In view of the fact that shallow artificial neural networks (ANNs) rely on prior knowledge for artificial extraction of features, while shallower network structures limit the ability of neural networks to learn complex nonlinear relationships, this paper applies deep neural networks (DNN) to the study of inversion of multi-component volatile organic compounds (VOCs) by leaf-transformed infrared spectroscopy (FTIR), and the effectiveness of the algorithm was verified by simulation experiments. Eight VOCs including benzene, toluene, 1,3-butadiene, ethylbenzene, styrene, o-xylene, m-xylene, and p-xylene were selected from the US Environmental Protection Agency (EPA) database. In the wavelength range of 8~12 μm, each gas has four different concentration lines, and the absorbance spectrum at one concentration is selected from each VOCs gas according to Beer-Lambert's law to obtain 65 536 different kinds. Samples of VOCs mixed gas absorbance spectra. The absorbance spectra of 5 000 groups of mixed gases were randomly selected, of which 4 000 were used as training samples and 1000 were used as prediction samples. The dimensional reduction of the spectral matrix was performed by integral extraction and principal component extraction, and the spectral dimension was reduced from 3457 to 30 dimensions. The new matrix obtained by preprocessing the spectral matrix was used as the network input, and the concentration matrix of the eight VOCs was used as the output. A deep neural network regression prediction model of 30-25-15-10-8 was established, and multiple groups were realized by using spectral data. Inversion of VOCs concentration, the root mean square error of the sample obtained by inversion was 0.002 7×10-6, which was obvious compared with the accuracy of previous methods using nonlinear partial least squares fitting and artificial neural network. improve. The root mean square error of each VOCs gas does not exceed 0.005×10-6, and the root mean square error of each sample does not exceed 0.006×10-6, which proves that the deep neural network prediction model has good nonlinear fitting ability. And good stability. When the training sample is insufficient (typical value: less than 500), the deep neural network cannot fully learn, the network error is larger, and the accuracy is lower than that of the single hidden layer artificial neural network, but as the number of training samples increases, the deep neural network accuracy is continuously improved. When the number of training samples is sufficient, the deep neural network has stronger nonlinear relation learning ability than the shallow artificial neural network, and the prediction accuracy is higher and the model is more stable. At the same time, due to the dimensionality reduction of the spectral matrix before training, the complexity of the algorithm is greatly reduced, and the inversion efficiency is effectively improved. The analysis shows that the deep neural network prediction model has good nonlinear fitting ability and good stability. It can fully learn the data features without manual extraction of features, and at the same time, the concentration inversion of multi-component VOCs can achieve higher precision.