Water quality is a key factor affecting aquaculture. Currently, the adjusting and controlling of aquaculture water quality depends on real-time monitored data increasingly. During the online real-time water quality data collecting, abnormal data are usually obtained as the instability of water quality sensor devices and wireless sensor network devices, as well as the changing of external conditions. The abnormal water quality perception data will have a serious impact on the decision-making of water adjusting and controlling, which must be eliminated in time. In this paper, based on the analysis to the limitation of conventional methods to exclude abnormal data, an abnormal aquaculture water quality perception data elimination approach based on least square fitting algorithm is proposed, by which weighted sequence is constructed using fitting residuals, the suitable fitting curve coefficient is obtained through iterating. Combining with Laiyite criterion, the abnormal data are detected and eliminated correctly. Finally, the method is verified by the real-time collected data from the aquaculture ponds.