The invention provides a transformer partial discharge fault diagnosis method based on hierarchical threshold synchrosqueezed wavelet. The method is characterized by comprising the following steps of: 1, subjecting a transformer partial discharge signal f(t) to SWT decomposition to obtain an SWT decomposition coefficient S<y><a(i)>(t), wherein the i = 1, 2, L, I; 2, estimating a noise standard deviation [sigma]n by using a formula described in the specification; 3, assigning values to ca(i) step by step at a fixed step length for the ith coefficient of the SWT, calculating the minimum mean square error by a mean square error formula, and determining the optimal value of the ca(i); 4, calculating the hierarchical threshold [gamma]a(i) = ca(i) [sigma]n, and achieving the ith SWT coefficient noise reduction through the ith SWT coefficient formula; 5, using a noise reduction formula to reconstruct de-noised layer coefficients to obtain a de-noised de-noising transformer local discharge signal y(t); and 6, performing feature extraction on the de-noised de-noising transformer local discharge signal y(t), and performing fault diagnosis according to extracted features.