Abstract
Rockmass classification systems are an integral part of today’s geotechnical design process. Many of these classification systems are however based on subjective or semiquantitative assessments which leads to a call for more objective classification systems.
One way to achieve this goal in mechanized (TBM) tunneling is to use the TBM operational data – or computed parameters thereof - as a basis to decide whether or not “regular advance” is at hand. We support this data driven approach by presenting MSAC (Multivariate sequence Segmentation, Abstraction and Classification) which is a computational framework that ultimately tells how “regular” the present TBM data is. MSAC utilizes several techniques of unsupervised machine learning and consists of multiple steps: 1. Computation of parameters like the specific penetration; 2. Preprocessing of the data (e.g. filtering, noise reduction); 3. Segmentation of the multivariate sequence into subsequences; 4. Abstraction/feature engineering from the subsequences; 5. Classification of the sequences by applying a statistical threshold on the Mahalanobis distances of each abstracted feature.
We compare MSAC to another TBM data driven rockmass classification system which is based on computation of the torque ratio and discriminates “hindered” from “regular” advance via a fixed threshold. The comparison shows that MSAC yields more comprehensible results, is less afflicted by construction site specific specialties and is less sensitive towards outliers, sensory malfunctions or data noise.
One way to achieve this goal in mechanized (TBM) tunneling is to use the TBM operational data – or computed parameters thereof - as a basis to decide whether or not “regular advance” is at hand. We support this data driven approach by presenting MSAC (Multivariate sequence Segmentation, Abstraction and Classification) which is a computational framework that ultimately tells how “regular” the present TBM data is. MSAC utilizes several techniques of unsupervised machine learning and consists of multiple steps: 1. Computation of parameters like the specific penetration; 2. Preprocessing of the data (e.g. filtering, noise reduction); 3. Segmentation of the multivariate sequence into subsequences; 4. Abstraction/feature engineering from the subsequences; 5. Classification of the sequences by applying a statistical threshold on the Mahalanobis distances of each abstracted feature.
We compare MSAC to another TBM data driven rockmass classification system which is based on computation of the torque ratio and discriminates “hindered” from “regular” advance via a fixed threshold. The comparison shows that MSAC yields more comprehensible results, is less afflicted by construction site specific specialties and is less sensitive towards outliers, sensory malfunctions or data noise.
Original language | English |
---|---|
Article number | 103466 |
Journal | Tunnelling and Underground Space Technology |
Volume | 103 |
DOIs | |
Publication status | Published - Sept 2020 |
Keywords
- TBM tunneling
- Unsupervised learning
- Data driven classification
- Mahalanobis distance
ASJC Scopus subject areas
- Geotechnical Engineering and Engineering Geology
- Building and Construction
Fields of Expertise
- Information, Communication & Computing