The study of causality has always been the theme of human exploration of the world. The traditional study of only the correlation between things can no longer meet the needs of current economic and social development. This paper combines self-supervised learning and adversarial learning, and uses Bayesian network model and system dynamics inversion model to mine its own supervised information from large-scale unsupervised data. Based on comparative constraints, we build a unified framework of physical drive and data drive, and then use minimax game strategies to learn consistent representations of different causal modes. Therefore, the true causality can be approximated. It provides a powerful analysis tool for revealing the inherent laws hidden behind the data. This paper integrates the idea of non-random causal learning into the machine learning framework, which has important guiding significance for overcoming the shortcomings of existing deep learning in abstraction, reasoning and neural network interpretability.