Learning to Synthesize Noise: The Multiple Conductor Power Line Case

The performance of communication systems is strongly dependent on noise. Modeling and reproducing noise patterns play an important role in the development of enhanced communication algorithms. This article exploits Machine Learning (ML) techniques to analyze the Power Line Communication (PLC) noise distribution and synthetically reproduce unseen traces. The generation method takes as input a dataset consisting of noise measurements and processes them into spectrograms, represented as images. A Deep Convolutional Generative Adversarial Network (DCGAN) is trained to generate new spectrograms with the same statistical distribution. Lastly, the Griffin-Lim algorithm converts the synthesized spectrograms into new noise traces. The scalability of the proposed approach allows to incorporate the mutual dependence of multi-conductor noise traces and replicate them. The presented method is evaluated through qualitative and quantitative metrics: the generated noise traces are perceived indistinguishable from the measured ones, and at the same time, their statistical properties are preserved as proven by numerical results.