A methodology for data-driven risk analysis based on virtual-reality-generated information and generative adversarial network

To improve the safety of complex systems, it is essential to analyze and maintain the risk at an acceptable level. However, risk analysis is usually encountered with the difficulty of data deficiency, particularly for complex systems and unusual operations. In this paper, we proposed a methodology for data-driven risk analysis based on virtual-reality-generated information and a generative adversarial network (GAN). First, the concerned accident scenario for risk analysis is formulated. Second, the virtual reality (VR) model of the corresponding accident scenarios and operations is constructed. The experiment data, containing operation failure information, is subsequently collected. Third, to effectively support the data-driven risk analysis, the scale of the experiment data is augmented through GAN. Based on the augmented data, risk analysis is carried out in the form of data-driven Bayesian networks (BN). Eventually, the feasibility of the proposed methodology is validated with the case study of risk analysis of emergency operations in deepwater blowout. Our results show that the proposed methodology is beneficial to deal with the data deficiency in the domain of risk analysis.