Due to the inevitable presence of anomalies in the data collected by the structural health monitoring system, it is impossible to obtain the true health status of the structure. Therefore, anomaly data detection is crucial for structural analysis and state evaluation. A multi-channel data anomaly detection method based on a combination prediction model is proposed. Firstly, the structural health monitoring data is divided into two sections. The first section only includes intermittent anomalies caused by the environment, while the second section includes intermittent anomalies and data anomalies caused by sensor failures. Secondly, by estimating the local density of each data point based on cosine kernel density and adaptively selecting parameter radii, and improving the density based spatial clustering of applications with noise (DBSCAN) algorithm, the improved model is used to remove intermittent anomalies in the previous data and obtain clean data (i.e., normal data without problems). Next, based on the correlation between multiple sensors, combined with the spatial features of convolutional neural networks (CNN) and the temporal features of long short term memory networks (LSTM), a mathematical model representing normal data features is trained to clean the data. Then, the predicted data is obtained by inputting the later stage data into the mathematical model, and the predicted data is compared with the later stage data to obtain the prediction error. The extreme value theory (EVT) algorithm is used to fit the distribution of the prediction error and set a threshold to detect abnormal conditions in the data. Finally, analyzing the acceleration monitoring data of Dowling Hall pedestrian overpass shows that this method can effectively improve the detection ability of abnormal data in structural health monitoring.