Abstract: In order to improve the recognition accuracy of partial discharge signals in high-voltage cable accessories, a multi-scale time-frequency analysis method based on wavelet transform is proposed to study the denoising processing and feature extraction mechanism of signals. Based on the db4 wavelet decomposition, key features such as energy spectrum, wavelet packet entropy, and zero crossing rate are extracted, and principal component analysis is used to achieve dimensionality reduction and fusion, constructing a composite feature vector for classification and recognition. The results demonstrate that the hierarchical threshold wavelet denoising significantly improved signal fidelity, and the fused features improved the classification accuracy of the SVM model to 96.2%, effectively distinguishing three typical discharge types and verifying the practical value and robustness of the method.