Electronic fetal heart monitoring is a common method to detect fetal abnormalities used by obstetricians. Effective analysis and diagnosis of cardiotocography during labor not only helps to solve the problem of neonatal cerebral palsy caused by fetal distress, but also greatly reduces neonatal mortality. Among the existing analysis algorithms, most of them are based on machine learning to extract and classify the characteristics of cardiotocography. The results which depend on the recognition of features are always unstable. The baseline of fetal heart rate is the most basic characteristic. In this paper, the baseline characteristics of fetal heart rate are firstly extracted, and then the Long Short-Term Memory network is used for segmental classification of fetal heart rate. The results of the experiment show the superiority and efficiency of deep learning in feature extraction, and make it possible for fetal distress detection in computer-aid diagnosis which has greatly reduced the burden on doctors.