Bearings are the components that are widely used in various machinery, whose working condition is crucial for safe operation. Running under time-varying speed conditions in most case, fault features are submerged in background, and traditional analysis tools such as frequency spectrum is not sufficient to process the bearing fault signal with characteristic of varying speed and strong background noise. In this paper, the bearing faults under time-varying rotation speed are diagnosed based on dual-tree complex wavelet transforms (DTCWT) and order analysis (OA) is carried out. DTCWT is used for angle domain resampled signal denoising for better fault characteristic extraction since there is strong background noise in the measured vibration signal. The effectiveness and performance of proposed method are verified by numerical simulation and bearing fault dataset collected by Ottawa university, and the fault-related feature order of resampled stationary signal is well extracted. This paper studied the diagnosis of bearing fault under time-varying speed conditions, which provides an effective way for nonstationary bearing fault diagnosis.