Abstract: In order to identify abnormal state of large-span arch bridge, an improved novelty detection technique based on BP neural algorithm was used. The method used BP neural to train a large number of measured data and got the novelty index when bridge state was normal. Then the threshold and whether the bridge abnormal or not was determined. Combined with concrete situation and data analysis, the method can accurately identify the abnormal region. The method, which avoids the effect of the model error, can greatly improve the practical value, decrease false alarm, make the identification result more accurate and more practical.