In many scientific applications, measured time series are corrupted by noise or distortions. Traditional denoising techniques often fail to recover the signal of interest, particularly when the signal-to-noise ratio is low. In this work, we demonstrate that deep learning-based denoising methods can outperform traditional denoising techniques and exhibit greater robustness to variation in noise and signal characteristics. Our motivating example is the detection of materials with Nuclear Quadrupole Resonance (NQR) spectroscopy. Specifically, we aim to detect the presence of radio frequency signals emitted as a result of interactions between the nuclear quadrupole moments and the electric field gradients in fentanyl hydrocholoride and eventually other narcotics. These are short-duration and low-amplitude signals often obscured by strong interference, making detection with traditional methods difficult. We explore real-valued and complex-valued deep learning architectures to process and denoise inherently complex-valued NQR signals. Our findings demonstrate that complex-valued neural networks are superior to their real-valued counterparts and more effective than traditional denoising methods, offering a powerful new approach for reliable fentanyl detection using NQR spectroscopy.