Model Performance Improvement by Accumulated Application of Machine Data in Machine Learning Model for TBM Advance Rate Prediction

This study quantitatively verified the impact of applying accumulated data on the model’s prediction accuracy, overfitting, and adaptive learning ability by using a method that accumulates and retrains machine data of a TBM generated whenever excavation progresses at regular intervals. To achieve this, the performance of five machine learning algorithms was evaluated on two field datasets. The best-performing gradient boosting model was selected as the preliminary model. The performance results of the preliminary model and the cumulative model were then compared using another field dataset. The field data for the performance comparison were divided into 14 steps based on ground information, and the performance of the two models was compared sequentially at each step. The results showed that the preliminary and cumulative models exhibited similar predictive performance in the initial intervals. However, the cumulative model more closely matched actual measurements as new data was added than the preliminary model. Consequently, the preliminary model, based on past data, has clear limitations in adapting to the diverse variables encountered in real-world situations. On the other hand, cumulative models are essential for improving real-time prediction performance of processes with constantly changing environments, such as TBM, by continuously increasing relevant field data.