Abstract: In order to identify and classify the abnormal structures in coal-bearing strata more accurately, a POA−ELM model based on the pelican optimization algorithm (POA) and the extreme learning machine (ELM) is proposed. The performance of extreme learning machine is unstable because the input weights and hidden layer bias are generated randomly. The POA can be used to optimize the input weights and hidden layer bias of extreme learning machine, so as to improve the performance of extreme learning machine model. The POA−ELM model is applied to identify and classify the abnormal structures in coal-bearing strata. Firstly, three coal-bearing strata simulation models of small fault, scour zone and collapse column are established with the COMSOL Multiphysics5.5. The Ricker wave is the source signal. The in-seam wave signals are collected by wave transmission method, and the in-seam wave data set is established. Then the z-score method is used to standardize the in-seam wave data and the principal component analysis (PCA) is used to reduce the dimension. Secondly, the POA is used to optimize the extreme learning machine, and the POA−ELM classification model is constructed with MATLAB. The POA−ELM model is used to classify small fault, scour zone and collapse column. The classification performance of ELM and POA−ELM is evaluated and compared by cross-validation method and evaluation indices such as accuracy, precision and recall rate. The results show that the POA can effectively optimize the ELM, and the POA−ELM model has higher classification accuracy and better stability. The classification accuracy of POA−ELM for abnormal structures can reach more than 99%. Thirdly, in order to verify the classification effect of POA−ELM in practical applications, after wavelet de-noising, z-score standardization and PCA dimensionality reduction, the real fault in-seam wave data are used as the test set and imported into the POA−ELM model for classification. The results show that the identification accuracy of POA−ELM model for real fault can reach more than 97%. Finally, based on the same data set, the classification effects of POA−ELM, ELM, support vector machine (SVM) and BP neural network are compared. The results show that the identification and classification accuracy of POA−ELM model is the highest. Through research and analysis, the POA can effectively optimize the ELM, and the POA−ELM model can accurately classify different geological structures and effectively identify real faults, which is better than other methods.