The rapid identification of natural unknown minerals in the field is limited by three challenges：differences in the resolution of different spectral devices，weak model generalization ability due to insufficient sample size，and limited ability to extract high-dimensional complex spectral features. To address these three challenges，this paper designs and implements a Raman spectroscopy classification model that combines a multi-scale convolutional neural network with spectral sample generation. This model is integrated with a portable Raman spectrometer to enable rapid identification of unknown minerals in the field. First，a cubic spline curve fitting algorithm was used to match the dimensions of the spectra collected by different devices，thereby eliminating differences in sampling resolution between different spectral devices. Second，5 668 spectral samples of 1 648 mineral types in the global mineral spectral database were fed into a generative adversarial network for training，producing 15 000 augmented samples to alleviate the constraints of data scarcity on model classification performance. Finally，a new multi-scale deep convolutional network was used to synchronously extract the broad and narrow peak features of Raman spectra，thereby enhancing the characterization capability of complex spectra. In addition，the model proposed in this paper was compared with several classic machine learning models，such as k-nearest neighbor（k-NN），support vector machine（SVM），and random forest（RF），to evaluate their performance in identifying unknown minerals. The experimental results demonstrate that the proposed multi-scale convolutional neural network combined with spectral sample generation achieves significantly higher accuracy in classifying Raman spectra of unknown minerals compared to other traditional machine learning models，with top-1 and top-3 accuracy rates of 93.26% and 98.94%，respectively. The proposed model was applied to identify 50 types of unknown natural mineral samples using a portable Raman spectroscopy system，achieving an accuracy rate of 100%. The identification time for a single sample was only 1-2 min，demonstrating the advantages of the method proposed in this paper，which includes rapid，accurate，and sample-free analysis.