Subcutaneous interstitial fluid continues to be the preferred site for glucose sensing due to its easy access and lower risk of infection than that of the blood stream. But changes in subcutaneous interstitial fluid glucose are delayed with respect to changes in blood glucose. Besides, the sampling signals are inevitably influenced by noise in the measurement process. For the reasons above, a neural network soft-sensing method based on wavelet denoising is put forward to accurately predict blood glucose levels. In this method, some auxiliary variables associated with blood glucose are denoised and then used to train the neural network to establish the blood glucose soft-sensing model. The methodology is tested using the simulation data of NO.1 and NO.2 adult. Testing result shows that the blood glucose values obtained by this model has smaller root mean square error, better signal-to-noise ratio, and smaller measurement delay than subcutaneous interstitial fluid glucose values.