•A correlation analysis algorithm for IMFs reconstruction is proposed.•The CorAA can extract main feature of fault modes and eliminate useless information.•The proposed method can diagnose accurately under different working conditions.•The fusion method is superior to peer methods under identical working conditions. The effective and accurate diagnosis of the fault of a gearbox is crucial. However, differences in working condition significantly affect the energy of the original vibration signals of a gearbox, which makes it difficult to distinguish the faulty signals from normal signals. To solve this problem, this paper proposes an integrated method based on complementary ensemble empirical mode decomposition (CEEMD), sample entropy (SampEn) and the correlation analysis algorithm (CorAA) for the fault diagnosis of a gearbox under different working conditions. In this method, CEEMD is used to decompose the raw vibration signals into sets of finite intrinsic mode functions (IMFs). Then, the correlation coefficients between the raw signal and each IMF are calculated using the CorAA. Subsequently, the IMFs with large correlation coefficients are selected for a probabilistic neural network (PNN) to classify the fault patterns. Finally, two cases are studied based on experimental gearbox fault diagnosis data, and the integrated method achieves classification rates of 97.50% and 95.16%. The proposed approach outperforms all other existing methods considered, thus validating its effectiveness and superiority.