For the actual abnormal data identification and diagnosis of bridge health monitoring system, the abnormal feature library construction method is presented. Firstly, the monitoring data are denoised and extracted to the signal features in time domain, frequency domain and AR model. Then, using KPCA feature extraction method, abnormal data feature library is built, including short-circuit, beat -wave and ship-collision. Finally, the abnormal state of bridge monitoring system is identified by BP neural network to verify the effectiveness of feature library construction. The vibration data analysis for Jiangyin Yangtze River Bridge shows that the method can be used to diagnose abnormalities.