Cracks are common and potentially hazardous failures in rotating machinery, and their timely detection is critical for ensuring the safe and reliable operation of such equipment. The early crack induces only slight local stiffness changes in the rotor shaft, resulting in extremely weak crack-related superharmonic components in the vibration response. Meanwhile, strong noise in actual operating environments further obscures early crack features, posing challenges to crack diagnosis. This paper proposes an optimized maximum correlation kurtosis deconvolution (MCKD) method to enhance the early crack characteristics of the rotor in a noisy environment. First, the vibration signal is denoised based on complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), and the multi-scale permutation entropy (MPE) is used to reconstruct the signal to filter out high-frequency noise. Subsequently, the minimum average multi-scale permutation entropy (MAMPE) is used as the objective function, and the nondominated sorting genetic algorithm-III (NSGA-III) is used to adaptively optimize the MCKD parameters (filter length L and deconvolution period T). Finally, based on the optimized MCKD, the reconstructed signal is filtered to enhance the crack characteristics in the spectrum. The experimental results demonstrate that the optimized MCKD method can effectively enhance early crack features of the rotor under strong noise conditions with a signal-to-noise ratio (SNR) of − 8 to − 10 dB, offering a reliable foundation for early fault diagnosis.