Transformer-Autoencoder Model for Accelerated Multi-Symbol Optoacoustic Demodulation

Exploiting the optoacoustic effect for communication across the air-water interface has attracted attention in recent years. Yet, modulation/demodulation remains a major challenge due to the involvement of two signal types, laser in the air and acoustics underwater. Particularly, the possible creation of acoustic pressure by different laser beams, the underwater noise and the multipath propagation complicate the demodulation of the received acoustic signals and limit the achievable data rate and power efficiency. To address these issues, the paper proposes a novel Transformer-Autoencoder based Demodulation (TAD) scheme that enables increasing the number of symbols by mitigating the inter-symbol interference. Incorporating such a transformer-autoencoder reduces the demodulation latency through simultaneous inference of multiple symbols rather than the traditional sequential approach, which constitutes a novel concept in the realm of demodulation methodologies. In addition, no signal preprocessing is needed. The effectiveness of TAD is validated using a lab-collected dataset. The results show that TAD achieves up to 97.74 % accuracy, while reducing inference time by 55% than ResNet for processing 1000 symbols.