Abstract: Convolutional neural networks(CNN)are a classification method widely used for analyzing visual images. Because of the presence of complex spatial relationships such as nonlinearity and coupling, studies on CNN-based numerical data are limited. The purpose of this paper is to find a feasible way to enable the extension of CNN application to address numerical data. A new CNN classification method based on radar chart(Radar-CNN)is thus proposed. The numerical data are first expressed as radar charts which are fed into CNN to build a classification model. To further investigate the influence of features' scale and sequence on the performance, Rank Radar-CNN and SFS Radar-CNN are also proposed as two variants. Furthermore, to verify the validity of the proposed algorithm, Tennessee Eastman chemical process dataset is introduced for experimental tests. The results show that Radar-CNN and its improved algorithms have excellent performance.