<p>PointNet++ has the functions of object recognition and semantic segmentation, and performs well in the recognition and classification of similar objects. It has been widely used in outdoor 3D scenes. The functional space layout of subway station buildings has the characteristics of similarity and replicability, so it is of great significance to adopt intelligent algorithm to modular design of functional space. In this study, plane data of subway stations in several cities were collected. Build 3D model and export cloud point data (X, Y,Z, rgbC), then enhance the data. The data set is divided into training set, verification set and test set according to the ratio of 8:1:1. The PointNet++ is used to train the 3D data set. Results show that the 3D data set derived from the building model runs well in the PointNet++ network, and can realize the effective transmission of information. Firstly, Pointnet++ model’s recognition and classification results of training set meet expectations, and the stable values of Training mean loss and Training accuracy are 0.42 and 0.76, respectively. Secondly, the training model is used to predict the test set. The Eval avg class Accuracy and Eval Accuracy are 0.80 + and 0.75 + respectively, and the mean intersection over union is about 60%. Thirdly, the prediction results of the test set in the training model show that the Ground Truth and Predict are highly matched, and the model has a strong self-learning ability and can actively optimize the scheme space. This paper proposes a modular design method based on deep learning of 3D spatial data, which can identify and classify architectural space efficiently and accurately.</p>