Monitoring the fetal heartbeat during pregnancy is essential for identifying early signs of distress and ensuring healthy development. However, mining the fetal electrocardiogram (FECG) from abdominal recordings is difficult because the fetal signal is very small and easily hidden beneath the mother’s stronger ECG and other noise. Many existing methods struggle because they need extra reference channels. Also, they depend on assumptions that do not match real recordings, or change the fetal waveform while trying to remove the maternal signal. Hence, this paper introduces a Deep Arrhythmia Classification Framework Using Structural Similarity–Wavelet (DAC-SSW) to overcome these issues. The proposed method first applies a structural-similarity suppression strategy that selectively removes only those abdominal segments that resemble maternal QRS morphology, ensuring minimal distortion of the fetal waveform. Haar wavelet decomposition is then used to enhance the fetal QRS components by isolating frequency bands where fetal energy is most prominent. An adaptive detection mechanism identifies fetal R-peaks, from which a stable Fetal Heart Rate (FHR) trace is computed. Then, the cleaned fetal ECG is augmented and divided into short analysis windows. Then, a hybrid CNN-LSTM is employed for feature extraction. Finally, the SoftMax layer converts the learned representation into class probabilities for fetal rhythm classification. To improve prediction reliability, an uncertainty-aware classification loss is employed to jointly optimize classification performance and prediction confidence. Using the recordings from the NI-FECGDB, the proposed method achieved a mean sensitivity of 96.88%, a PPV of 98.17%, and an accuracy of 97.24% in detecting fetal R-peaks.