CNN-BiLSTM-based Learnable Viterbi Signal Detection without Channel State Information

Due to the huge cost of acquiring the channel model and its parameters in the field of communication signal detection, the application of model-based algorithms are limited when the channel knowledge is unknown. At the same time, pure data-driven deep learning methods require a lot of training and are unexplainable. We propose a joint data and model-driven learnable Viterbi signal detection method based on convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM), which can realize the signal detection of Viterbi algorithm in channel-free model scenarios. We learn the channel-unknown state on which traditional model-based methods rely, via a trained CNN-BiLSTM neural network and either parametric estimation (i.e. Gaussian mixture model with Akaike information criterion) or nonparametric estimation (i.e. adaptive kernel density). The simulation results show that proposed method can still guarantee the accuracy of signal detection without considering channel state information (CSI), in addition, different from the traditional Viterbi detection, the method has higher robustness to CSI uncertainty.