<p>Machine learning algorithms (MLAs) are emerging as a powerful tool for f... more <p>Machine learning algorithms (MLAs) are emerging as a powerful tool for forecasting complex and nuanced rock mass behaviour, particularly when large, multivariate datasets are available. In engineering practice, it is often difficult for geomechanical professionals to investigate all available data in detail, and simplifications are necessary to streamline the engineering design process. An MLA is capable of processing large volumes of data quickly and may uncover relationships that are not immediately evident when manually processing data. This research compares two algorithms developed for two mines representing end member behaviours of rock failure mechanisms: squeezing ground with high radial convergence, and spalling ground with high in situ stresses and seismicity. For the squeezing ground case study, a Convolutional Neural Network is used to forecast the yield of the tunnel liner elements using tunnel mapping images as the input. For the high stress case study, a Long Short Term Memory network is used to forecast the in-situ stresses that takes time series microseismic events and geomechanical properties as inputs. The two case studies are used to compare input data requirements and pre-processing techniques. Ensemble modelling techniques used to quantify MLA uncertainty for both case studies are presented. The development of the two MLAs is discussed in terms of their complexity, generalizability, performance evaluation, verification, and practical applications to underground rock engineering. Finally, best practices for MLA development are proposed based on the two case studies to ensure model interpretability and use in engineering applications.</p>
As underground instrumentation improves and the storage of large volumes of data becomes more cos... more As underground instrumentation improves and the storage of large volumes of data becomes more cost effective, the rock engineering community has access to larger datasets than ever before. Machine learning algorithms (MLAs) present an opportunity to uncover nuanced rock mass deformation mechanics more efficiently than conventional data analysis tools, resulting in increased reliability of underground excavations. MLAs require appropriate pre-processing of inputs as well as ground truth validation of outputs. Convolutional Neural Networks (CNNs) are an MLA that allow for the preservation of spatial and temporal dependencies within a dataset. CNNs were developed for image recognition and segmentation, such as video processing, and are efficient at analyzing sequential snapshots of an excavation as the environmental and in-situ factors change. Herein a CNN is developed for Cigar Lake Mine, Saskatchewan, Canada, to predict tunnel liner yield. The mine experiences a complex time-dependent ground squeezing behaviour resulting from the poor geological conditions and the artificial ground freezing implemented to stabilize the ore cavities and to control ground water during the ore extraction process. A sensitivity analysis of the CNN training parameters, called hyperparameters, is completed to optimize the final CNN performance. Hyperparameters analyzed include: the amount of training data, the convolutional filter size, and the error weighting scheme. Two final models are developed, one balanced model able to accurately predict tunnel liner yield across all classes of severity, and one targeted model that is calibrated to predict the higher classes of tunnel liner yield particularly well. Model results demonstrate that the CNN is a promising tool for preserving the spatial and temporal dependencies between input variables, and for predicting tunnel liner yield. This is a novel approach for geomechanical datasets. In combination, the two final CNNs achieve a prediction precision of > 87% across all classes and a recall of up to 99.9% for the higher yield classes. The activation strengths of the inputs were studied, and it was determined that the primary installed support class is the most dominant predictor of tunnel liner yield.
<p>Machine learning algorithms (MLAs) are emerging as a powerful tool for f... more <p>Machine learning algorithms (MLAs) are emerging as a powerful tool for forecasting complex and nuanced rock mass behaviour, particularly when large, multivariate datasets are available. In engineering practice, it is often difficult for geomechanical professionals to investigate all available data in detail, and simplifications are necessary to streamline the engineering design process. An MLA is capable of processing large volumes of data quickly and may uncover relationships that are not immediately evident when manually processing data. This research compares two algorithms developed for two mines representing end member behaviours of rock failure mechanisms: squeezing ground with high radial convergence, and spalling ground with high in situ stresses and seismicity. For the squeezing ground case study, a Convolutional Neural Network is used to forecast the yield of the tunnel liner elements using tunnel mapping images as the input. For the high stress case study, a Long Short Term Memory network is used to forecast the in-situ stresses that takes time series microseismic events and geomechanical properties as inputs. The two case studies are used to compare input data requirements and pre-processing techniques. Ensemble modelling techniques used to quantify MLA uncertainty for both case studies are presented. The development of the two MLAs is discussed in terms of their complexity, generalizability, performance evaluation, verification, and practical applications to underground rock engineering. Finally, best practices for MLA development are proposed based on the two case studies to ensure model interpretability and use in engineering applications.</p>
As underground instrumentation improves and the storage of large volumes of data becomes more cos... more As underground instrumentation improves and the storage of large volumes of data becomes more cost effective, the rock engineering community has access to larger datasets than ever before. Machine learning algorithms (MLAs) present an opportunity to uncover nuanced rock mass deformation mechanics more efficiently than conventional data analysis tools, resulting in increased reliability of underground excavations. MLAs require appropriate pre-processing of inputs as well as ground truth validation of outputs. Convolutional Neural Networks (CNNs) are an MLA that allow for the preservation of spatial and temporal dependencies within a dataset. CNNs were developed for image recognition and segmentation, such as video processing, and are efficient at analyzing sequential snapshots of an excavation as the environmental and in-situ factors change. Herein a CNN is developed for Cigar Lake Mine, Saskatchewan, Canada, to predict tunnel liner yield. The mine experiences a complex time-dependent ground squeezing behaviour resulting from the poor geological conditions and the artificial ground freezing implemented to stabilize the ore cavities and to control ground water during the ore extraction process. A sensitivity analysis of the CNN training parameters, called hyperparameters, is completed to optimize the final CNN performance. Hyperparameters analyzed include: the amount of training data, the convolutional filter size, and the error weighting scheme. Two final models are developed, one balanced model able to accurately predict tunnel liner yield across all classes of severity, and one targeted model that is calibrated to predict the higher classes of tunnel liner yield particularly well. Model results demonstrate that the CNN is a promising tool for preserving the spatial and temporal dependencies between input variables, and for predicting tunnel liner yield. This is a novel approach for geomechanical datasets. In combination, the two final CNNs achieve a prediction precision of > 87% across all classes and a recall of up to 99.9% for the higher yield classes. The activation strengths of the inputs were studied, and it was determined that the primary installed support class is the most dominant predictor of tunnel liner yield.
Uploads
Papers by J M