GFSK Demodulation Using Learned Sequence Correlation

Abstract

Artificial neural networks (ANNs) are increasingly being considered for physical layer communications, an area where traditional statistical signal processing and decision theory have been successful. In addition to being able to learn these traditional solutions as a special case, ANNs can also learn more general solutions based on measured or simulated radio data. Furthermore, ANNs enable flexible, reconfigurable implementations that can support multiple standards, can be updated in the field, and can adapt to new or changing field conditions using online learning techniques. This paper examines several ANN topologies inspired by generative probabilistic modeling. All are suitable for noncoherent demodulation of power-efficient modulations such as FSK and OOK, and simulation results are presented for the specific case of GFSK demodulation in Bluetooth LE. When trained in AWGN conditions, the proposed ANN topologies are capable of learning the traditional MAP and ML solutions as well as new solutions with softer trade-off between performance and complexity. When trained with other radio impairments, the demodulators learn additional capabilities for improved robustness. The paper also provides analytical results necessary for training the proposed network topologies with noncoherent input data.