AbstractBridge modal frequency is an important parameter reflecting its overall property change and widely used for bridge condition assessment. However, the effects of multiple environmental conditions on the modal frequency will mask the variation induced by structural damage. Traditional single regression models cannot quantify measurable and unmeasurable environmental effects simultaneously, resulting in poor prediction and separation performance. Therefore, a two-stage elimination model (TSEM) integrating regression analysis and trend decomposition technique was developed. Environmental principal components (PCs) sensitive to the single-order modal frequency were extracted based on partial least-squares analysis. To quantify the nonlinear effects of measurable environmental factors, the baseline predictor with respect to modal frequency and environmental PCs was constructed through relevance vector machine technology. An error compensation model based on singular spectrum analysis was established to extract trend-related components and remove the part of residual modal variability unknot considered by the baseline model. On this basis, exponential weighted moving average control chart was established to highlight slight abnormal changes in modal frequency. A cable-stayed bridge case verified its validity and accuracy. The results indicate that the proposed TSEM has better modeling, generalization, and separation performance than the baseline model, and the variation of normalized frequency tends to be more stable. Additionally, the significant differences of damage sensitivity of different orders were determined.