The output PDF (probability density function) shape of surface bubble size in froth flotation is believed to be closely related to the operational statuses of reagent additions. An online operational status recognition method for quality evaluation of reagent addition is presented based on adaptive learning of the dynamic distribution features of the surface bubble size. We avoid the bubble over-segmentation problem by exploring an improved image segmentation algorithm to get the accurate bubble size statistics taking account of the local regional distribution of the image brightness value. By utilizing the kernel density estimation, we obtain the PDF and CDF (cumulative distribution function) of the bubble size statistics effectively. The distribution features of the bubble size statistics under the PDROS (pre-defined reagent operation statuses) are learned by FNC (furthest neighbor clustering), successively, the current health status of the reagent addition in the test time period is inferred by Bayesian inference according to the dynamic change of the bubble size PDFs; what is more, the statistical distribution features under PDROS are updated online according to the disturbance of the process operation conditions. This status recognition method is tested and practically applied in a copper ore beneficiation plant. The experimental results on the real production data demonstrate that this method outperforms other quasi machine vision based production condition recognition methods with much lower error recognition rate. It paves the way for the realization of an optimal control and fault diagnosis for reagent addition in the flotation process operation. (C) 2013 Elsevier Ltd. All rights reserved.