Based on the back-propagation (BP) neural network prediction method and combined with the infrared spectrometer, Raman spectrometer and polarimetry analysis system through the optical fiber, a multi-spectral blood glucose measurement system is developed and a processing method of data fusion is proposed. 30 human blood samples were measured to obtain the optical rotatory dispersion spectrum, infrared spectrum and Raman spectrum, respectively. Spectral data was preprocessed and normalized. The BP neural network model was established to predict the blood glucose content. We use the Clarke error grid analysis to compare the blood glucose content obtained by the three measurement methods and by data fusion. Results show that the fitting precision of the fusion data is 0.9992, and the prediction error is lower than 0.2 mmol/L, which can meet the accuracy of clinic medicine. This method also has high robustness and strong tolerance.