Abstract: Due to harsh environments, to sensor aging and failures, health monitoring data of civil engineering structures are prone to outliers, leading to difficulties in objectively reflecting the actual service status of the structures. To address critical problems, such as poor detection efficiency, requirements for subjective thresholds and challenges in handling multi-type anomaly, encountered in traditional offline detection methods, an online diagnosis method based on a Bayesian filtering framework for multi-type anomaly monitoring data is proposed. The method uses a Bernoulli probabilistic model to describe the multi-type outliers in the monitoring data. The parameters governed in the probabilistic model are then extended to the structural state vector. A Bayesian filtering framework is utilized to online update the extended state vector, the online diagnosis and correction of the multi-type abnormal data can be thusly realized. A single-degree-of-freedom system, a three-span bridge model and a shaking table test on a scaled-down model of a 12-storey reinforced concrete frame are used to validate the effectiveness and applicability of the method proposed. The results show that the method proposed can online diagnose and correct multi-type abnormal data for structural health monitoring. Meanwhile, it improves the reconstruction accuracy of structural dynamic responses, and significantly enhances the time efficacy of anomaly diagnosis for structural health monitoring.