Physical information fused data enhancement collaborative increment ore grade online detection method and system

Abstract

The invention provides a data enhancement collaborative increment ore grade online detection method and system fused with physical information, and the method comprises the steps: S1, collecting a fluorescence intensity characteristic matrix of elements in an ore sample by using an X-ray fluorescence analyzer, and carrying out preprocessing; s2, performing physical consistency enhancement on the original small sample data set; s3, dividing a matrix effect generated in the X-ray fluorescence analysis XRF using process into a linear main effect and a nonlinear secondary effect, and respectively obtaining a linear feature set and a nonlinear feature set; s4, obtaining a linear-nonlinear collaborative prediction model; and S5, performing online prediction on new ore sample data by using the linear-nonlinear collaborative prediction model, and dynamically updating model parameters according to a prediction result. The method can realize rapid, continuous and lossless measurement in the ore grade detection process, has the characteristics of high stability and outstanding analysis efficiency, and meets the development requirements of green detection and intelligent mines.

Description

Technical Field

[0001] This invention belongs to the field of industrial ore grade prediction technology, specifically relating to a data-enhanced collaborative incremental ore grade online detection method and system that integrates physical information. Background Technology

[0002] As a core raw material for the steel industry, the demand for ore is closely related to the industrialization process. Traditional ore composition analysis mainly relies on chemical methods (such as ICPAES, EDTA titration, and iodine-fluorine method). Although these methods are highly accurate, they have long testing cycles and complex pretreatment processes, making it difficult to meet the needs of rapid on-site testing.

[0003] In recent years, X-ray fluorescence spectroscopy (XRF) has been widely used in industrial testing due to its rapid and non-destructive characteristics. However, this technology faces two major challenges in practical applications: First, complex matrix effects (such as the absorption and enhancement effects of high atomic number elements on low atomic number elements) lead to a significant nonlinear relationship between the intensity of elemental characteristic peaks and grade, severely limiting the prediction accuracy of traditional linear models. Second, labeled data suitable for model calibration is extremely scarce. In industrial settings, preparing standard samples with homogeneous composition and known chemical values ​​is costly and time-consuming, while chemical analysis itself is also time-consuming and labor-intensive. This collectively makes it extremely difficult to obtain high-quality labeled data, making many advanced calibration methods that rely on large amounts of data difficult to apply in practice. Third, in industrial settings, the mapping relationship between spectra and ore grade is time-varying due to changes in ore source and equipment status drift. Static models built based on limited initial samples will continuously decrease in prediction accuracy under this dynamic environment, making it difficult to meet long-term online requirements. Therefore, it is necessary to introduce an online learning mechanism for the model.

[0004] Against this backdrop, XRF matrix effect correction methods based on neural networks have attracted widespread attention due to their powerful nonlinear mapping capabilities, aiming to overcome the bottlenecks of traditional methods. However, existing shallow neural networks suffer from insufficient expressive power, while deep neural networks, although offering high prediction accuracy, are computationally complex and difficult to tune. Furthermore, existing methods are lacking in feature selection and model simplification, failing to fully characterize the complex interactions of multiple elements in matrix effects, resulting in high model complexity and difficulty in meeting the dual requirements of real-time performance and lightweight design in industrial online detection. Therefore, the core of this research lies in developing suitable online learning algorithms that can balance prediction accuracy and lightweight architecture during model updates, thereby simultaneously meeting the requirements of long-term accuracy and real-time computation in industrial settings. Summary of the Invention

[0005] Objective: This invention aims to overcome the shortcomings of existing XRF techniques in predicting ore grade, which are susceptible to matrix effects and limited by small sample sizes. It proposes a data-enhanced, collaborative, incremental online ore grade detection method and system that integrates physical information. This system effectively expands the training dataset and rapidly extracts robust features from equipment data. Furthermore, it constructs a linear-nonlinear collaborative online prediction model using recursive least squares and a stacked interpretable construction algorithm. This invention provides reliable technical support for achieving high-precision, adaptive prediction of ore grade in complex industrial environments.

[0006] This invention first discloses a data-enhanced collaborative incremental online detection method for ore grade that integrates physical information, comprising the following steps:

[0007] Step S1: Use an X-ray fluorescence analyzer to acquire the fluorescence intensity feature matrix of elements in the ore sample, and preprocess the fluorescence intensity feature matrix to obtain the dataset. The ore sample was ilmenite, the element to be tested was titanium, and the interfering elements were iron, calcium, chromium and manganese. This indicates the ore grade value corresponding to the energy dispersive spectroscopy (EDS) data. This represents the normalized fluorescence intensity characteristic data of the ore;

[0008] Step S2 involves employing a data simulation method based on matrix effect-based constraint reconstruction to simulate the original small sample dataset. Perform physical consistency enhancements and build an enhanced dataset. This method generates synthetic data that conforms to both geostatistical laws and spectral physical constraints by simulating the physical processes of matrix effects in XRF.

[0009] Step S3, based on the stepwise feature selection method that integrates XRF mechanism and statistical information, divides the matrix effects generated during the use of XRF in X-ray fluorescence analysis into linear main effects and nonlinear secondary effects, thereby enhancing the dataset. Hierarchical feature decoupling is performed by sequentially dividing the features into linear feature sets. and nonlinear feature sets The linear feature set reflects the dominant linear response characteristics, such as the fluorescence yield of the target element itself, while the nonlinear feature set reflects the matrix interference effects of other elements on the target element. Fluorescence feature data obtained for conditional mutual information selection It is a linear feature;

[0010] Step S4, Model pre-training stage, based on linear feature set With ore grade label First, a linear baseline model is established using the least squares method, and the residuals between the labels and the predicted values ​​are calculated. And based on nonlinear feature sets and A nonlinear compensation model is established using the Deep Interpretable Construction Algorithm (DICA), thereby obtaining a linear-nonlinear collaborative prediction model. The Deep Interpretable Construction Algorithm (DICA) is a stacked learning algorithm specifically designed for the characteristics of XRF data (significant nonlinear relationships, scarce samples, and difficulty in fitting). It effectively extracts nonlinear features through a multi-layer abstraction mechanism.

[0011] Step S5: When new ore sample data (i.e., the X-ray fluorescence intensity feature matrix acquired in real time and preprocessed) arrives in the industrial data stream, the new ore sample data is predicted online using a linear-nonlinear collaborative prediction model, and the model parameters are dynamically updated based on the prediction results.

[0012] Step S1 includes the following steps:

[0013] Step S11: Using an X-ray fluorescence analyzer, the sample is excited by a high-pressure X-ray tube to generate characteristic X-rays. The sensor detects and converts these X-rays into pulsed height energy spectra, thus obtaining the dataset. ,in This represents the obtained energy spectrum data of the ore;

[0014] Step S12, for the dataset After removing outliers and noise values, normalization is performed to form the dataset. .

[0015] Step S2 includes the following steps:

[0016] Step S21: Calculate the similarity between different ore samples in the original dataset based on the Euclidean distance of the spectral feature vectors; obtain a standardized similarity score, and set a similarity threshold. , The value ranges from 0 to 1. A baseline sample is randomly selected from the dataset, and samples that meet the similarity score greater than a threshold are searched. Similar samples with similar physicochemical backgrounds;

[0017] Step S22: For the selected pair of benchmark samples and similar samples, a synthesis operation based on a fusion physical mechanism is performed: First, linear interpolation is performed to simulate elemental mixing. Then, by introducing a nonlinear physical compensation term characterizing matrix effects, the correction process of the target element fluorescence intensity being affected by the dominant matrix element is reconstructed. Finally, a physically consistent and information-enhanced synthetic sample is generated, constituting an enhanced dataset. .

[0018] Step S21 includes the following steps:

[0019] Step S211, let the feature vector of the benchmark sample and the feature vector of the corresponding similar sample be respectively... and Use formula Calculate the Euclidean distance between two samples Where T represents transpose, This represents the nth feature vector of the benchmark sample. This represents the nth feature vector of a corresponding similar sample;

[0020] Step S212, normalization processing is performed. This converts the distance metric into a standardized similarity score. ;in, This represents the maximum Euclidean distance between all pairs of samples in the dataset.

[0021] Step S22 includes the following steps:

[0022] Step S221, with a random mixing ratio For similar samples selected based on similarity and Perform linear interpolation. Calculate the linear part of the concentration of each element. , and Each of the following is a sample The eigenvalues ​​of the target element and The feature values ​​of the target element;

[0023] Step S222: Identify the element with the highest fluorescence intensity in the dataset (excluding the target element) as the dominant matrix element; identify the dominant matrix element that most significantly interferes with the target element signal; calculate the standardized feature intensity difference between the dominant matrix element and the target element in sample pairs, and construct the product term. , as a nonlinear physical compensation component; where, Indicates the normalized feature strength of the dominant element of the benchmark sample; The standardized feature strength of the dominant element representing similar samples; This represents the normalized feature strength of the target element in the benchmark sample; The normalized characteristic intensity of the target element in the benchmark sample is represented; in the ilmenite sample, the dominant matrix element is Fe, and the target element is Ti.

[0024] Step S223: Combine the linear interpolation result with the nonlinear physical compensation term to generate enhanced target element feature values ​​that better reflect physical reality. The calculation formula is: ;in, It is a scale factor hyperparameter that can be optimized through data-driven methods, used to correct model simplification errors; symbol The dominant type of matrix effect (mass absorption effect or secondary fluorescence enhancement effect) in the specific mineral system is determined. In the ilmenite sample, the absorption of Ti by Fe is dominant, and the sign is negative.

[0025] Step S3 includes the following steps:

[0026] Step S31: Based on X-ray fluorescence (XRF) analysis of fluorescence physical mechanisms and distance correlation, select linear feature sets. ;

[0027] Step S32, next, select the nonlinear feature set based on conditional mutual information (CMI). This allows for the precise selection of elements that exhibit strong nonlinear interactions with the core linear features under given conditions and specific matrix environments.

[0028] Step S31 includes the following steps:

[0029] Step S311: Select the element with the highest fluorescence intensity based on the XRF fluorescence characteristic intensity. ;

[0030] Step S312, combined with formula Filter the element feature with the highest distance relevance value ,in , They are respectively distance variance and distance variance Let the distance covariance be... The calculated distance correlation value, The feature matrix representing all samples, A label matrix representing all samples;

[0031] Step S313, element features and Merging as a linear feature set ;

[0032] Step S32 includes the following steps:

[0033] Step S321, using conditional mutual information Identify features that significantly affect linear features, thereby filtering input features for the nonlinear component. , and It is a random variable. It is a joint probability distribution. , and It is a conditional probability distribution;

[0034] Step S322, for the two linear features Each selects one nonlinear feature. Constructing a nonlinear feature set ;

[0035] Step S4 includes the following steps:

[0036] Step S41, using linear feature sets and ore grade labels A linear baseline model was established using the least squares method to obtain the prediction results. ;

[0037] Step S42, using nonlinear feature sets And the difference between the linear calculation result and the actual ore grade value. Based on the Deep Interpretable Construction Algorithm (DICA), a nonlinear model for predicting ore grade is constructed. The Deep Interpretable Construction Algorithm (DICA) is a stacked learning algorithm specifically designed for the characteristics of XRF data (significant nonlinear relationships, scarce samples, and difficulty in fitting). It effectively extracts nonlinear features through a multi-layer abstraction mechanism.

[0038] Step S41 includes the following steps:

[0039] S411, establish a linear relationship between elemental fluorescence intensity and grade; comprehensively consider the contributions of major elements, linear model. ,in, and Indicates the first Two fluorescence intensity characteristics of a sample , They are respectively Corresponding linear coefficients and Corresponding linear coefficient (% / kcps) For the intercept term (%);

[0040] S412, the model parameters are determined by least squares estimation to minimize the sum of squared predicted residuals for all samples; objective function Where n is the number of samples. For each sample, the true ore grade is given; construct the following design matrix X and parameter vector. :

[0041] , ,

[0042] The optimal estimate of the parameter vector is then determined by... Calculation, where For grade vector, The prediction result is calculated for the second linear feature corresponding to the nth sample. ;

[0043] Step S42 includes the following steps:

[0044] S421, based on nonlinear feature sets And the difference between the linear calculation result and the actual ore grade value. A nonlinear model for ore grade prediction was constructed using Deep Interpretable Construction Network (DICA).

[0045] Initialize the following parameters: Maximum number of hidden layer nodes Number of current hidden layer nodes Maximum number of iterations for random parameters Maximum number of hidden layers The allocation range of hidden layer parameters and expected error ,in, Set minimum values ​​for hidden layer parameters. Set the step size for the parameters. Set the maximum value for the hidden layer parameters;

[0046] S422, nonlinear feature set As input, As the expected output; first construct the first hidden layer, the predicted output of the first hidden layer. Represented as The first hidden layer input is a non-linear feature set. L represents the number of nodes in the first hidden layer. Activation function; hidden layer parameters and Based on dynamic stacking geometry control strategy Randomly generated, ;in threshold , For Singapore The angle between the output vector of each node and the current residual vector Add the residual before the node in the first hidden layer. Add the first hidden layer The residuals corresponding to each node; output weights From the formula Calculated, where Represent the Moore-Penrose (M–P) generalized inverse; then calculate the residual of the first hidden layer. If the residual does not meet the preset tolerance error If the first hidden layer is not found, then the next hidden layer will be constructed; for the first hidden layer... layer, ≥2: Output the previous layer As input of the current layer The expected output is adjusted to the current cumulative residual. The hidden parameters in the prediction output of the i-th layer and The generated data is still generated according to the dynamic stacking geometry control strategy; the accumulated residuals are then updated. Repeat step S422, the interlayer construction process, until the cumulative residuals meet the tolerance error ℓ or the number of hidden layers reaches the preset maximum value. Stop at this time;

[0047] In step S422, a new hidden layer internal node is added using a dynamic stacking geometry control strategy, as follows:

[0048] Automatically generated within a random interval 1 candidate node parameter pool is constructed. Candidate nodes and Angle between Through formula Calculation, where For the generated first One candidate node; The angle between each candidate node and the residual is expressed as follows: , for A set of angles between candidate nodes and residuals For the first The angles between candidate nodes and residuals; then select those that meet the conditions. Add the node to the hidden layer. These are dynamically adjusted parameters, intermediate parameters. Typically set to , For Singapore The angle between the output vector of each node and the current residual vector Add the first layer to the hidden layer The residuals corresponding to each node are then updated, and the hidden layer output matrix is ​​updated. The output weights of the network are calculated using a global optimization method. , For the first The output weights of each node For activation function, and For the first The two hidden layer parameters of each node are used to calculate the network residual simultaneously. , The output vector of the hidden layer is used; if the residual still does not meet the training objective, the hidden layer nodes are generated and new hidden layer internal nodes are added using a dynamic stacking geometry control strategy until the training objective is met or the maximum number of node additions is reached.

[0049] When industrial data stream samples arrive, step S5 uses the model obtained in S4 to perform real-time online prediction for each sample, obtains the prediction results, and dynamically updates the model parameters based on the results, including the following steps:

[0050] Step S51, set a data window The fixed capacity is N, where For the current moment, for Time sample, express The input feature vector of the sample at time step 1. express True value of ore grade chemical analysis for a given sample at a given time; when a new sample... Upon arrival: If the window is not full, proceed directly. Add samples to the window; if the window is full, remove the oldest sample first, then add the new sample. Add window;

[0051] In step S52, when the window is not full, the new sample features use the feature column indices selected during the initial model training phase (for example, if the first and third columns of input features are selected as linear features during the initial training process, then the new sample also uses the first and third columns as linear features). Once the window is full, feature selection is immediately performed again based on all samples within the window to select new linear and non-linear features. Subsequent arriving samples will be filtered for features based on the results of this feature selection until the set update interval is reached, at which point features are reselected. This process is executed cyclically as the window scrolls, achieving dynamic feature selection and thus obtaining linear features. Nonlinear characteristics ,in and Let t be two linear features of the new sample. and These are two nonlinear features obtained by calculating the mutual information of the conditions for using the t-th new sample;

[0052] Step S53: The linear model selects the results based on the linear dynamic characteristics of each sample. and the true value of chemical analysis of ore grade Linear baseline predictions are performed on each sample to obtain the prediction results and... residual Furthermore, a recursive least squares method with a forgetting factor is used to update the model parameters in real time based on the prediction results.

[0053] Step S54: The nonlinear model predicts the residuals online based on the linear model for each sample. Results of nonlinear dynamic feature selection The nonlinear part is predicted, and the parameters are updated online based on the prediction results while the network structure is fixed. Finally, the linear prediction results and the nonlinear prediction results are added together to obtain the final grade prediction result of the sample.

[0054] Step S52 includes the following steps:

[0055] Step S521: Initialize the sliding window with the first batch of samples. Before the sliding window reaches its saturation capacity, the feature extraction of all new samples uses the features corresponding to the feature column indices selected during the initial model training phase.

[0056] Step S522: After the window is filled, perform the first online feature selection, selecting linear and nonlinear dynamic feature sets based on all the data in the current window. and Subsequent samples will be filtered for features based on the feature selection results until the set update interval is reached, at which point features will be reselected. Step S522 periodically updates the dynamic feature set to ensure that the feature selection capability remains synchronized with the current working conditions.

[0057] Step S53 includes the following steps:

[0058] Step S531: Calculate the initial parameter vector based on the model pre-training samples. With covariance matrix ;

[0059] Step S532, for each newly arrived real-time sample The linear part of the input feature vector is Calculate the linear prediction value of the current sample. and the resulting prediction error As input to the Deep Interpretable Construction Algorithm (DICA);

[0060] Step S533, using a forgetting factor Online parameter updates are performed using the recursive least squares method: First, the gain matrix is ​​calculated. Where 0 < λ ≤ 1, the forgetting factor Used to adjust the weights of historical data; then the parameter vector at time t is updated. Simultaneously update the covariance matrix at time t. This mechanism utilizes only the current sample information and the model state at the previous moment, and can achieve real-time adaptive adjustment of parameters without storing historical data.

[0061] Step S54 includes the following steps:

[0062] S541, for each newly arrived sample and its corresponding linear prediction error The dynamically selected nonlinear eigenvector is Calculate the nonlinear prediction value of the current sample. ,in The DICA algorithm, designed for deep interpretability, ultimately determines a nonlinear function based on the fitted data and uses the linear portion to obtain the predicted values. The predicted value obtained from the nonlinear part Add them together to get the final prediction result for the sample;

[0063] Step S542, in incremental learning mode, the Deep Interpretable Construction Algorithm (DICA) network adopts a hierarchical recursive residual propagation mechanism: the target residual of the l-th layer is... ;

[0064] For the layer, ≥2, the target residual is calculated recursively. ,in It is the target residual corresponding to the t-th node in the k-th layer. It is the output weight corresponding to the t-th node in the k-th layer. For the first Hidden layer output, the first The prior output error of the layer is ;

[0065] Step S543, update the output weights of each layer using a hierarchical independent recursive least squares algorithm: for the... Layer, calculate the gain matrix Then update the output weights. And update the covariance matrix simultaneously. .

[0066] This invention also provides a data-enhanced collaborative incremental online ore grade detection system based on the method described above, incorporating physical information, comprising:

[0067] Data acquisition module: X-ray fluorescence analyzer is used to acquire characteristic spectral lines of ore samples to obtain the fluorescence intensity characteristic matrix of elements in the sample;

[0068] Data augmentation module: Uses a random interpolation method based on similarity constraints and physical constraints to obtain an augmented dataset;

[0069] Data processing module: performs stepwise feature selection on the augmented dataset to provide input features for model pre-training;

[0070] Model pre-training module: The least squares method is used to establish a baseline linear model, and the Deep Interpretable Construction Algorithm (DICA) is applied to construct a nonlinear compensation model with dynamic residual propagation architecture to obtain a collaborative initial model;

[0071] Online prediction and update module: When industrial data streams arrive, the collaborative model is used to make online predictions for each data point and the model parameters are dynamically updated based on the prediction results.

[0072] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0073] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0074] This invention offers the following advantages: First, it proposes a data-enhanced, collaborative, incremental online ore grade detection method that integrates physical information, achieving high-precision, real-time online detection of ore grade in industrial environments. This method effectively replaces the traditional offline testing process that relies on manual sampling, significantly improving detection efficiency and response speed. Compared to fitting methods based on empirical formulas, this method provides more accurate predictions; compared to the basic parameter method, it exhibits stronger adaptability to complex mineral processing environments. Furthermore, the model structure possesses good interpretability, clearly demonstrating the construction process of the hidden layers and the mechanism of feature interaction, enhancing the confidence in the prediction results and facilitating its widespread application in industrial practice.

[0075] Secondly, addressing the core challenge of scarce labeled samples in industrial settings, this invention proposes a data augmentation method based on physical mechanisms. This method ensures the geological plausibility of the generated samples by introducing similarity constraints and innovatively constructs a nonlinear compensation model simulating "matrix effect reconstruction," overcoming the physical limitations of traditional linear interpolation. This not only expands the training set quantitatively but also improves the physical realism and distribution consistency of the data, laying a reliable data foundation for building high-precision, highly generalizable prediction models under small sample conditions.

[0076] Third, this method is particularly suitable for energy dispersive X-ray fluorescence (EDXRF) analysis scenarios, enabling rapid, continuous, and non-destructive measurements during ore grade detection. It features high stability and outstanding analytical efficiency, meeting the development needs of green detection and smart mines.

[0077] Fourth, the Deep Interpretable Construction Algorithm (DICA) employed in this invention possesses numerous advantages in model architecture and learning mechanism. This algorithm constructs a multi-layered hidden structure and introduces a dynamic control strategy based on spatial geometric relationships to achieve high-precision modeling of complex nonlinear relationships in XRF spectra through layer-by-layer abstraction. In terms of training mechanism, DICA employs a forward construction approach, requiring only the least squares method to solve for the output weights. Hidden layer parameters are randomly generated and fixed within a preset range, effectively avoiding common problems of traditional gradient descent neural networks, such as sensitivity to parameter initialization, susceptibility to local minima, and strong dependence on learning rate adjustment. Furthermore, compared to deep models that rely on massive amounts of labeled data for stable optimization, this algorithm exhibits excellent generalization ability even in small-sample scenarios, significantly suppressing the risk of overfitting. In addition, the algorithm integrates a node pooling strategy with geometric constraints, enhancing interpretability while ensuring network expressiveness. Based on the physical mechanism analysis of the XRF excitation process in ore, this invention decomposes the system response into two parts: a linear baseline and a nonlinear matrix effect compensation, modeled by a linear regression model and a data-driven DICA model, respectively. This hybrid architecture, guided by both physical mechanisms and data-driven approaches, enables accurate estimation of grade and significantly improves the environmental adaptability of the detection model. Attached Figure Description

[0078] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0079] Figure 1 This is a flowchart of an online method for detecting ore element grade according to an embodiment of the present invention;

[0080] Figure 2 The distance correlation calculation results are used to select the linear characteristics of ore element grade according to an embodiment of the present invention;

[0081] Figure 3 This is a diagram showing the calculation results of conditional mutual information in nonlinear feature selection according to an embodiment of the present invention; Detailed Implementation

[0082] like Figure 1 As shown in the figure, this embodiment of the invention provides a data-enhanced collaborative incremental online detection method for ore grade that integrates physical information, comprising the following steps:

[0083] S1: The fluorescence intensity feature matrix of elements in the ore sample was acquired using an X-ray fluorescence analyzer. After preprocessing, a dataset was obtained. The ore sample was ilmenite, the element to be measured was titanium, and the interfering elements were iron, calcium, chromium, and manganese. The specific steps of S1 are as follows:

[0084] S11 enables the X-ray fluorescence analyzer to excite the sample using a high-pressure X-ray tube to generate characteristic X-rays, which are then detected by a sensor and converted into pulsed high-energy spectra, yielding a dataset. , This represents the obtained energy spectrum data of the ore. This indicates the ore grade value corresponding to the energy dispersive spectroscopy (EDS) data;

[0085] S12, Remove outliers and noise values ​​from the dataset collected in step S11, and then normalize the resulting dataset to form a new dataset. , This represents the normalized fluorescence intensity characteristic data of the ore;

[0086] S2: An augmented dataset is obtained based on a random interpolation method that integrates physical mechanisms and similarity constraints. The specific steps of step S2 are as follows:

[0087] S21, calculate the similarity between different samples in the dataset based on Euclidean distance, obtain a similarity score, and set a similarity threshold. Randomly select a baseline sample and find samples that meet the similarity threshold. Similar samples;

[0088] First, let the vectors of the benchmark sample and the corresponding similar sample be respectively... , Using the formula Calculate the Euclidean distance between samples .

[0089] Secondly, normalization is adopted. This converts the distance metric into a standardized similarity score. In the formula, This represents the maximum Euclidean distance between all pairs of samples in the dataset. A similarity threshold is set during data augmentation. Randomly select a baseline sample and randomly find a sample with a similarity greater than a threshold. Similar samples.

[0090] S22, interpolation is performed on the selected benchmark samples and corresponding similar samples, and the nonlinear reconstruction process of matrix effects is simulated. Nonlinear compensation is applied to the results of random interpolation of Ti elements to obtain the augmented dataset. ;

[0091] First, with random mixing ratio For similar samples selected based on similarity and Perform linear interpolation to calculate the linear component of the target element concentration. , and Samples and The feature values ​​of the target element.

[0092] Secondly, the element with the highest fluorescence intensity in the dataset, excluding the target element, is identified as the dominant matrix element; the dominant matrix element that most significantly interferes with the target element signal is identified; the standardized feature intensity difference between this matrix element and the target element in sample pairs is calculated, and their product term is constructed. This serves as a nonlinear physical compensation component. Among them, and The normalized characteristic intensities represent the dominant matrix element and the target element, respectively. In the ilmenite sample, the dominant matrix element is Fe, and the target element is Ti.

[0093] Finally, the linear and nonlinear parts are combined to generate the enhanced target element eigenvalues. : .in, It is a scaling factor hyperparameter that can be optimized through data-driven methods to correct for model simplification errors. (Symbol) The dominant type of matrix effect (mass absorption effect or secondary fluorescence enhancement effect) in the specific mineral system is determined. In the ilmenite sample, the absorption of Ti by Fe is dominant, and the sign is negative.

[0094] S3: XRF-based physical mechanisms and statistical methods will enhance the dataset. Stepwise feature selection yields a linear feature set. and nonlinear feature sets Step S3 is as follows:

[0095] S31: XRF-based fluorescence physical mechanism analysis and distance correlation selection of linear feature sets ;

[0096] First, the element with the highest fluorescence intensity is selected based on the XRF fluorescence characteristic intensity. ;

[0097] Secondly, consider distance correlation. The element feature with the highest filter value

[0098] Finally, element characteristics and Merging as a linear feature set ;

[0099] S32: Selecting nonlinear feature sets based on conditional mutual information (CMI) CMI measures the mutual information between two variables given one or more variables. This allows for the precise selection of elements that exhibit strong nonlinear interactions with core linear features under given conditions and a specific matrix environment.

[0100] First, use conditional mutual information. Identify features that significantly affect linear features, thereby filtering input features for the nonlinear component. , and It is a random variable. It is a joint probability distribution. , and It is a conditional probability distribution.

[0101] Secondly, for each of the two linear features, one nonlinear feature is selected. Constructing a nonlinear feature set ,in Fluorescence feature data obtained for conditional mutual information selection It is a linear feature.

[0102] S4, Model pre-training stage, based on the linear feature set With ore grade label A linear baseline model is established using the least squares method, and based on a nonlinear feature set... The residual between the predicted results of the linear benchmark model and the actual grade is used to establish a nonlinear compensation model using the Deep Interpretable Construction (DICA) algorithm, and the linear and nonlinear models are superimposed to form an initial co-prediction model; the specific steps of step S4 are as follows:

[0103] S41, using linear feature sets and ore grade labels A linear baseline model is established using the least squares method:

[0104] First, based on the selected features, a linear relationship between elemental fluorescence intensity and ore grade is established. Taking into account the contributions of major elements, the linear model is uniformly expressed as follows: in, The predicted total elemental grade (%). and These represent the two fluorescence intensity features of the i-th sample, , For the corresponding linear coefficients (% / kcps). For the intercept term (%);

[0105] Secondly, the model parameters are determined through least squares estimation to minimize the sum of squared predicted residuals for all samples. The objective function is defined as follows: Where n is the number of samples. and They represent the first Two fluorescence intensity characteristics of a sample Find the true value of its corresponding ore grade. Construct the following design matrix and parameter vector:

[0106] ,

[0107] The optimal estimate of the parameter vector is then determined by... Calculation, where Using the grade vector, the prediction result is calculated. .

[0108] S42, using nonlinear feature sets And the difference between the linear calculation result and the actual ore grade value. A nonlinear model for predicting ore grade is constructed based on Deep Interpretable Construction Network (DICA).

[0109] First, based on nonlinear feature sets And the difference between the linear calculation result and the actual ore grade value. A nonlinear model for predicting ore grade is constructed using Deeply Interpretable Construction Network (DICA). During this process, the following key parameters need to be initialized: the maximum number of hidden layer nodes. Number of current hidden layer nodes Maximum number of iterations for random parameters Maximum number of hidden layers The allocation range of hidden layer parameters and expected error ;

[0110] Secondly, the nonlinear feature set As input, As the expected output, we first construct the first hidden layer, whose predicted output is represented as... .in The input for this layer is L, which represents the number of nodes in this layer, and the hidden layer parameters are... and Based on dynamic stacking geometry control strategy Randomly generated, ,in threshold Output weights From the formula Calculated, where This represents the Moore-Penrose (M–P) generalized inverse. Then, the residual of the first hidden layer is calculated. If the residual does not meet the preset tolerance error If the first hidden layer is not found, then the next hidden layer will be constructed; for the first hidden layer... layer( ): Output the previous layer As input of the current layer The expected output is adjusted to the current cumulative residual. The predicted output of this layer is Implicit parameters and It is still generated according to the dynamic stacking geometry control strategy. The accumulated residuals are then updated. Repeat the above inter-layer construction process until the cumulative residual meets the tolerance error ℓ or the number of hidden layers reaches the preset maximum value. Stop when the time comes.

[0111] Step S5: When the industrial data stream arrives, use a linear-nonlinear collaborative model to make online predictions for each data point, and dynamically update the model parameters based on the prediction results. The specific steps of step S5 are as follows:

[0112] S51, Set a data window The fixed capacity is N, where For the current moment, for Time sample, express The input feature vector of the sample at time step 1. express True value of ore grade chemical analysis for a given sample at a given time; when a new sample... Upon arrival: If the window is not full, proceed directly. Add samples to the window; if the window is full, remove the oldest sample first, then add the new sample. Add window;

[0113] S52, when the window is not full, new sample features use the feature column indices selected during the initial model training phase (e.g., if the first and third columns of input features were selected as linear features during initial training, then new samples also use the first and third columns as linear features). Once the window is full, the system immediately re-selects features based on all samples within the window, choosing new linear and non-linear features. Subsequent arriving samples will be filtered for features based on this feature selection result until a set update interval is reached, at which point features are re-selected. This process is executed cyclically as the window scrolls, achieving dynamic feature selection and thus obtaining linear features. Nonlinear characteristics .

[0114] First, during the system startup phase, the sliding window is initialized and filled using the initial batch of samples. During system startup, before the sliding window reaches its saturation capacity, the feature extraction of all new samples is compared with the feature set ultimately determined by the initial training model. and Maintain consistency, that is, use the features corresponding to the feature column indices obtained from the initial training set.

[0115] Secondly, after the window is filled, the system performs the first online feature selection, selecting linear and nonlinear dynamic feature sets based on all the data in the current window. and Subsequent samples will be filtered for features based on the feature selection results until the set update interval is reached, at which point features will be reselected. This step periodically updates the dynamic feature set to ensure that its feature selection capability remains synchronized with the current operating conditions.

[0116] S53, the linear model selects results based on the linear dynamic characteristics of each sample. and the true value of chemical analysis of ore grade Linear baseline predictions are performed on each sample to obtain the prediction results and... The residuals are calculated, and a recursive least squares method with a forgetting factor is used to update the model parameters in real time based on the prediction results.

[0117] First, the initial parameter vector is calculated based on the model's pre-trained samples. With covariance matrix .in is the linear feature matrix of the initial batch stored, and f is the corresponding ore grade tag vector.

[0118] Secondly, for each newly arrived real-time sample The linear part of the input feature vector is Calculate the linear prediction value of the current sample. and the resulting prediction error As input to the Deep Interpretable Construction Algorithm (DICA);

[0119] Finally, the system employs a forgetting factor. Online parameter updates are performed using the recursive least squares method: First, the gain matrix is ​​calculated. Among them, the forgetting factor Used to adjust the weights of historical data. Then the parameter vector is updated. Simultaneously update the covariance matrix. This mechanism utilizes only the current sample information and the model state at the previous moment, enabling real-time adaptive adjustment of parameters without storing historical data.

[0120] S54, the nonlinear model selects the result based on the residuals and nonlinear dynamic characteristics predicted online by the linear model for each sample. The nonlinear part is predicted, and the parameters are updated online based on the prediction results while keeping the network structure fixed. Finally, the linear prediction results and the nonlinear prediction results are added together to obtain the final grade prediction result of the sample.

[0121] First, for each newly arrived sample and its corresponding linear prediction error... Its dynamically selected nonlinear eigenvector is Based on this, the nonlinear prediction value of the current sample is calculated. And the predicted values ​​obtained from the linear part. The predicted value obtained from the nonlinear part Add them together to get the final prediction result for the sample.

[0122] Secondly, in the incremental learning mode, the DICA network adopts a hierarchical recursive residual propagation mechanism. For the first... layer, The target residual is calculated recursively. ,in It is the target residual corresponding to the t-th node in the k-th layer. It is the output weight corresponding to the t-th node in the k-th layer. For the first Hidden layer output, the first The prior output error of the layer is ;

[0123] Finally, a hierarchical independent recursive least squares algorithm is used to update the output weights of each layer. For the th layer... Layer, calculate the gain matrix Then update the output weights. And update the covariance matrix simultaneously. .

[0124] To illustrate the performance of this invention in ore grade detection, 41 ilmenite XRF data samples were selected for data enhancement and grade detection experiments. The five statistical analyses of the initial 41 data samples are described in Table 1.

[0125] To evaluate the effectiveness of the data augmentation methods, four methods were selected for comparison: random interpolation, random interpolation with physical constraints, similarity-based interpolation, and the original data. Each data augmentation method first divided the 41 data points into a 33:8 ratio, using the same 33 data points for augmentation to generate 50 new samples. These 50 new samples were then combined with the 41 original data points to form an augmented dataset of 91 data samples. All five methods used the 8 un-augmented data points as the test set. Each experiment was run 30 times, and the average value was taken. Table 2 shows the diagnostic root mean square error (RMSE), variance, and coefficient of determination for the proposed method and several comparative algorithms. This table presents the results of 30 independent experiments in mean form. The experimental results show that the proposed method achieves better results in terms of RMSE, mean absolute error (MAE), and coefficient of determination. 2 The method achieves optimal values ​​across all metrics, demonstrating its effectiveness.

[0126] To verify the effectiveness of the feature selection method, Pearson correlation coefficient, maximum mutual information (MIC), and a method without feature selection were compared. This experiment used the XRF ilmenite augmentation dataset, randomly divided into training and test sets in a 7:3 ratio. Figure 2 The distance correlation calculation results are used to select the linear characteristics of ore element grade according to an embodiment of the present invention; Figure 3 This is a graph showing the results of conditional mutual information calculation in nonlinear feature selection according to an embodiment of the present invention. Table 3 shows the diagnostic root mean square error, variance, and coefficient of determination for the method of the present invention and several comparative methods. The table presents the results of 30 independent experiments in the form of mean values. The experimental results show that the method of the present invention achieves better results in terms of root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The method achieves optimal values ​​across all metrics, demonstrating its effectiveness.

[0127] To verify the effectiveness of the ore grade detection method, several widely used algorithms in XRF quantitative analysis were compared, including Partial Least Squares Regression (PLS), Principal Component Regression (PCR), Support Vector Regression (SVM), K-Nearest Neighbors (KNN), Extreme Gradient Boosting (XGBoost), Hierarchical Deep Neural Networks (HDNN), and Random Forest (RF). This experiment used the XRF-enhanced ilmenite dataset, randomly divided into training and test sets in a 7:3 ratio. Table 4 shows the diagnostic root mean square error (RMSE), variance, and coefficient of determination for the proposed method and several comparative algorithms. The table presents the results of 30 independent experiments in mean form. The experimental results show that the proposed method achieves better results in terms of RMSE, mean absolute error (MAE), and coefficient of determination. 2 The method achieves optimal values ​​across all metrics, demonstrating its effectiveness.

[0128] Table 1. Five basic descriptive statistical analyses of the sample.

[0129]

[0130] Table 2 Diagnostic results of different data augmentation methods

[0131]

[0132] Table 3 Diagnostic results of different feature selection methods

[0133]

[0134] Table 4 Diagnostic results of different feature selection methods

[0135]

[0136] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method of online detection of synergistic incremental ore grade with data augmentation incorporating physical information, characterized in that, Includes the following steps: In step S1, an X-ray fluorescence analyzer is used to collect a fluorescence intensity feature matrix of elements in the ore sample, and the fluorescence intensity feature matrix is preprocessed to obtain a data set , wherein the ore sample is ilmenite, the element to be measured is titanium element, and the interference elements are iron, calcium, chromium and manganese; , wherein the ore sample is ilmenite, the element to be measured is titanium element, and the interference elements are iron, calcium, chromium and manganese; , wherein the ore sample is ilmenite, the element to be measured is titanium element, and the interference elements are iron, calcium, chromium and manganese; Step S2 involves employing a data simulation method based on matrix effect-based constraint reconstruction to simulate the original small sample dataset. Perform physical consistency enhancements and build an enhanced dataset. ; Step S3, based on the stepwise feature selection method that integrates XRF mechanism and statistical information, divides the matrix effects generated during the use of XRF in X-ray fluorescence analysis into linear main effects and nonlinear secondary effects, thereby enhancing the dataset. Hierarchical feature decoupling is performed by sequentially dividing the features into linear feature sets. and nonlinear feature sets ; Fluorescence feature data obtained for conditional mutual information selection It is a linear feature; Step S4, Model pre-training stage, based on linear feature set With ore grade label First, a linear baseline model is established using the least squares method, and the residuals between the labels and the predicted values ​​are calculated. And based on nonlinear feature sets and A nonlinear compensation model is established using the Deep Interpretable Construction Algorithm (DICA), thereby obtaining a linear-nonlinear co-prediction model. Step S5: When new ore sample data arrives in the industrial data stream, the new ore sample data is predicted online using a linear-nonlinear collaborative prediction model, and the model parameters are dynamically updated based on the prediction results.

2. The method according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Using an X-ray fluorescence analyzer, the sample is excited by a high-pressure X-ray tube to generate characteristic X-rays. The sensor detects and converts these X-rays into pulsed height energy spectra, thus obtaining the dataset. ,in This represents the obtained energy spectrum data of the ore; Step S12, for the dataset After removing outliers and noise values, normalization is performed to form the dataset. .

3. The method according to claim 2, characterized in that, Step S2 includes the following steps: Step S21: Calculate the similarity between different ore samples in the original dataset based on the Euclidean distance of the spectral feature vectors; obtain a standardized similarity score, and set a similarity threshold. A baseline sample is randomly selected from the dataset, and samples that meet the similarity score greater than a threshold are searched. Similar samples with similar physicochemical backgrounds; Step S22: For the selected pair of benchmark samples and similar samples, a synthesis operation based on a fusion physical mechanism is performed: First, linear interpolation is performed to simulate elemental mixing. Then, by introducing a nonlinear physical compensation term characterizing matrix effects, the correction process of the target element fluorescence intensity being affected by the dominant matrix element is reconstructed. Finally, a physically consistent and information-enhanced synthetic sample is generated, constituting an enhanced dataset. .

4. The method according to claim 3, characterized in that, Step S21 includes the following steps: Step S211, let the feature vector of the benchmark sample and the feature vector of the corresponding similar sample be respectively... and Use formula Calculate the Euclidean distance between two samples Where T represents transpose, This represents the nth feature vector of the benchmark sample. This represents the nth feature vector of a corresponding similar sample; Step S212, normalization processing is performed. This converts the distance metric into a standardized similarity score. ;in, This represents the maximum Euclidean distance between all pairs of samples in the dataset.

5. The method according to claim 4, characterized in that, Step S22 includes the following steps: Step S221, with a random mixing ratio For similar samples selected based on similarity and Perform linear interpolation. Calculate the linear part of the concentration of each element. , and Each of the following is a sample The eigenvalues ​​of the target element and The feature values ​​of the target element; Step S222: Identify the element with the highest fluorescence intensity in the dataset (excluding the target element) as the dominant matrix element; identify the dominant matrix element that most significantly interferes with the target element signal; calculate the standardized feature intensity difference between the dominant matrix element and the target element in sample pairs, and construct the product term. , as a nonlinear physical compensation component; where, Indicates the normalized feature strength of the dominant element of the benchmark sample; The standardized feature strength of the dominant element representing similar samples; This represents the normalized feature strength of the target element in the benchmark sample; The normalized characteristic intensity of the target element in the benchmark sample is represented; in the ilmenite sample, the dominant matrix element is Fe, and the target element is Ti. Step S223: Combine the linear interpolation result with the nonlinear physical compensation term to generate enhanced target element feature values ​​that better reflect physical reality. The calculation formula is: ;in, It is a scale factor hyperparameter that can be optimized through data-driven approaches.

6. The method according to claim 5, characterized in that, Step S3 includes the following steps: Step S31: Based on X-ray fluorescence (XRF) analysis of fluorescence physical mechanisms and distance correlation, select linear feature sets. ; Step S32, next, select the nonlinear feature set based on conditional mutual information. .

7. The method according to claim 6, characterized in that, Step S31 includes the following steps: Step S311: Select the element with the highest fluorescence intensity based on the XRF fluorescence characteristic intensity. ; Step S312, combined with formula Filter the element feature with the highest distance relevance value ,in , They are respectively distance variance and distance variance Let the distance covariance be... The calculated distance correlation value, The feature matrix representing all samples, A label matrix representing all samples; Step S313, element features and Merging as a linear feature set ; Step S32 includes the following steps: Step S321, using conditional mutual information Identify features that significantly affect linear features, thereby filtering input features for the nonlinear component. , and It is a random variable. It is a joint probability distribution. , and It is a conditional probability distribution; Step S322, for the two linear features Each selects one nonlinear feature. Constructing a nonlinear feature set ; Step S4 includes the following steps: Step S41, using linear feature sets and ore grade labels A linear baseline model was established using the least squares method to obtain the prediction results. ; Step S42, using nonlinear feature sets And the difference between the linear calculation result and the actual ore grade value. A nonlinear model for predicting ore grade is constructed based on the Deep Interpretable Construction Algorithm (DICA). Step S41 includes the following steps: S411, establish a linear relationship between elemental fluorescence intensity and grade; comprehensively consider the contributions of major elements, linear model. ,in, and Indicates the first Two fluorescence intensity characteristics of a sample , They are respectively Corresponding linear coefficients and Corresponding linear coefficients, For the intercept term; S412, the model parameters are determined by least squares estimation to minimize the sum of squared predicted residuals for all samples; objective function Where n is the number of samples. For each sample, the true ore grade is given; construct the following design matrix X and parameter vector. : ， ， The optimal estimate of the parameter vector is then determined by... Calculation, where For grade vector, The prediction result is calculated for the second linear feature corresponding to the nth sample. ; Step S42 includes the following steps: S421, based on nonlinear feature sets And the difference between the linear calculation result and the actual ore grade value. We used the Deep Interpretable Construction Network (DICA) to construct a nonlinear model for ore grade prediction. Initialize the following parameters: Maximum number of hidden layer nodes Number of current hidden layer nodes Maximum number of iterations for random parameters Maximum number of hidden layers The allocation range of hidden layer parameters and expected error ,in, Set minimum values ​​for hidden layer parameters. Set the step size for the parameters. Set the maximum value for the hidden layer parameters; S422, nonlinear feature set As input, As the expected output; first construct the first hidden layer, the predicted output of the first hidden layer. Represented as The first hidden layer input is a non-linear feature set. L represents the number of nodes in the first hidden layer. Activation function; hidden layer parameters and Based on dynamic stacking geometry control strategy Randomly generated, ;in threshold , For Singapore The angle between the output vector of each node and the current residual vector Add the residual before the node in the first hidden layer. Add the first hidden layer The residuals corresponding to each node; output weights From the formula Calculated, where Represent the Moore-Penrose generalized inverse; then calculate the residual of the first hidden layer. If the residual does not meet the preset tolerance error If the first hidden layer is not found, then the next hidden layer will be constructed; for the first hidden layer... layer, ≥2: Output the previous layer As input of the current layer The expected output is adjusted to the current cumulative residual. The hidden parameters in the prediction output of the i-th layer and The generated data is still generated according to the dynamic stacking geometry control strategy; the accumulated residuals are then updated. Repeat step S422, the interlayer construction process, until the cumulative residuals meet the tolerance error ℓ or the number of hidden layers reaches the preset maximum value. Stop at this time; In step S422, a new hidden layer internal node is added using a dynamic stacking geometry control strategy, as follows: Automatically generated within a random interval 1 candidate node parameter pool is constructed. Candidate nodes and Angle between Through formula Calculation, where For the generated first One candidate node; The angle between each candidate node and the residual is expressed as follows: , for A set of angles between candidate nodes and residuals For the first The angles between candidate nodes and residuals; then select those that meet the conditions. Add the node to the hidden layer. These are dynamically adjusted parameters, intermediate parameters. Typically set to , For Singapore The angle between the output vector of each node and the current residual vector Add the first layer to the hidden layer The residuals corresponding to each node are then updated, and the hidden layer output matrix is ​​updated. The output weights of the network are calculated using a global optimization method. , For the first The output weights of each node For activation function, and For the first The two hidden layer parameters of each node are used to calculate the network residual simultaneously. , The output vector of the hidden layer is used; if the residual still does not meet the training objective, the hidden layer nodes are generated and new hidden layer internal nodes are added using a dynamic stacking geometry control strategy until the training objective is met or the maximum number of node additions is reached. When industrial data stream samples arrive, step S5 uses the model obtained in S4 to perform real-time online prediction for each sample, obtains the prediction results, and dynamically updates the model parameters based on the results, including the following steps: Step S51, set a data window The fixed capacity is N, where For the current moment, for Time sample, express The input feature vector of the sample at time step 1. express True value of ore grade chemical analysis for a given sample at a given time; when a new sample... Upon arrival: If the window is not full, proceed directly. Add samples to the window; if the window is full, remove the oldest sample first, then add the new sample. Add window; In step S52, when the window is not full, the new sample features use the feature column indices selected during the initial model training phase. Once the window is full, feature selection is immediately performed again based on all samples within the window to select new linear and nonlinear features. Subsequent arriving samples will be filtered for features based on the results of this feature selection until the set update interval is reached, at which point features are reselected. This process is executed cyclically as the window scrolls, achieving dynamic feature selection and thus obtaining linear features. Nonlinear characteristics ,in and Let t be two linear features of the new sample. and These are two nonlinear features obtained by calculating the mutual information of the conditions for using the t-th new sample; Step S53: The linear model selects the results based on the linear dynamic characteristics of each sample. and the true value of chemical analysis of ore grade Linear baseline predictions are performed on each sample to obtain the prediction results and... residual Furthermore, a recursive least squares method with a forgetting factor is used to update the model parameters in real time based on the prediction results. Step S54: The nonlinear model predicts the residuals online based on the linear model for each sample. Results of nonlinear dynamic feature selection The nonlinear part is predicted, and the parameters are updated online based on the prediction results while the network structure is fixed. Finally, the linear prediction results and the nonlinear prediction results are added together to obtain the final grade prediction result of the sample. Step S52 includes the following steps: Step S521: Initialize the sliding window with the first batch of samples. Before the sliding window reaches its saturation capacity, the feature extraction of all new samples uses the features corresponding to the feature column indices selected during the initial model training phase. Step S522: After the window is filled, perform the first online feature selection, selecting linear and nonlinear dynamic feature sets based on all the data in the current window. and Subsequent samples will be filtered for features based on the feature selection results of this time, until the set update interval is reached and features are reselected. Step S522 updates the dynamic feature set periodically. Step S53 includes the following steps: Step S531: Calculate the initial parameter vector based on the model pre-training samples. With covariance matrix ; Step S532, for each newly arrived real-time sample The linear part of the input feature vector is Calculate the linear prediction value of the current sample. and the resulting prediction error As input to the Deep Interpretable Construction Algorithm (DICA); Step S533, using a forgetting factor Online parameter updates are performed using the recursive least squares method: First, the gain matrix is ​​calculated. Where 0 < λ ≤ 1, the forgetting factor Used to adjust the weights of historical data; then the parameter vector at time t is updated. Simultaneously update the covariance matrix at time t. ; Step S54 includes the following steps: S541, for each newly arrived sample and its corresponding linear prediction error The dynamically selected nonlinear eigenvector is Calculate the nonlinear prediction value of the current sample. ,in The DICA algorithm, designed for deep interpretability, ultimately determines a nonlinear function based on the fitted data and uses the linear portion to obtain the predicted values. The predicted value obtained from the nonlinear part Add them together to get the final prediction result for the sample; Step S542, in incremental learning mode, the Deep Interpretable Construction Algorithm (DICA) network adopts a hierarchical recursive residual propagation mechanism: the target residual of the l-th layer is... ; For the layer, The target residual is calculated recursively. ,in It is the target residual corresponding to the t-th node in the k-th layer. It is the output weight corresponding to the t-th node in the k-th layer. For the first Hidden layer output, the first The prior output error of the layer is ; Step S543, update the output weights of each layer using a hierarchical independent recursive least squares algorithm: for the... Layer, calculate the gain matrix Then update the output weights. And update the covariance matrix simultaneously. .

8. A data-enhanced collaborative incremental online ore grade detection system based on the method described in any one of claims 1 to 7, characterized in that, include: Data acquisition module: X-ray fluorescence analyzer is used to acquire characteristic spectral lines of ore samples to obtain the fluorescence intensity characteristic matrix of elements in the sample; Data augmentation module: Uses a random interpolation method based on similarity constraints and physical constraints to obtain an augmented dataset; Data processing module: performs stepwise feature selection on the augmented dataset to provide input features for model pre-training; Model pre-training module: The least squares method is used to establish a baseline linear model, and the deep interpretable construction algorithm DICA is applied to construct a nonlinear compensation model with dynamic residual propagation architecture to obtain a collaborative initial model; Online prediction and update module: When industrial data streams arrive, the collaborative model is used to make online predictions for each data point and the model parameters are dynamically updated based on the prediction results.

9. A storage medium, characterized in that, It stores a computer program or instructions that, when run on a computer, perform the steps of the method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

Images