Exercise electrocardiogram (ECG) recordings are strongly affected by motion artefacts and muscle noise, while true clean reference signals are often unavailable in real exercise conditions. This study presents a morphology-preserving denoising approach for exercise ECG using a DSP-supervised one-dimensional convolutional neural network autoencoder. In the proposed framework, noisy ECG recordings are used as inputs, while pseudo-reference signals generated by a clinically assessed conventional filtering method are used as supervision targets. The model was validated on a private 48-patient treadmill ECG database sampled at 1 kHz using a patient-wise evaluation strategy to assess its ability to denoise previously unseen noisy ECG recordings. Relative to the DSP pseudo-reference, the proposed framework increased the median signal quality from 5.99dB to 16.22dB, while achieving a median Pearson Correlation Coefficient of 0.988, and preserving clinically relevant beat structure with a median R-peak F1-score of 0.942. These findings indicate that the proposed DSP-supervised autoencoder effectively suppresses exercise-related noise while preserving ECG morphology, highlighting the potential of pseudo-reference-guided deep learning for real-world biomedical signal enhancement when true clean references are unavailable.