Differential optical absorption spectroscopy (DOAS) is widely used for online gas detection in industry. However, when the concentration of industrial gas is low, the spectral absorption is not obvious and the SNR is very low. So if the inversion of industrial gas concentration is carried out by using the traditional methods, it is very difficult to meet the requirements of industrial application. According to the differential absorption spectra of SO2, tritium lamp is used as the light source to collect the high-dimensional data of absorption spectra in 189.73～644 nm band. And after selecting and preprocessing the absorption spectra data, a deep belief network (DBN) model is established based on the training set data to extract the low-dimensional features of the test data. Furthermore, the extreme learning machine (ELM) is constructed by using the low-dimensional embedding characteristics of training data to realize the calculation of the SO2 concentration. The effectiveness of the proposed model is evaluated, and it seems that the method is more suitable for accurate on-line detection of SO2 concentration in industrial field.