It has been a hot spot in the area of material science that how to design an experiment to produce a kind of ceramic matrix composites (CMCs) which possesses ideal performances. An approach is to build models with data from previous experiments recorded in published papers and predict the experiment parameters needed by the ideal CMCs. According to the database of CMCs funded by the National Material Genome Engineering, 8 factors were considered to affect the tensile property of CMCs, which were the basis, the reinforcement fiber type, the reinforcement fiber volume content, the perform type, the porosity, the interface type, the interface thickness and the density. Among the data we collected from papers, however, only few of pieces contained all the 8 factors, most were incomplete, some of them even lacked multiple factors. This paper’s work mainly researched how to take advantage of the incomplete data to build an effective model to predict the tensile property of CMCs. We proposed a model to predict the tensile property of CMCs based on a 1-D convolution neural network (CNN), the training data of which were all from papers. To decrease the influences of the incompleteness of data, we tried several methods to process the missing data, such as the mean imputation, the K-Means clustering imputation, the Hot-Deck Imputation and the regression imputation. The results showed that the regression imputation with a dual-hidden-layer feedforward network performed better and improved the performance of the CNN tensile property prediction model.