Convolutional neural networks are widely used in the field of image processing due to their excellent properties such as local correlation and weight sharing, and have become one of the most popular neural network architectures. However, for non-image forms of data such as genomics, voice,and finance, traditional convolutional networks may not be fully applicable. In order to get rid of this dilemma, researchers have been trying to develop network structures such as recurrent neural networks and attention mechanism networks that can be used to analyze non-image data,and expand the application scope of neural networks. The research and development of new network architecture is undoubtedly difficult and costly. From another perspective, we propose a data preprocessing method for convolutional network structure. By processing the source data, it is converted into a specific one-dimensional feature vector or two-dimensional image matrix,and then sent to a custom convolution structure to observe its algorithm performance. In the experiment,the classical data sets on UCI and Kaggle platforms are used for testing, and traditional machine learning models such as SVM, decision tree, and random forest are used to compare the feasibility and effectiveness of the method.