A multi-objective optimisation method for correcting monitoring data and its engineering application

Engineering system evaluation relies on monitoring data and theoretical models. Effective integration of these data is essential for improving reliability and supporting engineering decision-making. This study proposes a multi-objective optimisation method for correcting monitoring data without fixed physical models. A physically interpretable indicator system consisting of deviation, physical similarity and smoothness is introduced to quantify correction quality. These indicators are integrated into a unified multi-objective formulation and solved using a neural network that fuses observations with model predictions. Numerical experiments using simulated data with artificially introduced noise show that the proposed method outperforms the Kalman filter and physics-informed neural networks in correction performance, robustness to model bias and resistance to observational noise. To further investigate the role of each objective, an ablation study is conducted by removing objective terms and by individually adjusting their weights. The results show that the physical similarity indicator improves the identification of critical transition regions, while the smoothness indicator suppresses spurious fluctuations. Optimal performance is achieved when the three objective terms are maintained at comparable orders of magnitude. A reservoir slope case study is performed to validate that the method can assess the relative reliability of individual observations, supporting subsequent engineering analysis and decision-making.