The failure of instances in virtual environment is always causing huge losses. This paper designs and implements a virtual-environment-oriented failure detection system. In consideration of hierarchical dependence in virtual environment, a hierarchical model is used to classify the failure of instances in different levels in order to improve the accuracy of instance failure recovery, and the mechanism of self-adaptive dynamic failure detection cycle is applied to the trade-off between the CPU resource cost and the timeliness of the virtual machine detector. This system has been implemented and verified in the OnceCloud platform.