For the problem that rolling bearing diagnosis is affected by the environmental noise so that extraction of characteristic frequency is difficult, a rolling bearing fault diagnosis method based on wavelet packet transform and complete ensemble empirical model decomposition adaptive noise (CEEMDAN) was proposed. The raw vibration signal collected by the sensor was split through CEEMDAN and the high-noise signal and low-noise signal were divided through the kurtosis-correlation coefficient screening criteria (K-C). The wavelet packet transform was used to split the high noise signal and then select appropriate component reconstruction to filter out the environmental noise and integrate with the low noise signal to generate a new vibration signal for envelope demodulation, and the actual fault characteristic frequency was extracted to achieve fault diagnosis of rolling bearings. After comparative experiments, the method proposed clearly extracted the rotational frequency, fault characteristic frequency and its frequency multiplier and modulation frequency of rolling bearings, and the signal-noise ratio after noise reduction was increased by 7.61 dB from the simulation signal calculation, which effectively optimized the effect of noise filtering.