Transformers’ mechanical properties and insulation performance are affected by pre-compression of winding. A new method for the diagnosis and localization of transformer winding looseness based on singular value spectrum entropy and SVM is proposed. First, 110 kV transformer winding looseness experiment is performed to measure vibration signals under different conditions. Then the time series of signals are reconstructed. Optimal feature sequence of reconstructing space is extracted by singular value decomposition and the features — singular value spectrum entropy is got by combining information to judge winding looseness, which are used as inputs of diagnostic model. The multi classification support vector machine is optimized by particle swarm optimization algorithm. The test result of the model is compared with BP and PNN neural network. The experimental result shows that this method can effectively determine whether the winding is loose and correctly identify the loose phase, which verifies this method’s feasibility and accuracy. This work is supported by National Natural Science Foundation of China (No. 51177039).