212
learning process. The weighting effects the strength of the signal at the connection. A node
might filter data by only allowing values above a threshold to continue. Nodes, or artificial
neurons, are arranged into layers with each layer typically performing a style of transformation (Bengio, 2015).
Convolutional Neural Network (CNN) is a neural network pattern that resembles the
organisation of an animal’s visual cortex. In the visual cortex individual neurons only respond
to stimuli within a restricted region (called the receptive field) in the field of view. The receptive fields overlap to give the animal a full coverage of the field of view. The major advantage
of this structure is the algorithm does not require human help with feature design. They
have been in wide use since early 1990s. One of the first use cases was zip code reading system for United States Postal Service developed by Bell Labs (Chollet, 2017). In recent years
CNNs have become a dominant computer vision technology which is enabling rapid growth
of automation and robotics.
A machine learning network called Mask R-CNN, developed by Facebook’s AI research team
(Region based Convolutional Neural Network mask regions with Convolutional Neural Networks),
is used for object detection and segmentation. Mask R-CNN is an approach to solve the problem of
instance segmentation where objects in an image are detected and delineated. Instance segmentation
is the combination of two problems: object detection and semantic segmentation. Object segmentation finds and classifies a variable number of objects in an image. While semantic segmentation is
understanding an image at the pixel level (https://research.fb.com/publications/mask-r-cnn/). For
example, which parts of a loaded haul truck image are the truck and which parts of the image are
the ore.
3 POTENTIAL APPLICATIONS
Many areas of the mining value chain could benefit from machine learning. This is particularly
true for the collection and initial analysis of geological observations and measurements.
Once a company has identified a resource target, they will invest large amounts of time
and money in discovering the physical distribution of the mineralization. The primary data
collection methods are outcrop or surface sampling, trench sampling and samples collected
from drill rigs. Samples from these methods are often chemically analysed and are described
by observations. The companies use this data to determine if the deposit is economically
viable using technical and business modelling techniques.
Samples are commonly photographed with a digital single lens and reflex mirror (DSLR)
camera as part of the collection procedure. A deposit commonly has a photo library of
images of thousands of meters of drill core. These photos are normally poorly annotated
with metadata and are simply used by the geologist as a visual reference to the sample to save
them the effort of looking at the physical sample (if it still exists).
The sample photos are normally photos of drill core or chips. Drill chips (Figure 1) are
the samples created by drilling techniques where a drill-bit fragments the rock and compressed air lifts these fragments to the surface. Core samples are cylinders of rock cut as
a diamond tipped steel sleeve (the drill-rod) cuts down into the rock. Photos of drill
cores (Figure 2) are normally photos of storage trays containing a length of the core (for
example, 5 meters of core). To associate the part of a core tray photo with a drill hole interval, the strips of core in the photograph must be extracted. This is a tedious manual process.
Core tray photos often contain metadata. The hole name and depth measurements are
often written on the tray or on blocks placed in the tray along with the core (Figure 3).
Extracting this data and storing it along with the core images allows the location of the
images in 3-dimensional space.
The logging of core is a time-consuming process and geologists can waste valuable time
logging core from uneconomic areas. Having a system that can identify zones of interest
within a core and direct a geologist to review those zones provides a more efficient utilisation
of the geological team’s time. By directing senior geological staff to view key zones of mineralisation, the overall quality of the logging will improve. Currently, the geologists logging
learning process. The weighting effects the strength of the signal at the connection. A node
might filter data by only allowing values above a threshold to continue. Nodes, or artificial
neurons, are arranged into layers with each layer typically performing a style of transformation (Bengio, 2015).
Convolutional Neural Network (CNN) is a neural network pattern that resembles the
organisation of an animal’s visual cortex. In the visual cortex individual neurons only respond
to stimuli within a restricted region (called the receptive field) in the field of view. The receptive fields overlap to give the animal a full coverage of the field of view. The major advantage
of this structure is the algorithm does not require human help with feature design. They
have been in wide use since early 1990s. One of the first use cases was zip code reading system for United States Postal Service developed by Bell Labs (Chollet, 2017). In recent years
CNNs have become a dominant computer vision technology which is enabling rapid growth
of automation and robotics.
A machine learning network called Mask R-CNN, developed by Facebook’s AI research team
(Region based Convolutional Neural Network mask regions with Convolutional Neural Networks),
is used for object detection and segmentation. Mask R-CNN is an approach to solve the problem of
instance segmentation where objects in an image are detected and delineated. Instance segmentation
is the combination of two problems: object detection and semantic segmentation. Object segmentation finds and classifies a variable number of objects in an image. While semantic segmentation is
understanding an image at the pixel level (https://research.fb.com/publications/mask-r-cnn/). For
example, which parts of a loaded haul truck image are the truck and which parts of the image are
the ore.
3 POTENTIAL APPLICATIONS
Many areas of the mining value chain could benefit from machine learning. This is particularly
true for the collection and initial analysis of geological observations and measurements.
Once a company has identified a resource target, they will invest large amounts of time
and money in discovering the physical distribution of the mineralization. The primary data
collection methods are outcrop or surface sampling, trench sampling and samples collected
from drill rigs. Samples from these methods are often chemically analysed and are described
by observations. The companies use this data to determine if the deposit is economically
viable using technical and business modelling techniques.
Samples are commonly photographed with a digital single lens and reflex mirror (DSLR)
camera as part of the collection procedure. A deposit commonly has a photo library of
images of thousands of meters of drill core. These photos are normally poorly annotated
with metadata and are simply used by the geologist as a visual reference to the sample to save
them the effort of looking at the physical sample (if it still exists).
The sample photos are normally photos of drill core or chips. Drill chips (Figure 1) are
the samples created by drilling techniques where a drill-bit fragments the rock and compressed air lifts these fragments to the surface. Core samples are cylinders of rock cut as
a diamond tipped steel sleeve (the drill-rod) cuts down into the rock. Photos of drill
cores (Figure 2) are normally photos of storage trays containing a length of the core (for
example, 5 meters of core). To associate the part of a core tray photo with a drill hole interval, the strips of core in the photograph must be extracted. This is a tedious manual process.
Core tray photos often contain metadata. The hole name and depth measurements are
often written on the tray or on blocks placed in the tray along with the core (Figure 3).
Extracting this data and storing it along with the core images allows the location of the
images in 3-dimensional space.
The logging of core is a time-consuming process and geologists can waste valuable time
logging core from uneconomic areas. Having a system that can identify zones of interest
within a core and direct a geologist to review those zones provides a more efficient utilisation
of the geological team’s time. By directing senior geological staff to view key zones of mineralisation, the overall quality of the logging will improve. Currently, the geologists logging