Abstract: With the wide application of health monitoring systems on various large-scale structures, the massive amount of monitoring data contains various system anomalies due to the failure of monitoring equipment. Such system anomalies due to monitoring equipment seriously hinder further mining and analysis of monitoring data, and thus anomaly identification of monitoring data is essential. Unfortunately, current anomaly identification methods often identify structural anomaly data due to changes in structural performance or external inputs as system anomalies; moreover, current anomaly identification methods have fixed time scales and are inflexible in application. To this end, this paper proposes an anomaly identification method based on multilayer perceptron: the method firstly identifies the system anomaly monitoring data through the intrinsic correlation of the multichannel monitoring data; then discerns whether the data are system anomalies through the reconstruction error of the monitoring data, thus avoiding the workload-intensive and cumbersome labelling of the dataset; and finally precisely locates the system anomaly data through the variable data sequence length of the input and output to accurately locate the system anomaly data. The effectiveness of the anomaly identification method proposed in this paper has been verified by MATLAB simulation data of simply supported girders and ship impact monitoring data of long bridges.