Y-Net: A Dual Path Model for High Accuracy Blind Source Separation
Huanzhuo Wu, Jia He, Máté Tömösközi, Frank H. P. Fitzek · 2020
Industrial IoT (IIoT) in conjunction with Ultra-Reliable Low-Latency Communications (URLLC) often struggles with data-rich, information-poor contexts. Blind Source Separation (BSS) is one of the key technologies which can obtain the desired high-value information from all of the observed raw sensory data. As shown by recent studies, BSS can be both fast enough for low-latency requirements and sufficiently accurate to be a reliable method in large IoT deployments. Nonetheless, the trade-off between signal context usage and data recovery accuracy often affects the separation quality of BSS. In this paper, we propose for the first time a novel dual path convolutional neural network model, called Y-Net, for high accuracy BSS. Specifically, the separation quality is improved by the parallel perception and joint combination of both high-and low-level features of input signals, which we demonstrated through extensive numerical evaluations. In particular, Y-Net improves the Source-to-Distortion Ratio by 2.70% to 35.32% for different target signals, while the model size is only slightly increased, compared to other current solutions.