A system and method for homogenizing and blending graphite flotation products based on AI visual recognition

Abstract

This invention discloses a homogenization and blending system and method for graphite flotation products based on AI visual recognition, relating to the field of graphite mineral processing technology. Specifically, it includes: installing an online sampling device and a hyperspectral visual imager at the end of the flotation production line to acquire hyperspectral and grayscale images of graphite concentrate; performing outlier removal and multi-scale denoising on the original images, and normalizing the grayscale images and spectral curves respectively; extracting gloss features, particle edge features, and spectral absorption peak features from the normalized data and fusing them into a joint feature vector; inputting the joint feature vector into a one-dimensional convolutional neural network to predict carbon content, particle size, and impurity index; finally, using the predicted quality label and inventory as input, employing an improved multi-objective particle swarm optimization algorithm to solve for the optimal blending ratio that meets the target specifications. This invention achieves online detection and intelligent blending of graphite concentrate quality, effectively improving product quality stability.

Description

Technical Field

[0001] This invention relates to the field of graphite mineral processing technology, specifically to a system and method for homogenizing and blending graphite flotation products based on AI visual recognition. Background Technology

[0002] In the graphite flotation production process, due to the inherent differences in the origin of the raw ore, the structure of the vein, the grade of the ore and the selectivity, even under the same physical flotation process parameters, the graphite concentrate (finished product) produced at different time periods still has significant differences in carbon content, particle size distribution, specific surface area and impurity composition.

[0003] Currently, the industry typically employs an offline model of "production first, testing later, and then mixing." This involves storing different batches of finished products in separate silos, and then manually calculating the blending ratio based on laboratory test results. This existing technology has the following drawbacks: long laboratory testing cycles prevent real-time reflection of batch differences, leading to significant fluctuations in the quality of the mixed product; the blending ratio relies on manual experience, lacking scientific basis and making it difficult to achieve optimal homogenization; and repeated silo transfers and mixing increase energy consumption and time costs, hindering continuous and stable industrial production. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a homogenization and blending system and method for graphite flotation products based on AI visual recognition. By analyzing images of the graphite flotation process, a multi-objective optimization algorithm is constructed to generate a scientific blending ratio, ensuring product quality and shortening production time.

[0005] This invention proposes a method for homogenizing and blending graphite flotation products based on AI visual recognition, comprising the following steps: S1. Collect raw data and preprocess the raw data to obtain normalized grayscale image and normalized spectral reflectance curve; S2. Calculate the mean and variance of grayscale values ​​for each block in the normalized grayscale image, and statistically analyze the mean and standard deviation of the gloss index for all blocks to obtain a gloss feature vector. Perform edge detection, morphological closing operation, and gradient direction quantization on the normalized grayscale image to obtain a particle edge feature vector. Smooth, correct the baseline, and locate the absorption peaks on the normalized spectral curve, extracting the graphite main peak depth, impurity absorption depth, and peak statistics to obtain a spectral absorption peak feature vector. Concatenate the gloss feature vector, particle edge feature vector, and spectral absorption peak feature vector to obtain a fused feature vector. S3. Construct a quality prediction model, using the fused feature vector as input, and obtain the predicted carbon content, particle size, and impurity distribution through multi-layer convolution and fully connected mapping. S4. An improved multi-objective particle swarm optimization algorithm is adopted to obtain the optimal mixing ratio by minimizing the weighted sum of squared deviations between the mixed mass and the target mass.

[0006] As a further technical solution, S1 includes: S101. Acquire spectral reflectance curves and grayscale images using a hyperspectral imager and a grayscale camera; S102. An outlier detection and median replacement method based on interquartile range was adopted to detect and correct outlier pixels in the original grayscale image, resulting in a corrected image. S103. A denoising method combining adaptive nonlocal mean filtering and guided filtering based on local noise standard deviation estimation was adopted to perform multi-scale denoising and edge restoration on the corrected image, resulting in denoised image data. S104. Normalize the denoised image data and adaptively normalize the spectral reflectance curve data to obtain a normalized grayscale image and a normalized spectral reflectance curve.

[0007] As a further technical solution, the acquisition of gloss characteristics in S2 includes: The normalized grayscale image is divided into non-overlapping image blocks, and the grayscale mean and variance of each block are calculated. A gloss index function is constructed based on the mean and variance; Calculate the mean and standard deviation of the gloss index for all blocks to form a gloss feature vector.

[0008] As a further technical solution, particle edge feature extraction in S2 includes: Gaussian smoothing is performed on the normalized grayscale image and the gradient magnitude and gradient direction are calculated. Non-maximum suppression and double threshold connection are performed using the high and low thresholds determined by the cumulative histogram to obtain a binary edge image. Edge density is calculated as a positive correlation index of granularity. Morphological closing operation is performed on the edge image to obtain the contour of connected particles. The median and standard deviation of the contour area are statistically analyzed. The gradient direction is quantized into multiple intervals at the edge points, the pixel frequency of each interval is counted, and the normalization is used to obtain the orientation histogram vector. The edge density, median area, standard deviation of area, and orientation histogram are concatenated to form the particle edge feature vector.

[0009] As a further technical solution, the spectral absorption peak characteristics of S2 include: Savitzky-Golay smoothing and asymmetric least squares baseline correction are applied to the normalized spectral curves. Local minima of the corrected spectrum are identified as absorption peaks, and peaks with depths greater than the threshold are retained. The depth of the main graphite peak is extracted within the characteristic absorption range of graphite. The maximum absorption depth is extracted within the iron ion absorption range and the hydroxyl absorption range, respectively. The mean depth, standard deviation of depth, and position variance of all absorption peaks are calculated to form the characteristic vector of spectral absorption peaks.

[0010] As a further technical solution, the quality prediction model in S3 includes: The quality prediction model consists of a one-dimensional convolutional neural network, which includes an input layer, three one-dimensional convolutional blocks, a global average pooling layer, and two parallel fully connected regression heads. The input layer reshapes the fused feature vectors into a tensor form that adapts to one-dimensional convolution; Each convolutional block contains, in sequence, a one-dimensional convolutional layer, a batch normalization layer, a ReLU activation function layer, and a max pooling layer; The global average pooling layer averages and flattens the output of the last convolutional block along the length dimension. One regression head outputs continuous predicted values ​​for carbon content and particle size, while the other regression head outputs the impurity index.

[0011] As a further technical solution, the improved multi-objective particle swarm optimization algorithm in S4 includes: Particle position encoding represents the usage of materials in each batch. The fitness function consists of a weighted multi-objective function and a penalty term. The penalty term includes the degree of violation of inventory constraints, total output constraints and non-negativity constraints. The penalty coefficient increases exponentially with the number of iterations. Maintain an external archive to store non-dominated solutions, select the global optimal position based on Pareto dominance, and use dynamic inertia weights that decrease linearly with the number of iterations; Gaussian mutation is applied to the particle position with a set probability; the process terminates when the maximum number of iterations is reached or the fitness value changes less than the threshold for multiple consecutive generations, and the solution with the smallest fitness in the archive is output as the optimal mixing scheme.

[0012] As a further technical solution, the multi-objective function in S4 includes: The multi-objective function is the weighted sum of squares of the deviations between the mixed carbon content, particle size, and impurity index and the target value, with carbon content having the largest weight, followed by particle size, and impurity index having the smallest weight. The target quality specifications include the target carbon content and permissible deviation, the target particle size and permissible deviation, the target impurity index and permissible deviation, and the planned total production weight.

[0013] This invention also proposes a graphite flotation product homogenization and blending system based on AI visual recognition, comprising: The online sampling and preprocessing module is used to acquire hyperspectral and grayscale images of graphite concentrate under flowing conditions, and to perform outlier processing, adaptive multi-scale denoising, and normalization on the original images. The feature extraction module is used to extract and fuse gloss features, grain edge features, and spectral absorption peak features from the preprocessed image. The quality prediction module uses a one-dimensional convolutional neural network to predict carbon content, particle size, and impurity index based on the fused feature vector. The intelligent blending calculation module uses a multi-objective particle swarm optimization algorithm to solve for the optimal blending ratio that meets the target quality specifications, based on the predicted quality value and inventory of each batch of materials. The automation control module is used to set the feeder flow rate and execute the mixing ratio according to the optimal blending ratio.

[0014] Beneficial effects of this invention: The traditional "post-production inspection" has been transformed into "real-time prediction." The quality prediction model has grasped the inherent quality differences between batches and achieved quality homogenization of large production lines through scientific proportioning, solving the industry pain point of large quality fluctuations between batches of graphite products. By using multi-objective optimization algorithms to replace manual experience, complex constraint optimization calculations for multiple batches and multiple objectives can be completed in seconds, quickly responding to inventory changes and making full use of low-grade materials to reduce inventory backlog. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0016] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 1. Data Acquisition and Preprocessing An online sampling device and a hyperspectral / multispectral visual imager are installed at the discharge chute at the end of the flotation production line. The hyperspectral imager uses a CMOS camera with a wavelength range of 400-1000 nm, coupled with a halogen lamp light source, to continuously acquire hyperspectral images of the graphite concentrate in its flowing state at a frequency of 30 frames per second. Simultaneously, a grayscale camera is set up to synchronously acquire surface morphology images. The acquired raw data includes: the spectral reflectance curve of each pixel. ( (wavelength) and grayscale image matrix To ensure data representativeness, each sampling period lasted 10 seconds, acquiring at least 300 image frames.

[0019] The original grayscale image suffers from defects such as saturated pixels, dead pixels, motion blur, and uneven illumination. Furthermore, the original spectral curve is affected by particle size and packing density, resulting in multiplicative scattering and baseline drift, which mask the true absorption peaks. These defects lead to distorted noise estimation, biased feature extraction, and decreased model prediction accuracy. Therefore, this embodiment employs outlier detection and repair, adaptive multi-scale denoising to suppress noise, linear normalization of the image to eliminate illumination differences, and standard normal transformation of the spectrum to correct scattering. This provides a clean, consistent, and high-quality input for subsequent feature extraction and quality prediction.

[0020] 1.1 Outlier Handling The original image may contain anomalous pixels caused by sensor defects, momentary dust occlusion, or overexposure due to surface specular reflection. Directly using these outliers for subsequent analysis would severely affect the accuracy of feature extraction. Therefore, this embodiment proposes a local anomaly detection and correction method based on interquartile range.

[0021] (1) For the original image Each pixel in Select a neighborhood window centered on it. (Size 5×5 pixels). Calculate the lower quartiles of the pixel grayscale values ​​within the window. and upper quartiles This leads to the interquartile range. : (2) Abnormal judgment conditions: In the formula, 1.5 is the standard boxplot threshold. This detection is sensitive to local outliers and is not affected by the overall grayscale distribution.

[0022] (3) Outlier correction: For pixels detected as anomalous, the median of its non-anomalous neighboring pixels is used for replacement. Let... For window Given the set of coordinates of all non-abnormal pixels within the range, the corrected pixel value is: In the formula, This indicates taking the median. Represents the original image in pixel coordinates grayscale value, Represents the original image in pixel coordinates The pixel grayscale value at that location, This indicates the pixel position within the neighborhood window.

[0023] 1.2 Adaptive Multi-Scale Denoising To address the noise issues caused by uneven surface reflection and dust interference in graphite concentrate during flow, an adaptive denoising strategy based on a combination of nonlocal mean and guided filtering is adopted.

[0024] (1) Estimation of local noise standard deviation right Each pixel Select neighborhood window (Size 15×15). Calculate the median of grayscale values ​​within the window. : For grayscale images First, calculate the local noise variance. ,in: In the formula, 1.4826 comes from the theoretical conversion relationship between the absolute deviation of the median and the standard deviation under a normal distribution. This indicates taking the median. Represents any pixel position in the neighborhood of the original grayscale image. The grayscale value at that location.

[0025] (2) Adaptive selection of smoothing parameters for nonlocal mean filtering Traditional nonlocal mean filtering uses globally fixed smoothing parameters, which are difficult to adapt to noise differences in different regions of an image. This embodiment adaptively determines the smoothing parameters for each pixel based on the local noise standard deviation.

[0026] use Calculate the adaptive smoothing parameters for each pixel: In the formula, These represent the weighting coefficients used in nonlocal mean filtering to calculate the similarity between pixel blocks. This is the proportionality coefficient, with a value of 0.6. The base smoothing parameter has a value of 2.

[0027] For each center pixel In the search window Traverse all neighboring pixels within a 31×31 pixel area. Define the image patch size as 7×7 (radius). ), relative offset Calculate the Gaussian-weighted Euclidean distance between two image patches: In the formula, For Gaussian kernel, , The sum of all Gaussian weights within the kernel, used for normalization. Indicates Intra-block offset centered The grayscale value at that location.

[0028] Weighting coefficient The calculation is as follows: Finally, the output image of the adaptive nonlocal mean filter. for: (3) Guided filtering edge repair While nonlocal mean filtering smooths noise, it inevitably weakens high-frequency information such as grain edges. To repair edges, this embodiment uses the original corrected image... As a guide graph, the nonlocal mean filter output is used. Guided filtering is applied to restore edges weakened by NLM smoothing. (In a local window) (Radius adaptive: at the edge) flat areas Within this framework, we assume a linear model: in, and The calculation formula is as follows: In the formula, This represents the mean of the guide graph within the window. This represents the variance of the guide graph within the window. This represents the mean of the NLM output within the window. This is a regularization parameter to prevent the denominator from being too small.

[0029] The output of each pixel is determined by all windows containing that pixel. , The weighted average is used to obtain the final denoised image, which is denoted as... .

[0030] 1.3 Normalization (1) Gray value normalization Grayscale images acquired in different batches or under different lighting conditions may exhibit overall brightness differences. To eliminate the impact of these differences on subsequent feature extraction, the denoised images are linearly normalized.

[0031] In the formula, This represents the minimum global grayscale value of the image. Represents the global maximum value of the image. This represents the normalized pixel.

[0032] (2) Spectral adaptive normalization Traditional SNV assumes each pixel is independent and does not consider the correlation of the spectra of neighboring pixels. This invention proposes a spatially constrained adaptive normalization method: for each pixel... Taking a 3×3 neighborhood centered on the pixel, the mean and standard deviation of the spectral curves of all pixels within the neighborhood are calculated. Then, SNV correction is performed on the pixel. The corrected spectral curve retains the scattering differences of the local microstructure. In the formula, The mean of the spectral curves within the neighborhood. The standard deviation of the neighborhood spectral curve, This represents the normalized spectral data.

[0033] Normalized image and normalized They are used together as input for subsequent feature extraction.

[0034] 2. Feature Extraction Normalized grayscale image (size (range [0,1]) and normalized spectral curve (wavelength The value range is real numbers, the mean is 0, and the standard deviation is 1. This module extracts three types of features: gloss features, particle edge features, and spectral absorption peak features, and merges them into a joint feature vector for use in subsequent quality prediction models.

[0035] 2.1 Gloss Feature Extraction Gloss reflects the specular reflectivity of graphite concentrate surfaces and is related to carbon content and the degree of graphitization. This embodiment uses a combination of local grayscale statistics and texture energy analysis to extract gloss features.

[0036] (1) Local gray mean and variance For normalized images Divide the image into non-overlapping blocks The block size is 16×16 pixels, where , Calculate the mean and variance of gray levels within each block: In the formula, .

[0037] Gloss is positively correlated with the mean and negatively correlated with the variance (high-gloss surfaces have uniform and bright grayscale). The gloss index is defined as: In the formula, This is the adjustment coefficient.

[0038] The gloss feature vector of the image is: (2) To further compress the dimensions, calculate the mean of all gloss indices. and standard deviation : The final gloss vector is: 2.2 Extraction of Particle Edge Features Particle edge features are used to evaluate the particle size distribution and shape of graphite concentrate, directly affecting subsequent particle size prediction models. This invention employs an improved method combining Canny edge detection with edge orientation histograms.

[0039] (1) For normalized images Gaussian smoothing is performed to obtain Calculate the gradient magnitude. and gradient direction : Take the cumulative histogram of gradient magnitudes, and set a high threshold. Set the cumulative probability to 0.75, the corresponding amplitude, and the low threshold. Non-maximum suppression and double thresholding are performed to obtain a binary edge image. , where 1 represents an edge point.

[0040] (2) Edge density and grain size proxy Edge density is the proportion of edge pixels to total pixels. Edge density is positively correlated with graphite particle fineness: the finer the particles, the more edges there are.

[0041] Furthermore, regarding Morphological closing operations are performed to obtain the contours of connected particles, and the area distribution of the contours is statistically analyzed. Let the area of ​​the k-th connected region be... (Number of pixels), then the granularity features include: In the formula, This represents the total number of connected components. This represents the average area.

[0042] (3) Edge histogram gradient direction At the edge point The data is quantized into 9 intervals (each interval is 40°, ranging from 0 to 180°), the pixel frequency in each interval is counted, and the result is normalized to obtain the orientation histogram vector. This feature reflects the degree of directional alignment of graphite particles.

[0043] The final particle edge feature vector is: 2.3 Extraction of Spectral Absorption Peak Features Spectral absorption peaks are key indicators of graphite and its impurities. This example uses normalized spectral curves. Peak detection and quantitative analysis were performed.

[0044] (1) Spectral smoothing and baseline correction To suppress residual noise, a Savitzky-Golay filter (window width 11, polynomial order 3) is used. Smoothing is performed to obtain a smoothed spectrum. Then, the baseline is estimated using the asymmetric least squares baseline correction method. The corrected spectrum is obtained by subtracting the baseline: (2) Absorption peak localization Search Local minima are retained, with depths greater than the threshold. The peak. Assume a total of [number] peaks were detected. The first absorption peak, the... The position of each peak is denoted as Depth is denoted as Half-height and width are denoted as .

[0045] (3) Feature construction Based on the characteristic absorption ranges of graphite and typical impurities (such as pyrite and quartz), the following features are extracted: Graphite main peak characteristics: Absorption peaks were detected in the wavelength range of 600~700nm; if present, the depth was taken. Otherwise, set it to 0.

[0046] Impurity index: Calculate the maximum absorption depth at 400~500nm (iron ion absorption) and 900~1000nm (water or hydroxyl absorption) respectively. and .

[0047] Absorption peak statistics: mean depth of all peaks Depth standard deviation Location variance .

[0048] The characteristic vector of the spectral absorption peak is: 2.4 Feature Fusion Concatenate the three feature vectors end-to-end to obtain the fused feature vector: 3. Quality Prediction Model In obtaining the fused feature vector Subsequently, this embodiment further constructs an improved one-dimensional convolutional neural network (1D-CNN) as a quality prediction module. This module uses... As input, the system uses multi-layer convolution and fully connected mapping to output the predicted carbon content of the current batch of graphite concentrate. Particle size prediction value (median particle size) ) and impurity distribution index .

[0049] The 1D-CNN model consists of an input layer, three one-dimensional convolutional blocks, a global pooling layer, and two fully connected regression heads. Compared with traditional fully connected networks, one-dimensional convolution can capture the local dependencies between different dimensions of the feature vector (such as the co-variation between adjacent features), thus improving prediction accuracy.

[0050] (1) The input layer is a standardized feature vector To accommodate one-dimensional convolution operations, it is reshaped into a 20×1 tensor, i.e., 20 in height and 1 in number of channels.

[0051] (2) Three convolutional blocks are stacked in sequence. Each convolutional block contains a one-dimensional convolutional layer (Conv1D), batch normalization (BN), ReLU activation function and max pooling (MaxPool1D).

[0052] (3) Perform global average pooling (GAP) on the output of convolution block 3 along the length dimension to obtain the mean of each channel: In the formula, This represents the output feature map of the third one-dimensional convolutional block; Then Flattened, it becomes a one-dimensional vector.

[0053] (4) Fully connected regression head Two parallel fully connected regression heads share These are used to predict carbon content and particle size (continuous values), as well as the impurity index (0~1 probability value), respectively.

[0054] Carbon content and granularity regression head: Contains two fully connected layers. The first layer outputs 32 dimensions using the ReLU activation function, and the second layer outputs 1 dimension using linear activation. Impurity Exponential Regression Head: The structure is the same as above, except that a Sigmoid activation function is added after the output layer to restrict the output to [0,1]. In the formula, Indicates the weights of the first layer. This indicates the weights of the second layer. Indicates the first layer bias. This indicates the second layer bias. This represents the Sigmoid activation function.

[0055] Final output quality label .

[0056] 4. Intelligent proportioning calculation This embodiment provides a specific implementation of an intelligent blending calculation module. The module takes the predicted quality labels and inventory levels of each batch of materials in the current inventory as input, and the target product quality specifications set by the user as constraints. It uses an improved Multi-Objective Particle Swarm Optimization (MOPSO) algorithm to solve for the optimal blending ratio.

[0057] 4.1 Optimization Objectives and Constraints (1) Set target product quality specifications Target carbon content: Allowable deviation (This embodiment takes) ); Target granularity: Allowable deviation (This embodiment takes) ); Target impurity index: Allowable deviation (This embodiment takes) ); The planned production weight is ton; The combined index is weighted and averaged. In the formula, Indicates the first The usage of materials in a batch meets the requirements. , Indicates inventory level. This indicates the carbon content after mixing. Indicates the particle size after mixing, Impurity index after mixing.

[0058] (2) Objective function Construct a weighted multi-objective function to minimize the weighted sum of squared deviations between the mixed quality and the target quality: In the formula, , , Let be the weighting coefficient, satisfying In this embodiment, It is 0.7. It is 0.2. The value is 0.1, prioritizing ensuring that the carbon content meets the standard.

[0059] (3) Constraints Inventory constraints: 4.2 Improved Multi-Objective Particle Swarm Optimization Algorithm This invention employs an improved multi-objective particle swarm optimization algorithm to solve the aforementioned constrained nonlinear programming problem. Standard particle swarm optimization is prone to getting trapped in local optima and requires maintaining the Pareto front for multi-objective problems. This embodiment enhances the model's search capability by incorporating dynamic inertia weights, mutation operations, and constraint handling.

[0060] (1) Particle encoding and initialization Let the particle swarm size be Maximum number of iterations The position vector of each particle. This indicates a blending scheme.

[0061] During initialization, for each particle, an initial position that satisfies the inventory constraint and the total production constraint is randomly generated.

[0062] (2) Fitness calculation and Pareto dominance For each particle, first calculate its mixed mass, then calculate the objective function value. At the same time, check the degree of constraint violation. : The fitness function is calculated using the following formula: In the formula, The penalty coefficient is initially set to... As the number of iterations increases exponentially to This forces the particles to gradually enter the feasible region.

[0063] Within a multi-objective optimization framework, each particle maintains its historical best position. (Individual optimality), and the globally optimal position selected through Pareto dominance. (Randomly selected from the non-dominated solution set). The non-dominated solution set is stored in an external archive with a capacity of 50.

[0064] (3) Velocity and position update The formulas for updating particle velocity and position are as follows: In the formula, In this embodiment, the weights are dynamic. , ; , In this embodiment, the learning factor is used as the learning factor. , ; , This represents a uniform random number in the range [0,1].

[0065] (4) Mutation operation To prevent premature convergence, after each iteration, a probability is used to determine the outcome. Apply Gaussian variation to the position of each particle: in, This indicates the mutation intensity; boundary constraints need to be reapplied after mutation.

[0066] (5) At the end of each generation, all non-dominated solutions (Pareto fronts) in the current population are merged with the external archive, and dominated solutions are deleted. If the archive exceeds its capacity, redundant solutions are deleted based on the crowding distance. Subsequently, a solution is randomly selected from the archive as the next generation. (Shared by all particles) to maintain diversity.

[0067] (6) When the maximum number of iterations is reached Or the change in the optimal fitness value over 30 consecutive generations is less than The algorithm terminates when the time is right. The solution with the lowest fitness in the output archive is taken as the final optimal blending scheme. .

[0068] 4.3 Output of blending ratio The blending ratio of each batch of materials is calculated from the optimal solution: Meanwhile, the prediction quality index after mixing is given: Will The corresponding feeder flow rate setting is sent to the automation control module for execution.

[0069] 5. Automated control execution In this embodiment, the automated control execution module is communicatively connected to the intelligent proportioning calculation module and the database module, respectively, to receive the optimal blending ratio scheme and execute the feeding, metering, and mixing operations of each batch of materials accordingly. This module specifically includes: multiple frequency converters, multiple high-precision electronic belt scales, a mixing device, and a programmable logic controller (PLC).

[0070] Each storage unit (i.e., silo) has a corresponding variable frequency feeder installed below its discharge port. The drive motor of each variable frequency feeder is electrically connected to the output terminal of the PLC. The optimal blending ratio scheme output by the intelligent proportioning calculation module includes the usage amount or mass percentage of each batch of materials. After the PLC analyzes the scheme, it calculates the target feeding flow rate of each variable frequency feeder and outputs the corresponding frequency control signal.

[0071] A high-precision electronic belt scale is installed below the discharge end of each variable frequency feeder. The signal output of the belt scale is electrically connected to the analog input of the PLC. During the feeding process, the belt scale detects the instantaneous feeding flow rate in real time and feeds it back to the PLC. The PLC has a built-in proportional-integral-derivative (PID) control algorithm, which compares the actual instantaneous flow rate with the target flow rate in a closed loop, dynamically corrects the operating frequency of the variable frequency feeder, so that the actual output of each batch of material accurately tracks the target value within the set error range (e.g., ±1%).

[0072] Each variable frequency feeder synchronously feeds material to the mixing device according to the aforementioned set ratio. The mixing device is preferably a double-spiral conical mixer, whose inlet is connected to the outlet of each variable frequency feeder via a sealed chute, used to receive and uniformly mix different batches of graphite concentrate. An online monitoring instrument can also be installed at the outlet of the mixing device. This online monitoring instrument is communicatively connected to a PLC or intelligent proportioning calculation module for real-time monitoring of the mixture's quality indicators, including carbon content, particle size, and impurity distribution index.

[0073] When the online monitoring instrument detects that the deviation between the instantaneous quality index of the mixture and the target product quality specification exceeds a preset threshold (e.g., the carbon content deviation is greater than 0.2%), the PLC automatically triggers a correction command based on the deviation signal: on the one hand, it directly makes a small compensation adjustment to the frequency of each variable frequency feeder; on the other hand, it sends the deviation information to the intelligent proportioning calculation module to request the recalculation of the blending ratio for the remaining feeding stage, so as to achieve dynamic closed-loop control.

[0074] Example 2 A graphite flotation product homogenization and blending system based on AI visual recognition, comprising: The online sampling and preprocessing module is used to acquire hyperspectral and grayscale images of graphite concentrate under flowing conditions, and to perform outlier processing, adaptive multi-scale denoising, and normalization on the original images. The feature extraction module is used to extract and fuse gloss features, grain edge features, and spectral absorption peak features from the preprocessed image. The quality prediction module uses a one-dimensional convolutional neural network to predict carbon content, particle size, and impurity index based on the fused feature vector. The intelligent blending calculation module uses a multi-objective particle swarm optimization algorithm to solve for the optimal blending ratio that meets the target quality specifications, based on the predicted quality value and inventory of each batch of materials. The automation control module is used to set the feeder flow rate and execute the mixing ratio according to the optimal blending ratio.

Claims

1. A method for homogenizing and blending graphite flotation products based on AI visual recognition, characterized in that, include: S1. Collect raw data and preprocess the raw data to obtain normalized grayscale image and normalized spectral reflectance curve; S2. Calculate the mean and variance of grayscale values ​​for each block in the normalized grayscale image, and statistically analyze the mean and standard deviation of the gloss index for all blocks to obtain a gloss feature vector. Perform edge detection, morphological closing operation, and gradient direction quantization on the normalized grayscale image to obtain a particle edge feature vector. Smooth, correct the baseline, and locate the absorption peaks on the normalized spectral curve, extracting the graphite main peak depth, impurity absorption depth, and peak statistics to obtain a spectral absorption peak feature vector. Concatenate the gloss feature vector, particle edge feature vector, and spectral absorption peak feature vector to obtain a fused feature vector. S3. Construct a quality prediction model, using the fused feature vector as input, and obtain the predicted carbon content, particle size, and impurity distribution through multi-layer convolution and fully connected mapping. S4. An improved multi-objective particle swarm optimization algorithm is adopted to obtain the optimal mixing ratio by minimizing the weighted sum of squared deviations between the mixed mass and the target mass.

2. The method for homogenizing and blending graphite flotation products based on AI visual recognition according to claim 1, characterized in that, S1 includes: S101. Acquire spectral reflectance curves and grayscale images using a hyperspectral imager and a grayscale camera; S102. An outlier detection and median replacement method based on interquartile range is used to detect and correct outlier pixels in the original grayscale image to obtain the corrected image. S103. A denoising method combining adaptive nonlocal mean filtering and guided filtering based on local noise standard deviation estimation is adopted to perform multi-scale denoising and edge restoration on the corrected image to obtain denoised image data. S104. Normalize the denoised image data and adaptively normalize the spectral reflectance curve data to obtain a normalized grayscale image and a normalized spectral reflectance curve.

3. The method for homogenizing and blending graphite flotation products based on AI visual recognition according to claim 1, characterized in that, The acquisition of gloss characteristics in S2 includes: The normalized grayscale image is divided into non-overlapping image blocks, and the grayscale mean and variance of each block are calculated. A gloss index function is constructed based on the mean and variance; Calculate the mean and standard deviation of the gloss index for all blocks to form a gloss feature vector.

4. The method for homogenizing and blending graphite flotation products based on AI visual recognition according to claim 1, characterized in that, Particle edge feature extraction in S2 includes: Gaussian smoothing is performed on the normalized grayscale image and the gradient magnitude and gradient direction are calculated. Non-maximum suppression and double threshold connection are performed using the high and low thresholds determined by the cumulative histogram to obtain a binary edge image. Edge density is calculated as a positive correlation index of granularity. Morphological closing operations are performed on the edge image to obtain the contours of connected particles. The median and standard deviation of the contour area are then calculated. The gradient direction is quantized into multiple intervals at the edge points, the pixel frequency of each interval is counted, and the normalization is used to obtain the orientation histogram vector. The edge density, median area, standard deviation of area, and orientation histogram are concatenated to form the particle edge feature vector.

5. The method for homogenizing and blending graphite flotation products based on AI visual recognition according to claim 1, characterized in that, The spectral absorption peak characteristics of S2 include: Savitzky-Golay smoothing and asymmetric least squares baseline correction are applied to the normalized spectral curves. Local minima of the corrected spectrum are identified as absorption peaks, and peaks with depths greater than the threshold are retained. The depth of the main graphite peak is extracted within the characteristic absorption range of graphite. The maximum absorption depth is extracted within the iron ion absorption range and the hydroxyl absorption range, respectively. The mean depth, standard deviation of depth, and position variance of all absorption peaks are calculated to form the characteristic vector of spectral absorption peaks.

6. The method for homogenizing and blending graphite flotation products based on AI visual recognition according to claim 1, characterized in that, The quality prediction models in S3 include: The quality prediction model consists of a one-dimensional convolutional neural network, which includes an input layer, three one-dimensional convolutional blocks, a global average pooling layer, and two parallel fully connected regression heads. The input layer reshapes the fused feature vectors into a tensor form that adapts to one-dimensional convolution; Each convolutional block contains, in sequence, a one-dimensional convolutional layer, a batch normalization layer, a ReLU activation function layer, and a max pooling layer; The global average pooling layer averages and flattens the output of the last convolutional block along the length dimension. One regression head outputs continuous predicted values ​​for carbon content and particle size, while the other regression head outputs the impurity index.

7. The method for homogenizing and blending graphite flotation products based on AI visual recognition according to claim 1, characterized in that, The improved multi-objective particle swarm optimization algorithm in S4 includes: Particle position encoding represents the usage of materials in each batch. The fitness function consists of a weighted multi-objective function and a penalty term. The penalty term includes the degree of violation of inventory constraints, total output constraints and non-negativity constraints. The penalty coefficient increases exponentially with the number of iterations. Maintain an external archive to store non-dominated solutions, select the global optimal position based on Pareto dominance, and use dynamic inertia weights that decrease linearly with the number of iterations; Gaussian mutation is applied to the particle position with a set probability; the process terminates when the maximum number of iterations is reached or the fitness value changes less than the threshold for multiple consecutive generations, and the solution with the smallest fitness in the archive is output as the optimal mixing scheme.

8. The method for homogenizing and blending graphite flotation products based on AI visual recognition according to claim 1, characterized in that, The multi-objective functions in S4 include: The multi-objective function is the weighted sum of squares of the deviations between the mixed carbon content, particle size, and impurity index and the target value, with carbon content having the largest weight, followed by particle size, and impurity index having the smallest weight. The target quality specifications include the target carbon content and permissible deviation, the target particle size and permissible deviation, the target impurity index and permissible deviation, and the planned total production weight.

9. A graphite flotation product homogenization and blending system based on AI visual recognition, characterized in that, include: The online sampling and preprocessing module is used to acquire hyperspectral and grayscale images of graphite concentrate under flowing conditions, and to perform outlier processing, adaptive multi-scale denoising, and normalization on the original images. The feature extraction module is used to extract and fuse gloss features, grain edge features, and spectral absorption peak features from the preprocessed image. The quality prediction module uses a one-dimensional convolutional neural network to predict carbon content, particle size, and impurity index based on the fused feature vector. The intelligent blending calculation module uses a multi-objective particle swarm optimization algorithm to solve for the optimal blending ratio that meets the target quality specifications, based on the predicted quality value and inventory of each batch of materials. The automation control module is used to set the feeder flow rate and execute the mixing ratio according to the optimal blending ratio.

Images