Raman spectroscopy is applied as an important method for material identification in field geology. However, analyzing the collected Raman spectroscopy results is time-consuming and labor-intensive, which arises a demand for labeling and sorting a large volume of in-situ Raman measurements automatically. In this study, we consider the spectral characteristics of mineral to develop a convolutional attention network for rapid and precise identification of mineral component. Moreover, we introduce Gradient-weight Class Activation Mapping Plus Plus(Grad-Cam++) to visualize the important region for predicting. Compared to pure Convolutional Neural Networks (CNN), our model is better at learning the details in characteristic peaks to distinguish minerals with similar Raman spectra. Overall, this study exhibits significance for automated process of labeling data collected by Raman instruments in field work and developing similar spectral recognition algorithms. PLAIN LANGUAGE SUMMARY: A deep-learning based model is proposed to identify specific mineral compoents from Raman spectra. The novel method accumulate experience from a vast amount of known data and perform rapid inference on unknown data as educated researchers. Futhermore, we show a technology named Grad-Cam++ to understand the reason of model's decisions in complex situations. It benefits researchers to build trust in intelligent systems and make continuous improvement on deep-learning based model. This study will provide reference and support for the development of artificial intelligence algorithms for observational instruments in field work.