Non-invasive fetal electrocardiogram (NI-FECG) is used to monitor the fetal heart health. Single-channel FECG extraction is a critical challenge for the development of long-term monitoring devices. Many approaches based on deep learning (DL) were proposed but impose prohibitive computational and memory requirements that limit the development of embedded devices. In this work, we propose a two-stage framework based on denoising autoencoder (DAE) models. Our approach sequentially isolates maternal and fetal components from abdominal signals. The models were trained and tested on simulated data and evaluated on real recordings. They achieved a median F1-score of 1.00 and 0.93 in QRS detection for simulated and real signals, respectively, and a median Pearson correlation coefficient of 0.89 for simulated data. Each model requires only 80,144 parameters (32-bit precision) occupying a total of 0.62 MB of memory -- representing a 5× reduction compared to the state-of-the-art approach while maintaining comparable detection accuracy.