Recording electrocardiogram (ECG) signals during treadmill exercise is strongly affected by electrode motion and muscle artefacts, and in many sce-narios, reference ECG signals are unavailable, making it difficult for super-vised algorithms to train. To address this issue, this approach proposes a hy-brid framework that integrates traditional DSP with deep learning techniques for ECG denoising and preserving the signal morphology. First, the autoen-coder uses the noisy exercise data as input signals, while the denoised signals are used as targets. The target signals are generated using our previously pub-lished hybrid adaptive filtering approach (Jasim et al., 2025). After a one-time training of the 1D CNN autoencoder, new ECG recordings can be de-noised by a single forward pass through the network, avoiding repeated adap-tive filtering. The proposed approach was evaluated using a private exercise ECG dataset from 48 subjects sampled at 1000 Hz. This database is collec-tion of patients suffering from ischaemic cardiomyopathy, who were under-going an implantable cardioverter defibrillator (ICD) implant and treadmill exercise testing. ECG recordings were segmented into fixed 5s windows and split subject-wise into training, validation, and test sets (38/4/6 subjects) to avoid data leakage. Regarding the test set, the proposed model achieved a mean normalized RMSE of 0.1206 ± 0.0566 and a mean normalized PRD of 14.90% ± 9.08%. In terms of signal quality, the average SNR increased from 4.22 ± 4.39 dB to 14.40 ± 3.74 dB, corresponding to an average SNR im-provement of 10.18 ± 3.89 dB. These results show that the proposed ap-proach rapidly and effectively learns ECG features and preserves the signal morphology, supporting effective noise removal for exercise ECG in the ab-sence of true ground truth.