Dynamic in-situ mechanical CT experiment is a method that uses advanced light sources to capture real-time 3D images of the internal structure evolution of materials under external fields.It is an non-destructive and non-contact measurement methods. It combines in-situ mechanical experiments and computed tomography (CT) technology, and has important application in the research of internal evolution mechanical mechanisms. However, during the in-situ experiment, there is a contradiction between the internal structure evolution rate and the time of CT scan, hence only sparse projections can be collected. In addition, the reconstruction noise caused by sparse angle sampling will affect the judgment and extraction of mechanical indicators. To address these issues, this paper proposes a dynamic in-situ mechanical CT sparse angle reconstruction method (Directed Total Variation, DTV) based on prior edge structure guidance. This method achieves high-quality dynamic in-situ mechanical CT characterization by introducing noise learning method in convolutional neural net to obtain prior edge structure information, which guides the direction of total variation denoising.