Joint Interference Cancellation and Signal Detection Using Latent Space Representations in VAE

Land Mobile Radios (LMRs) are a two-way consumer radio communication system, popularly used for public safety operations. An unintentional strong far-out interfering signal causes the LMR receiver to be overloaded and reduces the gain of the weak desired signal. The conventional non-learning based methods to mitigate the effects of interference require prior knowledge of the interferer or additional filtering components at the RF front-end of the receiver. In this paper, we propose a novel data-driven unsupervised Deep Learning-based approach for joint interference detection, interference cancellation and signal detection of narrowband LMR signals that we refer to as DeepLMR. The DeepLMR uses a Variational Autoencoder (VAE)-based framework known as Recovery VAE (Re-VAE), with a Gumbel-Softmax distribution that encodes the input to a lower dimensional representation as the latent space representations. The latent space representations are sampled from a categorical distribution and classified to the corresponding symbols of the transmitted signal. Experimental results with real-world signals distorted by a strong far-out interfering signal showed that our proposed DeepLMR architecture has bit error rate (BER) performance improvements as compared to the conventional frequency discriminator and other state-of-the-art Deep Learning-based architectures.