XXXI International Mineral Processing Congress 2024 Proceedings/Washington, DC/Sep 29–Oct 3 425
networks that correctly pass the overfitting, underfitting
and robustness assessment stages take part in the next pro-
cess of selecting the optimal network for the sorter system
(Figure 6).
Figure 7 shows example results of the overfitting and
underfitting study. The study was made on the basis of box
plots in accordance with the methodology presented in
(Dudzik, 2020).
It presents the top 10 structures from the optimal net-
work structure identification procedure. This procedure is
described in more detail in the publication (Progorowicz
et al., 2022). The assessment of the selection of a network
intended for a real facility mainly takes into account the
sample from the testing stage. For this stage, data is pre-
pared in such a way that data leakage does not occur. This
means that this data does not take part in both the training
and validation processes of the network.
After conducting analyzes based on box plots and net-
work performance results, the optimal network structure
is selected. The selection procedure is described in the ref-
erences (Dudzik, 2020) and (Progorowicz, et al., 2022).
Typically, after performance analyzes are performed, only a
few networks are selected from among over 50,000 neural
networks. These selected networks can be tested at a later
stage on the sorter software.
Figure 8 shows exemplary network regression (R)
results for the problem of identifying the zinc ore (Zn) and
lead ore content from a photo of a rock.
As can be seen, in the case of the testing stage, values
close to 0.9 (0.896 fof Zn and 0,884 do Pb), which is a
satisfying result.
For the network indicated in Figure 8, the mean abso-
lute errors calculated for training, test and validation data
sets are presented in Figure 9.
As shown in Figure 9, the vast majority of the network
results achieved error values close to 0.
SORTING RESULTS
The choice of sensors for sorting is very dependent on types
of sorted materials. Nevertheless, the XRT sensor can be
Figure 6. Convolutional AI models for final image analysis
Figure 7. Boxplots of the mean absolute error values from
tests stage (10 best structures)
networks that correctly pass the overfitting, underfitting
and robustness assessment stages take part in the next pro-
cess of selecting the optimal network for the sorter system
(Figure 6).
Figure 7 shows example results of the overfitting and
underfitting study. The study was made on the basis of box
plots in accordance with the methodology presented in
(Dudzik, 2020).
It presents the top 10 structures from the optimal net-
work structure identification procedure. This procedure is
described in more detail in the publication (Progorowicz
et al., 2022). The assessment of the selection of a network
intended for a real facility mainly takes into account the
sample from the testing stage. For this stage, data is pre-
pared in such a way that data leakage does not occur. This
means that this data does not take part in both the training
and validation processes of the network.
After conducting analyzes based on box plots and net-
work performance results, the optimal network structure
is selected. The selection procedure is described in the ref-
erences (Dudzik, 2020) and (Progorowicz, et al., 2022).
Typically, after performance analyzes are performed, only a
few networks are selected from among over 50,000 neural
networks. These selected networks can be tested at a later
stage on the sorter software.
Figure 8 shows exemplary network regression (R)
results for the problem of identifying the zinc ore (Zn) and
lead ore content from a photo of a rock.
As can be seen, in the case of the testing stage, values
close to 0.9 (0.896 fof Zn and 0,884 do Pb), which is a
satisfying result.
For the network indicated in Figure 8, the mean abso-
lute errors calculated for training, test and validation data
sets are presented in Figure 9.
As shown in Figure 9, the vast majority of the network
results achieved error values close to 0.
SORTING RESULTS
The choice of sensors for sorting is very dependent on types
of sorted materials. Nevertheless, the XRT sensor can be
Figure 6. Convolutional AI models for final image analysis
Figure 7. Boxplots of the mean absolute error values from
tests stage (10 best structures)