Traditional human blood glucose testing methods are invasive and have certain limitations. In this paper, a new technique is proposed by combining energy conservation method and spectroscopy to realize non-invasive, real-time and accurate detection of human blood glucose. Firstly, a body sign data collection device was designed to collect blood glucose-related data in real time and upload it to the upper computer. Then the data were analyzed and evaluated with such three different machine learning algorithms as multiple linear regressions, k-nearest neighbor regression and support vector regression, thus the optimal algorithm for non-invasive blood glucose detection could be confirmed by comparisons. Experimental results show that, the proposed technique for noninvasive blood glucose detection realizes high feasibility, accuracy and robustness, the measurement accuracy based on the support vector regression algorithm is the best, and the correlation coefficient reached as high as 0.862.