Considering flaw feature extraction and condition monitoring of a high-speed turnout, a turnout flaw detection method was proposed that was based on ensemble empirical mode decomposition (EEMD) singular entropy and least square support vector machine (LSSVM). First, turnout vibration signals with non-stationary characteristics were adaptively decomposed into a certain number of intrinsic mode functions (IMFs) using EEMD. Each IMF contained different feature scales of the original signal. Then, with correlation analysis, a certain number of IMFs that had the largest correlation coefficients with the original signal were sifted out. The singular entropy of these IMFs were computed and used as the feature vectors. Last, in order to classify the working state and flaw type of the turnout, the feature vectors fused with multi-point singular entropies were input into the LSSVM to train and test. The vibration signals on the turnout platform and contrast experiment were analyzed, and the results showed that this method can be effectively applied to turnout flaw detection. In addition, the proposed method was immune to noise and had stable performance when the signal-to-noise ratio was higher than 20 dB.