Since the metal magnetic memory (MMM) signal tends to be influenced by environmental noise, the testability of the detected signal is reduced. The traditional singular value decomposition (SVD) method is applied to decompose and reconstruct the MMM signal of a yard crane. It is effective in reducing the noise, but it is difficult to determine the singular value number of reconstruction. In order to solve this problem, the MMM signal matrix is constituted on the principles of dichotomy and recursion, and the multi resolution singular value decomposition is realized through the repeated SVD method to obtain approximate signals and detail signals of different resolutions, which correspond respectively to the effective components and the noise components. The experimental results of a certain yard crane magnetic memory signal processing show that the proposed method can effectively improve the signal-to-noise ratio of the reconstructed signal, and accurately judge the stress concentration area. This will lay the foundations for stress state evaluation and early fault diagnosis.