To address the challenges of ineffective harmonic component extraction and low fault classification accuracy caused by inter-turn short circuits in generator windings, this paper proposes a fault diagnosis method that combines Empirical Mode Decomposition (EMD) with an improved Backpropagation (BP) neural network. Voltage signals are decomposed using EMD to obtain Intrinsic Mode Functions (IMFs), from which feature parameters such as variance contribution, average period, and correlation coefficient are extracted to construct a multidimensional fault feature vector. A three-layer BP neural network with an adaptive learning rate is designed and optimized using a gradient descent strategy to improve convergence speed and classification accuracy. Experimental results show that the proposed method achieves an average diagnostic accuracy of 96% for four levels of inter-turn short circuit faults, significantly outperforming traditional spectral analysis methods. This study provides a theoretical basis and technical support for online monitoring and intelligent diagnosis of electrical faults in generators.