Unsupervised Blind SNR Region Estimation Using Prototype-Based Multi-Stage Deep Neural Network

This paper presents a novel unsupervised learning approach for signal to noise ratio (SNR) region estimation combined with supervised modulation classification. Unsupervised learning consists of training a network while the input data labels are not provided. Previous work has shown that knowing the SNR or being within some range of SNR improves performance when performing modulation classification by using multiple networks trained separately. Existing methods are either supervised or have very specific requirements of a dataset that might not be possible to obtain in the implementation environment. Current modulation classification methods perform poorly at low or negative SNR values which previous works have shown is due to the difference in frames' SNR. Our proposed method is a frame-level SNR region estimator which uses a custom prototype-based objective function that is minimized using a regression deep neural network. The estimator network partitions a dataset by estimating SNR ranges and each range is trained on a separate network for modulation classification. We explore using a different number of clusters of a hierarchical clustering method to evaluate the separability of the SNR ranges to determine an upper bound and feasibility of a multi-network approach. The performance of our method is evaluated using a weighted dataset partition accuracy on the modulation classification dataset Deep-Sig RadioML2016, which consists of multiple modulation types and SNR values. Results show the ability to effectively estimate and separate multiple SNR ranges in a dataset.