Deep learning based fault diagnosis has been investigated a lot via massive amount of data. In this paper, the fault identification problem is considered for hypersonic vehicles in the zero-shot case. ``Zero-shot'' indicates that target fault samples are not available in the training process. To tackle this challenge, the fault description by human-defined is used to characterize the unknown fault. Specifically, a relation network is utilized which learns to compare the known fault samples against the defined unknown fault description. In the feature extraction, a deep neural network structure is constructed by adding long and short term memory neural network to the convolution neural networks. The zero-sample fault identification experiments are carried out on a hypersonic vehicle with Winged-Cone configuration to diagnose actuator fault. The results show that it is indeed possible to diagnose target faults without their samples.