A VMD-Based Method for Outlier Correction in Structural Health Monitoring Data

Abstract

The quality of Structural Health Monitoring (SHM) data is paramount to the accuracy of structural condition assessment and service life prediction. However, monitoring data acquired in the field often exhibit significant non-stationarity and contain outliers due to environmental interference and other factors, posing severe challenges to subsequent data analysis. Traditional outlier detection methods often suffer from low accuracy and high false-positive rates when processing non-stationary signals, owing to interference from trend and periodic components. To address this issue, this study proposes a joint data correction framework based on Variational Mode Decomposition (VMD) and the Isolation Forest algorithm. The proposed method first utilizes the adaptive decomposition capability of VMD to decompose the original non-stationary signal into a series of Intrinsic Mode Functions (IMFs). The first component (IMF1) is extracted as the macroscopic trend of the signal, achieving efficient signal detrending. Subsequently, the Isolation Forest algorithm is applied to the detrended residual signal to accurately identify and locate outliers. Finally, linear interpolation is employed to correct the identified outliers. To validate the effectiveness of the proposed method, a synthetic dataset comprising trend, multi-periodic oscillations, and noise was constructed. Comparative experimental results demonstrate that the proposed VMD-Isolation Forest framework significantly enhances the accuracy and robustness of outlier detection compared to the direct application of the 3σ rule or the Isolation Forest algorithm. It effectively corrects anomalous disturbances while maximally preserving the intrinsic dynamic characteristics of the original signal. This research provides an efficient and reliable preprocessing paradigm for non-stationary SHM data, laying a solid data foundation for subsequent high-precision structural damage identification and performance prediction models.