Iron ore sorting method based on hyperspectral images and spectral decomposition
By using hyperspectral imaging technology and random forest regression model, the spectral characteristics of iron ore are extracted and its content is predicted, which solves the problems of insufficient spectral characteristics and models in iron ore sorting and realizes rapid and accurate sorting of iron ore.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA UNIV OF MINING & TECH (BEIJING)
- Filing Date
- 2023-09-30
- Publication Date
- 2026-06-30
AI Technical Summary
Existing technologies lack suitable spectral characteristics and high-precision prediction models for iron ore sorting, making it impossible to accurately assess the iron content of each piece of iron ore, and resulting in high water and electricity consumption.
Hyperspectral imaging technology was used to extract the spectral characteristics of iron ore through variational mode decomposition, and the iron content was predicted by random forest regression, and the ore was sorted into multiple grades.
It enables rapid and accurate sorting of iron ore, improves the accuracy of iron content prediction and calculation efficiency, and reduces energy consumption and economic losses.
Smart Images
Figure CN117324286B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of metal smelting, and more specifically, to a method for iron ore sorting based on hyperspectral imaging and spectral decomposition. Background Technology
[0002] Iron ore is a crucial pillar of global industry, serving as a raw material for steelmaking and supplying the trillion-dollar metal market annually. However, the mining and processing of iron ore generates a large amount of waste rock with low iron content. Processing this waste rock leads to significant energy consumption, dust pollution, and economic losses. Furthermore, some low-iron waste rock can be used in other industrial sectors. Therefore, rapidly determining the iron content of iron ore and pre-sorting it accordingly can ensure the rational utilization of ores of different grades, thereby greatly mitigating these problems.
[0003] Traditional iron ore sorting techniques, such as magnetic separation, flotation, and gravity separation, suffer from the major drawback of their inability to accurately assess the iron content of each individual piece of ore. They only provide general separation for batches of ore and consume significant amounts of water and electricity (Nwaila et al., 2022; Peukerte et al., 2022). Hyperspectral imaging technology, however, avoids these drawbacks and has been widely used for the rapid determination of nutrient content in food and pigments in plants. For ore sorting, the most significant advantage of introducing hyperspectral imaging technology lies in its ability to accurately determine the composition of each individual ore by acquiring the spectral distribution across its surface space.
[0004] Using hyperspectral imaging to sort iron ore requires first predicting the iron content of each piece of ore and then sorting it based on the prediction results. Two key issues need to be addressed in this process: First, it is necessary to extract spectral features suitable for iron content prediction, a requirement that has not yet been met. Extracting appropriate spectral features for predicting the content of a substance using hyperspectral imaging technology will help improve accuracy and computational efficiency (Arifetal., 2022; Guoetal., 2023; Junttilaetal., 2022). Many studies in the field of rock spectroscopy do not focus on the spectral characteristics of iron ore (Okadaetal., 2020; Yousefietal., 2020), or some studies only focus on petrographic analysis of the ore surface through spectral analysis (Haoetal., 2020). (2020), however, these analyses are of very limited help in predicting the iron content of iron ore; secondly, it is necessary to develop high-precision prediction models. The large number of features in hyperspectral data makes it difficult for traditional simple models such as linear models to achieve good prediction results, which means that it is necessary to introduce machine learning algorithms with a certain degree of complexity. Summary of the Invention
[0005] The technical problem to be solved by the present invention is to provide an iron ore sorting method based on hyperspectral images and spectral decomposition, so as to solve the problem of lack of suitable spectral features and suitable prediction models when using hyperspectral imaging technology for iron ore sorting.
[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:
[0007] The iron ore sorting method based on hyperspectral images and spectral decomposition of the present invention includes the following steps: S1. Preprocessing: acquiring hyperspectral images of iron ore and extracting the average spectral curve of each piece of iron ore; S2. Feature extraction: decomposing the average spectral curve of iron ore through variational mode decomposition and extracting features from the decomposed signal; S3. Predicting iron content: predicting the total iron content of each piece of iron ore by combining random forest regression; and S4. Sorting: dividing the iron ore into multiple grades according to the predicted total iron content.
[0008] Optionally, in the above-described iron ore sorting method based on hyperspectral images and spectral decomposition, step S1 further includes: a. Calibration: When acquiring the hyperspectral image of the iron ore, the reference plate and dark current are measured simultaneously, and the calibration conversion from the original detector signal intensity to reflectivity is achieved using the following formula:
[0009]
[0010] Where R is the relative reflectance after hyperspectral image correction, R∈[0,1], I D For dark current, I W For reference board value, I S This represents the pixel signal intensity value.
[0011] b. Smoothing: The spectral curve is smoothed using the SG smoothing filter;
[0012] c. Masking: The values of label and background pixels in the iron ore hyperspectral image are all masked to 0 to avoid the influence of irrelevant pixels;
[0013] d. Obtain the average spectrum: Average the pixel-level spectral curves of each ore to obtain the average spectral curve of each ore.
[0014] Optionally, in the above-mentioned iron ore sorting method based on hyperspectral images and spectral decomposition, in step S2, variational mode decomposition is used to decompose the average spectrum of the iron ore into two sets of modal signals: the first intrinsic mode function (IMF1) and the second intrinsic mode function (IMF2). A straight line is fitted to the first intrinsic mode function (IMF1) in the range of 1204 nm to 2404 nm, and the slope, intercept, mean absolute error of the fitted line, and the value at 2124 nm are extracted as features. The values at 1945 nm, 2124 nm, and 2293 nm in the second intrinsic mode function (IMF2) are extracted as peak and valley features. The mean absolute value (MAV) of the spectral curve is calculated using the following formula to measure its amplitude:
[0015]
[0016] Where V2 represents the value of the second intrinsic mode function, WN is the number of wavelengths, and after feature extraction, the slope, intercept, mean absolute error of the fitting line of the first intrinsic mode function, the value at 2124nm, the values at 1945nm, 2124nm and 2293nm in the second intrinsic mode function, and the mean absolute value of the spectral curve MAV are recorded as the variational mode decomposition features.
[0017] Optionally, in the above-mentioned iron ore sorting method based on hyperspectral images and spectral decomposition, in step S3, the extracted variational mode decomposition features are fed into a random forest regressor for prediction to predict the total iron content of each iron ore piece.
[0018] Optionally, in the above-mentioned iron ore sorting method based on hyperspectral images and spectral decomposition, in step S4, after obtaining the predicted value of iron content TFe, the ore is divided into three grades: iron ore with TFe < 0.48 is classified as Grade III ore, with low iron content; iron ore with TFe <= 0.48 is classified as Grade II ore, with moderate iron content; and iron ore with TFe >= 0.64 is classified as Grade I ore, with very high iron content.
[0019] Compared with the prior art, the beneficial effects of the present invention are:
[0020] The method of this invention takes the hyperspectral image of iron ore as the research object. By combining the features extracted after decomposing the average spectrum of iron ore with random forest, the iron content can be predicted and the iron ore can be sorted objectively, quickly and accurately. Attached Figure Description
[0021] Figure 1 This is a flowchart of the iron ore sorting method based on hyperspectral images and spectral decomposition of the present invention;
[0022] Figure 2 This is a graph showing the predicted iron content obtained using the method of this invention;
[0023] Figure 3 This is a diagram showing the iron ore sorting results obtained using the method of this invention. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.
[0025] The present invention provides an iron ore sorting method based on hyperspectral images and spectral decomposition, comprising: S1. Preprocessing: acquiring hyperspectral images of iron ore and extracting the average spectral curve of each piece of iron ore; S2. Feature extraction: decomposing the average spectral curve of the iron ore using variational mode decomposition and extracting features from the decomposed signal; S3. Iron content prediction: predicting the total iron content of each piece of iron ore using random forest regression; and S4. Sorting: dividing the iron ore into multiple grades based on the predicted total iron content. This invention uses hyperspectral images of iron ore as the research object, and by combining the features extracted after decomposing the average spectrum of iron ore with random forest, it achieves the prediction of iron content and further objectively, quickly, and accurately sorts the iron ore to achieve rapid and precise sorting of each piece of iron ore.
[0026] like Figure 1 As shown, the iron ore sorting method based on hyperspectral images and spectral decomposition of the present invention includes the following steps:
[0027] S1. Preprocessing: Acquire hyperspectral images of iron ore and extract the average spectral curve of each piece of iron ore.
[0028] a. Correction
[0029] When acquiring hyperspectral images of iron ore, the reference plate and dark current were measured simultaneously. The calibration conversion from the original detector signal intensity to reflectivity was achieved using the following formula:
[0030]
[0031] Where R is the relative reflectance after hyperspectral image correction, R∈[0,1], I D For dark current, I W For reference board value, I S This represents the pixel signal intensity value.
[0032] b. Smoothing
[0033] The spectral curve of iron ore contains noise, which can significantly affect subsequent data processing. Therefore, SG smoothing is used to smooth the spectral curve.
[0034] c. Mask
[0035] The values of label and background pixels in the hyperspectral image of iron ore are all masked to 0 to avoid the influence of irrelevant pixels.
[0036] d. Obtain the average spectrum
[0037] The average spectral curve of each ore is obtained by averaging all pixel-level spectral curves in each ore.
[0038] S2. Feature Extraction: The average spectral curve of iron ore is decomposed by variational mode decomposition, and features are extracted from the decomposed signal.
[0039] After obtaining the average spectral curve, variational mode decomposition (VMD) is used to decompose the average spectrum into two sets of modal signals: the first intrinsic mode function (IMF1) and the second intrinsic mode function (IMF2). A straight line is fitted to IMF1 in the range of 1204nm-2404nm, and the slope, intercept, mean absolute error of the fitting line, and the value at 2124nm are extracted as features.
[0040] The values at 1945 nm, 2124 nm, and 2293 nm in IMF2 were extracted as peak and valley characteristics. Furthermore, the mean absolute value (MAV) of the spectral curve was calculated to measure its amplitude, as follows:
[0041]
[0042] Where V2 represents the value of IMF2 and WN is the number of wavelengths. After feature extraction, the eight extracted features (slope, intercept, mean absolute error of the fit of the IMF1 straight line, value at 2124nm, values at 1945nm, 2124nm and 2293nm in IMF2, and mean absolute value of the spectral curve MAV) are denoted as variational mode decomposition (VMD) features.
[0043] S3. Predict iron content: Combine random forest regression to predict the total iron (TFe) content of each iron ore.
[0044] The extracted VMD features are fed into a random forest regressor to predict the total iron content of each iron ore.
[0045] S4. Sorting: Based on the predicted total iron content, these iron ores are divided into multiple grades.
[0046] After obtaining the predicted iron content, the ore was divided into three grades: iron ore with TFe < 0.48 was classified as Grade III ore (low iron content), iron ore with TFe < 0.48 was classified as Grade II ore (moderate iron content), and iron ore with TFe > 0.64 was classified as Grade I ore (very high iron content).
[0047] The present invention will be further described below with reference to specific embodiments.
[0048] Data Description: Hyperspectral image data of iron ore was used as the research object. The data collection process was as follows: Several hundred kilograms of low-grade and high-grade raw ore were purchased from an iron ore mining area in Shandong Province and crushed using a jaw crusher. The main research object was ore with a particle size of 20-40 mm. Therefore, 150 samples with a diameter of 20-40 mm were manually screened from the crushed stone of both low-grade and high-grade raw ore, for a total of 300 samples.
[0049] After sample selection, samples were numbered, and the mass of each sample was measured using an electronic balance. A HyspexSWIR-384 hyperspectral imager with 288 bands was used to acquire hyperspectral images in the 930-2500 nm range, with twelve 100-watt halogen lamps ensuring high-quality illumination. A matte black strip and a grid were placed on a conveyor belt, and ore samples were placed in each grid according to a pre-designed arrangement to ensure neatness in the ore images. After placement, the grid was removed, and labels were placed near the samples to indicate their numbers. The conveyor belt was then activated to scan and acquire hyperspectral images. Each ore sample was imaged three times, changing its placement orientation, resulting in a total of 900 rock images in the hyperspectral image data. After obtaining the hyperspectral images, 300 ore samples were ground into iron powder below 200 mesh for iron content measurement. The iron content was then determined using potassium dichromate volumetric titration (Huetal., 2014), and the iron content was used for subsequent validation.
[0050] The method for rapidly and accurately predicting the iron content of iron ore using Variational Mode Decomposition (VMD) and Random Forest (RF) based on the present invention, and sorting the iron ore according to the iron content, includes the following steps:
[0051] S1. Preprocessing
[0052] a. Correction
[0053] When acquiring hyperspectral images of iron ore, and simultaneously measuring the reference plate and dark current, the calibration conversion from the original detector signal intensity to reflectivity can be achieved using the following formula:
[0054]
[0055] Where R is the relative reflectance after hyperspectral image correction, R∈[0,1], I DFor dark current, I W For reference board value, I S This represents the pixel signal intensity value.
[0056] b. Smoothing
[0057] Smoothing filter (SG) is one of the commonly used preprocessing methods in spectral analysis. The spectral curve of iron ore contains noise, which can significantly affect subsequent data processing. Therefore, smoothing filter (SG) is used to smooth the spectral curve.
[0058] c. Mask
[0059] The values of label and background pixels in the hyperspectral image of iron ore are all masked to 0 to avoid the influence of these irrelevant pixels.
[0060] d. Obtain the average spectrum
[0061] The average spectra of iron ore with different iron contents are significantly different. Therefore, the average of all pixel-level spectral curves in each piece of ore is calculated to obtain the average spectral curve of each piece of ore. Finally, a total of 900 average spectra are obtained.
[0062] S2. Feature Extraction
[0063] Appropriate spectral features significantly improve prediction efficiency and accuracy. However, the spectrum of iron ore is complex, requiring initial decomposition to facilitate further feature extraction. Therefore, Variational Mode Decomposition (VMD) was used to decompose the average spectrum into Intrinsic Mode Functions (IMFs), IMF1 and IMF2. IMF1 in the 1204nm-2404nm range was fitted with a straight line. The slope, intercept, mean absolute error of the fitted line, and the value at 2124nm were extracted as features. The slope of the fitted line measures the trend and steepness of the spectrum, the intercept measures the reflectance, and the mean absolute error of the fitted line measures the effectiveness of the straight-line fitting.
[0064] Values at 1945 nm, 2124 nm, and 2293 nm were extracted from IMF2 to measure its peak and trough characteristics. Furthermore, the mean absolute value (MAV) of the spectral curve was calculated to measure its amplitude, as follows:
[0065]
[0066] Where V2 represents the IMF2 value and WN is the number of wavelengths. After feature extraction, these eight extracted features are denoted as VMD features.
[0067] S3. Predicting Iron Content
[0068] Random forest is an ensemble machine learning method that combines the predictions of multiple decision trees to arrive at a final prediction. It boasts high recognition accuracy and, moreover, does not require feature transformation. The VMD features extracted from the average spectrum of a particular ore are fed into a random forest regressor to predict the total iron (TFe) content of that ore.
[0069] S4. Sorting
[0070] According to the national standard GB / T32545-2016 (PRC, 2016), iron ore with TFe ≥ 0.64 has extremely high industrial value, while iron ore with TFe < 0.48 has lower industrial value. Therefore, after obtaining the predicted iron content, the ore is divided into three grades: iron ore with TFe < 0.48 is labeled as Class III ore (low iron content), iron ore with TFe ≥ 0.48 and TFe < 0.64 is labeled as Class II ore (moderate iron content), and iron ore with TFe ≥ 0.64 is labeled as Class I ore (very high iron content).
[0071] The predicted value of the iron content (TFe content) of the ore was compared with the actual value of the TFe content, and the results are as follows: Figure 2 As shown, the predicted R², RMSE, and MAE are 0.94, 0.07, and 0.03, respectively, indicating that the method of this invention has a good predictive effect on iron content.
[0072] For sorting, its accuracy rate of 0.91 means it can accurately determine the type of iron ore based on its iron content. For example... Figure 3 (Where, ABC represent three different arrangement methods for each piece of ore, and incorrectly sorted ore is marked with a white box in the figure.) As shown, the sorting results are not significantly different from the actual ore categories. Although samples of different categories are arranged alternately, the sorted ore categories can still be identified relatively accurately. The quantitative evaluation indicators obtained through this invention are shown in Table 1. In summary, the method of this invention can quickly and accurately predict the iron content of iron ore and sort it.
[0073] Table 1 Quantitative Evaluation Indicators
[0074]
[0075] The above embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and are not intended to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or improve the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in the present invention; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for iron ore sorting based on hyperspectral images and spectral unmixing, characterized in that, Includes the following steps: S1. Preprocessing: Acquire hyperspectral images of iron ore and extract the average spectral curve of each piece of iron ore; S2. Feature extraction: The average spectral curve of the iron ore is decomposed by variational mode decomposition, and features are extracted from the decomposed signal; S3. Predicting Iron Content: Predicting the total iron content of each iron ore piece using random forest regression; and S4. Sorting: Based on the predicted total iron content, the iron ore is divided into multiple grades. In step S2, variational mode decomposition (VMD) is used to decompose the average spectrum of the iron ore into two sets of modal signals: a first intrinsic mode function (IEM) and a second IEM. A linear fit is performed on the first IEM in the range of 1204 nm to 2404 nm, and the slope, intercept, mean absolute error of the fit, and the value at 2124 nm are extracted as features. The values at 1945 nm, 2124 nm, and 2293 nm in the second IEM are extracted as peak and valley features. After feature extraction, the eight features—the slope, intercept, mean absolute error of the fit, and value at 2124 nm of the first IEM, the values at 1945 nm, 2124 nm, and 2293 nm of the second IEM, and the mean absolute value (MAV) of the spectral curve—are denoted as variational mode decomposition features. In step S3, the extracted variational mode decomposition features are fed into a random forest regressor for prediction to predict the total iron content of each iron ore.
2. The method for iron ore sorting based on hyperspectral images and spectral unmixing according to claim 1, characterized in that, Step S1 also includes: a. Calibration: When acquiring hyperspectral images of iron ore, the reference plate and dark current are measured simultaneously. The calibration conversion from the original detector signal intensity to reflectivity is achieved using the following formula: where R is the corrected relative reflectance of the hyperspectral image, R ∈ [0, 1], is the dark current, is the reference plate value, is the pixel signal intensity value; b. Smoothing: The spectral curve is smoothed using the SG smoothing filter; c. Masking: The values of label and background pixels in the iron ore hyperspectral image are all masked to 0 to avoid the influence of irrelevant pixels; d. Obtain the average spectrum: Average the pixel-level spectral curves of each ore to obtain the average spectral curve of each ore.
3. The method for iron ore separation based on hyperspectral images and spectral decomposition according to claim 1, characterized in that, In step S2, the mean absolute value (MAV) of the spectral curve is calculated using the following formula to measure its amplitude: wherein represents a value of the second eigenmode function, is the number of wavelengths.
4. The method for iron ore separation based on hyperspectral images and spectral decomposition according to claim 1, characterized in that, In step S4, after obtaining the predicted iron content TFe, the ore is divided into three grades: iron ore with TFe < 0.48 is classified as Grade III ore, with low iron content; iron ore with TFe <= 0.48 <= 0.64 is classified as Grade II ore, with moderate iron content; and iron ore with TFe >= 0.64 is classified as Grade I ore, with very high iron content.