Loading...
Improved tunneling knowledge through robust machine learning
Maher, James I.
Maher, James I.
Citations
Altmetric:
Advisor
Editor
Date
Date Issued
2015
Date Submitted
Collections
Research Projects
Organizational Units
Journal Issue
Embargo Expires
2016-08-31
Abstract
Earth Pressure Balance Machines (EPBMs) are essential equipment for excavating and constructing underground tunnels in urban environments with soft ground conditions. As examples, EPBMs are used for subways, underground highways, and water conduits. Our work utilizes data collected from hundreds of sensors on an EPBM to understand which systems affect the EPBM's performance. The ultimate goal is to optimize these systems in future tunneling projects, reducing project costs and construction time. We apply machine learning techniques to two data sets from EPBM excavated tunnels in the Seattle, WA, University Link subway project (U230 Contract). Specifically, we apply ensemble feature selection to identify sensor readings that are correlated with changes in the EPBM's advance rate. We found that current ensemble feature selection methods are insufficient for our data sets; thus, we created a novel ensemble feature selection method, JENNA Ensemble Network Normalization Algorithm (JENNA). JENNA allows diversity in the configurations of Feature Selection Algorithms, allows for regression FSAs, and enables both subset and ranker FSAs to be used simultaneously. JENNA also introduces a novel ensemble feature selection aggregation function that weights each feature by predicted accuracy performance and feature stability, in addition to average feature selection algorithm ranking. During our initial work, we identified a time delay between changes to some EPBM machine parameters and when these changes affect the EPBM's advance rate. In order to account for the time delay, we trained Recurrent Neural Networks (RNNs) to the data set. We then created a novel anomaly detection algorithm Recurrent Neural Network Anomaly Detection Algorithm (ReNN AnD) based on the trained RNNs. ReNN AnD varies from traditional anomaly detection, because it accounts for time delays in the data set. We used ReNN AnD to successfully detect soil at the front of an EPBM. This soil type information could be used to improve an EPBM's performance in future tunneling projects.
Associated Publications
Rights
Copyright of the original work is retained by the author.