Abstract: In response to the difficulty in diagnosing fault types due to the complexity and variability of gearbox fault signals under variable working conditions, a fault diagnosis method based on the Harris hawk optimizer (HHO) was proposed to optimize the multi-layer perceptron (MLP) neural network. Firstly, the method employed the root mean square-mean (RMS-MEAN) method for data preprocessing of different fault vibration signals of the gearbox, reducing the impact of random working conditions on the vibration signals. Secondly, a variable working condition correction factor k was introduced into the HHO to automatically optimize the hyperparameters of the MLP under variable working conditions, enhancing the periodic characteristics in the vibration signals, and constructing the optimal structure of the variable working condition MLP model. Finally, the feature enhancement data was input into HHO-MLP for fault diagnosis. According to the MCC5-THU gearbox fault data set, the fault diagnosis performance of the proposed method is significantly better than other models under varying working conditions, and the accuracy rate of fault classification can reach 97.5%, which shows its effectiveness under varying working conditions.