Abstract: A compound fault feature extraction method based on improved feature mode decomposition (FMD) was proposed to address the problem of non-stationary, multi-component, and strong background noise in rolling bearing compound fault signals, which makes it difficult to extract fault features effectively. FMD was used to decompose the rolling bearings compound fault signal into a series of mode functions, the key parameter characteristics that affect the decomposition accuracy were studied, and relevant parameter selection methods were proposed. From the perspective of the degree of correlation and energy between signals, a comprehensive evaluation factor algorithm was used to select fault sensitive mode functions, and the sensitive mode functions envelope spectrum were obtained through envelope demodulation to extract fault feature frequencies, achieving the diagnosis of rolling bearing fault types. The effectiveness of improved FMD has been demonstrated through simulation analysis, engineering example of rolling bearings fault diagnosis. Compared with the variational mode decomposition (VMD) method, the proposed method has a more effective ability to suppress the impact of noise interference and extract fault features.