Temperature effects on strain retrieval using machine learning: A field study

Temperature variations are known to dominate the quasi-static response of bridge structures and complicate structural health monitoring (SHM) tasks. This study investigates the feasibility of retrieving bridge strain responses using temperature measurements through machine learning (ML) techniques. A field dataset collected from a 27.5 m footbridge, including ten strain gauges and eleven temperature sensors, is evaluated. A novel temperature–strain separation method, combining Empirical Mode Decomposition and Moving Average Filtering (EMD–MAF), is proposed to isolate temperature-induced strain components and mitigate measurement noise and operational effects. Five scenarios are designed to assess the influence of thermal information on strain retrieval, considering single/multiple temperature sensors, historical temperature features, strain information from neighboring strain gauges, and their combinations. Four ML regression models: ridge regression (RR), light gradient boosting machine, XGBoost, and artificial neural networks, are trained and optimized using Bayesian hyperparameter tuning. Results show that the proposed EMD–MAF method successfully extracts thermal strain trends and reveals strong temperature–strain correlations. For strain retrieval, RR consistently provides the most reliable performance when using temperature data alone, while nonlinear models overfit and generalize poorly. When strain measurements from neighboring strain gauges are available, RR again yields the best generalization performance, as other strain gauges implicitly encode temperature and operational effects. Incorporating multi-sensor temperature data enhances model robustness, but additional historical temperature features do not significantly improve predictions in field conditions. The findings highlight the critical role of temperature in strain interpretation and demonstrate the practical limitations and capabilities of ML-based strain retrieval for real-world SHM systems.