Coal gangue sorting method based on deep residual network and near infrared spectrum

Abstract

The invention relates to the technical field of coal gangue sorting, in particular to a coal gangue sorting method based on a deep residual network and a near infrared spectrum, which comprises the following steps: S1, constructing a coal gangue spectrum data set X; s2, removing abnormal values of the spectral data in the step S1; s3, constructing a first deep residual network model M1; s4, training the first deep residual network model M1 in the step S3; s5, testing the second deep residual network model M2 in the step S4; and S6, analyzing a prediction result in the step S5. According to the method, the spectral information of the coal gangue is collected, the deep residual network is constructed, the spectral characteristics of the coal and the gangue are analyzed, intelligent separation and recognition of the coal and the gangue are achieved, noise interference can be restrained through the high feature extraction capacity and robustness, and therefore the accuracy and reliability of the spectral data of the coal gangue can be improved; and therefore, the working intensity of coal gangue separation and recognition can be reduced, and the efficiency and precision of coal gangue separation and recognition can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of coal gangue sorting, and in particular to a coal gangue sorting method based on deep residual network and near infrared spectrum. Background Art

[0002] Gangue is a waste product from coal mining and production. It has a low carbon content and is not easy to burn. If it is mixed with coal without treatment, it will affect the heating efficiency of coal. Therefore, before coal flows to the market, gangue must be effectively identified and sorted.

[0003] At present, the commonly used methods for sorting coal gangue include: manual sorting, wet sorting and dry sorting;

[0004] 1. Manual picking: It is inefficient and cannot meet production needs. It is labor-intensive and harmful to workers’ health.

[0005] 2. Wet selection method: The process is complicated, consumes a lot of water resources, and is easy to pollute the environment;

[0006] 3. Dry sorting methods mainly include: crushing method, X-ray method and image recognition method; among them: the crushing method has a lower cost, but poor universality and is only applicable to specific types of coal gangue; although the X-ray method can accurately identify coal gangue, due to its radioactivity, isolation equipment needs to be installed, and the equipment cost is relatively high; the image recognition method can efficiently identify coal gangue, but it is easily affected by light and dust, and it is difficult to meet production requirements. Summary of the invention

[0007] The technical problem to be solved by the present invention is: in order to solve the technical problems of high intensity and low sorting efficiency of existing coal gangue sorting, the present invention provides a coal gangue sorting method based on deep residual network and near-infrared spectroscopy. By improving the coal gangue sorting method, the intensity of coal gangue sorting operations can be reduced and the sorting efficiency of coal gangue can be improved.

[0008] The technical solution adopted by the present invention to solve the technical problem is: a coal gangue sorting method based on deep residual network and near infrared spectroscopy, comprising the following steps:

[0009] S1. Constructing the spectral data set X of coal gangue: using near infrared spectrometer to collect the spectral data of coal gangue samples;

[0010] S2, removing outliers from the spectral data in step S1: using the Euclidean distance formula and an iterative method to gradually extract outliers from the spectral data, after the outliers in the spectral data are removed, the removed spectral data are divided into a training set A and a test set B;

[0011] S3, constructing a first deep residual network model M1;

[0012] S4. The first deep residual network model M1 in the training step S3: constructing a binary cross entropy loss function And using the spectral data in the training set A from which outliers have been removed in step S2 to train the first deep residual network model M1 to obtain the second deep residual network model M2;

[0013] S5, testing the second deep residual network model M2 in step S4: using the second deep residual network model M2 to test the spectral data of the test set B from which outliers have been removed in step S2, to obtain a prediction result;

[0014] S6, analyze the prediction results in step S5: use the accuracy ACC evaluation index to perform performance analysis on the prediction results. Thus, by collecting the spectral information of coal gangue and constructing a deep residual network, analyzing the spectral characteristics of coal and gangue, the intelligent sorting and identification of coal and gangue is realized. Compared with the existing sorting and identification methods, this method can suppress noise interference through powerful feature extraction capabilities and robustness, thereby improving the accuracy and reliability of coal gangue spectral data, thereby reducing the operating intensity of coal gangue sorting and identification and improving the efficiency and accuracy of coal gangue sorting and identification; externally, the deep residual network can capture the complex and subtle feature information in the coal gangue spectral data through its deep structure. These features are crucial to improving the classification, identification and prediction accuracy of spectral data. The deep residual network can effectively solve the gradient vanishing problem through its unique residual learning mechanism, thereby deeply mining the feature information of spectral data. This ability makes the deep residual network model have higher sensitivity and accuracy when processing complex spectral data.

[0015] Furthermore, the step S2 comprises the following steps:

[0016] S2-1, calculate the average spectral value of the gangue spectral data set X in step S1

[0017] S2-2. Calculate the average spectral value of each spectral data in the gangue spectral data set X according to the Euclidean distance formula The Euclidean distance

[0018] S2-3, calculate the mean μ and standard deviation σ of each Euclidean distance D;

[0019] S2-4, when When x i is an outlier, and x i Eliminate the gangue spectral dataset X;

[0020] S2-5, when spectral data is removed from the gangue spectral data set X, repeat steps S2-1 to S2-4 until no outliers appear, and divide the removed spectral data into a training set A and a test set B;

[0021] Where: X = {x1, x2, ..., x m}, x i ∈R n , i=1,2,…,m, m is the number of samples, n is the sample dimension. Therefore, in the process of sample collection, environmental changes or inappropriate operations may cause abnormal values ​​that are significantly different from normal spectra in the spectral data set. The existence of these abnormal values ​​will not only interfere with data analysis, but also mislead the training of the model, thus affecting the final test results. Therefore, before modeling, it is necessary to remove the abnormal data in the spectral data set; since the sample distribution of the collected spectral data follows the normal distribution, the abnormal values ​​in the sample spectral data set can be gradually removed through the iterative method and the 3-sigma criterion.

[0022] Further, in step S2, any two samples x in the gangue spectral dataset X i and x j The Euclidean distance D(x i ,x j ) is calculated as:

[0023]

[0024] Where: x j ∈R n , j=1,2,…,m.

[0025] Furthermore, in step S2-1, the average spectral value The calculation formula is:

[0026]

[0027] Furthermore, in step S2-3, the calculation formula of the average value μ is:

[0028]

[0029] The calculation formula of standard deviation σ is:

[0030]

[0031] Further, in step S3, the first deep residual network model M1 includes:

[0032] Convolutional layer, pooling layer, residual module, fully connected layer, and activation function ELU(x).

[0033] Further, in step S3, the convolution layer is used to receive input coal gangue spectral data and extract spectral features;

[0034] The pooling layer is used to compress the feature map and retain significant information;

[0035] The residual module is used to change the parameters of the convolution layer, increase the number of data channels, and reduce the size of the feature map to extract the characteristic information of the gangue spectral data;

[0036] The fully connected layer acts as a classifier and outputs two values.

[0037] Furthermore, in step S3, the calculation formula of the activation function ELU(x) is:

[0038]

[0039] Among them: x is the network input value, α is the parameter. Therefore, the activation function can reduce the impact of gradient explosion.

[0040] Further, in step S4, the binary cross entropy loss function The calculation formula is:

[0041]

[0042] Where: y is the true label value 0 or 1, is the predicted probability value, and log(·) represents the logarithm with base 2.

[0043] Furthermore, in step S6, the calculation formula of the accuracy rate ACC is:

[0044]

[0045] Among them: TP (True Positive) represents the number of samples correctly classified as coal, FP (False Positive) represents the number of samples misclassified as coal, TN (True Negative) represents the number of samples correctly classified as gangue, and FN (False Negative) represents the number of samples incorrectly classified as gangue.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] By collecting the spectral information of coal gangue and constructing a deep residual network, the spectral characteristics of coal and gangue are analyzed, and the intelligent sorting and identification of coal and gangue are realized. Compared with the existing sorting and identification methods, this method can suppress noise interference through its powerful feature extraction ability and robustness, thereby improving the accuracy and reliability of coal gangue spectral data, thereby reducing the workload of coal gangue sorting and identification and improving the efficiency and accuracy of coal gangue sorting and identification; externally, the deep residual network can capture the complex and subtle feature information in the coal gangue spectral data through its deep structure. These features are crucial to improving the classification, identification and prediction accuracy of spectral data. The deep residual network can effectively solve the gradient vanishing problem through its unique residual learning mechanism, thereby deeply mining the feature information of spectral data. This ability makes the deep residual network model have higher sensitivity and accuracy when processing complex spectral data. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0049] Figure 1 It is a flow chart of the coal gangue sorting method based on deep residual network and near infrared spectroscopy of the present invention;

[0050] Figure 2 is a flow chart of step S2 of the present invention;

[0051] Figure 3 is a near infrared spectrum of the sample of the present invention;

[0052] Figure 4 Schematic diagram of the structure of the residual module of the present invention;

[0053] Figure 5 It is a schematic diagram of the structure of ID-ResNet of the present invention;

[0054] Figure 6 It is a flowchart of the five-fold cross validation of the present invention;

[0055] Figure 7 It is a process diagram of removing abnormal values ​​of the present invention;

[0056] Figure 8 This is a loss change curve diagram during ID-ResNet training of the present invention. DETAILED DESCRIPTION

[0057] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0058] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In addition, features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0059] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0060] like Figures 1 to 8 As shown, it is the best embodiment of the present invention. The coal gangue sorting method based on deep residual network and near infrared spectroscopy in this embodiment includes the following steps:

[0061] S1. Constructing the spectral data set X of coal gangue: using near infrared spectrometer to collect the spectral data of coal gangue samples;

[0062] S2, removing outliers from the spectral data in step S1: using the Euclidean distance formula and an iterative method to gradually extract outliers from the spectral data, after the outliers in the spectral data are removed, the removed spectral data are divided into a training set A and a test set B;

[0063] S3, constructing a first deep residual network model M1;

[0064] S4. The first deep residual network model M1 in the training step S3: constructing a binary cross entropy loss function And using the spectral data in the training set A from which outliers have been removed in step S2 to train the first deep residual network model M1 to obtain the second deep residual network model M2;

[0065] S5, testing the second deep residual network model M2 in step S4: using the second deep residual network model M2 to test the spectral data of the test set B from which outliers have been removed in step S2, to obtain a prediction result;

[0066] S6. Analyze the prediction results in step S5: Use the accuracy evaluation index ACC to perform performance analysis on the prediction results. Therefore, by collecting the spectral information of coal gangue and constructing a deep residual network, analyzing the spectral characteristics of coal and gangue, the intelligent sorting and identification of coal and gangue is realized. Compared with the existing sorting and identification methods, this method can suppress noise interference through powerful feature extraction capabilities and robustness, thereby improving the accuracy and reliability of coal gangue spectral data, and then reducing the workload of coal gangue sorting and identification and improving the efficiency and accuracy of coal gangue sorting and identification; externally, the deep residual network can capture the complex and subtle feature information in the coal gangue spectral data (for example: multiple wavelength information such as ultraviolet to infrared, reflection, projection and emission information at different wavelengths, spectral line intensity, peak position, peak shape) through its deep structure. These features are crucial to improving the classification, identification and prediction accuracy of spectral data. The deep residual network can effectively solve the gradient vanishing problem through its unique residual learning mechanism, thereby deeply mining the feature information of spectral data. This ability makes the deep residual network model have higher sensitivity and accuracy when processing complex spectral data.

[0067] In other words, the spectral data of coal gangue provides information on multiple wavelengths of coal gangue from ultraviolet to infrared, reflection, projection and emission information at different wavelengths, spectral line intensity, peak position, peak shape and other complex and subtle characteristic information. These complex and subtle characteristic information can more deeply reflect the chemical composition, physical structure and mineral composition of coal gangue. In addition, the high resolution of spectral data enables the deep residual network model to capture the subtle differences in the spectral dimension of coal gangue, thereby providing a more comprehensive basis for the sorting and identification of coal gangue; the image data of coal gangue only contains the appearance, texture and shape information of coal gangue, but in comparison, the spectral data of coal gangue is more detailed and accurate in describing the physical and chemical properties.

[0068] In other words, the spectral data of coal gangue has the characteristics of large data flow and high dimension, which enables the deep residual network model to automatically extract useful feature information from the spectral data of coal gangue (for example: multiple wavelength information such as ultraviolet to infrared, reflection, projection and emission information at different wavelengths, spectral line intensity, peak position, peak shape) without human intervention, reducing the impact of human factors on data quality, thus improving the accuracy of coal gangue sorting and identification.

[0069] Specifically, the first deep residual network model M1 refers to an untrained model, and the second deep residual network model M2 refers to a trained model.

[0070] Specifically, in step S1, the samples are numbered according to their origin and category, and then the samples are placed one by one under the near-infrared spectrometer for scanning to obtain spectral information. In this experiment, coal and gangue samples are collected from three regions A, B, and C respectively, and the spectral data of the samples are collected using the near-infrared spectrometer of VIAVI. The number of samples in each region is shown in Table 1; in order to reduce measurement errors, each sample will be scanned 5 times, and the average value will be taken as the final measurement result; since coal and gangue are both solid samples, the near-infrared spectrometer is set to diffuse reflection mode during collection, and the spectral wavelength range is 908.1nm-1676.2nm, with a total of 125 wavelength points, and background correction is performed every ten minutes; the sample spectrum is shown in Table 1. Figure 3 As shown, the horizontal axis represents the wavelength point, and the vertical axis represents the absorbance of the sample at the wavelength point.

[0071] Table 1 Number of coal and gangue samples from different regions

[0072]

[0073] In this embodiment, step S2 includes the following steps:

[0074] S2-1, calculate the average spectral value of the gangue spectral data set X in step S1

[0075] S2-2. Calculate the average spectral value of each spectral data in the gangue spectral data set X according to the Euclidean distance formula The Euclidean distance

[0076] S2-3, calculate the mean μ and standard deviation σ of each Euclidean distance D;

[0077] S2-4, when When x i is an outlier, and x i Eliminate the gangue spectral dataset X;

[0078] S2-5, when spectral data is removed from the gangue spectral data set X, repeat steps S2-1 to S2-4 until no outliers appear, and divide the removed spectral data into a training set A and a test set B;

[0079] Any two samples x in the gangue spectral dataset X i and x j The Euclidean distance D(x i ,x j) is calculated as:

[0080]

[0081] Average spectral value The calculation formula is:

[0082]

[0083] The calculation formula for the average value μ is:

[0084]

[0085] The calculation formula of standard deviation σ is:

[0086]

[0087] Where: X = {x1, x2, ..., x m}, x i 、x j ∈R n , i, j = 1, 2, ..., m, m is the number of samples, n is the sample dimension. Therefore, in the process of sample collection, environmental changes or inappropriate operations may cause abnormal values ​​that are significantly different from normal spectra in the spectral data set. The existence of these abnormal values ​​will not only interfere with data analysis, but also mislead the training of the model, thus affecting the final test results. Therefore, before modeling, it is necessary to remove the abnormal data in the spectral data set; since the sample distribution of the collected spectral data follows the normal distribution, the abnormal values ​​in the sample spectral data set can be gradually removed through the iterative method and the 3-sigma criterion.

[0088] Specifically, in order to avoid the interference of outliers, this experiment iteratively removes outliers based on the Euclidean distance according to the sample origin and category (for example, as shown in Table 2, the 2 spectra in the gangue in area A and the 46 spectra in area C are judged as outliers; the outlier determination process, such as Figure 7 As shown in the figure, in the first iteration of the gangue samples in area A, the Euclidean distance between 2 samples and the average spectral value exceeded the threshold, so they were removed as outliers. In the second iteration, no outliers were found and the iteration was terminated. Similarly, after 46 samples of coal in area C were removed as outliers, no outliers were found in the remaining sample sets. After the removal, the total number of samples in the gangue spectral data set became 428).

[0089] Table 2 Outlier detection results

[0090] type A B C coal / / 46 Gangue 2 / /

[0091] In this embodiment, in step S3, the first deep residual network model M1 includes: a convolutional layer, a pooling layer, a residual module, a fully connected layer, and an activation function ELU(x); the convolutional layer is used to receive input coal gangue spectral data and extract spectral features; the pooling layer is used to compress the feature map and retain significant information; the residual module (such as Figure 5 As shown in the figure, the residual module has the same structure) is used to change the parameters of the convolution layer, increase the number of data channels, and reduce the size of the feature map to extract the characteristic information of the coal gangue spectral data; the fully connected layer is used as a classifier to output two values;

[0092] The calculation formula of the activation function ELU(x) is:

[0093]

[0094] Among them: x is the network input value, α is the parameter. Therefore, the activation function can reduce the impact of gradient explosion.

[0095] Specifically, the residual network is a deep neural network model with the residual module as the core, which can alleviate the gradient vanishing problem faced by the neural network when the depth increases; a typical residual module, such as Figure 4 As shown in Figure 1, the essence of its design lies in the introduction of identity mapping; ResNet was originally proposed to process two-dimensional image data, while near-infrared spectrum is a typical one-dimensional data, so it is necessary to adjust the network parameters to enable it to process one-dimensional spectral data; The structure diagram of the first deep residual network model M1 is shown in Figure 1. Figure 5 As shown in Figure 3, first, the two-dimensional convolution layer is adjusted to a one-dimensional convolution layer. Secondly, the specific parameters in each module (for example, the number of convolution layers, the number of neurons in the fully connected layer, and the type of activation function) are adjusted. After adjustment, the number of convolution layers in the residual module is 2, the convolution kernel size is 3, and the sliding step size and the number of padding are both 1; Among them: ELU, as an activation function, can effectively alleviate the gradient vanishing problem caused by the increase in depth of the neural network; The structure and data dimension changes of ID-ResNet are shown in Table 3; Taking the change process of a single sample in the output dimension as an example, the first dimension is the number of samples, the second dimension is the number of data points for each sample, and the third dimension is the number of channels. The initial dimension of a single sample spectrum is (1, 125), which is the input data of the network model.

[0096] Table 3. ID-ResNet network parameters

[0097] Serial number Network Layer Number Output Dimensions 1 Convolutional Layer 1 (1，63，4) 2 Batch normalization 1 (1，63，4) 3 ELU 1 (1，63，4) 4 Max Pooling Layer 1 (1，63，4) 5 Residual Module 2 (1，63，4) 6 Residual Module 2 (1，32，16) 7 Residual Module 2 (1，16，32) 8 Residual Module 2 (1，8，64) 9 Global Pooling Layer 1 (1，1，64) 10 Flatten 1 (1，64) 11 Fully connected layer 1 (1，16) 12 ELU 1 (1，16) 13 Fully connected layer 1 (1，2)

[0098] In this embodiment, in step S4, the binary cross entropy loss function The calculation formula is:

[0099]

[0100] Where: y is the true label value 0 or 1, is the predicted probability value, and log(·) represents the logarithm with base 2. Specifically, the loss change curve of ID-ResNet during training is as follows: Figure 8 As shown in the figure, each sub-graph represents one of the five-fold cross-validation folds; in the initial period of training (0-400 rounds), the loss of training set A and test set B drops rapidly, indicating that ID-ResNet has good fitting ability and can quickly and accurately adjust model parameters under the guidance of labels; at the end of training (400-600 rounds), the loss of training set A and test set B tends to stabilize and remain near a small value, indicating that the model has good convergence.

[0101] In this embodiment, in step S6, the calculation formula of the accuracy rate ACC is:

[0102]

[0103] Among them: TP (True Positive) represents the number of samples correctly classified as coal, FP (False Positive) represents the number of samples misclassified as coal, TN (True Negative) represents the number of samples correctly classified as gangue, and FN (False Negative) represents the number of samples misclassified as gangue. Specifically, Figure 6 As shown in the figure, the samples are divided into 5 parts on average, and 1 part is selected as the test set each time, and the remaining 4 parts are used as training sets. A total of 5 experiments are carried out to ensure that each piece of data will be tested as a test set. In order to avoid the influence of sample imbalance on the experimental results, the number of coal mines and gangue in each sample is roughly the same when dividing the samples. Finally, the average value of the 5 prediction accuracy rates ACC is calculated and used to measure the performance of the model. In order to objectively evaluate the performance of ID-ResNet, this experiment compares it with the commonly used classification models in the field of near-infrared spectroscopy, including: support vector machine, decision tree, random forest and K nearest neighbor. Each method obtains the classification accuracy ACC through five-fold cross validation. The comparison results are shown in Table IV.

[0104] Specifically, in the five-fold cross validation, the average recognition accuracy of the deep residual network model reached 96.26%, which is better than traditional machine learning methods such as support vector machines and K nearest neighbors.

[0105] Table 4 Comparison of classification accuracy ACC of different models

[0106]

[0107]

[0108] In summary, the present invention collects spectral information of coal gangue and constructs a deep residual network to analyze the spectral characteristics of coal and gangue, thereby realizing intelligent sorting and identification of coal and gangue. Compared with the existing sorting and identification methods, this method can suppress noise interference through powerful feature extraction ability and robustness, thereby improving the accuracy and reliability of coal gangue spectral data, and further reducing the workload of coal gangue sorting and identification and improving the efficiency and accuracy of coal gangue sorting and identification; externally, the deep residual network can capture the complex and subtle feature information in the coal gangue spectral data through its deep structure, and these features are crucial to improving the classification, identification and prediction accuracy of spectral data. The deep residual network can effectively solve the gradient vanishing problem through its unique residual learning mechanism, thereby deeply mining the feature information of spectral data. This ability enables the deep residual network model to have higher sensitivity and accuracy when processing complex spectral data.

[0109] The above description is based on the ideal embodiment of the present invention. Through the above description, relevant personnel can make various changes and modifications without departing from the technical concept of the present invention. The technical scope of the present invention is not limited to the content in the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A coal gangue sorting method based on deep residual network and near infrared spectroscopy, characterized in that: The following steps are involved: S1. Constructing the spectral data set X of coal gangue: using near infrared spectrometer to collect the spectral data of coal gangue samples; S2, removing outliers from the spectral data in step S1: using the Euclidean distance formula and an iterative method to gradually extract outliers from the spectral data, after the outliers in the spectral data are removed, the removed spectral data are divided into a training set A and a test set B; S3, constructing a first deep residual network model M1; S4. The first deep residual network model M1 in the training step S3: constructing a binary cross entropy loss function And using the spectral data in the training set A from which outliers have been removed in step S2 to train the first deep residual network model M1 to obtain the second deep residual network model M2; S5, testing the second deep residual network model M2 in step S4: using the second deep residual network model M2 to test the spectral data of the test set B from which outliers have been removed in step S2, to obtain a prediction result; S6. Analyze the prediction results in step S5: Use the accuracy evaluation index ACC to perform performance analysis on the prediction results.

2. The method for separating coal gangue based on deep residual network and near infrared spectroscopy according to claim 1, characterized in that: The step S2 comprises the following steps: S2-1, calculate the average spectral value of the gangue spectral data set X in step S1 S2-2. Calculate the average spectral value of each spectral data in the gangue spectral data set X according to the Euclidean distance formula The Euclidean distance S2-3, calculate the mean μ and standard deviation σ of each Euclidean distance D; S2-4, when When x i is an outlier, and x i Eliminate the gangue spectral dataset X; S2-5, when spectral data is removed from the gangue spectral data set X, repeat steps S2-1 to S2-4 until no outliers appear, and divide the removed spectral data into a training set A and a test set B; Where: X = {x1, x2, ..., x m }, x i ∈R n , i=1,2,…,m, m is the number of samples, and n is the sample dimension.

3. The method for separating coal gangue based on deep residual network and near infrared spectroscopy according to claim 1, characterized in that: In step S2, any two samples x in the gangue spectral dataset X i and x j The Euclidean distance D(x i ,x j ) is calculated as: Where: x j ∈R n , j=1,2,…,m.

4. The method for separating coal gangue based on deep residual network and near infrared spectroscopy according to claim 2, characterized in that: In step S2-1, the average spectral value The calculation formula is:

5. The method for separating coal gangue based on deep residual network and near infrared spectroscopy according to claim 2, characterized in that: In step S2-3, the calculation formula of the average value μ is: The calculation formula of standard deviation σ is:

6. The method for separating coal gangue based on deep residual network and near infrared spectroscopy according to claim 1, characterized in that: In step S3, the first deep residual network model M1 includes: Convolutional layer, pooling layer, residual module, fully connected layer, and activation function ELU(x).

7. The method for separating coal gangue based on deep residual network and near infrared spectroscopy according to claim 6, characterized in that: In step S3, the convolution layer is used to receive input coal gangue spectral data and extract spectral features; The pooling layer is used to compress the feature map and retain significant information; The residual module is used to change the parameters of the convolution layer, increase the number of data channels, and reduce the size of the feature map to extract the characteristic information of the gangue spectral data; The fully connected layer acts as a classifier and outputs two values.

8. The method for separating coal gangue based on deep residual network and near infrared spectroscopy according to claim 6, characterized in that: In step S3, the calculation formula of the activation function ELU(x) is: Among them: x is the network input value, α is the parameter.

9. The method for separating coal gangue based on deep residual network and near infrared spectroscopy according to claim 1, characterized in that: In step S4, the binary cross entropy loss function The calculation formula is: Where: y is the true label value 0 or 1, is the predicted probability value, and log(·) represents the logarithm with base 2.

10. The method for separating coal gangue based on deep residual network and near infrared spectroscopy according to claim 1, characterized in that: In step S6, the calculation formula of the accuracy rate ACC is: Among them: TP (True Positive) represents the number of samples correctly classified as coal, FP (False Positive) represents the number of samples misclassified as coal, TN (True Negative) represents the number of samples correctly classified as gangue, and FN (False Negative) represents the number of samples incorrectly classified as gangue.

Images