A new semi-blind ICA algorithm, named supervised ICA, is proposed for improving the performance of ICA in FECG extraction. A supervision signal is created for assisting the usual ICA algorithm to find the desired FECG unambiguously. It is different from other semi-blind ICA algorithms in that the a priori information about the temporal structure of the desired signal is encoded in the supervision signal, instead of the object function. Performances of the new algorithm are demonstrated by real data processing. High quality FECG can be extracted from the multi-channel abdominal signals.