Reducing the Subjectivity of CMRR Calculation Using Machine Learning

The Coal Mine Roof Rating (CMRR) is a key input parameter for roof support design. This work aims to reduce the subjectivity of CMRR calculation by integrating geophysical logging data and laboratory data through the application of advanced machine learning algorithms, and to apply the algorithms to predict other rock mechanical properties, which are needed for geotechnical analysis. The opportunity identified to apply machine learning in the CMRR calculation includes the uniaxial compressive strength (UCS), rock quality designation (RQD), fracture spacing (FS) and the Geotechnical Units (GTUs).

The UCS: The correlation between the geophysical log data and the UCS has been established for supervised machine learning modelling. In this project, the “prediction based on group classification” concept was adopted, and a novel Group-based Machine Learning (GML) method was introduced to predict the UCS, which has demonstrated better performance than the conventional machine learning methods. The data analysis procedure to achieve high-quality training input data was demonstrated, and the techniques of data cleaning before the training were recommended. The implementation of the GML method requires using unsupervised learning models to classify the group of rock types firstly, and then the UCS values are predicted by the machine learning models trained for different groups; for each group, multiple machine learning models are evaluated. Finally, the previous two steps were integrated for the automatic group-based UCS prediction from the geophysical logs.

The RQD and Fracture Spacing (FS): An advanced computer vision model (i.e., Mask RCNN) trained by transfer learning technique has been adopted to extract the dimensions of the core pieces for the automatic RQD and FS calculation based on the drill core images. The standard deep learning training, such as the supervised learning method for computer vision  tasks, requires a huge amount of training datasets (millions of labelled images) and a long training time on the high-end supercomputer equipped with a graphic process unit (GPU) or tensor process unit (TPU). In this project, we adopted the transfer learning to build the automatic RQD and FS algorithms. The benefits of the transfer learning include (i) it enables the knowledge gained while solving one task (e.g., ordinary object detection) to be transferred to a different but relevant task (e.g., drill core detection); (ii) it provided flexibility to use the pre-trained model as starting point for object detection and segmentation tasks on the customized images (e.g., drill core images). Specifically, a pre-trained computer vision model with pixel-level region-based deep convolutional neural networks (Mask-RCNN) was fine-tuned for the detection and instance segmentation of the intact core plugs. The essential algorithm was built to extract the dimensions of the individual core segments so that the rock quality designation (RQD) and fracture spacing (FS) can be automatically calculated with corresponding depth registration. In addition, we have also developed an algorithm to process the multiple core box photos with promising results. In the test, 111 core box photos were processed in 14 minutes, which took about 7.5 seconds on average to process one core box image.

The Geotechnical Units (GTUs) were classified in terms of the adjacent sedimentary layers, which share similar geotechnical properties. The geotechnically relevant parameters were used to perform the automatic clustering by using the unsupervised learning method. The unsupervised learning method K-means has been used for the GTU classification, where the geotechnically relevant data were selected as input data, including the sonic transit time (STT), bulk density, UCS, RQD, FS, lithofacies, and geophysical strata rating (GSR). The classification was achieved automatically based on the clustering for the nearest data points in the high dimension space. The optimal group numbers were selected by the elbow-curve model. Among the input datasets, the UCS and RQD were computed by the machine learning algorithms; the lithofacies, GSR were calculated from the geophysical log data. The algorithms for the quantitative lithology and qualitative lithofacies interpreted from the geophysical logs were established. The machine learning methods were also established for predicting lithofacies. The geophysical strata rating (GSR) calculation algorithms were revisited, and the details of the algorithms were explicitly reformulated. The corresponding machine learning modelling was also developed for the GSR prediction.  

CMRR: the predicted quantities (UCS, RQD, FS) by the machine learning methods were converted into the corresponding ratings, which were further used for the unit ratings and raw CMRR calculation. Our results demonstrate that raw CMRR values from machine learning models have a good linear correlation (R²= 0.78) with the manually calculated CMRR values by site geotechnical engineers, indicating that the machine learning approach has the potential to be applied as an alternative and benchmark to ensure the used CMRR values are reliable and objective.

All the functions described in this report have been achieved in python programming. However, it would require additional work to further enhance the algorithms with more training data and to package the functional modules into a user friendly interface to suit different technical background users.