Abstract: Aiming at the problems that traditional modal decomposition methods rely on manual parameter selection and mode mixing when processing magnetic flux leakage signals,a new signal processing method combining Harris Hawks Optimization and Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise is proposed,enhanced the signal processing performance and noise-suppression capability.According to the axial distribution law of the leakage magnetic field,a 4-dimensional physical feature vector is extracted,and a BP neural network classification model is adopted to identify the type and location of transformer incipient faults.Simulation results show that the overall detection and recognition accuracy of the proposed method reaches 98.63%, which outperforms traditional decomposition schemes.In addition,the classifier model exhibited simple training process and strong robustness,and was suitable for the online monitoring scenario of early-stage transformer faults.