A Transformer-Based Zero-Shot Learning Paradigm for Portable-to-Benchtop Raman Spectral Translation

Abstract Raman spectroscopy, as a nondestructive testing technique, is often limited by noise interference and feature distortion in portable instruments, which severely affect the accuracy of substance detection. To address this, this paper proposes a fast and effective portable-to-benchtop Raman spectral translation method (ETS) for portable devices. First, ETS employs empirical mode decomposition (EMD) to filter out instrument-specific noise through multiscale decomposition. Then, a simplified and lightweight transformer network is used to realize feature mapping between portable and benchtop Raman instruments. Finally, sparse Bayesian learning (SBL) further strengthens the sparse representation of the intrinsic peaks of substances. In the experiment, ETS demonstrated an excellent instrument transfer performance. The cosine similarity of the transferred spectra exceeds 99%. It also demonstrated a great instrument generalization performance. This allows the transferred data to achieve an accuracy of 100% in pure-substance classification in the database. In addition, ETS has an execution time of approximately 0.254 s per spectrum on an Arm CPU with a model weight of only 9.43 MB. In summary, ETS effectively solves the cross-instrument consistency problem of portable Raman instruments, providing key technical support for on-site, rapid substance detection.