A noise suppression method based on Wasserstein GAN model

Noise suppression in a receiver system, such as a signal analyser, is fundamental for ensuring accurate and reliable signal measurement and analysis. Effective methods play a pivotal role in enhancing the dynamic range and sensitivity of the instrument, which is crucial for the detection and characterisation of weak signals that might otherwise be obscured by noise and interference. This paper introduces an innovative noise suppression approach using a Wasserstein Generative Adversarial Network (WGAN) model. The model is trained with real noise data to generate synthetic noise samples that closely mimic the actual noise. A specially designed activation function addresses the tiny but high-frequency oscillations present in noise data, while Kernel Density Estimate is used to handle inconsistencies across different datasets and reduce the randomness in generated samples. Both techniques improve the robustness of the model. By employing these generated samples, the receiver system can perform real-time numerical noise suppression, significantly improving the accuracy and efficiency of signal measurements compared to the present method. The advancement is vital across various applications, particularly in fields requiring high precision, such as wireless communications and aerospace. It broadens the scope for more precise and reliable signal analysis in complex environments.