In the practical case, the early fault signals of bearings are weak which is difficult to extract from strong noise. So, when a mask method is introduced into local mean decomposition (LMD), a method for extracting weak fault of rolling bearings based on mask signal method and LMD has been proposed in this paper. Because there is a mode mixing phenomenon when LMD decomposes the product function (PF) components in the noise background, it is difficult to distinguish which fault frequency is true or false. Furthermore, the mask signal method introduced to the decomposed PF components, alleviates the mode mixing phenomenon, and it can extract the real fault frequency. According to the analysis of the actual fault signals of rolling bearings, the kurtosis value has increased 8 times at the fault frequency using the mask signal method and LMD to process the fault signals with noise, and the signal’s noise ratio has increased 19.1%. At the same time, the fault signals have been extracted.