Abstract: The bridge structural health monitoring system often encounters data anomalies due to various interferences during collection,making it hard to accurately reflect the bridge’s health.To address these anomalies,we propose a method called VMD-KPCA-LSTM,which combines Variational Mode Decomposition (VMD),Kernel Principal Component Analysis (KPCA),and Long Short-Term Memory neural networks (LSTM).First,the collected data undergoes wavelet denoising and 3σ anomaly removal.Then,VMD decomposes the data into stable strain components,KPCA performs nonlinear dimensionality reduction,and LSTM predicts these components.The LSTM predictions are integrated to reconstruct the total strain time series.Compared to BP,GRU,LSTM,and VMD-PCA-LSTM models,the VMD-KPCA-LSTM model reduces MAPE by 19.948%,13.621%,11.724%,and 7.238% respectively.Thus,the VMD-KPCA-LSTM model is more effective for predicting strain anomaly data in cable-stayed bridges,providing a reliable data foundation for bridge health assessment and analysis.