Development of an Online Vision Based Integrated Quality Monitoring Model for Mineral Industry

Abstract

The iron ores produced in the mines are considered to be the raw material for the steel industry. The policy of despatching the raw materials (iron ores) from mine is highly depended on the class or grade value of ores. Moreover, the ore reserves in most of the mines are heterogeneous. Thus, regular quality monitoring of iron ores (class and grade value) in the mine is the urgent need for appropriate assigning the destinations of raw materials.
The present study attempts to develop support vector machine (SVM)-based algorithms for both the models (classification and regression model). Furthermore, the effect of water absorption in ore samples in the model performances was also studied. Finally, an integrated system was developed for continuous monitoring of class and grade values of iron ores.
Although quality control (classification and grade estimation) of iron ores is one of the challenging tasks in the mining industry, it is essential for the future destination of the ores. The model development using SVM-based algorithm was conducted in four stages viz. capturing images of iron ores, features extraction, features selection or optimisation, and image analyses. Since the aim of the study is to develop a machine vision system for online quality monitoring of iron ores, and this requires to capture the image of ores in dynamic mode. For this, a laboratory scale conveyor belt transportation system was designed and fabricated with image capturing setup. To conduct the proposed study, the iron ore samples were collected from the Gua Iron Mine of the Steel Authority of India Limited (SAIL), India. A stratified sampling technique was followed for collecting the iron ore samples from mine. In stratified random sampling technique, the heterogeneous ore samples from different strata were collected to represent variation in ore types. The collected samples were classified into five classes based on their lithology. These include lateritic iron ore (LIO), massive iron ore (MIO), flaky friable and blue dust ore (FFBDO), banded hematite quartz (BHQ) and shale, available in the mining area. The presence of five lithology is confirmed by the XRD study results.
The samples collected from the mine were fed into the conveyor belt set up in the laboratory at the inlet point. The images of the ores were captured continuously during transportation through the conveyor belt. A total of 812 captured images were considered for classification model development. The images were captured in RGB colour space and converted into six other colour spaces for feature extractions. A total of 280 image features were extracted from each image. The feature includes ten statistical features for each component of the six colour spaces (RGB, HSI, xyz, CMYK, Lab, and Grey) and four frequency-transformed components [discrete cosine transform (DCT), discrete Fourier transform (DFT), discrete wavelet transform (DWT), and Gabor filter]. To select the optimum feature subset for the model development, a sequential forward floating selection (SFFS) algorithm was used. For the classification model, the optimum feature contains nearly 2% (6 out of 280) of the total number of features. The support vector machine (SVM) algorithm was used for the development classification model. The model considered the optimised features of the image as input and class of ore as output. The datasets were divided into training and testing in the ratio of 7:3. That is, out of 812 datasets, 569 (nearly 70%) datasets were used to train the model, and the rest 243 (nearly 30%) datasets were used to test the model. The performance of the SVM classification model was evaluated on testing datasets using the four confusion matrix parameters (sensitivity, specificity, misclassification, and accuracy) along with the other indices such as the relationship index, Q-statistics, and correlation -coefficients. The sensitivity, specificity, accuracy, and misclassification of the model were found to be 0.9792, 0.9948, 0.9918, and 0.0082, respectively. The relationship index, Q-statistics, and correlation coefficient were found to be 0.0110, 0.9999, and, 0.9695, respectively. The high value of sensitivity, specificity, and accuracy, and the low value of the misclassification indicate a good performance of the model. The performance of the model was also compared with other classification algorithms (k-nearest neighbours, classification tree, discriminant classifier, Naïve Bayes) using the same datasets. It was observed that the proposed algorithm performs reasonably well in predicting the class of iron ores.
The proposed study also developed a regression model using SVM algorithm. A total of 53 captured images were considered for the regression model development. The optimised image features were considered as the input and the grade value of iron ores were considered as the output in the regression model. The same number and types of image features (=280) similar to classification model, were extracted for examining its suitability in grade prediction. In this case, the derived optimised feature subset consists of 8 features. The grade values of the ore samples were estimated using XRF study. The results indicated that the grade values (Fe2O3 %) of iron ores are ranged from 19.3487% to 97.3607 %. These optimised feature set and their respective iron grades were used for training and testing the support vector regression (SVR) model. The model was trained using 70% of the data and tested using 30% of the sample data. The model performance was evaluated using five indices viz. the sum of squared errors (SSE), root mean squared error (RMSE), normalised mean squared error (NMSE), correlation coefficient (R2), and bias. All the indices were determined from the observed and predicted values of testing samples. The SSE, RMSE, NMSE and bias values of the model were obtained as 537.5367, 5.9863, 0.0063, and 0.8875 respectively. The R-squared value of the model was obtained 0.9402. A higher R2 value (close to 1) indicates that the model gives satisfactory performance in prediction of grade values of iron ores. The performance of the model was also compared with the other types of algorithms like artificial neural network (ANN) with PCA, Gaussian process regression (GPR), etc. used in the previous studies. It was observed that the model with proposed algorithm performs better than the model algorithms used in the previous studies.
A comparative performance study of the developed model using the proposed algorithm was carried out for the dry and wet iron ore samples. The images of the ore samples in wet condition were similarly captured in the laboratory as dry samples. The same number and type of features were extracted from each wet sample images and optimised using the SFFS algorithm. It was observed that the optimised subset contains different features in comparison to the above case (dry sample). The result indicated that only 5 (out of 243 testing samples) samples were misclassified for dry image samples, whereas for wet images, 12 (out of 185 testing sample) samples were misclassified by the classification model. The regression model results showed a high R-squared value (=0.9402) for the dry iron ore sample in comparison to that for wet ore samples. The R-squared value for wet ore sample was found to be 0.9085. Also, it has observed that the RMSE for the dry sample (=5.9863) was relatively low in comparison to the wet sample (=13.1382).
The study also designed and developed an integrated machine vision system for quality monitoring of iron ores. The system was designed and developed in MATLAB with a graphical user interface (GUI) for both the classification and grade prediction of iron ore based on the developed algorithms.