The fetal heart rate variation serves as an effective means of fetal health monitor. Although can be captured through abdominal electrocardiogram (AECG) of pregnant women, the AECG signal contains interference from maternal heart activity and other sources. Due to the higher-amplitude maternal ECG (MECG), the waveform of the fetal ECG (FECG) is obscured, making FECG separation challenging. However, the similarity of the frequency patterns of both types of ECG signals implies the possibility of reconstructing FECG using the frequency and the strength information of FECG. To this end, we employ the short-time Fourier transform (STFT) to convert the AECG signal into its time-frequency representation, and utilize a neural network model for extracting FECG from the time-frequency domain. The model can obtain high-quality FECG by capturing the amplitude differences between fetal and maternal heartbeats in different frequency components of AECG. Moreover, a combined time-domain and frequency-domain loss is introduced to enhance the model's separation performance. The proposed method is validated on public databases and demonstrates outstanding performance, providing an effective and feasible solution for non-invasive monitoring of fetal health.