Long-term hygrothermal performance assessment of on-site wood-framed walls based on sensor monitoring and machine learning

Abstract

Moisture damage poses a significant threat to building envelopes, leading to mould growth and reduced thermal insulation, thereby affecting indoor air quality and human health. Wood, as a promising eco-friendly building material, is highly sensitive to moisture, making hygrothermal monitoring of wooden walls essential. However, sensors used for monitoring are prone to failures, and the high maintenance and replacement costs make long-term monitoring challenging. Therefore, this study combines short-term monitoring data from sensors within the wall with outdoor climate data to predict long-term hygrothermal responses using machine learning (ML) models after monitoring ended. The ML model, which consists of a one-dimensional convolutional neural network (1D-CNN) and a long short-term memory (LSTM) network, was trained using two years of monitoring data and outdoor climate data from a four-story timber-framed office building. Subsequently, the SHapley Additive exPlanation (SHAP) method was employed to interpret the impact of each feature on the model’s predictions. Finally, the ML model was used to predict the hygrothermal responses inside the wall for three years after monitoring ended, and the mould growth risk for walls in different orientations was assessed using the predicted data. The study found that during the monitoring period, 97.9% of the test points showed no mould growth risk, and 2.1% showed low risk, indicating the wall assembly strong adaptability to Tianjin’s climate. The ML model performed excellently in predicting the temperature inside the wall, with an average R² of 0.952, and showed moderate accuracy in predicting relative humidity, with an average R² of 0.805. Predictions for three years after monitoring ends indicated that the maximum mould index for north-oriented walls reached 1.06 during heavy rainfall periods, posing a potential low risk. The method proposed in this study allows for long-term assessment by updating outdoor climate data, effectively utilizing data from sensors during short-term monitoring periods and serving as a cost-effective alternative after monitoring ends.