The feature extraction is difficult to conduct for weak vibration signal in strong noise background. Thus, a feature extraction algorithm is proposed based on Improved Dictionary Learning and Shift Invariant Component Filtering(IDLSICF). Firstly, vibration signal is segmented and smoothed to decrease the complexity. Then, improved dictionary learning algorithm as well as efficient coefficient solver is used for constructing adaptive filter based on shift invariant sparse coding. The shift invariant components constructed by dictionary atoms is filtered to obtain optimal basis function for representing inherent signal features. Finally, intensity of characteristic frequency in optimal basis function is utilized for evaluating performance in signal feature extraction. Experiments on both simulation data and practical data demonstrate that the proposed algorithm can realize better performance on feature extraction compared with the uptodate methods and is more feasible for the practical applications.