UGP system: A deep learning-driven platform for automated identification of ultrafine granular powders using chromatographic fingerprinting.

This study developed an intelligent identification system for ultrafine granular powder (UGP) by integrating high performance liquid chromatography (HPLC) fingerprinting with deep learning algorithms. A comprehensive HPLC fingerprint database encompassing 530 batches from 53 UGP varieties across 29 botanical families was established using a standardized 60-min, six-wavelength detection protocol (210, 230, 254, 280, 327, and 380 nm). Chromatographic reproducibility was ensured with quality control (QC) sample retention time relative standard deviations (RSDs) below 2%. A three-layer one-dimensional convolutional neural network (1D-CNN) was designed with 32, 64, and 128 filters in successive layers for species classification. Data augmentation techniques including noise interference, baseline drift, and retention time shifts (3.5-60 min) expanded the dataset sixfold and enhanced model generalization capabilities. The optimized model achieved excellent performance on test data with 97.62% accuracy, 97.97% precision, and 97.16% recall, demonstrating consistent reproducibility with mean accuracy of 97.2% ± 0.65% across ten independent training runs. External validation using 63 commercial samples yielded 95.24% identification accuracy, confirming practical applicability. The Flask-based web system enables automated workflows from data upload to species identification and is accessible to users without specialized expertise. This work establishes a standardized approach for intelligent authentication of food-medicine homologous Chinese medicinal UGPs, addressing regulatory and consumer requirements for product authenticity and safety in pharmaceutical and functional food industries.