utilizing PLS kernel algorithm

could improve soft sensor model’s ability of adaptation

which is more computationally efficient than NIPALS.And deleting one or more redundant samples according to a similarity-based criteria that takes both input and output information into consideration may build more effective training sample set.The foregoing scheme is applied to build an industrial polypropylene unit’s melt index model.The results indicate that the proposed method can efficiently track the change of melt index during grades transition.