Multi-task synergy and covariance modeling for multi-element grade prediction

Abstract

This invention relates to the field of intelligent ore processing technology and discloses a multi-element grade prediction method based on multi-task collaboration and covariance modeling. The method includes: constructing an ore grade prediction model, which comprises a feature encoding layer, a fusion processing layer, and an output layer. The feature encoding layer uses multiple encoders for multimodal ore data; the fusion processing layer uses fusion blocks based on an attention mechanism; and the output layer is constructed based on a multilayer perceptron. Multimodal data of the ore to be detected is acquired and input into the ore grade prediction model. Corresponding single-modal features are obtained through multiple encoders. The fusion processing layer obtains fused features based on the single-modal features. The output layer obtains the mean vector and covariance matrix based on the fused features and performs a predicted distribution. The ore grade data is analyzed based on the predicted distribution results, and the process flow is adjusted accordingly. This method solves the problem that existing methods using machine learning for ore grade prediction cannot effectively and accurately detect ore grade.

Description

Technical Field

[0001] This invention relates to the field of intelligent ore processing technology, and in particular to a multi-element grade prediction method based on multi-task collaboration and covariance modeling. Background Technology

[0002] Rapid and accurate prediction of multi-element grades in ores is a core element for mines to achieve efficient resource utilization and optimized control of the mineral processing process. Traditionally, the determination of element grades has relied heavily on laboratory chemical analysis. While this method is highly accurate, it suffers from drawbacks such as long analysis cycles, high costs, and significant time lags, failing to meet the needs of real-time decision-making at the production site. To overcome this bottleneck, in recent years, machine learning-based spectral analysis techniques (such as X-ray fluorescence spectroscopy, XRF) have been applied to grade prediction, achieving progress from offline analysis to online estimation. However, these existing technical solutions still have significant shortcomings, mainly in the following aspects:

[0003] The existing methods often employ a "single-element independent modeling" strategy, building a separate prediction model for each target element. This approach fails to utilize the inherent symbiotic or associated relationships between elements in the ore, resulting in model redundancy, high computational costs, and prediction results that may deviate from reality due to the neglect of synergistic or antagonistic effects between elements.

[0004] The data utilization is limited and the feature fusion capability is weak. Most technologies only focus on using spectral data (such as XRF or NIR) for prediction, failing to effectively integrate the large amount of time-series process data (such as grinding and flotation parameters) and sample metadata (such as sampling information and environmental conditions) generated during industrial production. This single data perspective limits the model from capturing key factors affecting grade changes from multiple dimensions, thus restricting further improvement in prediction accuracy.

[0005] The predictive output is singular and lacks uncertainty assessment; traditional models typically only output point estimates of grade, failing to provide information on the uncertainty of the predicted values ​​or the correlation between grade predictions of different elements. This reduces the reliability of the prediction results and makes it difficult to provide more in-depth quantitative evidence for subsequent process optimization and risk decision-making.

[0006] Insufficient robustness and difficulty in handling missing labels: In real-world production environments, due to testing costs or process limitations, some samples may have missing test values ​​(labels). Existing methods lack effective mechanisms to handle such incomplete labeled data, often resulting in a reduction in available training samples or the need for complex preprocessing, thus affecting the model's practicality and generalization ability in real-world industrial scenarios.

[0007] The aforementioned shortcomings indicate that existing technologies are insufficient to meet the demands of modern mines for intelligent grade prediction that is highly accurate, reliable, interpretable, and adaptable to complex industrial data. Therefore, there is an urgent need for a novel prediction method that can comprehensively utilize multi-source data, collaboratively model multi-element relationships, output uncertainty information, and exhibit strong robustness. Summary of the Invention

[0008] This invention provides a multi-element grade prediction method based on multi-task collaboration and covariance modeling, in order to solve the problem that existing methods for predicting ore grade using machine learning cannot effectively and accurately detect ore grade.

[0009] To achieve the above objectives, the present invention employs the following technical solution:

[0010] This invention provides a multi-element grade prediction method based on multi-task collaboration and covariance modeling, comprising the following steps:

[0011] Step 1: Construct an ore grade prediction model. The ore grade prediction model includes a feature encoding layer, a fusion processing layer, and an output layer. The feature encoding layer includes multiple encoders set for multimodal ore data. The fusion processing layer includes a fusion block set based on an attention mechanism. The output layer is constructed based on a multilayer perceptron.

[0012] Step 2: Obtain multimodal data of the ore to be tested. Input the multimodal data into the ore grade prediction model. First, obtain the corresponding single-modal features through multiple encoders in the feature encoding layer. The fusion processing layer obtains fusion features based on the single-modal features and the fusion block. The output layer obtains the mean vector and covariance matrix based on the fusion features, and performs prediction distribution based on the mean vector and covariance matrix to obtain the prediction distribution results. Analyze the ore grade data and adjust the process flow according to the prediction distribution results.

[0013] Based on the above design, a covariance modeling mechanism for cases where some elemental analysis results are missing in ore samples was devised. This mechanism dynamically masks the loss components corresponding to missing labels during training while retaining their contribution to the covariance matrix. This approach ensures effective model training on incomplete datasets and improves usability in real-world industrial scenarios.

[0014] Furthermore, the multimodal data includes ore spectral data, process time series data, and environmental metadata; the multimodal data is preprocessed and then input into the ore grade prediction model, the preprocessing including spectral resampling and normalization.

[0015] Furthermore, the ore spectral data includes X-ray fluorescence spectral data and near-infrared spectral data; the plurality of encoders includes a spectral data encoder configured for the ore spectral data, the spectral data encoder being constructed based on a pre-trained deep convolutional neural network.

[0016] Furthermore, the process timing data includes grinding data, flotation data, and classification data. The grinding data includes the current, voltage, power consumption, feed rate, discharge rate, and grinding product particle size of the grinding equipment during operation. The flotation data includes the air volume and stirring speed of the flotation cell of the flotation equipment. The classification data includes the overflow velocity, underflow velocity, and hydrocyclone pressure of the classification equipment.

[0017] The plurality of encoders includes a time-series data encoder configured for grinding data, flotation data, and classification process data, the time-series data encoder being constructed based on a pre-trained time-series model.

[0018] Furthermore, the environmental metadata includes sampling depth, sampling time, environmental humidity, and environmental temperature, and the plurality of encoders includes a metadata encoder set for the environmental metadata;

[0019] The metadata encoder is built based on a pre-trained embedding network.

[0020] Through the above design, by integrating ore spectral data, process time series data and environmental metadata, and utilizing multimodal attention mechanism and covariance modeling, the collaborative and accurate prediction of the grade of multiple elements is achieved. This not only significantly improves the prediction accuracy and efficiency, but also explicitly characterizes the correlation between elements and provides prediction uncertainty. At the same time, it has good adaptability to missing label data, effectively meeting the needs of complex industrial sites for intelligent and highly reliable grade prediction technology.

[0021] Furthermore, the fusion processing layer unifies the feature dimensions of all single-modal features, calculates the attention score of each single-modal feature through the fusion block, and constructs the fused weighted feature based on the attention score.

[0022] Furthermore, the steps of obtaining the mean vector and covariance matrix based on the fusion features in the output layer specifically include: the output layer obtains the mean vector after linear transformation based on the fusion features, then obtains the diagonal and off-diagonal elements based on the fusion features combined with a multilayer perceptron, obtains the lower triangular matrix based on the diagonal and off-diagonal elements, and obtains the covariance matrix based on the lower triangular matrix combined with a transpose operation.

[0023] Furthermore, the step of obtaining diagonal and off-diagonal elements based on fusion features combined with a multilayer perceptron specifically includes: obtaining several elements based on fusion features combined with a multilayer perceptron, selecting a predetermined number of elements from them and then using softplus activation to obtain diagonal elements, and filling the remaining elements into the off-diagonal part of the lower triangle of the matrix to obtain off-diagonal elements.

[0024] Furthermore, the loss function of the ore grade prediction model is constructed based on the multivariate Gaussian negative log-likelihood loss of the predicted distribution and the true label.

[0025] Through the above design, a multivariate Gaussian negative log-likelihood loss is introduced into the loss function, while simultaneously learning the predicted mean and the covariance matrix between elements. By constraining the covariance term, the model can explicitly capture the correlation between different metal element grades, improving prediction stability and adaptability to complex geological conditions, and better characterizing the statistical dependencies between elements.

[0026] Beneficial effects:

[0027] The multi-element grade prediction method with multi-task collaboration and covariance modeling provided by this invention proposes a multi-objective collaborative prediction architecture based on deep learning, which can achieve joint modeling and distributed prediction of multi-element grades for various complex ore systems.

[0028] First, this invention introduces multimodal information such as ore spectral data, process time series data, and environmental metadata into the model. It uses pre-trained spectral encoders, time series encoders, and metadata encoders to extract key features under different modalities. Combined with a soft fusion module with an attention mechanism, it achieves dynamic weighted fusion of multi-source features, thereby fully mining the multi-dimensional information of the ore and improving the completeness and effectiveness of feature representation.

[0029] Secondly, this invention introduces a covariance modeling mechanism in the output layer. By simultaneously learning the predicted mean vector and covariance matrix, it constructs a predicted distribution of multi-element grades, thus extending from traditional "point estimation" to "distribution estimation." This predicted distribution can explicitly reflect the correlation and constraints between different elements and quantify the uncertainty of the prediction results. This allows the model to not only provide the most probable predicted values ​​for multiple elements but also the prediction confidence interval and related structural information.

[0030] Furthermore, by constraining the covariance matrix, the model can capture the symbiotic, associated, or antagonistic features among multiple elements, improving its stability and robustness in scenarios with complex ore compositions, high volatility, and incomplete labels. Particularly in cases where some sample labels are missing, the covariance modeling mechanism designed in this invention can dynamically mask the loss components corresponding to the missing labels while retaining their contribution to the covariance matrix, ensuring effective training of the model on incomplete datasets.

[0031] In summary, the multi-task collaborative and covariance modeling method of the present invention can achieve collaborative high-precision prediction of multi-element grades, which can not only improve the accuracy and interpretability of prediction results, but also provide quantitative support for mineral processing control, ore blending optimization and intelligent decision-making. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the network structure of the ore grade prediction model according to an embodiment of the present invention;

[0033] Figure 2 This is a flowchart illustrating the implementation of the ore grade prediction model in an embodiment of the present invention. Detailed Implementation

[0034] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "connected" or "linked" and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship also changes accordingly.

[0036] Please see Figure 1 This application provides a multi-element grade prediction method based on multi-task collaboration and covariance modeling, comprising the following steps:

[0037] Step 1: Construct an ore grade prediction model. The ore grade prediction model includes a feature encoding layer, a fusion processing layer, and an output layer. The feature encoding layer includes multiple encoders set for multimodal ore data. The fusion processing layer includes a fusion block set based on an attention mechanism. The output layer is constructed based on a multilayer perceptron.

[0038] Among them, the feature encoding layer has multiple encoders, including a spectral data encoder set for ore spectral data, which is built based on a pre-trained deep convolutional neural network; a time-series data encoder set for grinding, flotation and classification processes, which is built based on a pre-trained time-series model; and a metadata encoder set for the mining environment, which is built based on a pre-trained embedding network. The targeted encoder settings effectively realize feature extraction, which can more effectively express the characteristics of the ore in different aspects, making the subsequent ore grade prediction more accurate.

[0039] Multimodal data includes ore spectral data, process time series data, and environmental metadata;

[0040] The ore spectral data includes X-ray fluorescence spectral data and near-infrared spectral data. The X-ray fluorescence spectral data is obtained by detecting the elemental composition and elemental content of the ore using X-ray fluorescence spectroscopy, while the near-infrared spectral data is obtained by detecting the mineral composition, structural water and hydroxyl groups of the ore using near-infrared trench detection.

[0041] The process time series data includes grinding data, flotation data, and classification data. Grinding data includes the current, voltage, power consumption, feed rate, discharge rate, and grinding product particle size of the grinding equipment during operation. Flotation data includes the air volume and stirring speed of the flotation cell of the flotation equipment. Classification data includes the overflow velocity, underflow velocity, and hydrocyclone pressure of the classification equipment. The time series data is acquired by window aggregation when measuring ore spectral data.

[0042] Environmental metadata includes sampling depth, sampling time, ambient humidity, and ambient temperature. This metadata is obtained through manual recording.

[0043] In this embodiment, multimodal data of tungsten ore were collected. 4096 energy points were obtained by X-ray fluorescence spectrometry, 2151 wavelength points were obtained by near-infrared spectrometry, and grinding data, flotation data and classification data were collected by process sensors, with 60 seconds of data collected.

[0044] After acquisition, the multimodal data is preprocessed, which specifically includes spectral resampling of the spectral data and normalization of the multimodal data. Here, a uniform sampling grid of size 1024 is defined by spectral resampling to unify the 4096 energy points detected by the X-ray fluorescence spectrometer and the 2151 wavelength points detected by the near-infrared spectrometer to the same number of data points, thereby achieving data alignment.

[0045] The loss function for constructing the ore grade prediction model based on the multivariate Gaussian negative log-likelihood loss of the predicted distribution and the true label is specifically expressed by the following formula:

[0046] ;

[0047] in, Represents the loss function. Indicates the first The authentic label of the ore sample, and ; This represents the mean in the predicted distribution; Indicates the transpose operation; This represents the covariance matrix in the predicted distribution.

[0048] Step 2: Obtain multimodal data of the ore to be tested. Input the multimodal data into the ore grade prediction model. First, obtain the corresponding single-modal features through multiple encoders in the feature encoding layer. The fusion processing layer obtains the fusion features based on the single-modal features and the fusion block. The output layer obtains the mean vector and covariance matrix based on the fusion features, and performs prediction distribution based on the mean vector and covariance matrix to obtain the prediction distribution results. Analyze the ore grade data and adjust the process flow according to the prediction distribution results.

[0049] Please see Figure 2 After preprocessing, the multimodal data is input into the ore grade prediction model. The preprocessing includes spectral resampling of the spectral data and normalization of the multimodal data. Then, the ore grade prediction model extracts features through multiple encoders to obtain multiple single-modal features. In this embodiment, the fusion processing layer unifies the feature dimensions of the extracted single-modal features to 128. The fusion block uses soft attention fusion to calculate the attention score of each single-modal feature and constructs the fused weighted feature based on the attention score.

[0050] After obtaining the fused features, the output layer obtains the mean vector based on the fused features and a linear transformation. It can be expressed by the following formula:

[0051] ;

[0052] in, Let represent the learnable weight matrix, and ; Indicates the first Fusion characteristics of individual ore samples; This indicates a learnable bias, and ; This indicates the number of elements to be predicted, which is 3 in this case.

[0053] The output layer then obtains several elements based on the configured multilayer perceptron, expressed by the following formula:

[0054] ;

[0055] in, Indicates the obtained first Elemental grade of each ore sample; This represents a multilayer perceptron; here .

[0056] Select the first 3 elements from the obtained elements and combine them with softplus activation to obtain the diagonal elements, as shown by the following formula:

[0057] ;

[0058] in, This represents the j-th component. , This indicates that the diagonal element is retrieved; This represents the numerical stability constant, which is a very small positive number to avoid the diagonal elements being calculated close to 0, leading to a singular covariance matrix.

[0059] The off-diagonal elements are obtained by filling the off-diagonal portion of the lower triangle of the matrix with the remaining selected elements, as expressed by the following formula:

[0060] ;

[0061] in, This indicates the non-diagonal element being retrieved; This represents the remaining elements after selection. ; and These represent the row and column indices of the matrix, respectively; in this embodiment... , ,and .

[0062] Here, to clarify the filling order of the off-diagonal elements in the lower triangle, let the total number of elements to be predicted be . The model outputs from the multilayer perceptron. scalars as vectors Store, before One element is activated by softplus to become a diagonal element, and the remaining elements are... The elements are filled into the non-diagonal positions of the lower triangle in row-major order.

[0063] Based on diagonal elements and off-diagonal elements Obtain the lower triangular matrix Based on the lower triangular matrix The covariance matrix is ​​obtained by combining the transpose operation; here, we first obtain the lower triangular matrix. The reason is that the covariance matrix must satisfy symmetry and positive semidefiniteness. To avoid the problem that directly predicting the covariance matrix is ​​difficult to satisfy symmetry and positive semidefiniteness, Cholesky decomposition is used here, expressed by the following formula:

[0064] ;

[0065] in, This represents the obtained covariance matrix;

[0066] Further from the mean vector With covariance matrix The obtained predicted distribution results are represented as follows:

[0067] ;

[0068] in, This represents the elemental grade predicted by the ore grade prediction model.

[0069] In this embodiment, six elements are acquired from the multilayer perceptron. The first three elements are selected and combined with softplus activation to obtain the diagonal elements. The remaining three elements are used as off-diagonal elements. The lower triangular matrix is ​​then obtained from the diagonal and off-diagonal elements. The covariance matrix is ​​further obtained through transpose operation. From the obtained covariance matrix, the uncertainty and correlation of each element prediction can be obtained. The diagonal elements reflect the single-element prediction variance and can be used to evaluate the prediction confidence interval. The off-diagonal elements reflect the synergistic or restrictive relationship between different elements and can be used to analyze the symbiotic characteristics of multiple elements.

[0070] Specifically, the predicted distribution can provide information not only about the tungsten content in tungsten ore (… ),copper( ) and sulfur ( The most likely grade estimates for the three elements can also reflect the statistical correlation and prediction uncertainty among the predicted values ​​of different elements. This output has a direct guiding role in the actual production and operation of tungsten ore beneficiation plants.

[0071] The mean vector represents the expected value of each element's grade and can be used to monitor the trend of changes in raw ore grade in real time. When predicting... When the mean value decreases, it indicates that the raw ore has become dilute. The concentrator can accordingly increase the grinding fineness, appropriately increase the amount of flotation reagents, or extend the flotation time to maintain concentrate quality and recovery rate. When the predicted mean increases, it indicates that the raw ore is enriched. The amount of reagent added or the flotation intensity can be appropriately reduced to reduce reagent and energy consumption.

[0072] at the same time, and The predicted average can also serve as an important reference for process adjustments: if An increase in the predicted mean indicates enhanced activity of copper minerals, suggesting the need to optimize the collector ratio and adjust the stirring intensity; if An increase in the predicted mean indicates an increase in the proportion of pyrite or sulfide ores. The impact of sulfides on scheelite flotation should be reduced by increasing the pH or adding inhibitors such as lime or sodium sulfide.

[0073] Furthermore, the mean prediction results can also provide a quantitative basis for ore blending in different mining areas, such as when a certain mining area... When the content is high but the sulfur content is also high, it can be blended with low-sulfur ore sources to achieve stabilization of the grade of the ore feed and stable production control.

[0074] On the other hand, the covariance matrix provides information on the correlation and prediction uncertainty among its elements. The diagonal elements reflect the variance of the model's predictions for each element, which can be used to assess the confidence level of the prediction results. When the prediction variance is large, it indicates that the model's prediction reliability for the current operating condition is reduced. The operating system can automatically trigger a "steady-state hold" strategy, temporarily refraining from adjusting process parameters and increasing the frequency of online monitoring. When the variance is small and stable, an adaptive control mode can be enabled to dynamically optimize flotation aeration rate, stirring speed, or reagent dosage. Off-diagonal elements reflect the synergistic or restrictive relationships between different elements. For example, when the prediction result... When the covariance is positive, it indicates a synergistic enrichment relationship between tungsten and copper, making a combined flotation reagent system suitable to improve the overall recovery rate; when... When the covariance is negative, it indicates that sulfides have an inhibitory effect on scheelite flotation, and the pH value should be adjusted appropriately or sulfide ore inhibition measures should be strengthened. A strong positive correlation in covariance indicates strong copper-sulfur symbiosis, which allows for optimization of the type and ratio of collectors to improve sorting selectivity.

[0075] Furthermore, by continuously monitoring changes in the structure of the covariance matrix, process status monitoring and anomaly diagnosis can also be achieved. When the structure of the covariance matrix changes significantly within a short period of time, such as... The correlation suddenly weakened or When the correlation reverses, it may indicate a sudden change in ore properties or abnormal equipment operation. The system can issue an alarm signal based on this, prompting the user to check the classification pressure, flotation air volume, or raw ore particle size distribution.

[0076] In summary, the predicted distribution output in this embodiment provides the following three aspects of decision support for the beneficiation plant's production:

[0077] 1. Real-time grade prediction and process control based on mean vector to achieve adaptive optimization of feed flow and reagent regime;

[0078] 2. Based on uncertainty analysis and element correlation characterization using the covariance matrix, guidance is provided for risk assessment and coupling relationships in the flotation process;

[0079] 3. An anomaly detection mechanism based on predictive distribution dynamic changes provides data support for production stability maintenance and equipment condition diagnosis.

[0080] Therefore, the mean vector and covariance matrix output by the ore grade prediction model can enable the beneficiation plant to transform from "single grade prediction" to "multi-element collaborative modeling and intelligent decision support", which can significantly improve the automation, stability and economic benefits of the tungsten ore beneficiation process.

[0081] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A multi-element grade prediction method based on multi-task collaboration and covariance modeling, characterized in that, Includes the following steps: Step 1: Construct an ore grade prediction model. The ore grade prediction model includes a feature encoding layer, a fusion processing layer, and an output layer. The feature encoding layer includes multiple encoders set for multimodal ore data. The fusion processing layer includes a fusion block set based on an attention mechanism. The output layer is constructed based on a multilayer perceptron. Step 2: Acquire multimodal data of the ore to be tested, input the multimodal data into the ore grade prediction model, first obtain the corresponding single-modal features through multiple encoders of the feature encoding layer, the fusion processing layer obtains fusion features based on the single-modal features and the fusion block, the output layer obtains the mean vector and covariance matrix based on the fusion features, and performs prediction distribution based on the mean vector and covariance matrix to obtain the prediction distribution results, and analyze the ore grade data and adjust the process flow according to the prediction distribution results; The multimodal data includes ore spectral data, process time series data, and environmental metadata; the multimodal data is preprocessed and then input into the ore grade prediction model, the preprocessing including spectral resampling and normalization. The ore spectral data includes X-ray fluorescence spectral data and near-infrared spectral data; the multiple encoders include a spectral data encoder designed for the ore spectral data, which is constructed based on a pre-trained deep convolutional neural network; The process timing data includes grinding data, flotation data, and classification data. The grinding data includes the current, voltage, power consumption, feed rate, discharge rate, and grinding product particle size of the grinding equipment during operation. The flotation data includes the air volume and stirring speed of the flotation cell of the flotation equipment. The classification data includes the overflow velocity, underflow velocity, and hydrocyclone pressure of the classification equipment. The plurality of encoders includes a time-series data encoder designed for grinding data, flotation data, and classification process data, the time-series data encoder being constructed based on a pre-trained time-series model; The environmental metadata includes sampling depth, sampling time, ambient humidity, and ambient temperature, and the multiple encoders include a metadata encoder set for the environmental metadata; The metadata encoder is built based on a pre-trained embedding network.

2. The multi-element grade prediction method based on multi-task collaboration and covariance modeling according to claim 1, characterized in that, The fusion processing layer unifies the feature dimensions of all single-modal features and calculates the attention score of each single-modal feature through the fusion block, and constructs the fused weighted feature based on the attention score.

3. The multi-element grade prediction method based on multi-task collaboration and covariance modeling according to claim 1, characterized in that, The steps of obtaining the mean vector and covariance matrix based on the fusion features in the output layer specifically include: the output layer obtains the mean vector after linear transformation based on the fusion features, then obtains the diagonal and off-diagonal elements based on the fusion features and a multilayer perceptron, obtains the lower triangular matrix based on the diagonal and off-diagonal elements, and obtains the covariance matrix based on the lower triangular matrix and a transpose operation.

4. The multi-element grade prediction method based on multi-task collaboration and covariance modeling according to claim 3, characterized in that, The steps for obtaining diagonal and off-diagonal elements based on fusion features and multilayer perceptron specifically include: obtaining several elements based on fusion features and multilayer perceptron, selecting a predetermined number of elements from them and then using softplus activation to obtain diagonal elements, and filling the remaining elements into the off-diagonal part of the lower triangle of the matrix to obtain off-diagonal elements.

5. The multi-element grade prediction method based on multi-task collaboration and covariance modeling according to claim 4, characterized in that, The loss function of the ore grade prediction model is constructed based on the multivariate Gaussian negative log-likelihood loss of the predicted distribution and the true label.

Images