424 XXXI International Mineral Processing Congress 2024 Proceedings/Washington, DC/Sep 29–Oct 3
the regular RGB image is given together with the response
image from the hyperspectral camera. The system was first
calibrated to recognize the copper ore (sulfides) from the
regular rocks (sandstone), based on reflectance differences.
In this example, the system recognized the areas being
chemically similar to the copper ores shown by yellow color
and where the waste particles were shown by blue color. It
can be noticed that some particles have a similar color in
the RGB image, while the hyperspectral camera recognizes
more different areas. Further chemical analysis showed that
green marked areas corresponded to particles with high Cu
concentrations.
Application of AI Models
Output signals from different sensors can be processed by
AI models to provide the best mineral or material identi-
fication. Identification in such a case works by detecting
features hidden in the convolutional layers of the neural
network. Due to the limited image analysis time, which is
on average 33 ms, these filters are appropriately optimized
in terms of the number of neurons in hidden layers, the
type of layers and the selection of hyperparameters. Then
they go to the Dence layers, through which the identified
features are inferred from the convolutional layers.
In case of the selection of hyperparameters it is recom-
mended to use the Random Search. This recommendation
results from experience gained while working with data
received. This selection can be also made using the opti-
mization of hyperparameters, e.g., Random Search, Grid
Search or Baysian optimization (Madsen, 2018). A detailed
description of classical hyperparameter optimization for
the presented problem is described in (Progorowicz et al.,
2022).
The correct selection of neural networks in the pre-
sented case takes into account the overfitting, underfitting
and robustness analysis. For the first two analyzes given,
the goal is to identify networks capable of inferring general
material patterns with minimal noise mapping ability. In
turn, robustness analysis aims to identify such a structure
that, in the event of changes in input data, is able to gener-
ate correct responses for the device’s input material evalua-
tion system.
The examination of robustness of neural network
structures, according to position (Dudzik, 2020), is carried
out for the network validation stage. In turn, overfitting
and underfitting study is performed based on the results
obtained from the neural network testing stage.
For the purposes of this studies, analyzes should
be performed using certain indicators, e.g., R (Pearson
Correlation Coefficient) (Philip,2012), R2 (Coefficient of
determination) (Dudzik, Stręk, 2020), MSE (Equation 1),
MAE (Equation 2), SAE (Equation 3), SSE (Equation 4),
MaxARE (Equation 5), MARE (Equation 6).
MAE n y y 1
NNi
i
n
i
1
=-
=
/(1)
SAE yi y
NNi
i
n
1
=-
=
/(2)
SSE y y
NNi
i
n
i
1
2 =-
=
/^h (3)
maxc MaxARE y
y yNNi
i
i =
-
m or
%MaxARE MaxARE100% =6 @(4)
MARE n y
y yNNi
n y
ei 1 1
i
n
i
i
i
n
i 1 1
=
-
=
==
//(5)
The approximate number of network training processes
performed in the process of network structure optimiza-
tion exceeds 50,000 structures. Thanks to such a complex
process, satisfactory results are obtained (Progorowicz et al.,
2022).
The network training process consists of the train-
ing, validation and testing stages of the network. Neural
Figure 5. RGB image (left) and hyperspectral camera response (right) indicating different zones of copper
ore (green color)
the regular RGB image is given together with the response
image from the hyperspectral camera. The system was first
calibrated to recognize the copper ore (sulfides) from the
regular rocks (sandstone), based on reflectance differences.
In this example, the system recognized the areas being
chemically similar to the copper ores shown by yellow color
and where the waste particles were shown by blue color. It
can be noticed that some particles have a similar color in
the RGB image, while the hyperspectral camera recognizes
more different areas. Further chemical analysis showed that
green marked areas corresponded to particles with high Cu
concentrations.
Application of AI Models
Output signals from different sensors can be processed by
AI models to provide the best mineral or material identi-
fication. Identification in such a case works by detecting
features hidden in the convolutional layers of the neural
network. Due to the limited image analysis time, which is
on average 33 ms, these filters are appropriately optimized
in terms of the number of neurons in hidden layers, the
type of layers and the selection of hyperparameters. Then
they go to the Dence layers, through which the identified
features are inferred from the convolutional layers.
In case of the selection of hyperparameters it is recom-
mended to use the Random Search. This recommendation
results from experience gained while working with data
received. This selection can be also made using the opti-
mization of hyperparameters, e.g., Random Search, Grid
Search or Baysian optimization (Madsen, 2018). A detailed
description of classical hyperparameter optimization for
the presented problem is described in (Progorowicz et al.,
2022).
The correct selection of neural networks in the pre-
sented case takes into account the overfitting, underfitting
and robustness analysis. For the first two analyzes given,
the goal is to identify networks capable of inferring general
material patterns with minimal noise mapping ability. In
turn, robustness analysis aims to identify such a structure
that, in the event of changes in input data, is able to gener-
ate correct responses for the device’s input material evalua-
tion system.
The examination of robustness of neural network
structures, according to position (Dudzik, 2020), is carried
out for the network validation stage. In turn, overfitting
and underfitting study is performed based on the results
obtained from the neural network testing stage.
For the purposes of this studies, analyzes should
be performed using certain indicators, e.g., R (Pearson
Correlation Coefficient) (Philip,2012), R2 (Coefficient of
determination) (Dudzik, Stręk, 2020), MSE (Equation 1),
MAE (Equation 2), SAE (Equation 3), SSE (Equation 4),
MaxARE (Equation 5), MARE (Equation 6).
MAE n y y 1
NNi
i
n
i
1
=-
=
/(1)
SAE yi y
NNi
i
n
1
=-
=
/(2)
SSE y y
NNi
i
n
i
1
2 =-
=
/^h (3)
maxc MaxARE y
y yNNi
i
i =
-
m or
%MaxARE MaxARE100% =6 @(4)
MARE n y
y yNNi
n y
ei 1 1
i
n
i
i
i
n
i 1 1
=
-
=
==
//(5)
The approximate number of network training processes
performed in the process of network structure optimiza-
tion exceeds 50,000 structures. Thanks to such a complex
process, satisfactory results are obtained (Progorowicz et al.,
2022).
The network training process consists of the train-
ing, validation and testing stages of the network. Neural
Figure 5. RGB image (left) and hyperspectral camera response (right) indicating different zones of copper
ore (green color)