The separation performance of fast independent component analysis (ICA) can be damaged by outliers. However, the traditional model only discusses the condition that ICA is free of outliers. In general, those measurements that significantly deviate from the normal pattern of sensed data are considered as outliers. In the observed signals of ICA, outliers include noise and errors. Whereas, the traditional techniques used to find the outliers can not be used in ICA directly due to the constraint of the independence of the sources signals. In this paper, a non-polynomial function-based method is proposed to detect the outliers in the signals of ICA. Simulations show the effectiveness of the proposed approach to find the outliers in the observed data.