Near-infrared non-invasive blood glucose detection technology still fails to meet the accuracy required for clinical application. The main reason on the one hand is that the human blood glucose signal is weak, and the near-infrared absorption band of some components in blood overlaps with the absorption band of glucose. Multivariate regression methods such as partial least squares (PLS) are usually used to extract the glucose concentration information from spectral data. On the other hand, in the measurement process, background interference such as light source drift and measurement condition changes is inevitable. The impact of background interference on the measurement is often stronger than the spectral response caused by the changes in blood glucose concentration. Therefore, these background disturbances must be effectively controlled and eliminated before establishing the blood glucose prediction model. Otherwise, there will likely be pseudo correlations exiting in the blood glucose prediction model established by a multivariate regression method. Therefore, in order to non-invasively detect the blood glucose signal even better, the measurement system itself should have high blood glucose detection capability, and under the premise of keeping the measurement conditions as stable as possible, appropriate data processing methods should be used to eliminate most of the background interference. To this end, this paper evaluated the blood glucose detection capability of the self-developed non-invasive blood glucose detection system, proved that the system could achieve high detection accuracy. Then oral glucose tolerance test (OGTT) and oral water tolerance test (OWTT) were performed on three healthy subjects, and the spectral data of OGTT and OWTT at two different source-detector distances were compared, and the analysis results showed that the variance of the OGTT absorbance change was much larger than the OWTT under the two source-detector distances, and the wavelength distribution characteristic of the absorbance change’s variance for the three subjects varied greatly. Then the spectra data from the two source-detector distances was differentially processed, and the differential spectral data were compared. The analysis results indicated that the variance of OGTT differential absorbance change was far larger than that of OWTT, and the wavelength distribution characteristics of differential absorbance change’s variance for the three subjects were consistent with the absorption characteristic of glucose solution, which proved that the self-developed non-invasive blood glucose detection system combining the differential processing method could effectively eliminate the background interference and extract the signal of blood glucose.