Abstract: fault warning results caused by insensitive capture of weak features in distribution transformer winding loosening fault warning. Firstly，the force energy of the winding is introduced as a modulation parameter to reconstruct the mother wavelet function， breaking through the limitations of traditional fixed time-frequency analysis and achieving multi-scale accurate extraction of fault signals；Then，based on the signal decomposition results，the feature frequency amplitude changes and weight coefficients are fused to construct a quantitative fault index and a three-level warning mechanism，solving the problem of delayed response to progressive faults with a single threshold；Finally，experiments have shown that this method can reduce the distance detection error to 0.01mm and the partial discharge detection error to 0pC，verifying its high sensitivity and reliability in fault warning.