Pre-trained transformer for photonic compressive sampling

A pre-trained Transformer network is proposed for the application in temporal photonic compressive sampling, which can address the neck-strangling issues in classical compressive sampling algorithms when using random or orthogonal measurement matrices. The Transformer network is pre-trained to accommodate a diverse array of needs, and specific application requirements can be addressed by fine-tuning the network parameters to learn prior information. In this paper, we preliminarily validated the algorithm model through simulation to address the waveform measurement performance of linear frequency modulated (LFM) signals. Using a photonic compressive sampling architecture with an average sampling rate of only 40 MSa/s, the Transformer accurately realized waveform reconstruction of LFM signals with a frequency range from 0.1 to 50 GHz and an instantaneous bandwidth as large as 10 GHz under strong interference. A frequency identification error of less than 0.3 GHz was achieved, corresponding to a compression ratio of 1500:1.