Differential Raman Spectroscopy Courier Face Sheet Detection Study Based on RCSC and 1DCNN‐Transformer

ABSTRACT

To address the common issues of significant fluorescence interference, high data dimensionality, and low classification efficiency in detecting filler components in thermal paper courier face sheets, this study proposes an analytical method based on differential Raman spectroscopy combined with an improved spectral clustering (RCSC) and a one‐dimensional convolutional neural network (1DCNN)‐Transformer fusion model. Differential Raman spectroscopy is employed to suppress fluorescence background noise, and characteristic peaks of inorganic fillers are extracted from 179 courier face sheet samples. Orthogonality‐constrained improved principal component analysis (PCA) reduces the original spectral data from 1912 to 84 dimensions, retaining 90.13% of the variance. The RCSC clustering algorithm, based on elemental ratio‐cosine similarity, is proposed to optimize the similarity metric. Finally, a 1DCNN‐Transformer fusion model is constructed for efficient classification of the reduced‐dimensional data. The RCSC clustering achieves a Fisher discriminant accuracy of 95.5%. The 1DCNN‐Transformer fusion model achieves a classification accuracy of 88.9%, representing an improvement of over 20% compared to traditional methods. The proposed method requires no sample preprocessing and enables rapid, nondestructive identification of filler composition differences in courier face sheets, providing an effective analytical approach for forensic evidence traceability.