An adaptive hierarchical decomposition-based method for multi-step cutterhead torque forecast of shield machine

Abstract

Cutterhead torque is generated by interaction between geological environment and shield machine, which is one of the main load parameters of shield machine during the tunneling process. Its accurate forecast helps shield machine drivers adjust important operational parameters ahead of time for achieving safe and efficient tunneling. This paper presents an adaptive hierarchical decomposition-based method (AHDM) for multi-step forecast of cutterhead torque, where only original cutterhead torque signal is utilized as input and is decomposed adaptively to reduce its complexity and improve the forecast performance under complex geological environment and working conditions. First, matrix wavelet packet decomposition (MWPD) is developed to decompose original cutterhead torque signals into low-frequency and high-frequency sub-sequences. Then, an improved adaptive variational mode decomposition (IDVMD) is proposed for further decomposing the low-frequency sub-sequence into several simple series and residual series. Simultaneously, empirical wavelet transform is employed for dealing with the high-frequency sub-sequence. After adaptive hierarchical decomposition and significant complexity reduction of signal complexity, we design gated recurrent unit (GRU) based multi-step forecast network for each subsequence. Eventually, multi-step forecast of cutterhead torque is fulfilled by summing forecasting results for all sub-signals. To verify the developed AHDM, comparison with existing forecast methods in different five datasets is conducted. It is shown that the highest forecast accuracy obtained by the proposed AHDM reaches 99.030%, 98.721%, 98.653%, 98.113% and 97.674% from first-step to fifth-step, respectively. Moreover, the proposed AHDM outperforms current forecast methods on forecast accuracy for each step in 5 datasets. Especially, in the fifth-step forecast, forecast accuracy of AHDM is 6.424% higher on average than current methods. Meanwhile, when compared with current forecast methods, the RMSE and MAE indexes of the proposed AHDM reduce on average 72.937% and 71.315%, respectively. The results of k-fold cross validation show that proposed AHDM obtains strong robustness to different training and test set distributions. Therefore, for complex geological and working conditions, AHDM achieves high multi-step forecast accuracy, generalization ability and superiority.