Rapid detection of drugs plays an important role in restraining the spread of drugs and cracking down on drug crimes. Surface Enhanced Raman Spectroscopy (SERS) technology has many advantages such as fingerprint identification, fast detection speed, less sample consumption, no damage, high sensitivity and so on, which has attracted much attention. Its characteristics are especially suitable for the rapid detection and law enforcement of public security organs on the spot. This paper used gold nanoparticle sol as the enhancement reagent to enhance the Raman spectrum. Six solutions of amphetamine, ketamine, fentanyl, heroin, cocaine and methamphetamine were prepared by 1 μg·mL^-1. The volume ratio of drug solution, enhancement reagent and NaCl solution was 20:6:5, and 30 μL of the mixed solution was dropped on the surface of the slide. Let dry in the air and wait for inspection. Five samples were made for each drug solution, and Raman spectral data of 10 points were randomly collected for each sample. 300 groups of Raman spectral data of 6 drug solutions were collected, and 60 groups of Raman data were randomly selected as the training set. The model was trained by using the training set data. The remaining 240 groups of data were used as test sets to test the classification effect of the model. After pre-experiment comparison, 785 nm laser was selected as the excitation light source in the experiment, 50× objective lens was used, the laser intensity was 3.0 mW, the exposure time was 0.2 seconds, and the scanning times were 1 000 times. The bands from 400 to 1 700 cm^-1 were selected for test and research. Savitzky-Golay method was used for smoothing and de-noising Raman data, and the airPLS method was used for baseline correction to complete 0-1 normalization of data. Principal component analysis (PCA), variance screening, genetic selection algorithm and mutual information method were used to process the dimensioning of the data. Modeling training was carried out by the four support vector machine algorithms, random forest, artificial neural network and nearest neighbor respectively. The classification effect of the model was tested by the test set data, and the average accuracy was obtained by repeating 10 times. The results show that the accuracy of all classifiers is more than 95% when the principal component is 5, after the dimension reduction of Raman spectral data by the PCA method. Among the three bands selection methods, the combination of genetic selection algorithm and SVM classifier has higher accuracy. The classification accuracy of the combination of 5 Raman bands screened by the genetic selection algorithm has reached more than 95%, and the accuracy of the combination of 25 Raman bands has reached 99%. As a band selection algorithm, genetic selection algorithm can reduce the dimension of Raman spectral data collection and have stronger interpretation and more important significance, which provides a reference for the rapid detection technology of drugs.