Deep Learning-Based Ore Sorting System and Method

Abstract

The present invention discloses an ore sorting system and method based on deep learning. The method includes: performing particle size grading and ore washing on the raw ore through several layers of vibrating screens to obtain preprocessed graded ore, and collecting image data and hyperspectral data of the preprocessed graded ore; using a first network and a second network constructed based on deep learning algorithms to extract the surface features of the image data and the spectral features of the hyperspectral data respectively, and fusing the surface features and the spectral features and inputting them into a classification network to obtain the ore type and the main components of the ore; calculating jet parameters based on the conveyor belt speed, the ore quality, and the air density, and determining the sorting instruction for the concentrate, and performing ore sorting according to the jet parameters and the sorting instruction. The present invention relates to the technical field of ores, and solves the technical problems of low processing efficiency and insufficient sorting accuracy of existing ore sorting methods.

Description

Technical Field

[0001] The present invention belongs to the field of mining, involves deep learning technology, and specifically relates to an ore sorting system and method based on deep learning. Background Art

[0002] Mineral resources, as the basic material for national economic construction, with the rapid development of the economy, the demand for them has increased rapidly, and efficient mining technology has become increasingly important. Ore sorting is a key link in the mining industry, and its accuracy and efficiency have a significant economic impact on the mining industry.

[0003] Traditional mechanical sorting mainly realizes separation based on the differences in physical properties (such as density, magnetism, conductivity, hydrophobicity, etc.) between ores and gangue (waste rocks). Common methods include gravity separation, flotation, magnetic separation, electrostatic separation, etc. There are generally problems such as strong dependence on the physical characteristics of ores, complex processes, large equipment investment, and insufficient environmental friendliness, and it is difficult to meet the efficient sorting requirements of low-grade and complex symbiotic ores.

[0004] With the development of computer technology and image processing technology, image recognition technology has gradually been applied to the field of ore sorting. For example, by collecting mineral images, using convolutional neural networks to realize mineral recognition and sorting, or using support vector machines to sort single ores, etc. However, such methods usually have strict requirements for the original image data, and can often only sort specific types of ores, with a narrow scope of application. Moreover, since the characteristic information of the samples to be sorted is sometimes difficult to distinguish, the sorting results lack interpretability and the accuracy is relatively low.

[0005] At the same time, when determining the processing flow of the concentrate, the traditional method is usually to conduct a composition analysis on the selected concentrate after ore sorting, and based on the characteristics such as the content of various elements and the mineral composition in the concentrate, combined with the requirements of the target product, to determine the subsequent processing flow. If the content of a certain valuable metal in the concentrate is relatively high and the impurities are few, it may be directly further processed such as smelting; if the concentrate contains multiple valuable components or the impurity content is high, it may need to go through multiple complex purification and separation processes, such as using chemical leaching, flotation and other methods for further processing to improve the purity and quality of the concentrate and meet the requirements of different industrial production. However, when the traditional method determines the processing flow of the concentrate, it often relies on manual experience and conventional analysis means, lacking a comprehensive and accurate grasp of the ore characteristics, resulting in the processing flow may not be optimized, affecting production efficiency and economic benefits. Summary of the Invention

[0006] The present invention aims to at least solve one of the technical problems existing in the prior art; for this purpose, the present invention proposes an ore sorting system and method based on deep learning, which is used to solve the technical problems of low processing efficiency and insufficient sorting accuracy of existing ore sorting methods.

[0007] To achieve the above object, a first aspect of the present invention provides an ore sorting system based on deep learning, including:

[0008] A data acquisition module: used to perform particle size grading and ore washing on the raw ore through several layers of vibrating screens to obtain preprocessed graded ore, and collect image data and hyperspectral data of the preprocessed graded ore;

[0009] An ore classification module: used to respectively extract the surface features of the image data and the spectral features of the hyperspectral data by using a first network and a second network constructed based on deep learning algorithms, and input the fused surface features and spectral features into a classification network to obtain the ore type and the main components of the ore; wherein, the ore types include concentrate, middling ore and tailings;

[0010] A sorting execution module: used to calculate jet parameters based on the conveyor belt speed, ore quality and air density, determine the sorting instruction for the concentrate, and perform ore sorting according to the jet parameters and the sorting instruction.

[0011] It should be noted that the data acquisition module, ore classification module and sorting execution module of the present invention are communicatively connected.

[0012] Furthermore, the collection of the image data and hyperspectral data of the preprocessed graded ore includes:

[0013] Using an electromagnetic vibrating feeder to spread the preprocessed graded ore flat on the conveyor belt, and setting the conveyor belt to run at a preset speed at a constant speed;

[0014] Using an industrial camera and a hyperspectral camera to respectively obtain the image data and hyperspectral data of the graded ore at a preset shooting interval.

[0015] Furthermore, the first network includes a multi-scale residual pyramid module, a dynamic convolution attention module and a Transformer encoder that are sequentially formed; wherein,

[0016] The input of the multi-scale residual pyramid module is the image data, which includes several parallel expansion convolution branches composed of dilated convolution and atrous spatial pyramid pooling layers, and aggregates the outputs of several levels of expansion convolution branches through bilinear interpolation upsampling channel splicing and a 1×1 convolution layer to obtain an aggregated feature F agg ;

[0017] The input of the dynamic convolution attention module is the aggregated feature, which includes a parameter generation network composed of a global average pooling layer, a 1×1 convolution layer and a ReLU function, and a dynamic kernel convolution module composed of depthwise separable convolution and residual connection, to obtain an attention-enhanced feature, and input it into the Transformer encoder to obtain surface features 。

[0018] Furthermore, the formula of the dynamic convolutional attention module is as follows:

[0019] The convolution kernel parameter K is obtained through the parameter generation network: K = f (1) 1×1 (ReLU(f (0) 1×1 (GAP(F agg ))))); where GAP( ) represents the global average pooling layer, and f (0) 1×1 ( ) represents the first 1×1 convolutional layer, and f (1) 1×1 ( ) represents the second 1×1 convolutional layer, and ReLU( ) represents the ReLU function;

[0020] Through the dynamic kernel convolution module, the convolution kernel parameter and the aggregated feature are subjected to depthwise separable convolution and residual connection to obtain the attention-enhanced feature: Y out = F agg + α × reshape(f g (reshape(F agg ), K)); where reshape( ) represents the dimension reshaping operation, and f g ( ) represents the depthwise separable convolution, and α represents the learnable parameter.

[0021] The first network adopts a multi-scale residual pyramid module, extracts multi-scale texture features through the parallel connection of dilated convolution and the atrous spatial pyramid pooling layer, and combines bilinear interpolation and 1×1 convolution to aggregate features, which can effectively retain the details of the ore surface; the dynamic convolutional attention module uses global average pooling to generate dynamic convolution kernel parameters, combines depthwise separable convolution and residual connection to achieve dynamic focusing on the key areas of the ore surface, and enhances the feature discriminability; the Transformer encoder further captures the global context dependence and improves the feature expression ability in the case of the complex surface of the ore.

[0022] Furthermore, the second network includes a spectral band grouping module, a local-global spectral attention module, and a 3D convolutional feature compression layer; where

[0023] The input of the spectral band grouping module is hyperspectral data, including dividing several spectral bands of the spectral data into a preset number of groups by using the K-means algorithm to obtain several waveband subsets, and sequentially inputting the several waveband subsets into a three-dimensional convolutional layer and a pooling layer, then performing splicing in the spectral dimension, and generating an intermediate feature F through a spectral normalization layer group ;

[0024] The local-global spectral attention module includes extracting local band correlations of the intermediate feature by using three-dimensional convolution and the GELU activation function to obtain a local feature F local , and inputting the local feature into a multi-head self-attention mechanism: , to obtain a global feature F global , where Q, K, and V respectively represent the query matrix, key matrix, and value matrix obtained by linearly projecting the local feature, T represents the transpose operation, d represents the scaling factor, and is determined according to the number of attention heads and the number of channels of the global feature F global ;

[0025] The three-dimensional convolutional feature compression layer includes processing the global feature by using two-dimensional convolution and three-dimensional convolution, and then sequentially inputting it into a deformable three-dimensional convolutional layer and a global max pooling layer to obtain a spectral feature .

[0026] The second network divides waveband subsets through K-means clustering, and combines three-dimensional convolution and pooling layers to reduce the dimensional redundancy of ore hyperspectral features; the local-global spectral attention module extracts local band correlations through three-dimensional convolution and models the cross-band global relationship by using the multi-head self-attention mechanism, which can enhance the semantic expression of spectral features; the three-dimensional convolutional feature compression layer adopts deformable convolution and global pooling, which can flexibly adapt to spectral response differences and compress the feature dimension, improving the processing efficiency of ore hyperspectral feature extraction.

[0027] Further, the inputting the surface feature and the spectral feature into the classification network after fusion includes:

[0028] S1, establishing a semantic association between the surface feature and the spectral feature through a bidirectional cross-attention mechanism to obtain a joint feature ; where the bidirectional cross-attention mechanism includes:

[0029] Taking the surface feature as the query vector, the spectral feature as the key vector and the value vector, and calculating to obtain a first fusion feature according to the cross-attention calculation formula;

[0030] Taking the surface feature as the query vector , the spectral feature as the key vector and the value vector , and calculating to obtain a first fusion feature according to the attention calculation formula : ;

[0031] Take the spectral features as the query vector , and the surface features as the key vector and the value vector , and calculate the second fusion feature according to the attention calculation formula : ;

[0032] Concatenate the first fusion feature and the second fusion feature in the channel dimension to obtain the joint feature ;

[0033] S2. Use the gated network to generate the gated weight G for the joint feature, the surface feature, and the spectral feature, and perform an element-wise product of the gated weight and the joint feature to obtain the optimized fusion feature ; Among them, the mathematical expression of the gated network is: , σ( ) represents the Sigmoid function, δ( ) represents the GELU activation function, represents the feature concatenation operation, and W1 and W2 represent preset learnable parameter matrices;

[0034] S3. Input the optimized fusion feature into the classification network to obtain the main components of the ore and the ore type; among them, the classification network includes:

[0035] Extract the shared features from the optimized fusion feature using the normalization layer, the fully connected layer, and the dropout layer , and the mathematical expression is: , LayerNorm( ) represents the normalization layer, Linear( ) represents the fully connected layer, and Dropout( ) represents the dropout layer;

[0036] Extract the classification features from the shared features using the gradient reversal layer and the domain adversarial network, and then input the classification features into the ore type classifier containing the fully connected layer to output the probability distribution P1 of the predicted ore type; among them, the ore types include concentrate, middlings, and tailings;

[0037] Calculate the attention weights of the optimized fusion feature using the multi-layer perceptron MLP and the Softmax function, and perform an element-wise product with the optimized fusion feature to obtain the enhanced main feature, and input the enhanced main feature into the ore composition classifier containing the fully connected layer to output the probability distribution P2 of the main components of the ore; among them, the main components of the ore include Fe2O3, Fe3O4, SiO2, Al2O3, CuFeS2, FeS2, etc.

[0038] The classification network establishes semantic associations between images and spectral features through a bidirectional cross-attention mechanism to generate joint spectral features; and uses a gating network to dynamically adjust the weights of the joint features, surface features, and spectral features to optimize the fusion results; finally, in the classifier, the data processing efficiency is improved by extracting shared features, and then a domain adversarial network is used to enhance the discriminability of the classification features, and a multi-layer perceptron is used to enhance the expression of the principal component features through attention weights, and the ore type and principal component prediction tasks are processed separately to achieve intelligent ore classification.

[0039] Further, the jetting parameters calculated according to the conveyor belt speed, ore quality, and air density include:

[0040] Using a Kalman filter to predict the movement trajectory of the ore to obtain the time when the ore reaches the jetting position ;

[0041] According to the formula Calculate the jetting delay time ; where, represents the physical distance from the detection position to the jet valve, v b represents the conveyor belt speed, represents the response time of the solenoid valve, represents the time-consuming of the ore classification module;

[0042] Subtract the time when the ore reaches the jetting position from the jetting delay time to obtain the jetting trigger time ;

[0043] Set the sorting channel according to the ore type, and calculate the target lateral velocity according to the sorting channel spacing , and the calculation formula is: ; where h represents the falling height and g represents the acceleration due to gravity;

[0044] According to the formula Calculate the jetting duration; where, represents the ore quality, represents the air density, represents the nozzle cross-sectional area, represents the nozzle flow coefficient.

[0045] Further, the determination of the sorting instruction for the concentrate includes:

[0046] Obtain the physical and chemical property value A of the main component i of the concentrate ore according to the ore property database i ; where the chemical properties include magnetism, density, hydrophobicity, conductivity, and dielectric constant;

[0047] Set the economic weight E of the main ore component i of the concentrate i ;

[0048] According to the formula CII i =P i ×[β×A i +(1 - β)×E i Calculate the physicochemical property influence index CII of the main ore component i i ; where P i represents the probability value of the main ore component i, obtained from the output of the classification network, and β represents the preset influence coefficient;

[0049] According to the formula Calculate the sorting method adaptation score S of the concentrate k ; where k represents the sorting method index, represents the applicability function of the sorting method k to the main ore component i, and n represents the total number of main ore components;

[0050] Take the method with the adaptation score S k greater than the preset threshold as the sorting method of the concentrate to obtain the sorting instruction of the concentrate.

[0051] Further, the applicability function of the sorting method k to the main ore component i includes:

[0052] Magnetic separation applicability function: ; where k1 represents the slope factor, represents the magnetic susceptibility of the main component i, represents the preset magnetic susceptibility threshold;

[0053] Flotation applicability function: ; where γ represents the hydrophobicity gain coefficient, represents the hydrophobic contact angle of the main component i, represents the surface potential of the main component i, represents the critical value of the surface potential of the main component i, represents the potential sensitivity factor;

[0054] Gravity separation applicability function: ; where, represents the density difference between the main component i and the sorting medium, represents the density distribution dispersion of the ore, represents the density difference gain coefficient;

[0055] Electrostatic separation applicability function: ; where, represents the conductivity of the main component i, represents the dielectric constant of the main component i, 、 respectively represent the reference conductivity and the reference dielectric constant, and x represents the nonlinear coefficient.

[0056] Using Kalman filtering to predict the ore trajectory, and dynamically calculating the jet delay time and duration in combination with the conveyor belt speed, ore mass and air density, can ensure the accuracy of ore sorting actions; and in the subsequent sorting process of further determining the concentrate, based on the physical and chemical properties and economic weights of the main components of the concentrate, calculating the adaptation score through applicability functions (such as magnetic separation, flotation, etc.), and selecting the optimal sorting method, can improve the resource recovery rate and sorting pertinence.

[0057] The second aspect of the present invention provides an ore sorting method based on deep learning, including:

[0058] Carrying out particle size classification and ore washing treatment on the raw ore through several layers of vibrating screens to obtain pre-treated classified ore, and collecting image data and hyperspectral data of the pre-treated classified ore;

[0059] Using a first network and a second network constructed based on deep learning algorithms to extract the surface features of the image data and the spectral features of the hyperspectral data respectively, and fusing the surface features and the spectral features and inputting them into a classification network to obtain the ore type and the main components of the ore; among them, the ore type includes concentrate, middlings and tailings;

[0060] Calculating jet parameters according to the conveyor belt speed, ore mass and air density, and determining the sorting instruction of the concentrate, and performing ore sorting according to the jet parameters and the sorting instruction.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] The present invention improves the accuracy and efficiency of ore sorting through multi-modal data fusion and deep feature extraction technologies. First, collecting the image and hyperspectral data of the ore, constructing a dual-network structure to extract the surface texture and spectral component features respectively, and realizing cross-modal semantic association through a bidirectional cross-attention mechanism to enhance the classification ability for complex ores (such as multi-mineral mixtures); in addition, the dynamic convolution attention and gating network mechanism dynamically focus on key areas and adjust feature weights, effectively suppressing noise interference, especially showing stronger robustness in scenarios with light changes or ore surface contamination; in addition, the multi-scale residual pyramid and spectral band grouping technologies can retain key information while reducing the computational complexity, making the ore classification process have the advantages of high accuracy and low latency in industrial scenarios;

[0063] In terms of ore separation strategies, the Kalman filter is used to predict the ore movement trajectory. By combining parameters such as conveyor belt speed and air density, the jet delay time and duration are accurately calculated to ensure the precise triggering of separation actions. At the same time, based on the physical and chemical properties of the main components of the ore, an applicability function is constructed to dynamically select the optimal concentrate separation method, which can not only improve the separation efficiency, reduce the separation cost, but also specifically recover high-value minerals and enhance the resource utilization rate. Brief Description of the Drawings

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0065] Figure 1 It is a schematic diagram of the technical process of the ore separation system based on deep learning provided by the present invention;

[0066] Figure 2 It is a schematic diagram of the framework of the ore separation system based on deep learning provided by the present invention;

[0067] Figure 3 It is a schematic diagram of the technical process for determining the concentrate separation instruction provided by the present invention. Detailed Embodiments

[0068] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0069] Please refer to Figures 1-3 , the first aspect embodiment of the present invention provides an ore separation system based on deep learning, including:

[0070] Data acquisition module: used to perform particle size classification and ore washing on the raw ore through several layers of vibrating screens to obtain preprocessed classified ore, and collect the image data and hyperspectral data of the preprocessed classified ore;

[0071] Ore classification module: used to extract the surface features of the image data and the spectral features of the hyperspectral data respectively by using the first network and the second network constructed based on deep learning algorithms, and fuse the surface features and spectral features and input them into the classification network to obtain the ore type and the main components of the ore; among them, the ore type includes concentrate, middlings, and tailings;

[0072] Sorting execution module: It is used to calculate jet parameters based on conveyor belt speed, ore quality, and air density, determine the sorting instruction for concentrate, and perform ore sorting according to the jet parameters and sorting instruction.

[0073] To achieve an intelligent ore sorting process, it is first necessary to perform hierarchical preprocessing on the ore to eliminate impurity interference and ensure the uniformity of ore particles. Specifically,

[0074] In the data acquisition module, the raw ore needs to be physically screened through multiple layers of vibrating screens. Each layer of sieve mesh separates ore particles in different particle size ranges according to a preset aperture. At the same time, combined with high-pressure water flow flushing, impurities such as soil and dust attached to the surface are removed to form clean and uniformly sized graded ore;

[0075] Next, the preprocessed graded ore is spread evenly on the conveyor belt through an electromagnetic vibrating feeder. The conveyor belt runs steadily at a preset constant speed to ensure that the ore is continuously transported in a single layer without overlap. The industrial camera and the hyperspectral camera are synchronously triggered to take pictures at preset intervals: The industrial camera obtains two-dimensional grayscale or color image data of the ore surface to capture visual features such as texture and shape; The hyperspectral camera collects continuous spectral data covering hundreds of bands to record the spectral response information of the ore composition.

[0076] Next, in the ore classification module, the first network (image feature extraction) and the second network (hyperspectral feature extraction) respectively extract the surface features of the image data and the spectral features of the hyperspectral data, and input the fused surface features and spectral features into the classification network (feature fusion and classification decision) to achieve high-precision identification of ore types and main components; Among them,

[0077] The first network is composed of a multi-scale residual pyramid module, a dynamic convolution attention module, and a Transformer encoder, and is used to extract rich surface features from the ore image. Specifically:

[0078] 1. Multi-scale residual pyramid module:

[0079] The input is the preprocessed image data (with dimensions of H×W×C, where H represents length, W represents width, and C represents the number of channels). By paralleling multiple dilated convolution branches, each branch contains a dilated convolution with a different dilation rate (dilation rate r = 2, 4, 8) and an atrous spatial pyramid pooling layer (ASPP) to extract multi-scale texture features; Then, the outputs of each branch are upsampled by bilinear interpolation and concatenated along the channels, and then reduced to aggregated feature F through a 1×1 convolution agg : F agg =Conv1×1(Concat(Upsample(F (1) aspp ),…,Upsample(F(L) aspp ))); where, Conv1×1( ) represents a 1×1 convolution, Concat( ) represents channel concatenation, Upsample( ) represents a bilinear interpolation upsampling operation to upsample the outputs of each branch of ASPP to the same resolution, F (1) aspp represents the first dilated convolutional branch, F (L) aspp represents the L-th dilated convolutional branch;

[0080] The combination of dilated convolutions with different dilation rates (r = 2, 4, 8) and the multi-level dilation rates of ASPP enables each branch to capture multi-scale features from local textures (small dilation rates) to global structures (large dilation rates). For example, the branch with a dilation rate of r = 2 can detect fine cracks on the ore surface, while the branch with r = 8 can perceive the outlines of large mineral blocks; at the same time, traditional downsampling (such as pooling) reduces the image resolution and loses edge details, while the multi-scale residual pyramid module replaces pooling with dilated convolutions to expand the receptive field while maintaining the spatial resolution; subsequent bilinear interpolation upsampling restores the outputs of each branch to the same resolution, which can avoid information loss caused by downsampling and is especially suitable for the fine analysis of ore surface textures;

[0081] 2. Dynamic Convolutional Attention Module

[0082] Taking the aggregated feature F agg as the input, dynamic convolution kernel parameters K = f (1) 1×1 (ReLU(f (0) 1×1 (GAP(F agg )))) are generated through global average pooling (GAP) and 1×1 convolution; where, GAP( ) represents the global average pooling layer, f (0) 1×1 ( ) represents the first 1×1 convolutional layer, f (1) 1×1 ( ) represents the second 1×1 convolutional layer, ReLU( ) represents the ReLU function;

[0083] Subsequently, K and F agg are subjected to depthwise separable convolution and residual connection with F agg to generate the attention-enhanced feature: Y out = F agg + α × reshape(f g(reshape(F agg ),K)); Among them, reshape( ) represents an operation to reshape the dimensions of the data, and f g ( ) represents depthwise separable convolution, and α represents a learnable parameter;

[0084] By dynamically adjusting the convolution kernel parameters, the convolution operation can adaptively focus on key regions such as ore boundaries and texture differences. For example, when there is metallic luster on the ore surface, the dynamic kernel will strengthen the response to the highlight area; then adding the dynamic convolution result to the original features can strengthen the feature response in important regions and suppress background noise;

[0085] 3. Transformer Encoder

[0086] Input the attention-enhanced features into the Transformer, capture global context dependencies through the multi-head self-attention mechanism, and output surface features , which strengthens the global feature expression of the ore under complex backgrounds.

[0087] The second network consists of a spectral band grouping module, a local-global spectral attention module, and a three-dimensional convolutional feature compression layer, and is used to extract spectral features of hyperspectral data; the specific network structure is as follows:

[0088] 1. Spectral Band Grouping Module

[0089] Divide the bands of the input hyperspectral data into G groups through the K-means algorithm; each group of bands is concatenated along the spectral dimension after three-dimensional convolution (kernel size 3×3×1) and pooling processing to generate intermediate feature F group : F group =Concat(Pool(Conv3D(Group1)),…,Pool(Conv3D(Group K ))); Among them, Pool( ) represents the pooling layer, Group i represents the i-th group of bands, and Conv3D( ) represents three-dimensional convolution;

[0090] Hyperspectral data usually contains hundreds of bands, and direct processing will lead to the curse of dimensionality; dividing the bands into G groups through K-means, with the number of bands in each group reduced to B / G (B represents the number of hyperspectral bands), can significantly reduce the computational complexity of subsequent three-dimensional convolution; after three-dimensional convolution for each group, downsampling is performed through the pooling layer, and then concatenation along the spectral dimension to generate intermediate features can integrate the features of different groups in spectral order. For example, concatenating the features of metal oxide sensitive bands and silicate sensitive bands can promote cross-group feature interaction and improve the effective utilization of band correlations;

[0091] 2. Local-Global Spectral Attention Module

[0092] First, local band correlations are extracted from the intermediate feature F through three-dimensional convolution and the GELU activation function to obtain the local feature F group : F local : F local =GELU(Conv3D(F gruop ));

[0093] Subsequently, the local feature is input into the multi-head self-attention mechanism to generate the global feature:

[0094] , obtaining the global feature F global , where Q, K, and V represent the query matrix, key matrix, and value matrix obtained by linearly projecting the local feature through the linear projection layer respectively, T represents the transpose operation, and d represents the scaling factor, which is determined according to the number of attention heads and the number of channels of the global feature F global ; Through the cooperation of three-dimensional convolution and self-attention, both local spectral details and global component correlations are taken into account, enhancing the semantic expression of spectral features;

[0095] 3. Three-Dimensional Convolution Feature Compression Layer

[0096] After two-dimensional and three-dimensional convolution processing of the global feature, the deformable three-dimensional convolution is used to dynamically adjust the sampling position of the convolution kernel to adapt to the spectral feature, and the dimension is compressed through global max pooling to obtain the spectral feature .

[0097] The classification network realizes feature fusion and decision-making through the bidirectional cross-attention mechanism, gating network, and multi-task classifier; its specific processing process is as follows:

[0098] S1. Establish the semantic association between the surface feature and the spectral feature through the bidirectional cross-attention mechanism to obtain the joint feature F joint ; Among them, the bidirectional cross-attention mechanism includes:

[0099] Take the surface feature as the query vector , the spectral feature as the key vector and the value vector , and calculate the first fusion feature according to the attention calculation formula: ;

[0100] Take the spectral feature as the query vector , the surface feature as the key vector and the value vector , and calculate the second fusion feature according to the attention calculation formula: ;

[0101] Concatenate the first fusion feature and the second fusion feature in the channel dimension to obtain the joint feature F joint ;

[0102] S2. Generate the gating weight G by using the gating network for the joint feature, the surface feature, and the spectral feature, and perform an element-wise multiplication of the gating weight and the joint feature to obtain the optimized fusion feature F opt ; where the mathematical expression of the gating network is: , σ( ) represents the Sigmoid function, δ( ) represents the GELU activation function, represents the feature concatenation operation, and W1, W2 represent the preset learnable parameter matrices;

[0103] S3. Input the optimized fusion feature into the classification network to obtain the main components of the ore and the ore type; where the classification network includes:

[0104] Extract the shared feature from the optimized fusion feature by using the normalization layer, the fully connected layer, and the dropout layer, and the mathematical expression is: F shared =Dropout(δ(Linear(LayerNorm(F opt )))), LayerNorm( ) represents the normalization layer, Linear( ) represents the fully connected layer, and Dropout( ) represents the dropout layer;

[0105] Extract the classification feature from the shared feature by using the gradient reversal layer and the domain adversarial network, and then input the classification feature into the ore type classifier containing the fully connected layer to output the probability distribution P1 of the predicted ore type; where the ore types include concentrate, middlings, and tailings;

[0106] Calculate the attention weight of the optimized fusion feature by using the multi-layer perceptron MLP and the Softmax function, and perform an element-wise multiplication of the attention weight and the optimized fusion feature to obtain the enhanced main feature, and input the enhanced main feature into the ore component classifier containing the fully connected layer to output the probability distribution P2 of the main components of the ore; where the main components of the ore include Fe2O3, Fe3O4, SiO2, Al2O3, CuFeS2, FeS2, etc.

[0107] Through the first network (image feature extraction) and the second network (hyperspectral feature extraction) constructed based on the deep learning algorithm in the ore classification module, the system can efficiently fuse the multi-modal information of the ore surface texture and chemical composition, and output the probability distributions of the ore type (concentrate, middlings, tailings) and the main components (such as Fe2O3, CuFeS2, etc.).

[0108] It should be noted that the first network, the second network, and the classification network constructed based on the deep learning algorithm can be applied to this system only after being trained and verified. The training data of the first network are several ore images, and the labels include concentrate, middlings, and tailings. The training data of the second network are several hyperspectral data of ores, and the labels are predefined main ore component categories (such as Fe2O3, Fe3O4, etc.), covering generalization categories of known and unknown main components. As the downstream task of multimodal fusion, the classification network takes as input the fusion features of the image surface features extracted by the first network and the spectral features extracted by the second network, and the output includes the predicted labels of ore types and main ore components.

[0109] Next, in the sorting execution module, by calculating the jet parameters of different types of ores, using the Kalman filter to predict the ore movement trajectory, and combining information such as the conveyor belt speed and the distance from the detection position to the jet valve, the jet delay time, trigger time, and jet duration are accurately determined, so as to blow the ore to the corresponding channel to achieve the preliminary separation of different types of ores. And further analysis is carried out on the concentrate among them. According to the numerical values of the physical and chemical properties and economic weights of the main components of the concentrate ore, the physical and chemical property influence index is calculated, combined with the sorting method adaptation score, and the subsequent sorting method reference for the concentrate is given to achieve the refined sorting of the concentrate, improve the resource recovery efficiency and sorting accuracy, and make the ore sorting process more scientific and efficient.

[0110] Specifically, in one implementation, the jet parameters calculated according to the conveyor belt speed, ore mass, and air density may include the following steps:

[0111] The sorting execution module undertakes the key task of converting the classification result into an actual sorting operation in the entire ore sorting process, and its specific operation steps are as follows:

[0112] Set the sorting channels for concentrate, middlings, and tailings according to the ore type respectively;

[0113] Use the Kalman filter to predict the movement trajectory in combination with the current movement state of the ore, so as to obtain the time when the ore reaches the jet position ;

[0114] According to the formula Calculate the jet delay time ; where represents the physical distance from the detection position to the jet valve, v b represents the conveyor belt speed, represents the solenoid valve response time, represents the time consumption of the ore classification module;

[0115] Then the time when the ore reaches the jet position and the jet delay time Take the difference to obtain the jet trigger time , ensuring that the jet device can perform jet operations on the ore at the optimal time;

[0116] Calculate the target lateral velocity based on the sorting channel spacing The calculated target lateral velocity , and the calculation formula is: ; where h represents the falling height and g represents the acceleration due to gravity;

[0117] Calculate the jet duration according to the formula ; where represents the ore mass, represents the air density, represents the nozzle cross-sectional area, represents the nozzle flow coefficient, so that the jet duration can meet the requirement of blowing the ore to the specified channel.

[0118] In one implementation, determining the sorting instruction for the concentrate may include the following operation steps:

[0119] Obtain the physico-chemical property value A of the main ore component i of the concentrate from the ore property database i ; where the chemical properties include magnetism, density, hydrophobicity, conductivity, and dielectric constant;

[0120] Set the economic weight E of the main ore component i of the concentrate i ;

[0121] Calculate the physico-chemical property influence index CII of the main ore component i according to the formula CII i =P i ×[β×A i +(1-β)×E i ; where P i represents the probability value of the main ore component i, output by the classification network, and β represents the preset influence coefficient; i

[0122] Calculate the sorting method adaptation score S of the concentrate according to the formula ; where k represents the sorting method index, k represents the applicability function of the sorting method k to the main ore component i, including:

[0123] Magnetic separation applicability function: ; where k1 represents the slope factor, represents the magnetic susceptibility of the main component i, represents the preset magnetic susceptibility threshold;

[0124] Flotation applicability function: ; where γ represents the hydrophobicity gain coefficient, represents the hydrophobic contact angle of the main component i, represents the surface potential of the main component i, represents the critical value of the surface potential of the main component i, represents the potential sensitive factor;

[0125] Gravity separation applicability function: ; where represents the density difference between the main component i and the separation medium, represents the density distribution dispersion of the ore, represents the density difference gain coefficient;

[0126] Electrostatic separation applicability function: ; where represents the conductivity of the main component i, represents the dielectric constant of the main component i, and represent the reference conductivity and the reference dielectric constant respectively, and x represents the non-linear coefficient;

[0127] Finally, the method with the adaptation score S k greater than the preset threshold is used as the separation method for the concentrate, and a separation instruction for the concentrate is obtained, which provides a pretreatment basis for the subsequent processing (such as smelting) of the concentrate;

[0128] It should be noted that the concentrate is the part with the highest economic value in the ore and usually contains target metals (such as Fe2O3, CuFeS2, etc.). After preliminary separation directly through the jet parameters (such as blowing to the concentrate channel), further refined separation is still required to maximize the resource recovery rate; moreover, since the concentrate may contain multiple mineral components, it is difficult to completely separate them by a single separation method. Therefore, a combined separation strategy needs to be selected according to the physical and chemical properties of the main components to reduce ineffective separation steps and reduce energy consumption and the use of chemical reagents.

[0129] The second aspect of the embodiments of the present invention provides an ore separation method based on deep learning, including:

[0130] Performing particle size classification and ore washing on the raw ore through several vibrating screens to obtain the pretreated classified ore, and collecting the image data and hyperspectral data of the pretreated classified ore;

[0131] Using the first network and the second network constructed based on the deep learning algorithm to extract the surface features of the image data and the spectral features of the hyperspectral data respectively, and fusing the surface features and the spectral features and inputting them into the classification network to obtain the ore type and the main components of the ore; where the ore type includes concentrate, middlings, and tailings;

[0132] The jetting parameters are calculated based on the conveyor belt speed, ore quality, and air density, and the sorting instruction for the concentrate is determined. Ore sorting is performed according to the jetting parameters and the sorting instruction.

[0133] Some of the data in the above formula are taken as numerical values after removing the dimension. The formula is the one closest to the actual situation obtained through software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

[0134] The working principle of the present invention:

[0135] First, after the raw ore is subjected to particle size classification by a vibrating screen and ore washing treatment, it is spread evenly on a belt moving at a constant speed by an electromagnetic vibrating feeder, and image and spectral data are synchronously collected using an industrial camera and a hyperspectral camera. Then, the first network and the second network respectively extract features from the image and the hyperspectral data, and then feature fusion is performed through a bidirectional cross-attention mechanism and a gated network. The optimized features after fusion are input into a classification network to output the ore type and the probability distribution of the main components. Then, the jetting parameters are calculated based on information such as the conveyor belt speed, ore quality, and air density, and different types of ore sorting are completed using a jetting device. The subsequent sorting instruction for the concentrate is further determined in combination with the main component attributes and economic weights of the concentrate.

[0136] The above embodiments are only used to illustrate the technical method of the present invention and not to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical method of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical method of the present invention.

Claims

1. The ore sorting system based on deep learning is characterized by: include: Data acquisition module: used to classify and wash the raw ore through several layers of vibrating screens to obtain pre-treated classified ore, and to collect image data and hyperspectral data of the pre-treated classified ore; Ore classification module: used to extract the surface features of image data and the spectral features of hyperspectral data respectively by using the first network and the second network constructed based on the deep learning algorithm, and then fuse the surface features and the spectral features and input them into the classification network to obtain the ore type and the main components of the ore; wherein the ore type includes concentrate, middlings and tailings; Sorting execution module: used to calculate the jet parameters according to the conveyor belt speed, ore quality and air density, and determine the sorting instructions for the concentrate, and perform ore sorting according to the jet parameters and sorting instructions; The first network includes a multi-scale residual pyramid module, a dynamic convolutional attention module and a Transformer encoder, which are sequentially constructed; The input of the multi-scale residual pyramid module is image data, which includes several levels of dilated convolution branches composed of dilated convolution and dilated spatial pyramid pooling layers in parallel, and the outputs of several levels of dilated convolution branches are aggregated through bilinear interpolation upsampling channel splicing and 1×1 convolution layer to obtain the aggregated feature F agg ; The input of the dynamic convolution attention module is the aggregated feature, which includes a parameter generation network composed of a global average pooling layer, a 1×1 convolution layer, and a ReLU function, and a dynamic kernel convolution module composed of a depth-separable convolution and a residual connection, to obtain the attention enhancement feature, which is input into the Transformer encoder to obtain the surface feature. .

2. The deep learning-based ore sorting system according to claim 1, characterized in that: The collected pre-processed image data and hyperspectral data of the graded ore include: The pre-treated graded ore is spread on the conveyor belt by using an electromagnetic vibrating feeder, and the conveyor belt is set to convey at a uniform speed at a preset speed; Industrial cameras and hyperspectral cameras are used to obtain image data and hyperspectral data of graded ores at preset shooting intervals.

3. The deep learning-based ore sorting system according to claim 1, characterized in that: The formula of the dynamic convolution attention module is: The convolution kernel parameter K is obtained through the parameter generation network: K=f (1) 1×1 (ReLU(f (0) 1×1 (GAP(F agg ))))); Among them, GAP( ) represents the global average pooling layer, f (0) 1×1 ( ) represents the first 1×1 convolutional layer, f (1) 1×1 ( ) represents the second 1×1 convolutional layer, ReLU( ) represents the ReLU function; The dynamic kernel convolution module is used to perform deep separable convolution and residual linking on the convolution kernel parameters and aggregate features to obtain the attention enhancement feature: Y out =F agg +α×reshape(f g (reshape(F agg ),K)); where reshape( ) represents the dimension reshaping operation, f g ( ) represents a depth-wise separable convolution, and α represents a learnable parameter.

4. The deep learning-based ore sorting system according to claim 3, characterized in that: The second network includes a spectral band grouping module, a local-global spectral attention module and a three-dimensional convolutional feature compression layer; wherein, The input of the spectral band grouping module is hyperspectral data, including using the K-means algorithm to divide several spectral bands of the spectral data into a preset number of groups to obtain several band subsets, and then inputting the several band subsets into the three-dimensional convolution layer and the pooling layer in turn to perform spectral dimension splicing, and generating intermediate features F through the spectral normalization layer. group ; The local-global spectral attention module includes extracting the local band correlation of the intermediate features using three-dimensional convolution and GELU activation function to obtain the local feature F local , and input the local features into the multi-head self-attention mechanism: , and obtain the global feature F global , where Q, K, and V represent the query matrix, key matrix, and value matrix obtained by the linear projection layer of local features, T represents the transposition operation, and d represents the scaling factor. global The number of channels is determined; The three-dimensional convolution feature compression layer includes processing the global features using two-dimensional convolution and three-dimensional convolution, and then inputting them into the deformable three-dimensional convolution layer and the global maximum pooling layer in sequence to obtain the spectral features. .

5. The deep learning-based ore sorting system according to claim 4, characterized in that: The step of fusing the surface features and the spectral features and inputting them into the classification network comprises: S1, establishes semantic associations between surface features and spectral features through a bidirectional cross-attention mechanism to obtain joint features ; Among them, the bidirectional cross attention mechanism includes: Using surface features as query vectors , the spectral features serve as key vectors Sum value vector , the first fusion feature is calculated according to the attention calculation formula : ; Use spectral features as query vectors , surface features as key vectors Sum value vector , the second fusion feature is calculated according to the attention calculation formula : ; The first fusion feature and the second fusion feature are concatenated in the channel dimension to obtain the joint feature ; S2, the joint features, surface features, and spectral features are combined using a gating network to generate a gating weight G, and the gating weight is element-wise multiplied with the joint features to obtain the optimized fusion feature ; Among them, the mathematical expression of the gating network is: ,σ( ) represents the Sigmoid function, δ( ) represents the GELU activation function, represents the feature concatenation operation, W1 and W2 represent the preset learnable parameter matrices; S3, inputting the optimized fusion features into the classification network to obtain the main components and ore types of the ore; wherein the classification network includes: Use normalization layers, fully connected layers, and random dropout layers to extract shared features from optimized fusion features , the mathematical expression is: ,LayerNorm( ) represents the normalization layer, Linear( ) represents the fully connected layer, Dropout( ) represents a random dropout layer; The classification features in the shared features are extracted using the gradient reversal layer and the domain adversarial network, and then the classification features are input into the ore type classifier containing the fully connected layer to output the probability distribution P1 of the predicted ore type; the ore types include concentrate, middlings and tailings; The multi-layer perceptron MLP and Softmax function are used to calculate the attention weight of the optimized fusion feature, and the element-by-element product is performed with the optimized fusion feature to obtain the enhanced main feature. The enhanced main feature is input into the ore component classifier containing a fully connected layer, and the probability distribution P2 of the ore main component is output; among which, the ore main components include Fe2O3, Fe3O4, SiO2, Al2O3, CuFeS2, and FeS2.

6. The deep learning-based ore sorting system according to claim 1, characterized in that: The jet parameters calculated according to the conveyor belt speed, ore mass and air density include: Use Kalman filter to predict the movement trajectory of ore and get the time when ore reaches the jet position ; According to the formula Calculate the jet delay time ;in, Indicates the physical distance from the detection position to the injection valve, v b Indicates the conveyor belt speed, Indicates the solenoid valve response time, Indicates the time consumption of the ore classification module; Time for the ore to reach the jet position With jet delay time Subtract and get the jet trigger time ; Set up sorting channels according to ore type and set the sorting channel spacing according to Calculate the target lateral velocity , the calculation formula is: ; Where h represents the falling height and g represents the acceleration due to gravity; According to the formula The jet duration is calculated; where, Indicates the quality of the ore, represents the air density, represents the nozzle cross-sectional area, Indicates the nozzle flow coefficient.

7. The deep learning-based ore sorting system according to claim 5, characterized in that: The determination of the separation instructions for the concentrate comprises: Obtain the physical and chemical property value A of the main component i of the concentrate according to the ore characteristics database i ; chemical properties include magnetism, density, hydrophobicity, conductivity and dielectric constant; Set the economic weight E of the main component i of the concentrate i ; According to the formula CII i =P i ×[β×A i +(1-β)×E i ] Calculate the physical and chemical property influence index CII of the main component i of the ore i ; Among them, P i represents the probability value of the main component i of the ore, which is obtained by the output of the classification network, and β represents the preset influence coefficient; According to the formula Calculate the concentration separation method adaptation score S k ; where k represents the sorting method index, represents the applicability function of sorting method k to ore principal component i, and n represents the total number of ore principal components; The adaptation score S k The method that is greater than the preset threshold is used as the concentrate sorting method to obtain the concentrate sorting instruction.

8. The deep learning-based ore sorting system according to claim 7, characterized in that: The applicability function of the separation method k to the main component i of the ore includes: Magnetic separation suitability function: ; where k1 represents the slope factor, represents the magnetic susceptibility of principal component i, Indicates the preset magnetic susceptibility threshold; Flotation suitability function: ; where γ represents the hydrophobicity gain coefficient, represents the hydrophobic contact angle of principal component i, represents the surface potential of the principal component i, represents the critical value of the surface potential of the principal component i, represents the potential sensitivity factor; Reselect the suitability function: ;in, represents the density difference between the principal component i and the separation medium, Indicates the density distribution dispersion of the ore, represents the density difference gain coefficient; Electroselection suitability function: ;in, represents the conductivity of the main component i, represents the dielectric constant of principal component i, , represent the reference conductivity and reference dielectric constant respectively, and x represents the nonlinear coefficient.

9. A method for ore sorting based on deep learning, applied to an ore sorting system based on deep learning as claimed in any one of claims 1 to 8, characterized in that: include: The raw ore is subjected to particle size classification and ore washing treatment through several layers of vibrating screens to obtain pre-treated classified ore, and image data and hyperspectral data of the pre-treated classified ore are collected; The first network and the second network constructed based on the deep learning algorithm are used to extract the surface features of the image data and the spectral features of the hyperspectral data respectively, and the surface features and the spectral features are fused and input into the classification network to obtain the ore type and the main components of the ore; wherein the ore type includes concentrate, middlings and tailings; The jet parameters are calculated based on the conveyor belt speed, ore quality and air density, and the sorting instructions for the concentrate are determined. The ore sorting is performed based on the jet parameters and sorting instructions; The first network includes a multi-scale residual pyramid module, a dynamic convolutional attention module and a Transformer encoder, which are sequentially constructed; The input of the multi-scale residual pyramid module is image data, which includes several levels of dilated convolution branches composed of dilated convolution and dilated spatial pyramid pooling layers in parallel, and the outputs of several levels of dilated convolution branches are aggregated through bilinear interpolation upsampling channel splicing and 1×1 convolution layer to obtain the aggregated feature F agg ; The input of the dynamic convolution attention module is the aggregated feature, which includes a parameter generation network composed of a global average pooling layer, a 1×1 convolution layer, and a ReLU function, and a dynamic kernel convolution module composed of a depth-separable convolution and a residual connection, to obtain the attention enhancement feature, which is input into the Transformer encoder to obtain the surface feature. .

Images