A green low-cost industrial-grade bauxite ore sorting method and system

Abstract

The present application relates to the technical field of ore dressing, in particular to a green low-cost industrial-grade bauxite ore separation method and system. The method comprises the following steps: based on the aluminum-silicon ratio constraint, the hyperspectral data is subjected to hierarchical coupling feature extraction to obtain an aluminum-silicon sensitive coupling feature set; the coupling feature set is input into a double-branch identification model, and an aluminum-silicon ratio grading self-adaptive weighted Focal-BCE loss function is used to train to obtain a bauxite grade identification model; the hyperspectral data of the bauxite to be measured is identified by using the grade identification model to obtain the corresponding grade classification result; and the operation parameters of the crushing and screening equipment are linked and controlled according to the grade classification result to complete the bauxite separation. The present application realizes green and accurate identification of different aluminum-silicon ratios of bauxite, and solves the problems of high cost, serious pollution and low precision of traditional methods.

Description

Technical Field

[0001] This invention relates to the field of ore beneficiation and metallurgy, specifically to a green and low-cost industrial-grade bauxite ore beneficiation method and system. Background Technology

[0002] Bauxite is a core strategic mineral resource for the production of electrolytic aluminum and high-end aluminum-based new materials. Aluminum, its product, is the second most consumed metal after steel, playing an irreplaceable role in aerospace, construction, chemical, and electronics industries. Globally, bauxite is mainly distributed in Guinea, Australia, Vietnam, and China, primarily in the lateritic and karst types (sedimentary and accumulation types). my country's bauxite is mainly karst-type, with generally poor endowment, limited reserves, and an import dependency rate exceeding 75%. Proven reserves account for only about 3.5% of global reserves, with a reserve-to-production ratio of less than 7 years, resulting in significant resource security pressure. Furthermore, domestic bauxite is mainly monohydrate gibbsite, generally characterized by low alumina content, high silica impurity content, and a low aluminum-silicon ratio (average only 4-7). The ore composition is complex, with significant overlap in grade ranges and mixed critical grade samples, making sorting and smelting significantly more difficult than overseas trihydrate gibbsite bauxite, leading to high smelting energy consumption and costs. Traditional manual and conventional testing methods are insufficient for the rapid and accurate grading of bauxite of different grades, suffering from high subjectivity, high error rates, and low efficiency, failing to meet the industrial demands of intelligent mine sorting, ore cascade utilization, and low-cost smelting. Currently, hyperspectral technology, with its refined spectral characteristics, has achieved significant results in mineral identification, mineral abundance estimation, and elemental content inversion. However, research on how to combine hyperspectral technology with deep learning methods to achieve accurate and stable grading of bauxite across multiple grades is limited, especially research capable of meeting the industrial needs of intelligent mine sorting, ore cascade utilization, and low-cost beneficiation. Therefore, a high-precision, low-cost intelligent grade sorting technology is urgently needed to improve the refining of high-grade bauxite.

[0003] Therefore, based on the shortcomings of existing research, this invention develops a green and low-cost industrial-grade bauxite ore sorting method and system based on imaging hyperspectral remote sensing, which can quickly, stably and non-destructively sort bauxite of different grades, providing technical reference for mining-related industries. Summary of the Invention

[0004] To address the shortcomings of existing methods and the needs of practical applications, this invention provides a green and low-cost industrial-grade bauxite ore beneficiation method, comprising the following steps: Based on the aluminum-silicon ratio constraint, hierarchical coupling feature extraction is performed on hyperspectral data to obtain an aluminum-silicon sensitive coupling feature set. The coupling feature set is input into a two-branch recognition model, and a bauxite grade recognition model is obtained by training with an aluminum-silicon ratio graded adaptive weighted Focal-BCE loss function. The grade recognition model is used to identify the hyperspectral data of the bauxite to be tested to obtain the corresponding grade classification result. The operating parameters of the crushing and screening equipment are adjusted in conjunction with the grade classification result to complete the bauxite sorting.

[0005] Optionally, the hierarchical coupling feature extraction of hyperspectral data based on the aluminum-silicon ratio constraint to obtain an aluminum-silicon sensitive coupling feature set includes: Three types of features are extracted from the preprocessed hyperspectral data: differential trend features, aluminum-silicon component-specific spectral index features, and fractal dimension features. The three types of features are then screened using the aluminum-silicon correlation coefficient threshold to remove noise bands that are irrelevant to the aluminum-silicon grade. The three types of features are then weighted and fused to obtain the aluminum-silicon sensitive coupling feature set.

[0006] Optionally, the three types of features extracted from the preprocessed hyperspectral data—differential trend features, aluminum-silicon component-specific spectral index features, and fractal dimension features—include: The first and second derivatives of the preprocessed hyperspectral data were calculated, and bands with an absolute correlation coefficient greater than 0.6 with the aluminum-silicon ratio were retained to obtain differential trend characteristics. Bands with an absolute correlation coefficient greater than 0.7 with alumina and silica were extracted respectively, and normalized spectral indices were constructed to obtain aluminum-silicon component-specific spectral index characteristics. The fractal dimension of the preprocessed hyperspectral data was calculated using the variation dimension method to obtain fractal dimension characteristics.

[0007] Optionally, the three types of features after weighted fusion and filtering yield the aluminum-silicon sensitive coupling feature set, including: Based on the correlation coefficient between the aluminum-silicon component-specific spectral index and the aluminum-silicon ratio, a band weight matrix is ​​constructed. The band weight matrix is ​​then used to weight the differential trend feature and the fractal dimension feature band by band. The weighted features are then concatenated with the aluminum-silicon component-specific spectral index feature to obtain the aluminum-silicon sensitive coupling feature set.

[0008] Optionally, the step of inputting the coupled feature set into the dual-branch recognition model and training it using the aluminum-silicon ratio graded adaptive weighted Focal-BCE loss function to obtain the bauxite grade recognition model includes: The coupled feature set is divided into local feature subsequences and global trend subsequences, which are input into the CNN convolutional branch and the Transformer coding branch, respectively. Cross-layer information interaction between the two branches is realized through a bidirectional feedback mechanism. The output features of the two branches are dynamically weighted and fused according to the aluminum-silicon ratio range. Grade classification logits are generated based on the fused features. The loss is calculated using the aluminum-silicon ratio graded adaptive weighted Focal-BCE loss function and the model parameters are updated by backpropagation to obtain the trained grade recognition model.

[0009] Optionally, dividing the coupled feature set into local feature subsequences and global trend subsequences includes: Extract the aluminum-silicon spectral index features and wavelet multi-scale features from the coupled feature set, and combine them to form a local feature subsequence, which is used as the input to the CNN convolution branch; extract the first and second order differential trend features from the coupled feature set, and combine them to form a global trend subsequence, which is used as the input to the Transformer coding branch.

[0010] Optionally, the model is trained using an aluminum-silicon ratio-graded adaptive weighted Focal-BCE loss function, including: Based on the aluminum-silicon ratio range corresponding to the sample, calculate the adaptive focusing factor and the class weight with rarity correction; calculate the total loss based on the adaptive focusing factor, class weight, and binary cross-entropy loss; and update all parameters of the bi-branch model through backpropagation using the total loss.

[0011] Optionally, the calculation of the total loss based on the adaptive focusing factor, class weights, and binary cross-entropy loss satisfies:

[0012]

[0013] in, Indicates the total loss. Indicates the number of samples in the batch. Indicates the number of grade categories, Indicates single-label loss. This represents the balance coefficient between sample size and the aluminum-silicon ratio rarefaction. Indicates the focus factor. This represents the binary cross-entropy loss term.

[0014] Optionally, the step of adjusting the operating parameters of the crushing and screening equipment in conjunction with the grade classification results to complete the bauxite separation includes: The aluminum-silicon ratio range corresponding to the grade classification result is analyzed; crusher control instructions, vibrating screen control instructions, and ore diversion instructions are generated according to the aluminum-silicon ratio range; the control instructions are sent to the corresponding equipment to perform the corresponding operations.

[0015] Secondly, to efficiently execute the green and low-cost industrial-grade bauxite ore sorting method provided by this invention, this invention also provides a green and low-cost industrial-grade bauxite ore sorting system, including: an input device, an output device, a processor, and a memory. The input device, output device, processor, and memory are interconnected. The memory stores program instructions used for the green and low-cost industrial-grade bauxite ore sorting method. This green and low-cost industrial-grade bauxite ore sorting system of the present invention has a compact structure and stable performance, and can stably execute the green and low-cost industrial-grade bauxite ore sorting method provided by this invention, further enhancing the overall applicability and practical application capability of this invention.

[0016] This invention constructs a three-layer feature subset using spectral smoothing, first / second derivatives, aluminum-silicon specific spectral indices, and fractal dimension. A layered coupled feature set is obtained through weighted fusion based on the aluminum-silicon ratio correlation coefficient. A CNN+Transformer dual-branch recognition network is built, and cross-layer bidirectional feature feedback units and segmented dynamic fusion weights based on the aluminum-silicon ratio are designed. An adaptive weighted Focal-BCE loss function based on the aluminum-silicon ratio is used to dynamically adjust the focusing factor and rarity correction weights, solving problems such as confusion of critical ores, missed detection of rare ores, and imbalanced sample distribution. After the model identifies the ore grade through online inference, the PLC-linked crusher and vibrating screen adaptively adjust the crushing and screening parameters, and the critical ores are diverted for re-inspection.

[0017] This invention requires no chemical reagents, performs non-destructive online detection, significantly improves the differentiation and classification accuracy of critical grade ores, and significantly reduces crushing energy consumption and red mud solid waste. It achieves green, low-cost, and high-precision industrial-grade intelligent sorting of bauxite, and is suitable for large-scale production of low-alumina-silica karst monohydrate gibbsite in China. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a green and low-cost industrial-grade bauxite ore sorting method provided in this embodiment of the invention; Figure 2 This is a framework diagram of a green and low-cost industrial-grade bauxite ore sorting system provided in an embodiment of the present invention. Detailed Implementation

[0019] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0020] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0021] Please see Figure 1 This invention provides a green and low-cost industrial-grade bauxite ore beneficiation method, comprising the following steps: S1. Based on the aluminum-silicon ratio constraint, hierarchical coupling feature extraction is performed on hyperspectral data to obtain an aluminum-silicon sensitive coupling feature set.

[0022] First, spectral data of karst bauxite ore were obtained and elemental content was determined.

[0023] In one embodiment, bauxite ore is crushed indoors using a Leich BB200 jaw crusher. The crushed ore is then passed through a 2YKN1860 vibrating screen from Liming Heavy Industry Technology Co., Ltd., retaining particles with a size of 10-50mm, setting aside particles smaller than 10mm, and allowing particles larger than 50mm to pass through, until all particles meet the standard. The crushed ore is then evenly and non-overlappingly placed on a black substrate and placed on a self-made imaging hyperspectral detection platform. An imaging spectrometer is used at a distance of 190cm from the sample to acquire image data of the ore in the 960-2500nm range, with a spectral resolution of 5.45nm, resulting in a total of 288 bands of hyperspectral image data. The elemental content of the ore after image data acquisition is then determined to obtain the aluminum-silicon ratio of each ore.

[0024] The calculation of the aluminum-silicon ratio includes the following steps: Each ore was analyzed using X-ray fluorescence spectroscopy (XRF). and The content; Based on the measured elemental content, using The aluminum-silicon ratio index of the ore was calculated.

[0025] Furthermore, the obtained spectral data were smoothed using a Savitzky-Golay filter, with the window size and order set to 11 and 2, respectively, to reduce the impact of noise while maintaining the overall trend of the spectrum. Then, the first / second derivatives of the spectral data are calculated, and a trend spectral dataset is constructed. This includes the following steps: First, calculate the first / second derivative spectrum. The difference calculation formula satisfies:

[0026]

[0027] In the formula, Indicates the wavelength of band i. and respectively wavelength The first and second derivatives, It is the interval between two adjacent wavelengths.

[0028] Then, the correlation coefficient between the first / second derivative spectra and the aluminum-silicon ratio index was calculated. ,reserve The band values ​​were used to remove noise bands unrelated to aluminum and silicon grades, resulting in the dataset. .

[0029] Furthermore, specific spectral index subsets for aluminum and silicon components were constructed. This includes the following steps: First calculate the spectral data and The correlation coefficient between them was calculated, and bands with an absolute correlation coefficient greater than 0.7 were retained; Recalculate spectral data and The correlation coefficient between them was calculated, and bands with an absolute correlation coefficient greater than 0.7 were retained; Next, the normalized spectral index of the spectral values ​​is calculated, and the normalized spectral index is then calculated with... The correlation coefficients between the components were sorted in descending order of their absolute values, and the top 288 values ​​were taken as the aluminum component-specific spectral indices. Data set; and calculation of normalized spectral index and The correlation coefficients between them were sorted in descending order of absolute value, and the top 288 values ​​were taken as the silicon-specific spectral indices. Dataset; Finally, the aluminum component's specific spectral index was determined. Data sets and silicon component-specific spectral indices Data sets were concatenated and combined to obtain specific spectral index subsets for aluminum and silicon components. .

[0030] In the embodiment, the correlation coefficient ,satisfy:

[0031] In the formula, This represents the spectral value of the i-th sample in the j-th band. This represents the element content in the j-th band of the i-th sample.

[0032] The normalized spectral index satisfies:

[0033] In the formula, Represents the normalized spectral index. and Let represent the spectral values ​​of the i-th and j-th bands, respectively. .

[0034] Furthermore, the fractal dimension of the original spectrum and the first / second order spectral data is calculated, and an analytical dimension dataset is constructed. .

[0035] The fractal dimension uses the variation dimension and includes the following steps: First, set the lag distance sequence. ,in n is the number of bands; Then calculate each Variation function And record all point pairs ( ); The variation function satisfies:

[0036]

[0037] In the formula, Let h be the total number of sample point pairs with lag of h. Let be the spectral value of the i-th band, and h be the hysteresis step size.

[0038] Next, a logarithmic transformation is performed to filter the fractal scaling interval, including the following steps: First, set the minimum interval length to 15. , ; Then, the sliding window iterates through all intervals and calculates the corresponding ( ) for each group. ); The curvature of each set is obtained by using least squares linear regression on the selected scaling intervals. and coefficient of determination The dimension of variation is calculated based on the curve corresponding to the maximum value of the coefficient of determination. The variation dimension ,satisfy:

[0039] Based on this, a geochemically constrained hierarchical coupling feature dataset of aluminum-silicon ratio was constructed. This includes the following steps: First extract the dataset The correlation coefficients between the spectral indices of aluminum and silicon and the aluminum-silicon ratio were used to construct a band weight matrix W. Then with the dataset The first / second derivatives are used to locate the characteristic absorption bands of aluminum and silicon, and this is used to preserve the dataset. Wavelet coefficients for the corresponding bands, excluding wavelet components without mineral response; Next, the selected differential features and wavelet features are multiplied by the weight matrix W band by band to achieve adaptive enhancement of the sensitive features of mineral components; Finally, the two weighted features are fused to obtain the aluminum-silicon ratio-constrained aluminum-silicon sensitive coupling feature set. S2. Input the coupled feature set into the dual-branch recognition model, and train the bauxite grade recognition model using the aluminum-silicon ratio graded adaptive weighted Focal-BCE loss function.

[0040] First, the dual-branch input channels are divided, and the aluminum-silicon sensitive coupling feature set obtained after weighted fusion based on aluminum-silicon ratio geochemical constraints is... As a unified data source for the network; extract Feature channels corresponding to aluminum-silicon spectral indices and wavelet multi-scale features are combined to form local feature subsequences, which are then used as inputs to CNN convolution branches; extraction The feature channels corresponding to the first / second-order differential trend features are combined to form a global trend subsequence, which is used as the input of the Transformer encoding branch.

[0041] Furthermore, a CNN feature extraction branch module and a Transformer feature encoding branch module are constructed respectively.

[0042] Specifically, the one-dimensional convolutional local feature extraction branch is constructed as follows: using wavelet + aluminum-silicon index local subsequence as one-dimensional CNN input, a multi-layer one-dimensional convolutional stacked structure is constructed. Each layer is configured with a one-dimensional convolutional layer, a ReLU activation layer, and a one-dimensional max pooling layer. The convolutional kernel size is set to 3, the stride is 1, and the padding is 1. The local absorption valleys and subtle spectral abrupt changes in the aluminum and silicon bands are extracted layer by layer. Each layer outputs a local intermediate feature map at the corresponding scale, and the intermediate output features of each convolutional layer are completely preserved.

[0043] Transformer global feature encoding branch construction: The differential trend global subsequence is used as the input sequence of Transformer; firstly, learnable band position encoding is added to compensate for spectral order information, and multiple TransformerEncoder layers are stacked. The number of multi-head attention heads is set to 8, and the feedforward network dimension is twice the input feature dimension; the long-distance dependence of the entire band is modeled by multi-head self-attention to explore the overall continuous change trend of aluminum-silicon ratio; the intermediate output features of each encoder layer are fully preserved.

[0044] Cross-layer feature dimension alignment processing: Unify the channel dimension and sequence length of the local intermediate features output by each convolutional layer of CNN and the global intermediate features output by each encoder layer of Transformer: Use 1×1 one-dimensional convolution to complete the channel number matching and linear interpolation to achieve sequence length alignment, eliminate the problem of mismatch between the dimensions of the two types of heterogeneous features, and provide a computational basis for subsequent cross-layer bidirectional mutual feedback interaction.

[0045] Multi-layer intermediate feature cache output: caches and outputs multi-layer intermediate features of each convolutional layer of CNN and each encoder layer of Transformer, without directly performing single feature concatenation at the end of the network, providing a multi-layer, multi-scale feature data source for bidirectional feature feedback and hierarchical dynamic weighted fusion.

[0046] Then, an aluminum-silicon ratio adaptive bidirectional feature feedback unit is constructed to realize real-time cross-layer information interaction in two branches, including: Forward information injection: The local absorption details of aluminum and silicon minerals output from each convolutional layer of the CNN are input in parallel to the multi-head self-attention layers of each Transformer layer. The attention weights are corrected based on the band features of mineral components, thereby increasing the attention response intensity of sensitive bands of aluminum oxide and silicon dioxide and weakening irrelevant noise bands. Reverse feature backpropagation: The global aluminum-silicon ratio evolution trend features output by each encoder layer of the Transformer are backpropagated to each convolutional layer of the CNN. The correlation between the convolutional channel and the aluminum-silicon ratio is calculated. Suppression weights are applied to convolutional channels with weak grade discrimination ability and only capture ore scattering noise, thereby quantitatively controlling the layered dynamic fusion coefficient.

[0047] Furthermore, the attention weights modified based on mineral component band features satisfy:

[0048] In the formula, This represents the correlation coefficient between aluminum and silicon.

[0049] Furthermore, the fusion weight coefficient is quantitatively controlled by the hierarchical dynamic fusion coefficient. ,satisfy:

[0050] Among them, each layer integrates comprehensive features , This indicates that the Transformer extracts features. This indicates that CNN extracts features.

[0051] Based on this, a hierarchical adaptive feature fusion and grade classification system is constructed. In this embodiment, instead of simply concatenating the output features of the last layer of the network, the global features output by each encoder layer of the Transformer and the local features output by the corresponding convolutional layer of the CNN are weighted and fused layer by layer, taking into account the aluminum-silicon ratio fusion weight coefficient. After layer-by-layer fusion, all hierarchical fused features are concatenated to obtain a comprehensive feature vector. A two-layer fully connected classification head is built, with Dropout set in the middle to suppress overfitting. Finally, the original logits of multi-grade classification are output without adding Sigmoid activation. The output dimension is determined according to the number of grade categories to achieve simultaneous discrimination of multi-label bauxite grades.

[0052] Furthermore, an adaptive weighted Focal-BCE loss function based on aluminum-silicon ratio is introduced for network gradient backpropagation updates, including: First, calculate the pre-adaptive hyperparameters and the balanced weights of each category to solve the problem of uneven distribution of sample quantity for each grade: calculate the dynamic focusing factor to obtain the true aluminum-silicon ratio of a single sample, and adaptively allocate the focusing coefficient according to the segmentation rule. The adaptive allocation of the focusing coefficient satisfies:

[0053] Then, the class weights with rarity correction are calculated. The basic weights are obtained by counting the number of positive and negative samples for each grade in the training set, and an aluminum-silicon ratio rarity correction coefficient is introduced. The double-constraint weights are obtained, satisfying:

[0054] In the formula, This represents the total number of positive samples in class c. This represents the total number of negative samples in class c. The constant 1e-8 is used to prevent the denominator from being zero. The intervals for low-grade and high-grade aluminum-silicon are defined. Standard intermediate grade range setting .

[0055] Furthermore, the class prediction probability is obtained through the Sigmoid mapping, satisfying:

[0056] In the formula, This represents the probability that the model determines the i-th sample to belong to grade c, and its value ranges from [0,1].

[0057] Then, calculate the discrimination confidence level, satisfying:

[0058] In the formula, Samples closer to 1 are very easy to distinguish, while samples closer to 0.5 are more difficult to distinguish.

[0059] Next, simple samples of modulation factor suppression are constructed based on adaptively assigned focusing coefficients; Furthermore, the dynamic focusing factor satisfies:

[0060] Furthermore, the binary cross-entropy loss term is calculated, satisfying:

[0061] In the formula, This is to prevent the logarithmic input from being 0 and to avoid computational overflow.

[0062] Next, the focus factor and class weights are added together to obtain the single-label loss, which satisfies:

[0063] In the formula, It balances the sample size with the aluminum-silicon ratio of rarity. It is based on adaptive adjustment of the aluminum-silicon ratio of a single sample to enhance the learning of critically overgraded samples.

[0064] Based on this, the global loss mean is calculated. The global loss represents the average of the summations of the losses of all samples and all grade labels within each batch, satisfying the following:

[0065] In the formula, B represents the number of batch samples, and C represents the number of grade categories.

[0066] Finally, the calculated total loss is substituted into the Adam optimizer, and the gradients are backpropagated to synchronously update all network weight parameters of the dual-branch CNN, Transformer encoder, and classification head, including: The data is divided into sample sets, and forward multi-label logits are obtained. The loss value between the prediction and the real multi-label is calculated according to the weighted Focal-BCE loss function. The Adam optimizer is used to backpropagate and update all network parameters of the double-branch convolution, Transformer encoder, and classification head. After each iteration, the classification precision, recall and F1 score are calculated using the validation set. An early stopping strategy is used to prevent the model from overfitting. The optimal recognition model is saved after the index converges.

[0067] It is understood that the adaptive weighted Focal-BCE loss for aluminum-silicon ratio grading in this invention includes two sets of mineral-specific adaptive constraints: First, the focusing factor is dynamically adjusted based on the aluminum-silicon ratio range, significantly amplifying the loss gradient for ores with critical overlap of 4-7, forcing the model to learn the weak spectral differences at grade boundaries, and reducing the misclassification rate of critical samples; Second, an aluminum-silicon rarity correction coefficient is added on the basis of sample balancing weights. This approach doubles the penalty for losses in rare high- and low-grade ores, reducing the under-detection of valuable bauxite. The dynamic focusing mechanism and the rare weight correction mechanism are coupled to simultaneously address three major industry pain points: sample distribution imbalance, critical grade confusion, and under-detection of rare ores. Furthermore, it forms a complete and complementary optimization chain with the aforementioned aluminum-silicon constraint characteristics and bidirectional feedback network. Ordinary fixed hyperparameter Focal-BCE and basic weighted BCE can only achieve single optimization and cannot achieve the technical effect of multiple simultaneous cost improvements.

[0068] S3. Use the grade identification model to identify the hyperspectral data of the bauxite to be tested, and obtain the corresponding grade classification results.

[0069] In this embodiment, a self-made imaging hyperspectral detection platform is used to acquire real-time imaging hyperspectral data of bauxite ore. The collected ore spectral dataset is then analyzed, the trained model is applied to the spectral dataset, and the grade of each ore is obtained. The results are labeled, and the location of each ore is recorded to provide a basis for subsequent automatic classification equipment to distinguish them.

[0070] S4. Adjust the operating parameters of the crushing and screening equipment according to the grade classification results to complete the bauxite separation.

[0071] In this embodiment, the predicted aluminum-silicon ratio range of the ore to be tested is first recorded. If it is identified as a high-alumina-silicon ore, the vibrating screen is controlled to reduce the screening aperture to further refine the particles and improve smelting efficiency. If it is identified as a low-alumina-silicon-high-silicon ore, the jaw crusher is linked to increase the crushing gap to reduce over-crushing and reduce the generation of red mud. Critical grade ore is separately diverted to the secondary hyperspectral re-examination channel.

[0072] It should be noted that the specific implementation methods described above, such as image processing, numerical simulation, and the construction and training of machine learning models, can all be accomplished by the processor by calling the corresponding computer program instructions stored in memory. Those skilled in the art can implement the above functions using algorithms and tools known in the prior art, according to actual needs.

[0073] Please see Figure 2In this embodiment, to efficiently execute the green and low-cost industrial-grade bauxite ore sorting method provided by this invention, the present invention also provides a green and low-cost industrial-grade bauxite ore sorting system, comprising: an input device 1, an output device 2, a processor 3, and a memory 4. The input device 1, output device 2, processor 3, and memory 4 are interconnected. The memory 4 stores program instructions used to execute the steps of the green and low-cost industrial-grade bauxite ore sorting method. The green and low-cost industrial-grade bauxite ore sorting system of this invention has a compact structure and stable performance, and can stably execute the green and low-cost industrial-grade bauxite ore sorting method of this invention, further enhancing the overall applicability and practical application capability of this invention.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the present invention.

Claims

1. A green and low-cost industrial-grade bauxite ore beneficiation method, characterized in that, Includes the following steps: Based on the aluminum-silicon ratio constraint, hierarchical coupling feature extraction is performed on hyperspectral data to obtain an aluminum-silicon sensitive coupling feature set; The coupled feature set is input into the dual-branch recognition model, and the bauxite grade recognition model is trained using the aluminum-silicon ratio graded adaptive weighted Focal-BCE loss function. The grade identification model is used to identify the hyperspectral data of the bauxite under test, and the corresponding grade classification results are obtained. Based on the grade classification results, the operating parameters of the crushing and screening equipment are adjusted in a coordinated manner to complete the bauxite separation.

2. The green and low-cost industrial-grade bauxite ore beneficiation method according to claim 1, characterized in that, The method of extracting hierarchical coupling features from hyperspectral data based on the aluminum-silicon ratio constraint yields an aluminum-silicon sensitive coupling feature set, including: Three types of features were extracted from the preprocessed hyperspectral data: differential trend features, aluminum-silicon component-specific spectral index features, and fractal dimension features. The three types of features were screened using the aluminum-silicon correlation coefficient threshold to eliminate noise bands that are unrelated to aluminum-silicon grade. The three types of features after weighted fusion and screening are used to obtain the aluminum-silicon sensitive coupling feature set.

3. The green and low-cost industrial-grade bauxite ore beneficiation method according to claim 2, characterized in that, The extracted preprocessed hyperspectral data contains three types of features: differential trend characteristics, aluminum-silicon component-specific spectral index characteristics, and fractal dimension characteristics. The first and second derivatives of the preprocessed hyperspectral data were calculated, and bands with an absolute value of correlation coefficient with aluminum-silicon ratio greater than 0.6 were retained to obtain differential trend characteristics. Bands with absolute correlation coefficients greater than 0.7 with alumina and silica were extracted respectively, and normalized spectral indices were constructed to obtain the specific spectral index characteristics of aluminum-silicon components. The fractal dimension of the preprocessed hyperspectral data was calculated using the variation dimension method, and the fractal dimension characteristics were obtained.

4. The green and low-cost industrial-grade bauxite ore beneficiation method according to claim 2, characterized in that, The three types of features after weighted fusion and filtering yield the aluminum-silicon sensitive coupling feature set, including: Based on the correlation coefficient between the aluminum-silicon component-specific spectral index and the aluminum-silicon ratio, a band weight matrix is ​​constructed; The band weighting matrix is ​​used to weight the differential trend characteristics and fractal dimension characteristics band by band. The weighted two types of features are concatenated with the aluminum-silicon component-specific spectral index features to obtain the aluminum-silicon sensitive coupling feature set.

5. The green and low-cost industrial-grade bauxite ore beneficiation method according to claim 1, characterized in that, The process of inputting the coupled feature set into the dual-branch recognition model and training it using the aluminum-silicon ratio graded adaptive weighted Focal-BCE loss function to obtain the bauxite grade recognition model includes: The coupled feature set is divided into local feature subsequences and global trend subsequences, which are then input into the CNN convolutional branch and the Transformer encoding branch, respectively. Cross-layer information interaction between the two branches is achieved through a two-way feedback mechanism, and the output characteristics of the two branches are dynamically weighted and fused according to the aluminum-silicon ratio range. Grade classification logits are generated based on fused features. The loss is calculated using the aluminum-silicon ratio graded adaptive weighted Focal-BCE loss function and the model parameters are updated by backpropagation to obtain the trained grade recognition model.

6. The green and low-cost industrial-grade bauxite ore beneficiation method according to claim 5, characterized in that, The step of dividing the coupled feature set into local feature subsequences and global trend subsequences includes: Extract the aluminum-silicon spectral index features and wavelet multi-scale features from the coupled feature set, combine them to form a local feature subsequence, and use it as the input to the CNN convolution branch; The first and second-order differential trend features in the coupled feature set are extracted and combined to form a global trend subsequence, which is used as the input of the Transformer encoding branch.

7. The green and low-cost industrial-grade bauxite ore beneficiation method according to claim 5, characterized in that, The model is trained using an adaptive weighted Focal-BCE loss function based on aluminum-silicon ratio, including: Based on the aluminum-silicon ratio range corresponding to the sample, calculate the adaptive focusing factor and the category weight with rarity correction; The total loss is calculated based on the adaptive focus factor, class weights, and binary cross-entropy loss. The total loss is then backpropagated to update all parameters of the bi-branch model.

8. The green and low-cost industrial-grade bauxite ore beneficiation method according to claim 7, characterized in that, The calculation of the total loss based on the adaptive focus factor, class weights, and binary cross-entropy loss satisfies: in, Indicates the total loss. Indicates the number of samples in the batch. Indicates the number of grade categories. Indicates single-label loss. This represents the balance coefficient between sample size and the aluminum-silicon ratio rarefaction. Indicates the focus factor. This represents the binary cross-entropy loss term.

9. The green and low-cost industrial-grade bauxite ore beneficiation method according to claim 1, characterized in that, The process of adjusting the operating parameters of the crushing and screening equipment based on the grade classification results to complete bauxite beneficiation includes: Analyze the aluminum-silicon ratio range corresponding to the grade classification results; Based on the aluminum-silicon ratio range, crusher control commands, vibrating screen control commands, and ore diversion commands are generated respectively. The control command is sent to the corresponding device to perform the corresponding operation.

10. A green, low-cost industrial-grade bauxite ore sorting system, characterized in that, The green, low-cost industrial-grade bauxite ore sorting system includes: an input device, an output device, a processor, and a memory. The input device, output device, processor, and memory are interconnected. The memory stores program instructions, which are used to execute the green, low-cost industrial-grade bauxite ore sorting method according to any one of claims 1-9.

Images