Aiming at the problem of insufficient timeliness of fault diagnosis of track circuit in steady-state environment， a method based on Gath-Geva （GG ） fuzzy clustering to divide the degraded state of track circuit is proposed， and the fault prediction of track circuit is carried out by using convolutional neural network （CNN） and bi-directional gated recurrent unit （BIGRU）. Firstly， through the centralized monitoring equipment， the normal working data of each fault type of ZPW-2000 track circuit within a certain time before the fault occurs are obtained. Then， the performance degradation states of track circuit are divided into stages by feature reduction and GG fuzzy clustering based on kernel principal component analysis， and different degradation states are identified. Finally， CNN-BIGRU hybrid neural network is used to mine the data characteristics of different fault types of track circuit， predicting the fault types corresponding to the degraded state of track circuit. Experimental results show that the algorithm can accurately divide the degraded state of track circuit and realize its fault prediction. The classification accuracy of CNN-BIGRU prediction model can reach 97.62% and the running time is only 13.18 s. It can provide an effective method for multi-mode health state recognition of track circuit.