A deep learning approach for rock fragmentation analysis

Highlights

• Deep neural network trained to predict sizes of rock fragments from an image.

• Data set composed of 61,853 labelled images of rock pile fragments.

• Percent error for coarse size prediction ranges within $\pm$25% for test set.

• 50% of the test set has a prediction percent error of $\pm$10%.

• Validation on sieved piles shows accurate prediction compared to image labelling.

Abstract

In mining operations, blast-induced rock fragmentation affects the productivity and efficiency of downstream operations including digging, hauling, crushing, and grinding. Continuous measurement of rock fragmentation is essential for optimizing blast design. Current methods of rock fragmentation analysis rely on either physical screening of blasted rock material or image analysis of the blasted muckpiles; both are time consuming. This study aims to present and evaluate the measurement of rock fragmentation using deep learning strategies. A deep neural network (DNN) architecture was used to predict characteristic sizes of rock fragments from a 2D image of a muckpile. The data set used for training the DNN model is composed of 61,853 labelled images of blasted rock fragments. An exclusive data set of 1,263 labelled images were used to test the DNN model. The percent error for coarse characteristic size prediction ranges within $\pm$25% when evaluated using the test set. Model validation on orthomosaics for two muckpiles shows that the deep learning method achieves a good accuracy (lower mean percent error) compared to manual image labelling. Validation on screened piles shows that the DNN model prediction is similar to manual labelling accuracy when compared with sieving analysis.

Introduction

The main objective of blasting in mines is to break in-situ rock mass to smaller rock fragments. More specifically, the goal is to achieve a specific fragment size distribution that eases handling, while minimizing damage to the final pit wall.¹ Fragmentation can affect the productivity and efficiency of downstream operations including digging, crushing, and grinding. To manage downstream effects, blast designs can be optimized through monitoring, analysis and modelling. Optimizing for cost, there are a range of close-to-optimal blast designs, but good blast design should adapt to the different rock mass conditions encountered at a mine site.^2,3

Fragmentation as one of the important blast outcomes has been the focus of numerous studies because it plays an important role in creating downstream benefits during blasting. Both prediction and measurement of rock fragmentation have been used as a basis for blast optimization. To model the effect of rock mass condition and blast design on fragmentation, many empirical models have been developed. Notable fragmentation models include the Kuznetsov,⁴ Kuz–Ram,⁵ extended Kuz–Ram,⁶ KCO,⁷ and $x_p$-frag⁸ models. More recently, fragmentation prediction has been reviewed in detail by Ouchterlony and Sanchidrián.⁹. The focus of fragmentation prediction includes characteristic sizes such as: $x_{50}$ (median, 50% weight passing), $x_{80}$, $x_{20}$ and $x_{\text{max}}$ (maximum size), uniformity factor ($n$) for the Rosin–Rammler distribution, and curve-undulation parameter ($b$) for the Swebrec function. The $x_p$-frag model proposed by Ouchterlony et al.⁸ emphasizes being distribution-free and only predicts characteristic sizes ($x_p$) so that the limitations of being fit to a specific distribution are reduced. Regardless of the model used, predicted parameters are commonly used to describe rock fragmentation with respect to fines generation, mid-range sizes, and oversize fraction. These studies acknowledge that predicted parameters will conform with trends, not absolute measures. To obtain evidence of an optimized blast, actual measurement is required.

Numerous techniques have been developed to measure fragmentation. Common methods include: qualitative visual observation, sieving, digital image analysis, and equipment monitoring. Visual observation and equipment monitoring methods provide inaccurate, qualitative and imprecise results. In the case of sieving, results are accurate but it is expensive and time-consuming. While digital image analysis methods have their own limitations, they have emerged as a common technique to measure fragmentation.¹⁰ Many image analysis approaches have been developed using different sensors and data processing techniques to estimate the rock size distribution of a captured rock pile surface. These include photography,¹¹ stereo photography,¹¹ and laser scanning.¹² Raina¹³ suggests that these methods can be grouped together as digital image analysis methods because they share similar limitations.

Major treatises have been published by the research community to describe digital image analysis methods and their limitations, namely those by Franklin and Katsabanis¹⁴ and Sanchidrián and Singh.¹⁵ Sanchidrián et al.¹⁰ suggests that image analysis techniques generally share four main sources of error: only sampling a surface to estimate internal characteristics, image quality, delineation of fragments, and estimation of fines. The most persistent limitation is poor/wrong rock segmentation which can result in disintegration and fusion of rock fragments.¹⁶ Due to this, extensive manual editing is usually required to correctly delineate fragments in captured rock images, a process that is time-intensive. As Ramezani et al.¹⁷ noted, the main challenge in rock segmentation is being robust when there are variations in lighting, image contrast, and complex rock texture and shape. The studies using stereo photogrammetry ^11,18 and laser scanning ¹⁹ present techniques to improve automated rock segmentation. While these techniques have improved rock segmentation, a number of other limitations still remain. For example, Sanchidrián et al.¹⁰ and Thurley¹⁹ both find that fines estimation still remains a major source of error in image analysis. To reduce this error when measuring muckpile fragmentation, Ouchterlony and Sanchidrián⁹ splices results from digital image analysis (＋10 cm) or in-pit sorting (＋2.5 cm) with laboratory sieving results. However, implementing sieve sampling or in-pit sorting methods are expensive and can disrupt production.

To increase the measurement frequency, area covered, and resolution of fragmentation measurement for muckpiles, Unmanned Aerial System (UAS) photography has been proposed by a number of studies.[20], [21], [22], [23] Through frequent data collection by UAS methods, the statistical reliability of the fragmentation measurement can be improved, as more samples are collected to understand population characteristics. However, these benefits are significantly hindered because poor/wrong automated rock segmentation has to be corrected through extensive manual editing. Ramezani et al.¹⁷ and Schenk et al.²⁴ used deep neural networks (DNNs) as a first step in fragment segmentation to improve automated delineation. Their methods and results are discussed in more detail in Section 2. The results from Ramezani et al.¹⁷ and Schenk et al.²⁴ have enabled fast and automated measurement but their strategy requires further investigation to better understand the accuracy and limitations of using DNNs for fragmentation measurement. Also, it should be noted that any limitations presented by data used to train DNNs are transferred to their results. For example, a DNN trained using 2D images will only be able to sample the surface of the pile.

This study presents the results of using deep learning strategies for rock fragmentation analysis. A convolution neural network architecture has been trained to predict scaled characteristic sizes of blasted rock fragments directly from a 2D image using an end-to-end deep learning strategy. The study evaluates the accuracy and performance of the DNN model as a tool for automated and fast rock fragmentation analysis. The outcomes of this evaluation demonstrate $\pm$25% percent error for coarse size prediction on the test set where 50% of the test set has a percent error of $\pm$10%. Validation of the DNN model on sieved piles shows accurate prediction compared to manual image labelling.