Mining Goes Digital – Mueller et al. (Eds)
© 2019 Taylor & Francis Group, London, ISBN 978-0-367-33604-2
455
A deep learning approach for automated quality control of iron ores
A.K. Gorai & B.C. Balusa
Department of Mining Engineering, National Institute of Technology, Rourkela, India
U. Sameer
Department of Computer Science and Engineering, National Institute of Technology, Rourkela, India
ABSTRACT: The present study aims to design a machine vision system using deep learning
algorithm for quality monitoring of iron ores. A total of 53 image samples were used for
model calibration and testing. The model was trained using 45 image samples and tested
using 9 image samples. The model parameters like the number of nodes and number layers
were optimized based on the Root Mean Squared Error (RMSE) values. It was observed
that the RMSE was lowest for the network architecture having 5-nodes and 3-hidden layers.
The performance of the optimized model was evaluated using four indices including RMSE,
Normalized Mean Square Error (NMSE), R-squared, and bias. The RMSE, NMSE,
R-squared, and bias of the optimized model were obtained as 8.77, 0.0026, 0.87, and −1.14
respectively. The results indicate that the model gives satisfactory performance in quality
predictions of iron ores.
1 INTRODUCTION
The quality of ore in terms of physical (size, shape, strength, etc.) and chemical (composition, grade, etc.) properties specifically defined for the extraction of valuable material
through particular treatment process (Ivanov, 1986). The underlying task of predicting a
continuous value falls under the category of regression in the field of machine learning. It
is the process of studying various characteristics of the data and fitting all the data points
into one model. This was performed first by identifying various distinct characteristics
of the data known as the features. Then, a regression algorithm such as linear regression
(Neter, 1996) or a logistic regression (Hosmer, 2013) was applied, and the regression value
of the test data was predicted. The success of these machine learning techniques is hugely
dependent on the choice of the features extracted, thus limiting the effectiveness of the
regression algorithms. Thus a deep learning algorithm has been designed for prediction
of continuous data without any specific feature extraction from the image samples. Deep
Learning is an emerging technique for performing representation learning. That is, information from the data is learned using artificial intelligence in the form of a deep neural
network. Unlike machine learning techniques which are largely dependent on the features
extracted from the data, deep learning techniques learn the features by themselves and perform the necessary prediction task using the learned features. It is the independent nature
of deep learning that makes it more effective and superior to machine learning based models when identification and extraction of the feature are not easy. In this work, we perform
the task of regression using a deep learning multi-layer network using the raw images of
minerals as the input.
Deep learning approach of neural network has been used in many different applications
for classification and regression analysis. Baccouche et al. (2011) proposed sequential deep
learning for human action recognition. This study developed a fully automated deep learning
model, which learns to classify human actions without using any prior knowledge. Yu and
© 2019 Taylor & Francis Group, London, ISBN 978-0-367-33604-2
455
A deep learning approach for automated quality control of iron ores
A.K. Gorai & B.C. Balusa
Department of Mining Engineering, National Institute of Technology, Rourkela, India
U. Sameer
Department of Computer Science and Engineering, National Institute of Technology, Rourkela, India
ABSTRACT: The present study aims to design a machine vision system using deep learning
algorithm for quality monitoring of iron ores. A total of 53 image samples were used for
model calibration and testing. The model was trained using 45 image samples and tested
using 9 image samples. The model parameters like the number of nodes and number layers
were optimized based on the Root Mean Squared Error (RMSE) values. It was observed
that the RMSE was lowest for the network architecture having 5-nodes and 3-hidden layers.
The performance of the optimized model was evaluated using four indices including RMSE,
Normalized Mean Square Error (NMSE), R-squared, and bias. The RMSE, NMSE,
R-squared, and bias of the optimized model were obtained as 8.77, 0.0026, 0.87, and −1.14
respectively. The results indicate that the model gives satisfactory performance in quality
predictions of iron ores.
1 INTRODUCTION
The quality of ore in terms of physical (size, shape, strength, etc.) and chemical (composition, grade, etc.) properties specifically defined for the extraction of valuable material
through particular treatment process (Ivanov, 1986). The underlying task of predicting a
continuous value falls under the category of regression in the field of machine learning. It
is the process of studying various characteristics of the data and fitting all the data points
into one model. This was performed first by identifying various distinct characteristics
of the data known as the features. Then, a regression algorithm such as linear regression
(Neter, 1996) or a logistic regression (Hosmer, 2013) was applied, and the regression value
of the test data was predicted. The success of these machine learning techniques is hugely
dependent on the choice of the features extracted, thus limiting the effectiveness of the
regression algorithms. Thus a deep learning algorithm has been designed for prediction
of continuous data without any specific feature extraction from the image samples. Deep
Learning is an emerging technique for performing representation learning. That is, information from the data is learned using artificial intelligence in the form of a deep neural
network. Unlike machine learning techniques which are largely dependent on the features
extracted from the data, deep learning techniques learn the features by themselves and perform the necessary prediction task using the learned features. It is the independent nature
of deep learning that makes it more effective and superior to machine learning based models when identification and extraction of the feature are not easy. In this work, we perform
the task of regression using a deep learning multi-layer network using the raw images of
minerals as the input.
Deep learning approach of neural network has been used in many different applications
for classification and regression analysis. Baccouche et al. (2011) proposed sequential deep
learning for human action recognition. This study developed a fully automated deep learning
model, which learns to classify human actions without using any prior knowledge. Yu and