Abstract: In response to the problem that partial discharge signals of reactors are easily overwhelmed by power frequency and harmonic interference in complex electromagnetic environments, making it difficult to balance weak pulse retention and strong noise filtering with fixed threshold noise reduction, a dual frequency signal deep noise reduction and feature reconstruction method based on adaptive dynamic threshold is studied. Using variational mode decomposition to decouple the frequency domain of signals collected by dual core sensors, and combining the kurtosis correlation coefficient joint criterion to screen for effective modes. Based on the estimation of noise standard deviation using wavelet median absolute deviation, a dynamic threshold function adaptively adjusted with frequency and noise level is constructed to achieve precise noise reduction. Extract the multi domain features of the denoised signal and reconstruct the feature vector through local global weighted fusion. The test results show that the design method improves the signal-to-noise ratio by more than 9.26dB under different levels of interference, and the Euclidean distance of the feature vector remains stable within 0.06, effectively solving the problem of difficult partial discharge feature extraction in complex environments.