A novel methodology for multiple flaw detection is presented in this study. It combines the dynamic extended finite element method (XFEM) with machine learning for the first time. The extreme learning machine (ELM) is chosen as a learning rule for modeling and prediction. The XFEM is employed to overcome the issues associated with the large quantity of input data required for ELM network training, whereas the ELM itself is used to bypass the time-consuming repeated analyses ordinarily required for the detection of multiple flaws. A large amount of potential flaw data for a structure is quasi-randomly generated by a Sobol sequence. For each effective flaw datum, the dynamic XFEM with circular/elliptical void enrichment is used to compute the structural dynamic characteristics (i.e., frequencies and displacement mode shapes), which is possible because re-meshing is not required for each flaw. The available data generated from the results of XFEM analyses are used for ELM network training. According to the measured dynamic characteristics, the trained ELM is then utilized to predict the size and location of flaws. The results show that the proposed novel methodology can identify accurately the location and size of circular or elliptical structural flaws. The method also is more efficient than previously proposed approaches, as it can avoid time-consuming iterative analyses, and it is robust against noise.