To achieve the accurate and efficient classification of IFC components， an improved Multi-view Convolutional Neural Network （MVCNN） model is proposed， which introduces a self-attention module and a Long Short-term Memory （LSTM）network. To address the limitations of MVCNN model feature fusion， an LSTM_ATT module is designed. By adaptively adjusting the feature relationships of each view data and fusing the input data of each view according to the training attention weights， a more discriminative 3D shape descriptor is obtained. In turn， this improves the classification and detection performance of the model for similar IFC components. Finally， the improved MVCNN model is experimentally compared with the MVCNN using the IFCNet database， which comprises 20 classes of IFC components commonly used in the construction industry. The experimental results show that the overall accuracy of the proposed model for classification and recognition as well as the F1 value reach 88.27% and 86.72%， respectively， which is an improvement of 9.46% compared with the classification accuracy before the improvement， and the classification recognition between similar components is obvious.