The purpose of this study is to make use of machine learning theory to deeply study the mechanical performance of steel reinforced concrete special-shaped column joints by data driven instead of the traditional mechanism driven method. In this study, the node data set of reinforced concrete special-shaped column containing several key parameters was established, and the improved neural network algorithm was used to construct the prediction model, and a high precision of the bearing capacity prediction was obtained. Through the combination of experimental data and computer vision model, the formation process and failure form of joint cracks are analyzed, and an image processing method without human intervention is proposed, which can effectively reduce the error. Based on explainable machine learning theory, we explain and analyze the factors that affect the bearing capacity of special-shaped column joints, and reveal the role of key factors in the bearing capacity of joints. This research combines neural network, computer vision and interpretable machine learning to provide comprehensive technical support and in-depth explanation for seismic design of reinforced concrete special-shaped column joints.