TBM big data preprocessing method in machine learning and its application to tunneling

Abstract

The big data generated by tunnel boring machines (TBMs) are widely used to reveal complex rock-machine interactions by machine learning (ML) algorithms. Data preprocessing plays a crucial role in improving ML accuracy. For this, a TBM big data preprocessing method in ML was proposed in the present study. It emphasized the accurate division of TBM tunneling cycle and the optimization method of feature extraction. Based on the data collected from a TBM water conveyance tunnel in China, its effectiveness was demonstrated by application in predicting TBM performance. Firstly, the Score-Kneedle (S-K) method was proposed to divide a TBM tunneling cycle into five phases. Conducted on 500 TBM tunneling cycles, the S-K method accurately divided all five phases in 458 cycles (accuracy of 91.6%), which is superior to the conventional duration division method (accuracy of 74.2%). Additionally, the S-K method accurately divided the stable phase in 493 cycles (accuracy of 98.6%), which is superior to two state-of-the-art division methods, namely the histogram discriminant method (accuracy of 94.6%) and the cumulative sum change point detection method (accuracy of 92.8%). Secondly, features were extracted from the divided phases. Specifically, TBM tunneling resistances were extracted from the free rotating phase and free advancing phase. The resistances were subtracted from the total forces to represent the true rock-fragmentation forces. The secant slope and the mean value were extracted as features of the increasing phase and stable phase, respectively. Finally, an ML model integrating a deep neural network and genetic algorithm (GA-DNN) was established to learn the preprocessed data. The GA-DNN used 6 secant slope features extracted from the increasing phase to predict the mean field penetration index (FPI) and torque penetration index (TPI) in the stable phase, guiding TBM drivers to make better decisions in advance. The results indicate that the proposed TBM big data preprocessing method can improve prediction accuracy significantly (improving R2s of TPI and FPI on the test dataset from 0.7716 to 0.9178 and from 0.7479 to 0.8842, respectively).