In a recent study published in the journal Scientific Reports, Japanese researchers introduced an innovative approach to optimize the performance of tunnel boring machines (TBMs) in soft ground conditions.
Study: Tunnel Boring Machine Efficiency: AI Predictions for Jack Speed and Torque. Image Credit: Oleg Totskyi/Shutterstock
They utilized different machine learning models to predict the optimal jack speed and torque settings for a micro slurry tunnel boring machine (MSTBM) based on the operator’s decisions and the machine’s logged data. Moreover, the research showed the potential of artificial intelligence (AI) to optimize MSTBM parameters and improve the performance of the machine.
Background
TBMs are advanced engineering machines designed for automated tunnel excavation. They can penetrate diverse geological formations, ranging from soft soil and clay to resilient rock and abrasive materials. Additionally, they offer several advantages over traditional tunneling methods, such as reducing manual labor, enhancing worker safety, and increasing construction speed.
However, TBMs face several challenges including adapting to varying ground conditions, optimizing operational parameters to ensure smooth excavation progress, and mitigating the risk of machine jamming or entrapment in complex geological formations. The unpredictable nature of subsurface conditions necessitates continuous monitoring and adjustment of operational parameters to maintain optimal performance and prevent potential setbacks.
Among these parameters, jack speed and torque play a pivotal role in controlling the machine's advancement through the tunnel. Therefore, accurate prediction and real-time monitoring of these parameters are essential for maximizing TBM excavation efficiency and ensuring project success.
About the Research
In the present paper, the authors focused on a specific type of TBM, namely the MSTBM, utilized for umbrella pipe support excavation in soft ground environments. The MSTBM incorporates a rotating cutter head, efficient cutting tools, and a slurry system that transports the excavated material via pipelines and provides soil treatment and pressure control at the cutter face. Operational parameters like jack speed and torque are traditionally adjusted manually by operators, leading to potential performance variations due to human error.
To mitigate this issue, the researchers proposed an AI-driven approach to synchronize the operator data and the MSTBM-logged data into a single data frame and then applied machine learning models to establish a robust correlation between the operator’s decision and the machine’s responses.
They collected data from an ongoing tunneling project in Japan, where the MSTBM was used to install 24 pipe holes with 18 pipes each at the entrance of a highway bypass tunnel. This data included operator input (jack speed and torque settings) for each 250 mm digging stroke, MSTBM's logged data (actual jack speed and torque values), rust force, hydraulic pressure, and other parameters at 1-s intervals. Furthermore, these data were divided into training and testing sets and utilized for model validation.
The study utilized Optuna to dynamically refine parameters like jack speed and torque settings. Optuna is an advanced hyperparameter optimization platform that employs Bayesian optimization with a pruning mechanism and supports both single and multi-objective optimization.
Various machine learning models, including K-nearest neighbors (kNN), random forest (RF), decision tree (DT), extreme gradient boosting (XGBoost), artificial neural network (ANN), and support vector machine (SVM), were compared using Optuna. These models were trained and tested on 80% and 20% of the data, respectively. Moreover, the best model was selected based on metrics such as correlation coefficient (R2), mean squared error (MSE), and mean absolute error (MAE) metrics.
Research Findings
The outcomes showed that the RF model outperformed other models in predicting both jack speed and torque. For jack speed prediction, the RF model achieved an R2 value of 96%, an MSE of 119.7, and an MAE of 4.42. Regarding torque prediction, the RF model attained an R2 value of 83%, an MSE of 0.62, and an MAE of 0.42. Additionally, the RF model exhibited a stable learning curve and a symmetrical error distribution, indicating effective learning and generalization ability. The study identified the max depth hyperparameter as the most influential factor in determining model performance across all model types.
The authors showcased the potential of using machine learning models to optimize the MSTBM’s operational parameters based on operator decisions and machine data. This approach can aid MSTBM operators in making informed decisions regarding jack speed and torque control, thereby preventing excessive or reduced tunneling speeds and enhancing overall performance and efficiency. Furthermore, the method's applicability extends to other TBM types, tunneling conditions, and engineering applications involving human-machine interaction.
Conclusion
In summary, the novel approach effectively enhanced operator monitoring and MSTBM data comprehension using machine learning models and Optuna-assisted hyperparameter optimization. The researchers emphasized the importance of data synchronization and hyperparameter tuning for achieving accuracy and efficiency in MSTBM control.
However, they acknowledged limitations including dependency on training data, interpretability challenges, adaptability to environmental changes, and ethical implications. Future research could mitigate these limitations and extend the model to diverse TBMs and tunneling projects.
Can AI Think Like Us?