Abstract: The monitoring methods for partial discharge signals typically rely on traditional signal processing and analysis techniques, judging partial discharge events by setting thresholds for signal amplitude values. However, due to the failure to capture the transient characteristics of discharges, the monitoring accuracy is relatively poor. Therefore, an online monitoring method for partial discharge signals in generators is proposed. This method involves collecting noisy discharge signals using multi-channel sensors and decomposing them into multi-scale approximation coefficients and detail coefficients. Nonlinear shrinkage is applied to the detail coefficients using a soft-thresholding function, and the partial discharge signal is reconstructed with the aid of inverse wavelet transform, highlighting the transient characteristics of discharges while smoothing the signal. Phase space reconstruction of the signal is performed within each window, and the reconstructed vectors are symbolized based on the magnitude of their elements. By statistically analyzing the probabilities of various symbol sequences, the permutation entropy of the signal is calculated and input into an extreme learning machine model for discharge pattern recognition and monitoring. Experimental test results demonstrate that when the proposed method is employed for online monitoring of partial discharge signals, the coverage of discharge pattern labels exceeds 96%, indicating a relatively ideal monitoring effect.