Magnesite is an important raw material for extracting magnesium metal and magnesium compound; how precise its grade classification exerts great influence on the smelting process. Thus, it is increasingly important to determine fast and accurately the grade of magnesite. In this paper, a method based on stacked autoencoder (SAE) and extreme learning machine (ELM) was established for the classification model of magnesite. Stacked autoencoder (SAE) was firstly used to reduce the dimension of magnesite spectrum data and then neutral network model of extreme learning machine (ELM) was adopted to classify the data. Two improved extreme learning machine (ELM) models were employed for better classification, namely, accuracy extreme learning machine (AELM) and integrated accuracy (IELM) to build up the classification models. The grade classification through traditional methods such as chemical approaches, artificial methods, and BP neutral network model was compared to that in this paper. Results showed that the classification model of magnesite ore through stacked autoencoder (SAE) and extreme learning machine (ELM) is better in terms of speed and accuracy; thus, this paper provides a new way for the grade classification of magnesite ore.
Magnesite is an important ore containing magnesium [
Since near-infrared (NIR) reflectance spectroscopy [
The collected NIR magnesite ore raw data is 973-high-dimensional data, which will increase the time of the algorithm model and reduce the accuracy of the model. Meanwhile the data of chemical factors which are normally contained in the NIR data did not influence the grade classification. Consequently, it is necessary to reduce the original data of the NIR original data before building the algorithm model.
That is why the dimension of these data should be reduced. Dimensionality reduction is one of the ways to deal with high-dimensional data [
We have discussed that magnesite ore can be divided into two types of which are super and nonspecial grade [
SVC HR-1024 portable field spectrum from Spectra Vista (US) was used to serve as the near-infrared spectrometer. Its spectral range was 350–2500 nm, internal memory was 500 scans, weight was 3 kg, port numbers were 1024, spectral resolution (FWHM ≦ 8.5 nm) was 1000–1850 nm, and the minimum integral time was 1 ms.
The 633 magnesite samples were from Dashiqiao Magnesite Mine in Liaoning Province, China. The spatial distribution of the ore body is controlled by the ore-controlling factors (such as tectonics, magmatic activity, stratigraphy, geochemical factors, and metamorphic factors) when the ore body is formed. As Figure
Magnesite ore body.
Temperature and humidity have some influence on the acquisition of spectral data. We have done a series of experimental verification. The same magnesite sample (third grade) was tested at different times, different temperatures, and humidity. And it is ensured that the test temperature is within the normal working range of the instrument. Temperature, humidity, and solar elevation angle are as shown in Table
Spectrum test time and corresponding temperature and humidity.
Testing time | Solar elevation angle | Temperature (°C) | Humidity (% rh) |
---|---|---|---|
10:48 | 61°15′ | 23.9 | 18.8 |
11:48 | 63°15′ | 27.8 | 14.3 |
12:48 | 59°50′ | 31.2 | 12.8 |
Different temperature and humidity corresponding to the spectral curve.
The observation was done before sunset on sunny day without much cloud on open air. The scanning time is 1 s/time and the probe is 300 mm far away from the surface of magnesite ore and perpendicular to the upper surface. To reduce the radiation quantity on spectrum test, the experimenters are not supposed to walk about or be dressed in dark. The two surfaces were tested for 3 times through SVC HR-1024 and the average value was used as the spectrum data. The spectral images of some samples are shown in Figure
The spectral images of part samples.
Liaoning Dashiqiao magnesite is an important source of magnesium in China, which has the characteristics of large reserves, thick layer, and shallow burial. After the spectrum test of 633 samples, contents of minerals in each sample were measured through chemical method. The samples were decomposed by hydrochloric acid, nitric acid, hydrofluoric acid, and perchloric acid before removing by HMT-copper the elements interfering the measurement including aluminium, iron, copper, zinc, and manganese; and then contents of calcium and magnesium were measured by EDTA standard solution complexometry. Industrial indicators of magnesite ore show that its grade classification is mainly determined by the contents of magnesium oxide, calcium oxide, and silicon dioxide. The physical and chemical test results of magnesite samples are presented in Table
Part of the physical and chemical test results of magnesite samples.
Number | Grade | Sample | Chemical composition% | |||||
---|---|---|---|---|---|---|---|---|
IgL | SiO2 | CaO | MgO | Fe2O3 | AL2O3 | |||
1 | Top grade | Magnesium stone | 51.79 | 0.17 | 0.32 | 47.41 | 0.25 | 0.06 |
2 | Top grade | Magnesium stone | 51.73 | 0.14 | 0.42 | 47.39 | 0.27 | 0.05 |
3 | First grade | Magnesium stone | 51.51 | 0.66 | 0.34 | 47.28 | 0.16 | 0.05 |
4 | First grade | Magnesium stone | 51.54 | 0.69 | 0.25 | 47.23 | 0.24 | 0.05 |
5 | Second grade | Magnesium stone | 51.02 | 1.01 | 0.81 | 46.93 | 0.18 | 0.05 |
6 | Second grade | Magnesium stone | 51.10 | 1.40 | 0.59 | 46.67 | 0.19 | 0.05 |
7 | Third grade | Magnesium stone | 48.83 | 4.16 | 0.83 | 44.65 | 2.07 | 0.08 |
8 | Third grade | Magnesium stone | 46.81 | 8.46 | 1.12 | 43.23 | 0.30 | 0.08 |
9 | Fourth grade | Magnesium stone | 50.03 | 1.49 | 3.52 | 44.75 | 0.16 | 0.05 |
10 | Waste ore | Magnesium stone | 49.55 | 3.57 | 0.48 | 46.05 | 0.30 | 0.05 |
According to the industrial indicators and chemical analysis results, the grades of 633 samples are 104 top grade, 109 first grade, 108 second grade, 110 third grade, 102 fourth grade, and 100 waste ores.
Autoencoder (AE) network is used to reduce the dimension of data and extract the features of data through the 3-layered neutral network. It can seek for the internal feature structure of data or encoding mode of data through self-learning. Its structure is shown in Figure
Autoencoder structure.
Autoencoder tries learning an identity function
From Layer 2 to Layer 3 is the decoding.
Parameters of
Gradient Descent and Backpropagation Algorithm (BP) were used for the iteration of
The first term
ELM [
For any given
The
From the autoencoder (AE) algorithm, we know that model’s output is equal to the input. The data dimension is reduced by controlling the number of hidden layer nodes less than the number of input nodes. SAE is to build multiple AE models to reduce the original data. SAE model’s parameters include the number of AE and the number of hidden nodes in each AE. The original spectral data is 633 × 973-high-dimensional data. In this paper, SAE constructs two AE models to achieve the purpose of reducing the original spectral data dimension; while the ELM is a single hidden layer feedforward neural network, the parameters we need to select are the number of hidden layer nodes, and we choose the hidden layer nodes through continuous testing which achieved good results in this paper.
NIR information of magnesite was collected through the spectrometer. These data need compression by SAE because of the high dimension and noise. And the data were classified through ELM to identify the magnesite: top grade and others. The procedures are displayed in Figure
Grade classification procedure based on SAE and ELM.
ELM in grade classification of magnesite can be concluded to 3 steps: Firstly, set the training set Secondly, the output matrix Finally, the output weight
With 633 samples in the experiment, the original spectral data collected is a 973-dimensional matrix. The redundancy of data makes the classification neutral network less accurate and consume longer time. The industrial grade of magnesite ore falls into 6 levels: top, first, second, third, fourth, and waste. The value of top, first, and second grade is well worth classifying. In this paper, SAE was used for the dimension reduction of pretreated spectral data. There are 2 layers of hidden layer, one of which contains 200 nodes and the other of which contains 100 nodes. So the original NIR data are supposed to reduce to 100 in the dimension reduction. SAE structure is displayed in Figure
Structure of SAE.
The spectral data treated by SAE is 633 × 100 and the training set and test set are randomly selected from the samples. There are 437 training samples, including 72 top grade, 73 first grade, 72 second grade, 80 third grade, 72 fourth grade, and 70 waste ores. Parameters of grade classification by ELM include activation function and number of nodes in the hidden layer. The activation function in ELM mainly includes Sigmoid function, sin function, and hardlim function. Number of nodes in the hidden layer exerts great influence on the learning and information processing of the network: excessive number makes the network more complex, prolongs the learning time, and leads to overfitting while smaller number may constrain the learning and processing ability of the network. In reality, empirical formula is conventionally used to roughly determine a range of the number before selecting an optimal one through repeated experiments. The optimal number in this paper is 45. Output weight is worked out through Formula (
Output results and expected scatter profiles of training set.
Output results and expected scatter profiles of test set.
Figures
Accuracy ELM training set of predicted and expected output scatter profile.
Accuracy ELM testing set of predicted and expected output scatter profile.
The predicted results show obvious improvement. The accuracy ELM in each group corresponds to a group of parameters and there may be differences in the results, so it is not stable. To make it more stable, an integrated accuracy ELM is proposed. The number of integrated groups is the quantity of selected ELM models that we choose. Group 11 is selected in this paper and the integrated model outputs the most grades among the 11 single models, which can help improve the prediction accuracy. The quantitative relation between predicted value and expected value in results of experimental simulation and statistical simulations reveals that the integrated accuracy ELM model is superior in its stability and accuracy in the grade classification of magnesite ore. Comparisons between simulation results in this model and those of the real samples are presented in Tables
Confusion matrix structure of the integrated accuracy ELM training group output.
The expected training set | The predicted training set | |||||
---|---|---|---|---|---|---|
Top grade | First grade | Second grade | Third grade | Fourth grade | Waste ore | |
Top grade | 72 | 0 | 0 | 0 | 0 | 0 |
First grade | 0 | 73 | 1 | 1 | 0 | 0 |
Second grade | 0 | 0 | 71 | 0 | 0 | 0 |
Third grade | 0 | 0 | 0 | 79 | 0 | 0 |
Fourth grade | 0 | 0 | 0 | 0 | 72 | 0 |
Waste ore | 0 | 0 | 0 | 0 | 0 | 70 |
Confusion matrix structure of the integrated accuracy ELM testing group output.
The expected training set | The predicted testing set | |||||
---|---|---|---|---|---|---|
Top grade | First grade | Second grade | Third grade | Fourth grade | Waste ore | |
Top grade | 32 | 0 | 0 | 0 | 0 | 0 |
First grade | 0 | 36 | 2 | 0 | 0 | 0 |
Second grade | 0 | 0 | 33 | 0 | 0 | 0 |
Third grade | 0 | 0 | 1 | 30 | 0 | 0 |
Fourth grade | 0 | 0 | 0 | 0 | 32 | 0 |
Waste ore | 0 | 0 | 0 | 0 | 0 | 32 |
ELM, accuracy ELM, and integrated accuracy ELM classification models were established for the grades of magnesite ore. The simulation results are displayed in Table
The simulation results of various models of magnesite ore grade classification.
Model types | Training set accuracy rate | Testing set accuracy rate | Model time consuming |
---|---|---|---|
ELM (%) | 84.897% | 75.0% | 0.3202 s |
Accuracy ELM (%) | 94.050% | 85.204% | 1.86 s |
Integrated accuracy ELM (%) | 99.542% | 98.469% | 59.7 s |
This table shows the accuracy of each model in the grade classification. The training set of conventional ELM model is not so accurate; the accuracy ELM model is more precise, reaching above 85% but it is not so stable. Both the accuracy and stability of integrated accuracy ELM model have been greatly improved and its accuracy reaches 98%.
Table
Different ways of detecting chart.
Experiment object | Accuracy of magnesite | Identification time-consuming | Expenses |
---|---|---|---|
Traditional artificial method | 65% | 2.5 days | $105 |
Chemical test method | 100% | 18 days | $1744 |
BP | 60% | 2.5 hours | $27 |
ELM | 98% | 2.0 hours | $27 |
This table presents that the traditional artificial method is not accurate enough though it is easy. Chemical medicine is needed in the chemical test method and the cost is about $1744, besides, some experimental apparatus and human cost will be above $435287, so this method is much more expensive. Comparatively, the investment on hardware devices including spectrometer and computer is no more than $43500. ELM is more accurate than BP, requiring less cost and shorter time. ELM is the most economical and accurate with considerable benefit.
In this paper, a new method of magnesite grade classification was put forward. Based on the nondestructive testing technology of NIR, the spectral data of magnesite ore was collected. And then, the dimension of these data was reduced with SAE. Finally, the models were established through ELM. Traditional ELM model is less accurate in the grade classification, because of which the accuracy ELM was proposed. To make up for the poor stability of accuracy ELM, the integrated accuracy ELM was proposed, which is superior to the two ELM models in terms of accuracy and stability. Its accuracy can reach as high as 98%. Compared with traditional methods of magnesite grade classification, integrated accuracy ELM has advantages in economic efficiency, accuracy, and rapidity; in addition, this method can achieve online testing of ores in large volume. Obviously, it is of great practical application value.
The authors declare no conflicts of interest.
This research is supported by National Natural Science Foundation of China (Grant nos. 41371437 and 61203214); Fundamental Research Funds for the Central Universities (N160404008); National Twelfth Five-Year Plan for Science and Technology Support (2015BAB15B01), China.