Deep learning shows superiority in transient security situational awareness of smart grids. However, it is hard to find optimal parameters for low-dimensional features and network structures in multi-layer network reconstruction. Thus an improved deep belief network (DBN) is proposed. In this method, the SMOTE algorithm is first adopted for oversampling to balance the proportion of data samples. This ensures mining of the deep structure of DBN. Then the binary neuron from Gibbs sampling is used in DBN to improve the noise immunity in the reconstruction process. Given the repeated manual debugging of network structure, a genetic algorithm (GA)-based GA-DBN model is finally employed to achieve global optimization. The low-dimensional features are abstracted accurately from the high-dimensional measurement data at the bottom layer and the classification precision is guaranteed. A test on the New England 39-bus system with imbalance and noise samples shows that the proposed method outperforms the existing methods in accuracy, F1 score and missing alarm. This work is supported by the National Natural Science Foundation of China (No. 51777126).