Data anomaly detection for structural health monitoring based on Markov transition matrix features and XGBoost

Structural health monitoring (SHM) systems produce enormous amounts of data over time. However, abnormal data caused by sensor breakdowns and environmental interference greatly reduce the Accuracy of structural state evaluations. This study develops and evaluates a low-dimensional feature representation strategy for anomaly detection in SHM data. First, a Markov transition matrix is constructed based on the first order differences of the raw time-series signals. The fluctuations patterns of the signal amplitude are effectively characterized in this step. Additionally, multidimensional features are directly extracted from the constructed matrices, including key characteristics such as information entropy, statistical properties, and steady-state distributions. Subsequently, the optimal feature subset is then chosen using the Permutation Feature Importance (PFI) approach. Finally, the XGBoost model is employed to identify and classify abnormal data. The results based on monitoring data obtained from a long-span arch bridge SHM system demonstrate that the proposed method achieves an overall classification Accuracy of 98.12%. Even in situations with little samples, the method has outstanding stability. Furthermore, monitoring data from offshore wind turbine SHM system is used to verify the robustness of the proposed approach. Finally, the cross-domain generalization capability of the proposed framework was evaluated by training on the bridge dataset and testing on the wind turbines dataset.