| Abstrakt: |
Such is the need for this work, since accurate concrete strength prediction at different curing conditions is critical to having structures that are both strong and long-lasting. Traditional methods for the prediction of concrete strength often lack the complexity that occurs in the interaction of the different environmental factors involved, hence leading to suboptimal practices in curing with potential structural weaknesses. Current research into this area has typically focused on disparate data sources and rather naïve modeling methods, further limiting predictive accuracy and creating a general lack of comprehensive knowledge of curing dynamics. Such limitations bring out the need for a more integrated, sophisticated predictive modeling approach to explain variability in concrete strength levels. This paper proposes a novel predictive modeling framework that will be powered by advanced machine learning techniques to take up these challenges. It will adopt a multimodal data integration approach driven by a combination of sensor data related to temperature, humidity, and strain gauges; environmental data related to weather conditions and atmospheric pressure; and historical records, such as mix design and curing duration, further leveraging techniques from data fusion, including the Kalman filter and Bayesian networks. This will be further integrated into a unified, enriched dataset, encapsulating the complex interaction of factors influencing concrete strength. In the present work, this is a chosen approach: hybrid modeling with ensemble learning using XGBoost for the prediction of static features, and Long Short-Term Memory (LSTM) networks for capturing temporal dependencies. In this case, a combination of these models via weighted averaging or stacking improves the accuracy of the predictions to a very great extent: the R² increased from 0.85 to 0.92, and MAE levels by 10–15%. In addition to that, AutoML with Feature Tools implements advanced feature engineering through the generation and selection of optimal features on transformation and aggregation primitives, further refining model performance and interpretability. This process at times reduces the Root Mean Squared Error levels by 5–10%. Finally, Bayesian Neural Networks together with Sobol sensitivity analysis can be used for handling uncertainty and uncovering key factors. BNN provides probabilistic predictions, therefore 95% confidence intervals, while Sobol analysis identifies those critical features that contribute more to variability and allows an in-depth understanding of the role each factor has in driving concrete strength. Indeed, the framework propounded in this work has made great strides in predictive abilities concerning concrete strength variability and will permit more efficient curing practices, thereby making construction outcomes more secure and long-lasting. |
Full text from SpringerLink