Time series clustering-enabled geological condition perception in tunnel boring machine excavation

Abstract

For automatic geology perception, this paper develops a DTW-Kmedoids-enabled time series clustering framework to gain a clear insight into the geological condition ahead of the tunnel boring machine (TBM) head. The core idea of DTW-Kmedoids is to incorporate dynamic time warping (DTW) as a distance metric into a popular clustering algorithm Kmedoids, aiming to aggregate time series data reflecting similar conditions of surrounding geology into the same cluster. It has been verified in a sequence dataset recording 377 operational rings from a TBM tunneling project in Singapore. Results indicate that the proposed DTW-Kmedoids algorithm outperforms Kmedoids and softDTW-Kmedoids in partitioning the whole dataset into four meaningful clusters on behalf of four geological conditions. Moreover, it shows merits in great robustness to handle varying-length temporal sequences and even fragmentary data instead of the full observations. Furthermore, an online clustering mechanism is well designed to adaptively update the cluster center, which can accomplish the dynamic recognition of the forward geological circumstances ring by ring. In short, the practical value behind the hybrid time series clustering is to automatically and quickly perceive lithology categories along the designed tunnel trail with no requirement of data labeling, possibly allowing for further inference of geological risk and guidance in safe excavation.