XRF ore element content prediction method based on SVD-CLSTM

Abstract

The invention belongs to the field of ore element content measurement, and particularly provides an XRF ore element content prediction method based on SVD-CLSTM, which is used for solving the problems of inaccurate calculation and lack of data test of a traditional X fluorescence spectrometer. According to the method, an LSTM layer is creatively applied to an ore element content prediction model to form a CLSTM model, SVD singular value noise reduction and the CLSTM model are combined to establish a relationship between peak value feature information and element content of sample elements, a convolution block carries out local feature extraction on XRF data, the LSTM layer carries out nonlinear modeling between channels to capture important features, and the content of the elements in the sample elements is predicted. The element content in the ore is accurately predicted, and the reliability and the accuracy of quantitative analysis of the target element are improved; in conclusion, the method has the advantages of non-destructive property, high efficiency, high detection precision, high prediction accuracy and the like, solves the problem of X-ray fluorescence energy spectrum peak value overlapping interference, and is of great significance to industrial production and environmental protection.

Description

Technical Field

[0001] The invention belongs to the field of ore element content (concentration) measurement, relates to an XRF element content prediction method, and in particular provides an XRF ore element content prediction method based on SVD-CLSTM, so as to realize accurate measurement of XRF ore element content. Background Art

[0002] The appropriate element concentration level in ore is crucial to industrial production and environmental protection, and is of great significance for the rational development and utilization of mineral resources; however, when the element concentration in the ore exceeds the standard, it will not only affect the processing and utilization efficiency of the ore, but may also cause pollution to the environment and affect human health. Traditional methods for detecting the element content of ores usually take a lot of time and resources and are destructive, which limits their effectiveness and feasibility in practical applications. In contrast, X-ray fluorescence spectroscopy (XRF), as a non-destructive testing technology, has the advantages of rapid, non-destructive, and multi-element analysis, and is widely used in the determination of ore element content. The sample spectrum and element content in the ore sample can be obtained by XRF technology; however, due to the complexity and nonlinear characteristics of ore samples, traditional statistical methods or machine learning methods alone often cannot accurately predict and analyze the content of elements in the ore. Therefore, a new non-destructive testing technology is urgently needed to accurately determine the element content in the ore, solve the problem of inaccurate calculation and lack of data verification of traditional X-ray fluorescence spectrometers, and provide an alternative measurement method for the quantitative analysis of elements in ores. Summary of the invention

[0003] The purpose of the present invention is to provide an XRF ore element content prediction method based on SVD-CLSTM to solve the problems of inaccurate calculation and lack of data verification of traditional X-ray fluorescence spectrometers; the present invention first adopts XRF non-destructive testing technology to obtain the spectrum of the ore, then denoises it through a singular value denoising (SVD) method, and finally uses a convolutional long short-time network model (CLSTM) to establish the relationship between the peak characteristic information of the element and the element content, thereby realizing the element content prediction, and effectively improving the reliability and accuracy of quantitative analysis.

[0004] To achieve the above object, the technical solution adopted by the present invention is:

[0005] An XRF ore element content prediction method based on SVD-CLSTM, characterized by comprising the following steps:

[0006] The sample data set is obtained by measuring the ore sample using a spectrometer (XRF), and then divided into a training set. Each training sample includes: the spectrum peak intensity information of the ore sample and the content of the target element therein;

[0007] The SVD singular value denoising method is used to denoise the peak intensity information of the spectrum of the training set to obtain the model input data;

[0008] A convolutional long short-term network model (CLSTM) is constructed and trained based on a training set; the convolutional long short-term network model includes: three convolutional blocks, an LSTM layer and a fully connected layer connected in sequence, the model input data is subjected to local feature extraction through the three convolutional blocks, the local features are subjected to nonlinear feature extraction through the LSTM layer, and the nonlinear features are subjected to regression prediction through the fully connected layer to obtain a content prediction result;

[0009] A spectrometer is used to measure the ore to be tested to obtain its spectrum peak intensity information, and the SVD singular value denoising method is used for denoising. The denoised data is sent to the trained convolutional long-short time network model, and the model outputs the predicted results of the target element content in the ore to be tested.

[0010] Furthermore, the specific process of the SVD singular value denoising method is as follows:

[0011] The peak intensity information of the spectrum of N sample data in the training set is used to form a peak intensity matrix X. First, the peak intensity matrix X is subjected to singular value decomposition, which is specifically expressed as:

[0012] X=U·Σ·V T

[0013] Where X is the peak intensity matrix, U is the left singular value matrix, V is the right singular value matrix, and Σ is the diagonal matrix of all eigencomponents;

[0014] Set the main feature division threshold M, select the first M main feature components of the full feature component diagonal matrix Σ, set the other feature components to 0, and obtain the main feature component diagonal matrix

[0015] By the main eigencomponent diagonal matrix Reconstruct the peak intensity matrix X after noise reduction d , complete SVD singular value denoising, peak intensity matrix X d It is expressed as:

[0016]

[0017] Among them, X d is the peak intensity matrix after denoising.

[0018] Furthermore, each convolution module consists of a convolution layer, an activation function layer (ReLu) and a maximum pooling layer (Max Pool).

[0019] Furthermore, in the LSTM layer, the calculation formula of the LSTM input gate is:

[0020] it =σ(W i ·[h t-1 ,X C ]+b i )

[0021] The calculation formula of the forget gate of LSTM is:

[0022] f t =σ(W f ·[h t-1 ,X C ]+b f )

[0023] The calculation formula of the output gate of LSTM is:

[0024] O t =σ(W O ·[h t-1 ,X C ]+b O )

[0025] The state update calculation formula of LSTM is:

[0026] u t =tanh(W u ·[h t-1 ,X C ]+b u )

[0027] The calculation formula of LSTM memory unit is:

[0028] C t =f t ·C t-1 +i t ·u t

[0029] The output calculation formula of LSTM is:

[0030] h t =O t tanh(C t )

[0031] Among them, i t is the output of the input gate, f t is the output of the forget gate, O t is the output of the output gate, U t is the cell state vector, C t is the output of the current memory unit, C t-1 is the output of the previous memory unit, h t is the hidden state of the LSTM unit at time step t, h t-1is the hidden state of the LSTM unit at time step t-1, X C is the local feature; σ is the sigmoid activation function, tanh is the hyperbolic tangent function; W i , W f , W O , W i Corresponding to the weight matrix, b i 、b f 、b O 、b u The corresponding offset.

[0032] Furthermore, the training process of the convolutional long short-term network model (CLSTM) is as follows: the model input data is sent to the convolutional long short-term network model (CLSTM) for training, the loss function is set to the MSE function, and the Adam optimizer is used to complete the training of the model.

[0033] Based on the above technical solution, the beneficial effects of the present invention are:

[0034] The present invention proposes an XRF ore element content prediction method based on SVD-CLSTM, creatively applies an LSTM layer to an ore element content prediction model to form a CLSTM model, combines SVD singular value denoising with the CLSTM model to establish a relationship between peak feature information of sample elements and element content, and finally realizes the content prediction of target elements; specifically, firstly, an XRF non-destructive testing technology is used to obtain spectrum peak intensity information, and then the SVD singular value denoising and the CLSTM model are combined to complete the element content prediction, local feature extraction is performed on XRF data through a convolution block, and then LSTM is used to perform nonlinear modeling between channels to capture important features, accurately predict the element content in the ore, and improve the reliability and accuracy of quantitative analysis; in summary, the present invention has the advantages of non-destructiveness, high efficiency, high detection accuracy, high prediction accuracy, etc., solves the problem of peak overlap interference of X-ray fluorescence energy spectrum, and is of great significance to industrial production and environmental protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the process of the XRF ore element content prediction method based on SVD-CLSTM in the present invention.

[0036] Figure 2 This is a structural principle diagram of the convolutional long short-term network model (CLSTM) in the present invention.

[0037] Figure 3 This is a diagram of the Fe element content prediction results based on SVD-CLSTM in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solution and beneficial effects of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0039] This embodiment provides a method for predicting the element content of XRF ore based on SVD-CLSTM. The target element is iron (Fe). The content prediction process is as follows: Figure 1 As shown, the specific steps include:

[0040] Step 1. Data collection;

[0041] In this embodiment, 174 pieces of ore are used as samples, and the ED-XRF fluorescence spectrometer is used to measure the ore samples, identify all elements in the ore samples, and form an element set, which is an element library of elements 1 to 100 in the periodic table; finally, 174 pieces of sample data are measured using the ED-XRF fluorescence spectrometer to form a data set, and each piece of sample data includes: a spectrum peak intensity vector of the ore sample and the target element content therein;

[0042] The above data set is divided into 5 fold cross validation for a total of 5 times; during the division process: the data set is divided into 5 subsets, 1 subset is used as the test set, and the remaining 4 subsets are used as training sets, and a different subset is selected as the test set each time;

[0043] Step 2. Use SVD singular value denoising method to denoise the training set;

[0044] The peak intensity vectors of the spectrum of N sample data in the training set are used to form a peak intensity matrix X. First, the peak intensity matrix X is subjected to singular value decomposition, which is specifically expressed as:

[0045] X=U·Σ·V T

[0046] Where X is the peak intensity matrix, U is the left singular value matrix, V is the right singular value matrix, Σ is the full eigencomponent diagonal matrix, the superscript “T” indicates the transpose of the matrix, and “·” indicates matrix multiplication;

[0047] In this embodiment, the first 6 main characteristic components of the full characteristic component diagonal matrix Σ are selected for the Fe element, and the others are set to 0 to obtain the main characteristic component diagonal matrix

[0048] By the main eigencomponent diagonal matrix Reconstruct the peak intensity matrix X after noise reduction d , complete SVD singular value denoising, peak intensity matrix X d It is expressed as:

[0049]

[0050] Among them, X d is the peak intensity matrix after noise reduction;

[0051] Step 3. Construct a convolutional long short-term network model (CLSTM) and complete the training based on the training set;

[0052] Convolutional Long Short-Term Network Model (CLSTM) Figure 2 As shown, it includes: 3 convolutional blocks, LSTM layer and fully connected layer connected in sequence;

[0053] Each convolution module consists of a convolution layer, an activation function layer (ReLu) and a maximum pooling layer (Max Pool). In this embodiment, the convolution layer in the three convolution blocks is represented as Conv(A×B,C:D), where A×B represents the convolution kernel size, C represents the number of input channels, and D represents the number of output channels;

[0054] Peak intensity matrix X after noise reduction d After the CNN module performs local feature extraction, the local feature X is obtained. C ; Local feature X C Input into the LSTM layer, and use the LSTM layer to extract nonlinear features in the data. LSTM includes the calculation of input gate, forget gate and output gate. The calculation formula of LSTM input gate is:

[0055] i t =σ(W i ·[h t-1 ,X C ]+b i )

[0056] The calculation formula of the forget gate of LSTM is:

[0057] f t =σ(W f ·[h t-1 ,X C ]+b f )

[0058] The calculation formula of the output gate of LSTM is:

[0059] O t =σ(W O ·[h t-1 ,X C ]+b O )

[0060] The state update calculation formula of LSTM is:

[0061] u t =tanh(W u ·[h t-1 ,X C ]+bu )

[0062] The calculation formula of LSTM memory unit is:

[0063] C t =f t ·C t-1 +i t ·u t

[0064] The output calculation formula of LSTM is:

[0065] h t =O t tanh(C t )

[0066] Among them, i t is the output of the input gate, f t is the output of the forget gate, O t is the output of the output gate, U t is the cell state vector, C t is the output of the current memory unit, C t-1 is the output of the previous memory unit, h t is the hidden state of the LSTM unit at time step t, h t-1 is the hidden state of the LSTM unit at time step t-1; σ is the sigmoid activation function, tanh is the hyperbolic tangent function; W i , W f , W O , W u Corresponding to the weight matrix, b i 、b f 、b o 、b u The corresponding is bias;

[0067] The output features of the LSTM layer are put into the fully connected layer and regression prediction is performed. The output layer outputs the predicted results of the target element content.

[0068] The training set is sent to the convolutional long short-term network model (CLSTM) for training, the loss function is set to the MSE function, and the Adam optimizer is used to complete the training of the model; in this embodiment, the Adam optimizer is used as the optimization algorithm to update the network parameters and minimize the loss function, and the learning rate is set to 0.001, the number of training iterations is 150 times, and the learning rate reduction factor is 0.5;

[0069] Step 4. Use the test set to test and use the determination coefficient R 2 The prediction effect is evaluated by the root mean square error RMSE; 2 It is used to describe the degree of fit between the regression line and the observed value.2 The closer it is to 1, the better the fitting effect of the surface model; RMSE reflects the degree of deviation between the predicted value and the true value. The smaller the RMSE, the higher the prediction accuracy of the surface model; the coefficient of determination R 2 The root mean square error RMSE is specifically expressed as:

[0070]

[0071] Where N is the number of samples, is the predicted value of the target element content in the sample, Y l is the measured value of the target element content in the sample, Y l The average value of .

[0072] To illustrate the beneficial effects of the present invention, the traditional random forest prediction method is used as a comparative example. The prediction results of the Fe element content based on SVD-CLSTM in this embodiment are as follows: Figure 3 As shown, further, the element content prediction results of the 5-fold cross validation of this embodiment and the comparative example are shown in Table 1;

[0073] Table 1

[0074]

[0075] As can be seen from the chart, compared with the traditional random forest prediction method, the present invention has a higher R 2 and lower RMSE, indicating that the XRF ore element prediction value of SVD-CLSTM in the present invention is closer to the true value of the element content, and also indicating that the XRF ore element prediction method of SVD-CLSTM in the present invention can effectively solve the problem of spectral line overlap, improve the accuracy of element quantitative analysis results, and embody the superiority of the present invention.

[0076] The above description is only a specific implementation mode of the present invention. Any feature disclosed in this specification, unless otherwise stated, can be replaced by other alternative features that are equivalent or have similar purposes; all the disclosed features, or all the steps in the methods or processes, except for mutually exclusive features and / or steps, can be combined in any way.

Claims

1. A method for predicting the element content of XRF ore based on SVD-CLSTM, characterized in that: The following steps are involved: The sample data set is obtained by measuring the ore sample using a spectrometer (XRF), and then divided into a training set. Each training sample includes: the spectrum peak intensity information of the ore sample and the content of the target element therein; The SVD singular value denoising method is used to denoise the peak intensity information of the spectrum of the training set to obtain the model input data; A convolutional long short-term network model (CLSTM) is constructed and trained based on a training set; the convolutional long short-term network model includes: three convolutional blocks, an LSTM layer and a fully connected layer connected in sequence, the model input data is subjected to local feature extraction through the three convolutional blocks, the local features are subjected to nonlinear feature extraction through the LSTM layer, and the nonlinear features are subjected to regression prediction through the fully connected layer to obtain a content prediction result; A spectrometer is used to measure the ore to be tested to obtain its spectrum peak intensity information, and the SVD singular value denoising method is used for denoising. The denoised data is sent to the trained convolutional long-short time network model, and the model outputs the predicted results of the target element content in the ore to be tested.

2. The XRF ore element content prediction method based on SVD-CLSTM according to claim 1, characterized in that: The specific process of the SVD singular value denoising method is: The peak intensity information of the spectrum of N sample data in the training set is used to form a peak intensity matrix X. First, the peak intensity matrix X is subjected to singular value decomposition, which is specifically expressed as: X=U·S·V T Where X is the peak intensity matrix, U is the left singular value matrix, V is the right singular value matrix, and Σ is the diagonal matrix of all eigencomponents; Set the main feature division threshold M, select the first M main feature components of the full feature component diagonal matrix Σ, set the other feature components to 0, and obtain the main feature component diagonal matrix By the main eigencomponent diagonal matrix Reconstruct the peak intensity matrix X after noise reduction d , complete SVD singular value denoising, peak intensity matrix X d It is expressed as: Among them, X d is the peak intensity matrix after denoising.

3. The XRF ore element content prediction method based on SVD-CLSTM according to claim 1 is characterized in that: Each convolution module consists of a convolution layer, an activation function layer (ReLu) and a maximum pooling layer (Max Pool). Furthermore, in the LSTM layer, the calculation formula of the LSTM input gate is: i t =σ(W i ·[h t-1 ,X C ]+b i ) The calculation formula of the forget gate of LSTM is: f t =σ(W f ·[h t-1 ,X C ]+b f ) The calculation formula of the output gate of LSTM is: The t =σ(W O ·[h t-1 ,X C ]+b O ) The state update calculation formula of LSTM is: u t =tanh(W u ·[h t-1 ,X C ]+b u ) The calculation formula of LSTM memory unit is: C t =f t ·C t-1 +i t ·u t The output calculation formula of LSTM is: h t =O t ·tanh(C t ) Among them, i t is the output of the input gate, f t is the output of the forget gate, O t is the output of the output gate, U t is the cell state vector, C t is the output of the current memory unit, C t-1 is the output of the previous memory unit, h t is the hidden state of the LSTM unit at time step t, h t-1 is the hidden state of the LSTM unit at time step t-1, X c is the local feature; σ is the sigmoid activation function, tanh is the hyperbolic tangent function; W i , W f , W O , W u Corresponding to the weight matrix, b i 、b f 、b O 、b u The corresponding offset.

4. The XRF ore element content prediction method based on SVD-CLSTM according to claim 1, characterized in that: The training process of the convolutional long short-term network model (CLSTM) is as follows: the model input data is sent to the convolutional long short-term network model for training, the loss function is set to the MSE function, and the Adam optimizer is used to complete the model training.

Images