Reconstructive Spectrum Analyzer with High‐Resolution and Large‐Bandwidth Using Physical‐Model and Data‐Driven Model Combined Neural Network

Abstract

Most neural networks (NNs) used for reconstructive spectrum analyzers (RSAs) rely on data-driven training strategies, which can be time-consuming due to the need for a large training dataset with a limited amount of output channels. Here, a specially designed NN is proposed for a reconstructive wavemeter based on temporal speckle obtained from a whispering gallery mode (WGM) resonator. By combining a physical model and data-driven model, it only takes 10 µs to obtain a reference speckle for the generation of a training dataset. The WGM resonator-based wavemeter assisted by the NN uses only one photo-detector to obtain a temporal speckle, achieving a spectral resolution of 3.2 fm. The number of output channels reaches 2300, which is the largest dynamic range achieved by an NN in RSA without the need for re-training. It demonstrates that the proposed NN has capability to resolve unseen spectrum with multi-tone wavelengths. Moreover, the proposed network exhibits better robustness in long-time measurement compared to data-driven model based networks. This opens up new possibilities for NN design methods in RSA, without the need for a large training dataset, by incorporating a physical model to achieve high-resolution, high-dynamic-range, and fast-speed spectrum measurement.

Mode division multiplexing reconstructive spectrometer with an all-fiber photonics lantern

“…During the calibration phase, the tunable laser wavelength was systematically adjusted in increments of 1 pm over a range of 2 nm. This scanning procedure resulted in a spectral sequence consisting of 2000 channels, keeping the order of magnitude the same as the other schemes focused on broadband operation [ 20 , 21 , 29 ]. The tunable laser for calibration was able to switch wavelengths at a rate of 10 kHz, and the camera used for capturing speckles operates at approximately 200 Hz.…”

A Triangular-Matrix-Based Spectral Encoding Method for Broadband Filtering and Reconstruction-Based Spectral Measurement

“…Compressed-sensing-based spectral encoding requires the spectral response matrix to obey the restricted isometric property, and the designed spectral encodings, like the Gaussian random matrix, are always quite complex and very hard to realize [ 15 , 23 , 24 , 25 ]. Machine learning methods have the advantages of high spectral encoding efficiency and much higher spectral reconstruction speed [ 15 , 26 ]; but the designed spectral encodings are also irregular and hard to accurately realize for almost all kinds of spectral filter devices. Thus, the complexity of the designed encoding schemes is a major bottleneck that limits the application of BFRSM.…”