Raman spectroscopy has been widely used in chemical industry, security,anti drug and other industries and research fields, but the traditional Raman spectroscopy analysis technology relies on the spectral database, through the spectral feature extraction for identification. Feature extraction is the key step of Raman recognition. Principal component analysis, factor analysis and other methods are usually used for feature extraction, and then KNN, SVM and random forest methods are used for qualitative identification of spectral features. When there is no undetermined substance in Raman database, it is easy to cause the wrong classification of the substance to be detected. In order to solve this problem, a method based on convolution neural network is proposed to identify the lack of substance spectrum in database. In the process of the experiment, we use nine categories, more than 200 kinds of psychotropic drugs Raman spectrum as the test object, through the construction of convolution neural network automatic feature extraction, and use softmax classifier to analyze more than 200 kinds of substances according to nine categories, such as amphetamine, cathinone, cannabinoids and so on. Compared with the traditional machine learning methods such as k-nearest neighbor and support vector machine, the accuracy of model recognition based on convolution neural network is significantly improved. This method can provide a new recognition method for Raman spectrum database.