A platform for research: civil engineering, architecture and urbanism
A platform for research: civil engineering, architecture and urbanism
On the Interpretability of Machine Learning Using Input Variable Selection: Forecasting Tunnel Liner Yield
https://orcid.org/0000-0002-2669-6101
Abstract To validate the application of machine learning (ML) to rock engineering practice, it is crucial that algorithm developers use appropriate methods to quantify how closely the ML reproduces the observed rock mass deformation. Input variable selection (IVS) is one approach that examines how ML uses the given data, or inputs, to forecast rock mass behavior. Three IVS methods were developed for two convolutional neural network (CNN) architectures that predict tunnel liner yield at the Cigar Lake Mine, which exhibits time-dependent squeezing deformation. One model architecture focused on accurately predicting the higher tunnel liner yield classes, while the second architecture prioritized prediction accuracy across all tunnel liner yield classes. The three IVS methods investigated herein were channel activation strength (CAS), input omission (IO), and partial correlation (PC). The IO and PC approaches proposed are novel approaches proposed for CNNs using a spatial and temporal geomechanical dataset. Performance of all models was compared using the corrected Akaike information criterion ($ AIC_{c} $), where lower values indicate better performance. Each IVS method was used to produce a unique ranking for each model architecture and training/testing data split: CAS produced an activation ranking, IO produced an Omission Ranking, and PC produced a correlation ranking. The activation rankings showed that the geology input had the lowest activation strength in the CNN relative to the other inputs (ground freezing, primary installed support class, and radial tunnel displacement). Geology had the highest omission ranking, resulting from it having the most negative impact on performance as compared to the other inputs when it was omitted from the models entirely. The PC approach, using the Correlation Rankings, found that the highest model performances were reached when the most recent radial tunnel displacement was added into the pool of candidate inputs. The three IVS approaches and their respective rankings proved to be useful for analyzing the CNN inputs in terms of importance and confirming underlying assumption about the deformation mechanics at Cigar Lake Mine. Collectively, the IVS analyses indicated that all of the available digitized inputs for the Cigar Lake Mine CNNs are needed to produce good model performances. Each IVS method revealed different insights into this CNN development. Undertaking IVS for ML developed using geomechanical datasets allows for verification of the algorithms and thereby a better understanding of the nuance of the rock mass deformation. At Cigar Lake Mine, these findings may be used to assist in forecasting the schedule and budget for ground support rehabilitation.
Highlights Convolutional neural networks presented to predict tunnel liner yield in squeezing ground at Cigar Lake Mine.Three input variable selection (IVS) methods are proposed to examine machine learning (ML) input saliency.Channel activation strength method is adapted for CNNs; novel input omission and partial correlation methods are developed and proposed.IVS reveals input dominance, whether to remove redundant data, and temporal effects of available inputs.The authors suggest that IVS become standard practice to validate ML for rock mechanics and rock engineering.
On the Interpretability of Machine Learning Using Input Variable Selection: Forecasting Tunnel Liner Yield
https://orcid.org/0000-0002-2669-6101
Abstract To validate the application of machine learning (ML) to rock engineering practice, it is crucial that algorithm developers use appropriate methods to quantify how closely the ML reproduces the observed rock mass deformation. Input variable selection (IVS) is one approach that examines how ML uses the given data, or inputs, to forecast rock mass behavior. Three IVS methods were developed for two convolutional neural network (CNN) architectures that predict tunnel liner yield at the Cigar Lake Mine, which exhibits time-dependent squeezing deformation. One model architecture focused on accurately predicting the higher tunnel liner yield classes, while the second architecture prioritized prediction accuracy across all tunnel liner yield classes. The three IVS methods investigated herein were channel activation strength (CAS), input omission (IO), and partial correlation (PC). The IO and PC approaches proposed are novel approaches proposed for CNNs using a spatial and temporal geomechanical dataset. Performance of all models was compared using the corrected Akaike information criterion ($ AIC_{c} $), where lower values indicate better performance. Each IVS method was used to produce a unique ranking for each model architecture and training/testing data split: CAS produced an activation ranking, IO produced an Omission Ranking, and PC produced a correlation ranking. The activation rankings showed that the geology input had the lowest activation strength in the CNN relative to the other inputs (ground freezing, primary installed support class, and radial tunnel displacement). Geology had the highest omission ranking, resulting from it having the most negative impact on performance as compared to the other inputs when it was omitted from the models entirely. The PC approach, using the Correlation Rankings, found that the highest model performances were reached when the most recent radial tunnel displacement was added into the pool of candidate inputs. The three IVS approaches and their respective rankings proved to be useful for analyzing the CNN inputs in terms of importance and confirming underlying assumption about the deformation mechanics at Cigar Lake Mine. Collectively, the IVS analyses indicated that all of the available digitized inputs for the Cigar Lake Mine CNNs are needed to produce good model performances. Each IVS method revealed different insights into this CNN development. Undertaking IVS for ML developed using geomechanical datasets allows for verification of the algorithms and thereby a better understanding of the nuance of the rock mass deformation. At Cigar Lake Mine, these findings may be used to assist in forecasting the schedule and budget for ground support rehabilitation.
Highlights Convolutional neural networks presented to predict tunnel liner yield in squeezing ground at Cigar Lake Mine.Three input variable selection (IVS) methods are proposed to examine machine learning (ML) input saliency.Channel activation strength method is adapted for CNNs; novel input omission and partial correlation methods are developed and proposed.IVS reveals input dominance, whether to remove redundant data, and temporal effects of available inputs.The authors suggest that IVS become standard practice to validate ML for rock mechanics and rock engineering.
On the Interpretability of Machine Learning Using Input Variable Selection: Forecasting Tunnel Liner Yield
https://orcid.org/0000-0002-2669-6101
Abstract To validate the application of machine learning (ML) to rock engineering practice, it is crucial that algorithm developers use appropriate methods to quantify how closely the ML reproduces the observed rock mass deformation. Input variable selection (IVS) is one approach that examines how ML uses the given data, or inputs, to forecast rock mass behavior. Three IVS methods were developed for two convolutional neural network (CNN) architectures that predict tunnel liner yield at the Cigar Lake Mine, which exhibits time-dependent squeezing deformation. One model architecture focused on accurately predicting the higher tunnel liner yield classes, while the second architecture prioritized prediction accuracy across all tunnel liner yield classes. The three IVS methods investigated herein were channel activation strength (CAS), input omission (IO), and partial correlation (PC). The IO and PC approaches proposed are novel approaches proposed for CNNs using a spatial and temporal geomechanical dataset. Performance of all models was compared using the corrected Akaike information criterion ($ AIC_{c} $), where lower values indicate better performance. Each IVS method was used to produce a unique ranking for each model architecture and training/testing data split: CAS produced an activation ranking, IO produced an Omission Ranking, and PC produced a correlation ranking. The activation rankings showed that the geology input had the lowest activation strength in the CNN relative to the other inputs (ground freezing, primary installed support class, and radial tunnel displacement). Geology had the highest omission ranking, resulting from it having the most negative impact on performance as compared to the other inputs when it was omitted from the models entirely. The PC approach, using the Correlation Rankings, found that the highest model performances were reached when the most recent radial tunnel displacement was added into the pool of candidate inputs. The three IVS approaches and their respective rankings proved to be useful for analyzing the CNN inputs in terms of importance and confirming underlying assumption about the deformation mechanics at Cigar Lake Mine. Collectively, the IVS analyses indicated that all of the available digitized inputs for the Cigar Lake Mine CNNs are needed to produce good model performances. Each IVS method revealed different insights into this CNN development. Undertaking IVS for ML developed using geomechanical datasets allows for verification of the algorithms and thereby a better understanding of the nuance of the rock mass deformation. At Cigar Lake Mine, these findings may be used to assist in forecasting the schedule and budget for ground support rehabilitation.
Highlights Convolutional neural networks presented to predict tunnel liner yield in squeezing ground at Cigar Lake Mine.Three input variable selection (IVS) methods are proposed to examine machine learning (ML) input saliency.Channel activation strength method is adapted for CNNs; novel input omission and partial correlation methods are developed and proposed.IVS reveals input dominance, whether to remove redundant data, and temporal effects of available inputs.The authors suggest that IVS become standard practice to validate ML for rock mechanics and rock engineering.
On the Interpretability of Machine Learning Using Input Variable Selection: Forecasting Tunnel Liner Yield
Morgenroth, J. (author) / Perras, M. A. (author) / Khan, U. T. (author)
2022
Article (Journal)
Electronic Resource
English
BKL:
38.58
Geomechanik
/
56.20
Ingenieurgeologie, Bodenmechanik
/
38.58$jGeomechanik
/
56.20$jIngenieurgeologie$jBodenmechanik
RVK:
ELIB41
Selection and design of corrosion protection liner for precast concrete segment tunnel liner
| British Library Conference Proceedings | 2002
Mapping and interpretability of aftershock hazards using hybrid machine learning algorithms
| Elsevier | 2024
| Engineering Index Backfile | 1955
| Engineering Index Backfile | 1956