A wavelet-domain data fusion model is proposed. Data from a set of sensors are decomposed into multiple scales. The details of all scales and the approximation of the most coarse scale are fused with local weights, and the signal is reconstructed from the fused result. This model is confirmed with the mathematical theory based on practical application. Multi-scale wavelet transform of a random sequence is analyzed, and the statistical relations between the smooth signal and detail signal at various scales are studied. Superiority of the wavelet multi-scale data fusion algorithm is shown mathematically. Experimental results show that bias stability of MEMS gyroscope can be improved after data fusion.