Power Efficient Joint Channel Coding and Frequency Modulation with Deep Learning

Achieving low-power and cost-effective wireless communication is essential for the Internet of Things (IoT) and massive machine-type communication. Common IoT protocols such as Bluetooth and LoRa adopt frequency modulation schemes that can be received non-coherently with significantly lower complexity and power consumption than in-phase and quadrature (IQ) modulation schemes with coherent receiving. In our paper, we propose a novel deep learning-based joint channel coding and modulation (JCM) method, tailored for non-coherent frequency modulation through digitally controlled oscillators (DCO). This approach features an encoder that converts sequences of information bits into DCO control samples, thereby controlling the instantaneous frequencies to modulate the radio frequency (RF) signal. The learned decoder is designed to extract information bits from the noise-corrupted received samples, eliminating the need for a preamble for time and frequency synchronization. This system is trained and evaluated under conditions influenced by phase noise and carrier frequency offset (CFO), using phase-noise statistics from a practical DCO circuit implementation. The test results show the ability of the proposed method in overcoming these common receiver challenges. Additionally, to minimize the encoder’s power consumption without substantially affecting the bit error rate (BER) performance, we integrated quantization-aware training, allowing the model to operate on fixed-point arithmetic. The power consumption of the encoder is assessed with a post-synthesis circuit simulation to evaluate the feasibility of the system in IoT applications.