7 standard substances and 26 mixtures made from 7 standard substances were selected as the research objects in this paper， with the accuracy and misjudgment rate as the evaluation indexes for recognition effect of the components of mixtures， and spectral peak matching algorithm， non-negative least squares matching algorithm and correlation coefficient matching coefficient algorithm were studied， in order to solve the problem of identifying the components of mixtures. Furthermore， a BP neural network model infrared spectroscopy technique based on multi-feature fusion was used to identify the components of the mixture， which was compared with the logistic regression model. The results showed that the recognition accuracies of three single matching algorithms are lower than 76.31%， while the recognition accuracies of multi-feature fusion logistic regression model and multi-feature fusion BP neural network model prediction set are 83.33% and 98.18%， respectively， and the misjudgment rates are 4.76% and 1.82%. The results showed that mid-infrared spectroscopy combined with BP neural network model could better identify the components of the mixture. In order to further explore the detection ability of the model towards the lowest concentration of mixed components， the detection limit for vanillin and ethanol solution was studied. The results showed that the accuracy rate was 100.00% and the misjudgment rate was 14.29% when the concentration of vanillin was 0.03 g/mL. When the concentration of vanillin was lower than 0.03 g/mL， the accuracy decreased by 20.00% and the misjudgment rate increased by 19.04%. Therefore， it may be considered that the minimum mass concentration of vanillin in vanillin and ethanol solution identified in this study is 0.03 g/mL.