The origin of the cosmic rays(CRs) in the knee region has been a controversial issue, and how to distinguish their composition for ground-based detectors is still very challenging. The most common method currently used to identify the components of CRs is based on calculation and extraction of composition sensitive variables from the reconstructed information. This method can be constrained by the limited computing power so that deeper relationships between the features of the CR components cannot be extracted. In this research, an updated deep neural network (DNN) model is developed to distinguish between primary CR components from extensive air showers detected by KM2A arrays. Our preliminary result shows that the DNN model with AUC value 0.900 in Proton-Helium identification is more effective for identification of CR in the knee region than the traditional method with AUC value 0.750, and similar result is obtained for Fe-MgAlSi identification.