Deep Learning Algorithms for the Monitoring and Control of Drilling and Grouting Works

Efficient logging and precise control of drilling and grouting operations are essential in construction of tunnels, dams, levees, and confinement of polluted areas. With the growing demands of digital solutions supporting the construction process, the integration of advanced logging and automated control systems of grouting parameters emerges as an innovative strategy to enhance quality and productivity in project execution. This study delves into the potential of employing machine learning models to accurately predict and regulate the pressure and flow rate within a grouting system. We introduce an innovative data logging system capable of producing extensive datasets on grouting parameters suitable for Deep Learning algorithms. Leveraging two distinct datasets gathered at the Mosul Dam (Iraq) and Rogun Dam (Tajikistan) Projects, we conduct a comparative analysis of two Artificial Neural Network architectures: one based on feedforward design, and the other on long short-term memory (LSTM). To the best of our knowledge, these datasets are larger than any other state-of-the-art (over 55,000,000 data points from approx. 7000 grout holes for Mosul Dam and 8,500,000 data points from approx. 3000 grout holes for Rogun Dam). Having such large datasets allowed us to obtain a realistic estimation of the model performance, without potential bias due to missing ground features. Furthermore, we investigate the influential features contributing to effective estimation of grouting parameters using SHAP (SHapley Additive exPlanations), a game theoretic approach aimed at explaining machine learning model outputs.