Technical Market Support » Metallurgical Coal
This project carried out an explorative study into the use of a 3D visualisation tool to help design spatial statistics related to the clustering of microstructure and high stress locations in coke. The points used in the analysis were generated from an existing stress model and from two AI approaches that highlighted locations of interest in the CT images. Overall the project explored the application of four novel approaches for working with coke volumes and to support better understanding of coke microstructure, including:
- Advanced visualisation platforms and tools to interact with coke volumes;
- Deep Learning approaches to generate points of interest from the coke volumes;
- Isomapping and clustering algorithms to generate points of interest from the coke volumes; and
- 3D statistical analysis to characterise the distribution of features in the coke volumes.
After evaluating a range of possible visualisation tools, the Drishti software platform (Limaye 2012) was selected for the project. Both a large touch screen and the iDome platform were successfully used with Drishti to support collaboration among the experts in this project. Used together these platforms provided extensive support for interactive volumetric visualisation of the 3D CT image stacks generated from scanning in the synchrotron.
The project was successful in demonstrating both the feasibility and effectiveness of applying immersive volumetric tools for interacting with coke micro CT images. The support for 3D, versus 2D visualisation was considered to provide significant assistance for the collaborative understanding of complex 3D microstructure. Drishti was found to be a relatively complex tool to learn to use and new users may require some time (~ 8 weeks) to become familiar with the work processes and features available. Some of the detailed functionality of the Drishti tool likely benefits from an appreciation of computer graphics. Overall, it is recommended that Drishti continues to be utilised in future projects using volumetric data. It should be noted that despite good support from the tool, developing good visualisations can still be a creative, time-consuming and iterative process that requires a visualisation expert and a domain expert to provide direction. While Drishti has excellent functionality and is currently well-supported by NCI group, there is always some risk with a non-commercial, specialist tool that relies on advanced hardware that the tool may not be maintained into the future.
Characterising the statistical distribution of locations in the coke volumes first requires the locations to be selected. Manual selection of these locations is too time consuming and prone to interpretation bias. With this in mind, two specific data-mining approaches (Convolution Neural Nets (CNN) and isomapping), both designed to automate this feature selection process were investigated in the project. One approach relied on recent advances in Deep Learning to automatically recognise features in large sets of images. The second approach developed an isomap and a feature-based distance metric that allowed clustering of structural features in the coke samples. Both approaches could be used to automatically identify features of interest in the coke samples and thus generate 3D point locations that could then be statistically characterised. The isomapping process had the advantage of being more readily linked to structural features and lower technical risk than the CNN.
To perform the statistical characterisation of features an existing Matlab tool, called 3DSpatialAnalysis was adapted for the project. This package generates a range of 3D statistical measures from a sample of point locations. This tool highlighted differences in the three cokes related to both the Deep Learning (CNN) heatmap locations and also the isomap class features. This is a promising result but needs to be validated further on a broader range of data.
Overall, it is recommended that selected outcomes from this project should be applied across a broader range of samples to validate findings and allow better correlation of the results with traditional coke strength measures. Of the two approaches, the isomapping technique contains less risk and also provides the most transparent link back to structural features in the coke. This in turn supports understanding and explaining the process at a fundamental level more seamless.