Ore element grade detection method

Abstract

The invention discloses an ore element grade detection method. The method comprises the following steps: respectively obtaining intensity values of X-fluorescence rays absorbed by interference elements and elements to be detected in an ore sample to be detected; inputting the intensity value of the X-ray absorbed by the to-be-detected element into a pre-trained preliminary grade value prediction model to obtain a preliminary grade value of the to-be-detected element; the preliminary grade value prediction model is constructed by using a unary linear regression method; inputting the intensity values of the X fluorescence rays absorbed by the interference elements and the to-be-detected elements into a pre-trained correction compensation model to obtain correction compensation values of the to-be-detected elements; the correction compensation model is constructed by using a dynamic regularization random weight neural network; and taking a sum value of the preliminary grade value and the correction compensation value as a final grade estimation value of the to-be-detected element, so as to obtain a detection result of the to-be-detected ore element grade. According to the method, errors caused by mutual interference among matrix elements are effectively analyzed in an ore dressing field environment, and the accuracy of a detection result is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of ore detection, and in particular to a method for detecting the element grade of an ore. Background Art

[0002] Against the backdrop of current social structures and economic expansion, the consumption of mineral resources in production and daily life is gradually increasing, and the number of mineral mining projects is also on the rise. To optimize the utilization efficiency of mineral resources, technicians urgently need to identify the types of metal elements contained in the ore and then specify appropriate resource utilization strategies based on the specific composition of the ore. To achieve this goal, it is necessary to strengthen research on metal element detection technology and promote the development of artificial intelligence-based detection algorithms and systems. These technologies must not only be scientific and efficient, but also be cutting-edge, thereby improving the efficiency and accuracy of the detection process. These measures can ensure the accuracy of metal element detection results, thereby providing solid technical support for the sustainable development of the production and construction sectors.

[0003] In current XRF testing research, the integration of artificial intelligence and machine learning algorithms is driving technological innovation. While widely used, traditional empirical coefficient and fundamental parameter methods are limited in accuracy and applicability. Leveraging advanced machine learning techniques such as neural networks, researchers are developing more precise and efficient quantitative analysis methods that can better correct for matrix effects and improve the accuracy of elemental content predictions. Furthermore, by introducing new correction parameters and optimizing algorithms, XRF testing technology is overcoming the shortcomings of traditional methods to meet the growing demands for complex sample analysis and real-time performance.

[0004] While past research has yielded some success in correcting matrix effects, many challenges remain. Instabilities can occur during neural network training and prediction. To achieve higher-precision results, further research is needed to optimize neural network algorithms. Typically, neural networks based on the back-propagation (BP) algorithm, after initializing weights and thresholds, use gradient descent to continuously adjust these weights and thresholds to minimize the loss function. However, this approach can easily lead to local optima. Summary of the Invention

[0005] The present invention aims to address, at least to some extent, one of the technical problems in the related art. To this end, a first object of the present invention is to provide a method for detecting elemental grade in ore, which addresses issues such as unstable predictions, overfitting of network models, and large prediction errors in existing XRF detection systems.

[0006] To achieve the above objectives, a first embodiment of the present invention provides a method for detecting the element grade of an ore, the method comprising:

[0007] S1, using an XRF detection system to scan the ore sample to be tested to obtain the X-ray fluorescence intensity values ​​absorbed by the interfering elements and the element to be tested in the ore sample to be tested, wherein the ore to be tested is magnetite, the element to be tested is iron, and the interfering elements are silicon, calcium, silicon dioxide and air absorption;

[0008] S2, inputting the X-ray fluorescence intensity value absorbed by the element to be measured into a pre-trained preliminary grade value prediction model to obtain a preliminary grade value of the element to be measured; the preliminary grade value prediction model is constructed using a univariate linear regression method;

[0009] S3, inputting the X-ray fluorescence intensity values ​​absorbed by the interfering element and the element to be measured into a pre-trained correction compensation model to obtain a correction compensation value of the element to be measured; the correction compensation model is constructed using a dynamic regularized random weight neural network;

[0010] S4, taking the sum of the preliminary grade value and the correction compensation value as the final grade estimation value of the element to be measured, so as to obtain the test result of the element grade of the ore to be measured.

[0011] In addition, the method for detecting the element grade of ore according to the above embodiment of the present invention may also have the following additional technical features:

[0012] According to one embodiment of the present invention, before training the preliminary grade value prediction model, it is necessary to obtain the data set required for training, including:

[0013] Scan multiple ore samples using an XRF detection system to obtain the X-ray intensity value x absorbed by the interfering elements and the element to be measured in each ore sample, and perform manual analysis on each ore sample to obtain the actual grade value y of each ore sample;

[0014] Construct a first data set D = {x, y} using the X-ray fluorescence intensity value x absorbed by the interfering element and the element to be measured in each ore sample and the actual grade value y of each ore sample;

[0015] Normalize the first data set D = {x, y} to obtain the input feature matrix A;

[0016] The ADMM algorithm is fused with the Lasso regression feature selection algorithm to select the main features of the feature matrix A, and the first feature set S = {x * ,y}, where x * As the main feature.

[0017] According to one embodiment of the present invention, the ADMM algorithm is used to fuse the feature selection algorithm of Lasso regression to perform main feature selection on the feature matrix A, including:

[0018] The goal of Lasso regression is to minimize the linear regression loss with L1 regularization:

[0019]

[0020] Where A represents the input feature matrix, i.e., the first standardized dataset D = {x, y}, with a dimension of m × n, where m is the number of samples and n is the number of features; θ is the weight coefficient vector to be solved, with a dimension of n × 1; z represents the auxiliary variable; b represents the output variable vector, with a dimension of m × 1; τ represents the L1 regularization parameter, which is used to control the sparsification strength;

[0021] In order to solve the non-smoothness caused by L1 regularization, ADMM introduces an auxiliary variable z and transforms the original problem into a constrained optimization problem:

[0022]

[0023] st:θ-z=0

[0024] Decompose the problem into three solvable subproblems and construct the augmented Lagrangian function:

[0025]

[0026] Step 1: Closed-form solution of the θ subproblem

[0027]

[0028] θ (k+1) =(A T A+ρI) -1 (A T b+ρ(z (k) -u (k) ))

[0029] Where ρ represents the penalty parameter of the ADMM algorithm; I represents the unit matrix with dimension n×n; z (k) represents the auxiliary variable of the kth iteration; u (k) represents the Lagrange multiplier of the kth iteration;

[0030] Step 2: Soft threshold function for the z subproblem

[0031]

[0032] z (k+1) =S λ / ρ (θ (k+1) +u(k) )

[0033]

[0034] Where v represents the input vector θ (k+1) +u (k) .

[0035] Step 3: Multiplier update formula

[0036] u (k+1) =u (k) +θ (k+1) -z (k+1)

[0037] Step 4: Feature selection results

[0038] θ selected =[13.92,2.65,3.19,0,0];

[0039] Thus, the first feature set S={x * ,y}.

[0040] According to one embodiment of the present invention, training a preliminary grade value prediction model includes:

[0041] The first feature set S={x * ,y} is divided into the first training feature set X r and the first test feature set X t , where the first training feature set contains k ore samples and the first test feature set contains nk ore samples;

[0042] Construct a univariate linear regression method, the expression is:

[0043] Y i =aX i +b,

[0044] Among them, Y i Represents the grade value of the element to be tested, X i represents the intensity of X-ray fluorescence absorbed by the element to be measured, and a and b are the parameters to be estimated;

[0045] In the first training feature set X r In the equation, the X-ray fluorescence intensity value x absorbed by the element to be measured is * As the input of the preliminary grade value estimation model, the actual grade value y of the element to be measured is taken as the expected output, and the model parameters a and b are obtained by least squares identification, thereby obtaining the preliminary grade value prediction model;

[0046] The first test feature set X tThe data of nk samples in the initial grade value prediction model are input to the first test feature set X t Make predictions on the test set samples.

[0047] According to one embodiment of the present invention, before training the correction and compensation model, constructing the required data set includes:

[0048] The first training feature set X r and the first test feature set X t The data is input into the preliminary grade value prediction model to train the first feature set X r and the first test feature set X t The test set samples are predicted to obtain the first training feature set X r and the first test feature set X t The preliminary grade value corresponding to each sample in;

[0049] Get the first training feature set X r and the first test feature set X t The difference between the preliminary grade value and the actual grade value corresponding to each sample in is constructed as a difference label Δy;

[0050] The difference label Δy replaces the actual grade value y in the first feature set, and constructs the second feature set G = {x * ,Δy}, and the second feature set G={x * ,Δy} is divided into the second training feature set and the second test feature set

[0051] According to one embodiment of the present invention, training the correction compensation model includes:

[0052] S21, construct a dynamic regularized random weight neural network, which consists of an input layer, a hidden layer, and an output layer. The hidden layer gradually adds nodes to continuously expand the network capacity, that is, each training process only adds one node to the hidden layer;

[0053] S22, perform data standardization preprocessing, and use the maximum-minimum normalization method to normalize the second training feature set. and the second test feature set Conduct standardization processing;

[0054] S23, parameter randomization initialization strategy, for the network parameters of the hidden layer candidate nodes, construct the random parameter space as follows: W ~ μ (-a, a) A*h ,B~μ(-b,b) h, where W is the input weight matrix, which obeys the uniform distribution in the interval [-a, a], A=m+1 is the input dimension, B is the bias vector, which obeys the uniform distribution in the interval [-b, b], and h represents the number of candidate nodes generated in the current iteration cycle;

[0055]

[0056] B=[b1 b2 b3 ... b k-1 b k ]

[0057] Where k represents the total number of random configurations of hidden layer parameters, and d represents the dimension of the input variable;

[0058] S24, assuming that the number of hidden layer nodes in the hidden layer of the current dynamic regularized random weight neural network is L-1, calculate the residual of the current network. The residual expression is:

[0059]

[0060] in, Represents the samples of the second training feature set, with a total of N groups of data, The value of is the intensity value of X-ray fluorescence absorbed by characteristic elements; e L-1 It represents the difference between the grade estimation value and the actual grade value of N groups of elements to be tested in the dynamic regularized random weight neural network algorithm model when the number of hidden layer nodes is L-1;

[0061] S25, calculate the output of the node when the Lth node is added according to the following formula:

[0062]

[0063] in, Indicates that when the input variable is When , the hidden layer output value of the newly added node of the dynamic regularized random weight neural network; w L Indicates the input value selected for the newly added node. The weight is selected from the W matrix in S23; represents the Nth group of input variables, and b L Indicates the bias of the newly added node, which is selected from the bias set B in S23; g L Represents the activation function, which satisfies the universal approximation ability theorem. Its output range is 0-1 and its shape is an S-shaped curve. The expression of the activation function is:

[0064]

[0065] Where, e represents a natural constant and υ is an input value;

[0066] S26, based on the dynamic evaluation framework based on the supervision mechanism of the following formula, the hidden layer parameters are screened layer by layer and the optimal neuron nodes are selected and retained to achieve adaptive optimization of the model structure:

[0067]

[0068] Where q = 1, 2, 3, ..., m-1, m, ξ L,q Dynamic discrimination parameters for newly added nodes in the hidden layer, used to evaluate the effectiveness of adding or removing neurons; and Represent the hidden layer residual function and candidate basis function respectively;

[0069] S27, for a given neural network training sample The objective cost function of the dynamic regularized random weight neural network model is:

[0070]

[0071] Where H is the hidden layer output matrix, ΔY is the expected output of the dynamic regularized random weight neural network, β is the output weight vector, and λ is l is the dynamic regularization coefficient, L is the number of nodes;

[0072] Dynamic regularization adaptively adjusts the regularization strength of each node, uses the bi-norm of the node weight to adjust the regularization coefficient, and combines ridge regression to suppress overfitting while retaining the key information of the model:

[0073]

[0074] Where c is the global regularization parameter; is a very small number to prevent the denominator from being divided by zero; ||β l ||2 refers to the bi-norm of the weight of the lth node;

[0075] Combined with the regularized least squares method, the output weight is directly calculated:

[0076]

[0077] H T Hβ+Λβ=H T ΔY

[0078] β=(H T H+Λ) -1 H T ΔY

[0079] Where Λ is a diagonal matrix, and its specific form is as follows:

[0080]

[0081] S28, when the incremental construction process of the dynamic regularized random weight configuration network reaches the preset stop condition, the system will terminate the parameter optimization and node addition and deletion operations.

[0082] The method for detecting the element grade of ore according to the embodiment of the present invention has the following technical advantages:

[0083] Through the collaborative modeling of the preliminary grade value estimation model and the correction compensation model, the quantitative analysis of the mutual interference errors between matrix elements is achieved. Compared with the traditional analysis method relying on empirical formulas, the fitting accuracy is improved; at the same time, compared with the instruments of the basic parameter method, its applicability in the complex environment of the mineral processing site is significantly improved.

[0084] The dynamic regularized random weight neural network proposed in the present invention is based on the analysis of the physical process mechanism of the X-ray fluorescence intensity value generated by the elements in the magnetite sample under excitation. The linear resolvable part is constructed as a linear regression model, and the nonlinear part is dynamically compensated by a data-driven dynamic regularized random weight neural network algorithm. The dynamic regularization proposed in the present invention achieves progressive sparsification by adaptively adjusting the regularization strength and suppressing redundant nodes according to the importance of node weights, effectively balancing the model complexity and generalization ability. Through the collaborative modeling of univariate linear regression and dynamic regularized random weight neural network algorithm, a significant improvement in the accuracy of iron grade detection in magnetite is finally achieved.

[0085] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 Flowchart of a method for detecting ore element grade according to an embodiment of the present invention;

[0087] Figure 2 Schematic diagram of a flow chart of a method for detecting ore element grade according to one embodiment of the present invention;

[0088] Figure 3 This is a graph showing the results of main feature selection using the ADMM algorithm combined with Lasso regression according to one embodiment of the present invention;

[0089] Figure 4 This is a diagram showing the prediction results of the iron grade training set in actual industrial data according to an embodiment of the present invention;

[0090] Figure 5 This is a graph showing the prediction results of an iron grade test set in actual industrial data according to an embodiment of the present invention. DETAILED DESCRIPTION

[0091] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0092] The following describes a method for detecting element grade of ore according to an embodiment of the present invention with reference to the accompanying drawings.

[0093] like Figure 1 As shown, the method for detecting the element grade of ore according to the embodiment of the present invention may include the following steps:

[0094] S1, use the XRF detection system to scan the ore sample to be tested to obtain the X-ray fluorescence intensity values ​​absorbed by the interfering elements and the element to be tested in the ore sample to be tested, wherein the ore to be tested is magnetite, the element to be tested is iron, and the interfering elements are silicon, calcium, silicon dioxide and air absorption.

[0095] Specifically, XRF technology, based on the characteristic fluorescent X-rays produced when matter interacts with X-rays, enables the simultaneous detection of multiple elements. When X-rays are irradiated on a mineral sample, the elements within the sample absorb X-rays of a specific energy and emit fluorescent X-rays with their own characteristic energy. By detecting the intensity of these fluorescent X-rays, the intensity of the X-rays absorbed by the target element and interfering elements can be determined. The presence of interfering elements can affect the accurate detection of the target element, so this step accurately distinguishes the two, laying the foundation for subsequent processing.

[0096] S2, inputting the intensity value of the X-ray fluorescence absorbed by the element to be measured into a pre-trained preliminary grade value prediction model to obtain the preliminary grade value of the element to be measured; the preliminary grade value prediction model is constructed using a univariate linear regression method.

[0097] Specifically, if Figure 2 As shown, the X-ray fluorescence intensity values ​​absorbed by the interfering elements and the elements to be measured obtained in step S1 are input into a pre-trained preliminary grade value prediction model. The model is constructed using a univariate linear regression method. Univariate linear regression is a simple and effective statistical analysis method. It establishes a linear relationship between the independent variable (here, the X-ray fluorescence intensity value) and the dependent variable (the grade of the element to be measured), and uses a known training data set to fit the best regression line, so that the preliminary grade value of the element to be measured can be quickly predicted based on the input intensity value. The advantage of this method is that the calculation process is relatively simple, easy to understand and implement, and it can quickly give an estimate that is closer to the true value in the preliminary stage, providing a reference basis for further precise correction.

[0098] S3, inputting the X-ray fluorescence intensity values ​​absorbed by the interfering element and the element to be measured into a pre-trained correction compensation model to obtain the correction compensation value of the element to be measured; the correction compensation model is constructed using a dynamic regularized random weight neural network.

[0099] Specifically, if Figure 2 As shown, the X-ray fluorescence intensity values ​​absorbed by the interfering elements and the element to be measured are input into a pre-trained correction and compensation model. This model is constructed using a dynamic regularized random weight neural network. The dynamic regularized random weight neural network is an advanced neural network architecture that optimizes the network training process by introducing a dynamic regularization mechanism. It can automatically adjust the network weights and structure to better adapt to complex data relationships. Compared with traditional regression models, this neural network model has stronger nonlinear fitting capabilities and adaptability, and can fully consider the impact of various factors on grade prediction, including the complex interactions of interfering elements. It can further analyze and calculate a correction and compensation value based on the X-ray fluorescence intensity values ​​absorbed by the interfering elements and the element to be measured. This is used to correct for deviations in the initial prediction and make the final prediction more accurate.

[0100] S4, taking the sum of the preliminary grade value and the correction compensation value as the final grade estimation value of the element to be measured, so as to obtain the test result of the element grade of the ore to be measured.

[0101] Specifically, the preliminary grade value is added to the correction compensation value to obtain the final grade estimate of the element to be measured, thereby obtaining the test result of the ore element grade. This entire process combines the high efficiency of XRF detection technology, the preliminary prediction ability of the univariate linear regression model, and the precise correction capability of the dynamic regularized random weight neural network, achieving rapid and accurate detection of ore element grade.

[0102] Therefore, the ore element grade detection method of the embodiment of the present invention, from the perspective of detection efficiency, the XRF detection system can complete the scanning of ore samples and obtain relevant data in a relatively short time, greatly shortening the detection cycle compared to traditional chemical analysis methods. This is of great significance for real-time monitoring and rapid decision-making in ore mining, mineral processing, and other links. It can provide timely feedback on ore grade information to adjust mining strategies or mineral processing process parameters, thereby improving production efficiency and resource utilization. In terms of accuracy, the synergistic effect of the univariate linear regression model and the dynamic regularized random weight neural network effectively overcomes the influence of interfering elements and other complex factors, making the final grade estimate closer to the true value. This helps to more accurately assess the value of ore resources, provide a reliable basis for the rational development and utilization of ore, and avoid resource waste or economic losses caused by inaccurate grade estimation. In addition, this detection method has a high degree of automation, reduces the interference of human factors, improves the stability and repeatability of the detection results, and is conducive to the standardization and normalization of ore grade detection operations, providing strong technical support for the digital transformation and intelligent development of the mining industry.

[0103] According to one embodiment of the present invention, before training the preliminary grade value prediction model, it is necessary to obtain the data set required for training, including:

[0104] Scan multiple ore samples using an XRF detection system to obtain the X-ray intensity value x absorbed by the interfering elements and the element to be measured in each ore sample, and perform manual analysis on each ore sample to obtain the actual grade value y of each ore sample;

[0105] Construct a first data set D = {x, y} using the X-ray fluorescence intensity value x absorbed by the interfering element and the element to be measured in each ore sample and the actual grade value y of each ore sample;

[0106] Normalize the first data set D = {x, y} to obtain the input feature matrix A;

[0107] The ADMM algorithm is fused with the Lasso regression feature selection algorithm to select the main features of the feature matrix A, and the first feature set S = {x * ,y}, where x * As the main feature.

[0108] Specifically, for each ore sample obtained from the industrial site, an X-ray fluorescence spectrometer is used to measure the characteristic X-ray intensity values ​​of four interfering elements (silicon, calcium, silica, and air absorption) and the X-ray fluorescence intensity value absorbed by the element to be measured, and the actual grade value of the element to be measured is measured to construct the first data set D = {x, y}.

[0109] Furthermore, the ADMM algorithm is used to iteratively update the variables, combined with L1 regularization to achieve feature selection, effectively screening out the feature subset that is significantly correlated with the target variable. In the experiment, the input features are silicon, iron, calcium, silicon dioxide and air absorption. The main features are silicon, iron and calcium selected by the feature selection algorithm, and the first feature set S = {x * ,y}.

[0110] For example, Figure 3 As shown in the figure, the X-ray absorption value of silicon is 13.92, the X-ray absorption value of SiO2 is 0, the X-ray absorption value of Fe is 2.65, the X-ray absorption value of air is 0, and the X-ray absorption value of Ca is 3.19. Therefore, the main features selected by the feature selection algorithm are the X-ray absorption values ​​corresponding to silicon, iron, and calcium.

[0111] According to one embodiment of the present invention, the ADMM algorithm is used to fuse the feature selection algorithm of Lasso regression to perform main feature selection on the feature matrix A, including:

[0112] The goal of Lasso regression is to minimize the linear regression loss with L1 regularization:

[0113]

[0114] Where A represents the input feature matrix, i.e., the first standardized dataset D = {x, y}, with a dimension of m × n, where m is the number of samples and n is the number of features; θ is the weight coefficient vector to be solved, with a dimension of n × 1; z represents the auxiliary variable; b represents the output variable vector, with a dimension of m × 1; τ represents the L1 regularization parameter, which is used to control the sparsification strength;

[0115] In order to solve the non-smoothness caused by L1 regularization, ADMM introduces an auxiliary variable z and transforms the original problem into a constrained optimization problem:

[0116]

[0117] st:θ-z=0

[0118] Decompose the problem into three solvable subproblems and construct the augmented Lagrangian function:

[0119]

[0120] Step 1: Closed-form solution of the θ subproblem

[0121]

[0122] θ (k+1) =(A T A+ρI) -1 (AT b+ρ(z (k) -u (k) ))

[0123] Where ρ represents the penalty parameter of the ADMM algorithm; I represents the unit matrix with dimension n×n; z (k) represents the auxiliary variable of the kth iteration; u (k) represents the Lagrange multiplier of the kth iteration;

[0124] Step 2: Soft threshold function for the z subproblem

[0125]

[0126] z (k+1) =S λ / ρ (θ (k+1) +u (k) )

[0127]

[0128] Where v represents the input vector θ (k+1) +u (k) .

[0129] Step 3: Multiplier update formula

[0130] u (k+1) =u (k) +θ (k+1) -z (k+1)

[0131] Step 4: Feature selection results

[0132] θ selected =[13.92,2.65,3.19,0,0];

[0133] Thus, the first feature set S={x * ,y}.

[0134] According to one embodiment of the present invention, training a preliminary grade value prediction model includes:

[0135] The first feature set S={x * ,y} is divided into the first training feature set X r and the first test feature set X t , where the first training feature set contains k ore samples and the first test feature set contains nk ore samples;

[0136] Construct a univariate linear regression method, the expression is:

[0137] Y i =aX i+b,

[0138] Among them, Y i Represents the grade value of the element to be tested, X i represents the intensity of X-ray fluorescence absorbed by the element to be measured, and a and b are the parameters to be estimated;

[0139] In the first training feature set X r In the equation, the X-ray fluorescence intensity value x absorbed by the element to be measured is * As the input of the preliminary grade value estimation model, the actual grade value y of the element to be measured is taken as the expected output, and the model parameters a and b are obtained by least squares identification, thereby obtaining the preliminary grade value prediction model;

[0140] The first test feature set X t The data of nk samples in the initial grade value prediction model are input to the first test feature set X t Make predictions on the test set samples.

[0141] According to one embodiment of the present invention, before training the correction and compensation model, constructing the required data set includes:

[0142] The first training feature set X r and the first test feature set X t The data is input into the preliminary grade value prediction model to train the first feature set X r and the first test feature set X t The test set samples are predicted to obtain the first training feature set X r and the first test feature set X t The preliminary grade value corresponding to each sample in;

[0143] Get the first training feature set X r and the first test feature set X t The difference between the preliminary grade value and the actual grade value corresponding to each sample in is constructed as a difference label Δy;

[0144] The difference label Δy replaces the actual grade value y in the first feature set, and constructs the second feature set G = {x * ,Δy}, and the second feature set G={x * ,Δy} is divided into the second training feature set and the second test feature set

[0145] According to one embodiment of the present invention, training the correction compensation model includes:

[0146] S21, construct a dynamic regularized random weight neural network, which consists of an input layer, a hidden layer and an output layer. The hidden layer gradually increases the nodes to continuously expand the network capacity, that is, each training process only adds one node to the hidden layer; illustratively, the number of nodes in the input layer is 5 (1 element to be tested and 4 interference elements); the hidden layer gradually increases the nodes to continuously expand the network capacity, that is, each training process only adds one node to the hidden layer, and the maximum number of nodes in the hidden layer is set to L during training. max =200, the maximum number of random configurations is T max =100, the expected tolerance error is ε=0.01; the number of nodes in the output layer is 1.

[0147] S22, perform data standardization preprocessing, and use the maximum-minimum normalization method to normalize the second training feature set. and the second test feature set Conduct standardization processing;

[0148] S23, parameter randomization initialization strategy, for the network parameters of the hidden layer candidate nodes, construct the random parameter space as follows: W ~ μ (-a, a) A*h ,B~μ(-b,b) h , where W is the input weight matrix, which obeys the uniform distribution in the interval [-a, a], A=m+1 is the input dimension, B is the bias vector, which obeys the uniform distribution in the interval [-b, b], and h represents the number of candidate nodes generated in the current iteration cycle;

[0149]

[0150] B=[b1 b2 b3 ... b k-1 b k ]

[0151] Where k represents the total number of random configurations of hidden layer parameters, and d represents the dimension of the input variable;

[0152] The input weight matrix W and bias vector B of the hidden layer are randomly assigned multiple times from the interval [-η, η], where η = {1, 3, 5, 10}. There are a total of a preset number of optional values, and the values ​​that meet the conditions are selected from the optional values;

[0153] S24, assuming that the number of hidden layer nodes in the hidden layer of the current dynamic regularized random weight neural network is L-1, calculate the residual of the current network. The residual expression is:

[0154]

[0155] in, Represents the samples of the second training feature set, with a total of N groups of data, The value is the intensity of X-ray fluorescence absorbed by the characteristic element. For example, the input value vector matrix is ​​1*(3+1), where 3 represents the three characteristic elements of silicon, calcium, and iron after being screened by the feature selection algorithm, 1 represents the element to be tested, and the total vector set of N training sets is N*(3+1), where the values ​​are the X-ray fluorescence intensity values ​​absorbed by the elements; e L-1 It represents the difference between the grade estimation value and the actual grade value of N groups of elements to be tested in the dynamic regularized random weight neural network algorithm model when the number of hidden layer nodes is L-1;

[0156] S25, calculate the output of the node when the Lth node (new candidate node) is added according to the following formula:

[0157] in, Indicates that when the input variable is When , the hidden layer output value of the newly added node of the dynamic regularized random weight neural network; w L Indicates the input value selected for the newly added node. The weight is selected from the W matrix in S23; represents the Nth group of input variables, and b L Indicates the bias of the newly added node, which is selected from the bias set B in S23; g L Represents the activation function, which satisfies the universal approximation ability theorem. Its output range is 0-1 and its shape is an S-shaped curve. The expression of the activation function is:

[0158]

[0159] Where, e represents a natural constant and υ is an input value;

[0160] S26, based on the dynamic evaluation framework based on the supervision mechanism of the following formula, the hidden layer parameters are screened layer by layer and the optimal neuron nodes are selected and retained to achieve adaptive optimization of the model structure:

[0161]

[0162] Where q = 1, 2, 3, ..., m-1, m, ξ L,q Dynamic discrimination parameters for newly added nodes in the hidden layer, used to evaluate the effectiveness of adding or removing neurons; and Represent the hidden layer residual function and candidate basis function respectively;

[0163] Specifically, if ξ L ≥0, it will be saved in the set Z, and the weight w of this configurationk and bias b k Saved in Ω, after multiple random configurations, the algorithm selects the optimal node from the candidate hidden layer node pool. The specific steps are as follows: 1. Select ξ in the set Z L 2. According to the parameters corresponding to the maximum value and Add the optimal candidate hidden layer nodes to the dynamic regularized random weight neural network algorithm model.

[0164] S27, for a given neural network training sample The objective cost function of the dynamic regularized random weight neural network model is:

[0165]

[0166] Where H is the hidden layer output matrix, ΔY is the expected output of the dynamic regularized random weight neural network, β is the output weight vector, and λ is l is the dynamic regularization coefficient, L is the number of nodes;

[0167] Dynamic regularization adaptively adjusts the regularization strength of each node, uses the bi-norm of the node weight to adjust the regularization coefficient, and combines ridge regression to suppress overfitting while retaining the key information of the model:

[0168]

[0169] Where c is the global regularization parameter; is a very small number to prevent the denominator from being divided by zero; ||β l ||2 refers to the bi-norm of the weight of the lth node;

[0170] Combined with the regularized least squares method, the output weight is directly calculated:

[0171]

[0172] H T Hβ+Λβ=H T ΔY

[0173] β=(H T H+Λ) -1 H T ΔY

[0174] Where Λ is a diagonal matrix, and its specific form is as follows:

[0175]

[0176] Add dynamic regularization to the objective function, apply weak regularization to nodes with large weight norm (to retain important features), and apply strong regularization to nodes with small weight norm (to suppress noisy features);

[0177] Specifically, in the dynamic regularized random weight neural network, the construction and solution process of the objective function are as follows: 1. Definition of the objective function: Based on the network model output weight H and output weight β, the second training set The objective function is to minimize the mean square error of the label value ΔY, which is expressed as: β = argmin||ΔY-Hβ||+β T Λβ. In the above objective function expression, ξ L is the hidden layer output weight matrix, ΔY is the expected output of the dynamic regularized random weight neural network; 2. Parameter optimization: The weight parameter β is iteratively optimized by the least squares method so that the objective function reaches the minimum value. The corresponding solution at this time is the optimal parameter. The expression of the optimal parameter is shown in the following formula: β=(H T H+Λ) -1 H T ΔY.

[0178] S28, when the incremental construction process of the dynamic regularized random weight configuration network reaches the preset stop condition, the system will terminate the parameter optimization and node addition and deletion operations.

[0179] Specifically, the training termination criteria for the dynamic regularized random weight neural network are defined as follows: training is terminated when the number of hidden layer nodes exceeds a preset maximum; training is also terminated if the current network residual norm is less than a preset tolerance error threshold. When the network meets the training termination criteria, it indicates that the neural network constructed with the current parameters has reached an optimal state. At this point, the dynamic regularized random weight neural network can be directly used to estimate the grade of the element to be measured in the test sample.

[0180] Based on G={x * ,Δy} uses a dynamic regularized random weight neural network to build a correction compensation model, and the estimated result of the model is the correction compensation value Δy of the element to be measured * The sum of the estimated results output by the preliminary grade value estimation model and the correction compensation model is the final grade estimate .

[0181] The coefficient of determination and root mean square error are used as evaluation indicators. The formulas of the two indicators are:

[0182]

[0183] Where N is the number of samples, y i Indicates the grade value of the element to be tested in the sample in the test set, Indicates the predicted grade value of the element to be tested after the network model, y m It represents the average value of the manual test values ​​of the elements to be tested.

[0184] For example, the prediction results of the iron grade training set in actual industrial data are as follows: Figure 4 As shown in the figure, the prediction results of the iron grade test set in the actual industrial data are as follows Figure 5 shown.

[0185] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0186] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0187] In the present invention, unless otherwise specified or limited, the terms "installed," "connected," "connect," "fixed," etc. should be understood in a broad sense. For example, they can refer to fixed connection, detachable connection, or integration; mechanical connection, electrical connection; direct connection, or indirect connection through an intermediate medium; internal communication between two components, or interaction between two components, unless otherwise specified. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0188] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A method for detecting the element grade of an ore, characterized in that: The method comprises: S1, using an XRF detection system to scan an ore sample to be tested to obtain X-ray fluorescence intensity values ​​absorbed by interfering elements and the element to be tested in the ore sample to be tested, respectively, wherein the ore to be tested is magnetite, the element to be tested is iron, and the interfering elements are silicon, calcium, air absorption, and silicon dioxide coexisting with the iron; S2, inputting the intensity value of the X-ray fluorescence absorbed by the element to be measured into a pre-trained preliminary grade value prediction model to obtain a preliminary grade value of the element to be measured; the preliminary grade value prediction model is constructed using a univariate linear regression method; S3, inputting the X-ray fluorescence intensity values ​​absorbed by the interfering element and the element to be measured into a pre-trained correction compensation model to obtain a correction compensation value of the element to be measured; the correction compensation model is constructed using a dynamic regularized random weight neural network; S4, taking the sum of the preliminary grade value and the correction compensation value as the final grade estimation value of the element to be measured, so as to obtain the detection result of the element grade of the ore to be measured.

2. The method for detecting the element grade of ore according to claim 1, characterized in that: Before training the preliminary grade value prediction model, it is necessary to obtain the data set required for training, including: Scan multiple ore samples using an XRF detection system to obtain the X-ray intensity value x absorbed by the interfering elements and the element to be measured in each ore sample, and perform manual analysis on each ore sample to obtain the actual grade value y of each ore sample; Construct a first data set D = {x, y} using the X-ray fluorescence intensity value x absorbed by the interfering element and the element to be measured in each ore sample and the actual grade value y of each ore sample; Normalize the first data set D = {x, y} to obtain an input feature matrix A; The ADMM algorithm is fused with the Lasso regression feature selection algorithm to select the main features of the feature matrix A, and the first feature set S = {x * ,y}, where x * As the main feature.

3. The method for detecting the element grade of ore according to claim 1, characterized in that: The ADMM algorithm is used to integrate the Lasso regression feature selection algorithm to select the main features of the feature matrix A, including: The goal of Lasso regression is to minimize the linear regression loss with L1 regularization: Where A represents the input feature matrix, i.e. the first dataset D = {x, y} after normalization, with a dimension of m × n, where m is the number of samples and n is the number of features; is the weight coefficient vector to be solved, with a dimension of n×1; z represents the auxiliary variable; b represents the output variable vector, with a dimension of m×1; τ represents the L1 regularization parameter, which is used to control the sparsification strength; In order to solve the non-smoothness caused by L1 regularization, ADMM introduces an auxiliary variable z and transforms the original problem into a constrained optimization problem: Decompose the problem into three solvable subproblems and construct the augmented Lagrangian function: Step 1: Closed-form solution to the subproblem Where ρ represents the penalty parameter of the ADMM algorithm; I represents the unit matrix with dimension n×n; z (k) represents the auxiliary variable of the kth iteration; u (k) represents the Lagrange multiplier of the kth iteration; Step 2: Soft threshold function for the z subproblem Where v represents the input vector Step 3: Multiplier update formula Step 4: Feature selection results Thus, the first feature set S={x * ,y}.

4. The method for detecting the element grade of ore according to claim 3, characterized in that: Training the preliminary grade value prediction model includes: The first feature set S = {x * ,y} is divided into the first training feature set X r and the first test feature set X t , wherein the first training feature set contains k ore samples, and the first test feature set contains nk ore samples; Construct a univariate linear regression method, the expression is: Y i =aX i +b, Among them, Y i Represents the grade value of the element to be tested, X i represents the intensity of X-ray fluorescence absorbed by the element to be measured, and a and b are the parameters to be estimated; In the first training feature set X r In the equation, the X-ray fluorescence intensity value x absorbed by the element to be measured is * As the input of the preliminary grade value estimation model, the actual grade value y of the element to be measured is used as the expected output, and the model parameters a and b are obtained by least squares identification, thereby obtaining the preliminary grade value prediction model; The first test feature set X t The data of nk samples in the initial grade value prediction model are input to the first test feature set X t Make predictions on the test set samples.

5. The method for detecting the element grade of ore according to claim 4, characterized in that: Before training the correction and compensation model, the required dataset is constructed, including: The first training feature set X r and the first test feature set X t The data of is input into the preliminary grade value prediction model to train the first feature set X r and the first test feature set X t The test set samples are predicted to obtain the first training feature set X r and the first test feature set X t The preliminary grade value corresponding to each sample in; Get the first training feature set X r and the first test feature set X t The difference between the preliminary grade value and the actual grade value corresponding to each sample in is constructed as a difference label Δy; The difference label Δy replaces the actual grade value y in the first feature set, and constructs the second feature set G = {x * ,Δy}, and the second feature set G={x * ,Δy} is divided into the second training feature set and the second test feature set 6. The method for detecting the element grade of ore according to claim 5, characterized in that: Training the correction compensation model includes: S21, construct a dynamic regularized random weight neural network, which consists of an input layer, a hidden layer, and an output layer. The hidden layer gradually adds nodes to continuously expand the network capacity, that is, each training process only adds one node to the hidden layer; S22, perform data standardization preprocessing, and use the maximum-minimum normalization method to normalize the second training feature set. and the second test feature set Conduct standardization processing; S23, parameter randomization initialization strategy, for the network parameters of the hidden layer candidate nodes, construct the random parameter space as follows: W ~ μ (-a, a) A*h ,B~μ(-b,b) h , where W is the input weight matrix, which obeys the uniform distribution in the interval [-a, a], A=m+1 is the input dimension, B is the bias vector, which obeys the uniform distribution in the interval [-b, b], and h represents the number of candidate nodes generated in the current iteration cycle; B=[b1 b2 b3 ... b k-1 b k ] Where k represents the total number of random configurations of hidden layer parameters, and d represents the dimension of the input variable; S24, assuming that the number of hidden layer nodes in the hidden layer of the current dynamic regularized random weight neural network is L-1, calculate the residual of the current network. The residual expression is: in, Represents the samples of the second training feature set, with a total of N groups of data, The value of is the intensity value of X-ray fluorescence absorbed by characteristic elements; e L-1 It represents the difference between the grade estimation value and the actual grade value of N groups of elements to be tested in the dynamic regularized random weight neural network algorithm model when the number of hidden layer nodes is L-1; S25, calculate the output of the node when the Lth node is added according to the following formula: in, Indicates that when the input variable is When , the hidden layer output value of the newly added node of the dynamic regularized random weight neural network; w L Indicates the input value selected for the newly added node. The weight is selected from the W matrix in S23; represents the Nth group of input variables, and b L Indicates the bias of the newly added node, which is selected from the bias set B in S23; g L Represents the activation function, which satisfies the universal approximation ability theorem. Its output range is 0-1 and its shape is an S-shaped curve. The expression of the activation function is: Where, e represents a natural constant and υ is an input value; S26, based on the dynamic evaluation framework based on the supervision mechanism of the following formula, the hidden layer parameters are screened layer by layer and the optimal neuron nodes are selected and retained to achieve adaptive optimization of the model structure: Where q = 1, 2, 3, ..., m-1, m, ξ L,q Dynamic discrimination parameters for newly added nodes in the hidden layer, used to evaluate the effectiveness of adding or removing neurons; and Represent the hidden layer residual function and candidate basis function respectively; S27, for a given neural network training sample The objective cost function of the dynamic regularized random weight neural network model is: Where H is the hidden layer output matrix, ΔY is the expected output of the dynamic regularized random weight neural network, β is the output weight vector, and λ is l is the dynamic regularization coefficient, L is the number of nodes; Dynamic regularization adaptively adjusts the regularization strength of each node, uses the bi-norm of the node weight to adjust the regularization coefficient, and combines ridge regression to suppress overfitting while retaining the key information of the model: Where c is the global regularization parameter; is a very small number to prevent the denominator from being divided by zero; ||β l ||2 refers to the bi-norm of the weight of the lth node; Combined with the regularized least squares method, the output weight is directly calculated: H T Hβ+Λβ=H T ΔY β=(H T H+L) -1 H T Y Where Λ is a diagonal matrix, the specific form is as follows: S28, when the incremental construction process of the dynamic regularized random weight configuration network reaches the preset stop condition, the system will terminate the parameter optimization and node addition and deletion operations.

Images