Adaptive Selection of Deterministic and Neural Receivers for Low-Power BPSK Demodulation on the MAX78000

Abstract

Small wireless devices need to decode signals accurately while using limited memory, energy, and processing time. Neural-network receivers can perform well under difficult channel conditions, but always running the strongest model can require unnecessary computation when a simpler receiver would be sufficient. This project developed AegisRx, an adaptive receiver that converts binary phase-shift keying (BPSK) in-phase and quadrature (I/Q) samples into bits on the MAX78000 microcontroller. A router chooses between a deterministic receiver, a small convolutional neural network (CNN), and a more robust CNN based on predicted decoding error and modeled computation cost. A simulator generated 47,500 packets with noise, timing error, phase and frequency offsets, multipath, I/Q imbalance, and DC offset. Compared with always using the robust INT8 CNN, adaptive routing reduced modeled computation by 20.80%, while the bit error rate (BER) increased slightly from 12.5887% to 12.8263%. On the main sweep across signal-to-noise ratios, packet success increased from 16.58% to 18.02%. Input-aware quantization-aware training reduced the robust INT8 model’s main-sweep BER from 24.8075% to 24.5844%, although it remained above the 22.4158% floating-point result. The robust INT8 model also passed a MAX78000 hardware known-answer test and completed CNN inference in 180 microseconds. Overall, these results suggest that adaptive receiver selection can reduce modeled computation with only a small accuracy tradeoff on low-power wireless hardware.