Abstract: Scenario generation is one of the key problems in digital twin (DT) technology, and the typicality of scenarios is the key to test effectiveness. The test scenarios of an intelligent networked vehicle are derived from real vehicle driving data. This paper proposes a new DT test scene generation method which extracts typical test scenes based on local road vehicle driving data collected by roadside radar, establishes a typical scene evaluation method of collision risk factors and traffic quality factors on the basis of the three typical applications of FCW, LCW and ICW, and builds an LSTM-AE-Attention model to identify these critical scenarios. The experimental results show that the constructed model can effectively evaluate and identify typical scenes, which provides effective support for the construction of the test scene library.