In industrial production, it is necessary to detect the running state of rolling bearings and diagnose their faults. When rolling bearing is damaged, the vibration signals collected often show the characteristics of non-stationary and modulation, and will inevitably be disturbed by strong noise, so it is very difficult to identify the fault features. How to effectively extract the components carrying fault feature information from complex non-stationary and modulated signals is the key of diagnosing bearing fault. Frequency slice wavelet transform (FSWT) uses frequency slice function based on the advantages of short-time Fourier transform (STFT) and wavelet transform (WT), which makes the traditional Fourier transform realize time-frequency analysis function. The traditional fault diagnosis method based on FSWT determines the most suitable center frequency and the faulty bandwidth by comparing the results of different frequency band processing, and realizes the reconstruction and description of arbitrary frequency band and local characteristics of the signal. However, this method is inefficient, non-adaptive and can not guarantee that the frequency band screened manually contains the required fault information. Aiming at the problem that traditional methods rely on manual operation and have no self-adaptability, an adaptive frequency slice wavelet transform (AFSWT) is proposed in this paper. Firstly, the signal spectrum is segmented continuously; spectrum segmentation covers the whole frequency band and avoids the process of manual selection of spectrum boundary. The method of equalization can improve the computational efficiency. Secondly, the spectral negative entropy is introduced as the evaluation basis to calculate the complexity of the signal in each frequency band in order to screen the cyclostationary information which may contain periodic shocks. Finally, the frequency band with the largest spectral negative entropy is selected and defined as the faulty center frequency and bandwidth. The signal components in the band are reconstructed and analyzed by envelope demodulation to realize fault diagnosis. The analysis results of a simulation signal show that the AFSWT method identifies the center frequency of 5 000 Hz and the bandwidth of 909 Hz, which is very close to the ideal result. Compared with fast spectral kurtosis, AFSWT has better applicability when the central frequency of signal is located in Fs/4, Fs/8 and Fs/16(Fs is the sampling frequency). Through the test of rolling bearing test-bench, the vibration signals of rolling bearing outer ring fault are collected and analyzed. After AFSWT analysis, the characteristic frequency and its 2-6 times frequency components can be clearly found in the envelope spectrum of the results. On the other hand, AFSWT takes 14.7 seconds to process test signals. The traditional FSWT needs repeated drawing of time-frequency distribution map, determination of central frequency band and selection of observation frequency, it often takes 5-10 minutes to determine the faulty center frequency and bandwidth. The above analysis shows that AFSWT can improve the calculation efficiency and screening accuracy by uniformly dividing the spectrum of the signal and screening the signal components according to the negative entropy of the spectrum. It is suitable for fault diagnosis of rolling bearings. [ABSTRACT FROM AUTHOR]