A Generative Digital Twin for Continually Enhancing the Intended Functional Safety of Cyber–Physical Systems

Cyber–physical systems (CPSs) used in safety-critical applications must safely operate even in unpredictable situations. The functional safety of their physical components, which have relatively low complexity, has been significantly assured through the safety standards. However, it is still difficult to ensure the intended functional safety (IFS) of autonomous control components driven by artificial intelligence models, as only a limited number of scenarios can be realized by engineers. This issue may cause the fatal injuries and the material damages under the unconsidered scenarios. To address this issue, we propose a generative digital twin (gDT) that continually enhances the IFS of the CPSs using a deep reinforcement learning algorithm and a retraining scheme. The workflow of the gDT involves 1) generating unknown hazardous scenarios that engineers cannot consider during the development phase and 2) feeding the generated scenarios back to the autonomous control components. Experimental results show that the proposed digital twin can discover the unknown hazardous scenarios against prototypes of the CPSs and can enhance the infraction avoiding rate to 90% against the discovered ones while maintaining the ability to operate in the existing known safe ones.