The reconstruction based on the singular value decomposition can effectively separate and suppress random noise components in monitoring signals

but its performance is limited by the construction of trajectory matrixes and effective component evaluation and selection. To solve this problem

an adaptive sum singular value pair (SSVP) optimization framework based on the min mutual information (MMI) is proposed and applied to the feature extraction of machine tool bearing fault signals. Firstly

singular value and singular value vector are calculated with the anti-angular average method

and singular value pairs (SVP) are obtained with the characterization ability of the SVP of sub-signal energy. Then

the optimal reconstructed components are obtained adaptively based on the MMI index

avoiding over-noise reduction or under-noise reduction. Meanwhile

the singular value ratio indexes of MMI are combined to determine the number of optimal decomposition dimensions of the Hankel matrix. Finally

the validity of the MMI-SSVP method is verified with the data of the faulty bearings of a spindle and the feeding system of a machining center in an industrial site respectively.