In the field of Prognostics and Health Management (PHM), deep learning methods have shown excellent remaining useful life (RUL) prediction ability because of their powerful data fitting ability. Such methods require a large amount of failure history to train, but for industrial assets, their life-cycle data are often insufficient, and effective repairs or replacements are frequently carried out before the failure occurs. On the one hand, these suspension history data cannot be directly used by traditional deep learning methods, resulting in the certain loss of degradation information. On the other hand, the internal structure of industrial assets and the types of features are complex and diverse. Traditional methods use Euclidean space to represent relevant data, ignoring the correlation between features. Therefore, we develop a semi-supervised adversarial graph prediction (SAGP) framework that combines adversarial learning, graph neural networks, and semi-supervised training, using both failure and suspension histories for RUL prediction. Based on graph neural networks, SAGP can effectively obtain the degradation information in the failure and suspension history by representing the degradation data in the form of graphs and setting multi-objective functions. Through semi-supervised learning, SAGP looks for failure and suspension histories similar to the current state and generates simulated data through adversarial training to improve the robustness of the model and finally achieve RUL prediction. Tool, engine, and bearing datasets are used to verify the accuracy and reliability of the SAGP model.