A method for on-line multi-component detection of sinter mixture based on hyperspectral imaging

Abstract

The application discloses a kind of based on sinter mixture multi-component online detection method of hyperspectral imaging, first, standardization hyperspectral dataset is constructed, then SG smoothing filtering, multivariate scattering correction and minimum noise separation denoising are executed, eliminate the spectral interference brought by industrial environment;Again, the neural network model for prediction including feature extraction network and inversion network is constructed, the feature extraction network is formed by multiple groups of same unit module, each group of unit module includes multi-scale convolution block and SE channel attention module;The neural network model for prediction can be used for actual sinter composition detection after training.Process of training introduces the joint loss function of multi-task regression loss and maximum mean difference loss to optimize.The application achieves excellent detection accuracy and running stability under cross-date variable industrial conditions, realizes multi-component contactless, global synchronous quantitative detection, adapts to the demand of production line online detection.

Description

Technical Field

[0001] This invention relates to mineral composition detection, specifically to an online detection method for multiple components of sintered ore mixtures based on hyperspectral imaging. This detection method integrates multi-scale attention and domain adaptive transfer, and belongs to the fields of intelligent detection of hyperspectral minerals and machine vision technology in the steel industry. Background Technology

[0002] Iron sinter is a core raw material for blast furnace smelting. The content of seven components in the mixture—TFe, SiO2, CaO, MgO, Al2O3, TiO2, and MnO—directly determines the sintering quality, blast furnace permeability, and energy consumption. Real-time, non-contact online detection is essential for intelligent manufacturing and low-carbon production in steel enterprises. Currently, mainstream detection methods and traditional analytical schemes in the industry have several inherent shortcomings:

[0003] The mainstream testing methods in the industry are: (1) Wet chemical analysis and XRF are both offline detection methods, which require manual sampling, crushing and grinding, and tablet preparation. A single test takes tens of minutes to several hours, and the data is seriously delayed, making it impossible to adjust the sintering process in real time; manual sampling under high temperature and dust conditions poses safety hazards, and uncontrollable human error is introduced into the entire sampling and preparation process. (2) LIBS laser online detection only collects data at a single point and cannot reflect the composition of the entire material; the optical window is easily contaminated in high temperature and high dust environments, and calibration and maintenance are frequent. The core optical components rely on imports and the equipment is expensive; the sensitivity for detecting low-content components such as Mn and Ti is insufficient, and the measurement repeatability is poor when the operating conditions fluctuate.

[0004] The existing traditional hyperspectral detection technology has the following shortcomings: (1) The preprocessing adopts a static algorithm, which is only suitable for single-day and single-batch data. It cannot eliminate the baseline drift and dust mixing noise caused by high-temperature thermal radiation across dates and working conditions, resulting in poor spectral consistency; (2) The unsupervised PCA is used to reduce the dimension and extract features. It relies solely on variance to select bands, which easily loses the sensitive and subtle bands of trace elements. The linear PLS model has weak resistance to matrix effects and spectral line coupling interference, and the prediction accuracy across working conditions drops significantly; (3) The acquisition, preprocessing, feature extraction, and modeling modules are independent of each other. There is no integrated automated production line, and the component anomaly visualization function is lacking, making it difficult to implement online quality inspection on the production line.

[0005] In summary, existing technologies cannot simultaneously meet industrial demands such as real-time performance, full-domain detection, adaptability to complex working conditions, low cost, and fully automated operation. There is an urgent need for a complete hyperspectral intelligent quantitative detection solution that integrates spectrum acquisition and multi-element output.

[0006] In summary, the traditional detection and hyperspectral analysis methods currently used in the field of sinter composition detection all have their own insurmountable technical shortcomings. None of the existing technical solutions can simultaneously meet the core industrial needs. Therefore, it is urgent to develop a standardized intelligent detection method that covers the entire chain from the acquisition of raw hyperspectral images to the synchronous quantitative output of multiple components, to make up for the comprehensive performance defects of existing detection technologies and meet the practical application requirements of intelligent manufacturing and stable production in sintering workshops. Summary of the Invention

[0007] To address the shortcomings of existing technologies, such as data lag, high detection costs, and poor stability, the present invention aims to propose an online detection method for multiple components of sintered ore mixtures based on hyperspectral imaging. This method aims to achieve contactless, synchronous quantitative detection of multiple components across the entire range, with low hardware investment, high automation, and suitability for online detection needs of production lines.

[0008] The technical solution of this invention is implemented as follows:

[0009] A method for online detection of multiple components in sinter mixture based on hyperspectral imaging, comprising the following steps:

[0010] 1) Establishment of hyperspectral dataset: Obtain different sintered ore mixture samples. All sintered ore mixture samples contain all the components to be detected and the content of each component is known. The content of any one component to be detected is different in different batches. Use a hyperspectral camera to collect hyperspectral data of all sintered ore mixture samples.

[0011] 2) Data preprocessing: All hyperspectral data were preprocessed separately, including SG smoothing filtering, MSC multivariate scattering correction and MNF minimum noise separation and denoising performed sequentially; the preprocessed hyperspectral data were then divided into training set and test set;

[0012] 3) Constructing a predictive neural network model: The predictive neural network model includes a feature extraction network and an inversion network; the feature extraction network is composed of multiple sets of structurally identical unit modules connected in series; each set of unit modules includes a multi-scale convolutional block and an SE channel attention module. The multi-scale convolutional block first uses one-dimensional convolutional branches of different sizes to perform parallel feature extraction. The multi-scale features output by each branch are concatenated in the channel dimension. The concatenated mixed features are fused across scales through 1×1 one-dimensional convolution; the fused features are fed into the SE channel attention module for adaptive feature weighting, and then downsampled by max pooling; the original spectral features are sequentially abstracted layer by layer through multiple sets of unit modules to achieve multi-granularity feature abstraction, and finally output a fixed-dimensional deep feature vector and provide it to the inversion network;

[0013] 4) Model training: Input the hyperspectral data corresponding to the training set into the prediction neural network model, continuously train and adjust the parameters of the neural network model using the hyperspectral data corresponding to the training set, establish the mapping relationship between the hyperspectral data and the content of the component to be detected, and verify the model performance through the test set. The trained prediction neural network model is the component prediction model.

[0014] 5) Use a hyperspectral camera to collect hyperspectral data of the sinter mixture sample to be tested and preprocess it according to step 2). Then input it into the component prediction model trained in step 4). The component prediction model outputs the content of the component to be detected in the sinter mixture.

[0015] Further, in step 2), the SG smoothing filter is specifically as follows: assuming a single original spectral vector... Window size 2m + 1, polynomial order p, local wavelength coordinates within the window ; Window Center Point smoothing value :

[0016]

[0017] In the formula, The convolution weight coefficients at position t within the filtering window, with a local offset of t. x i+t is the original spectral reflectance value at a position offset t with the i-th wavelength sampling point as the window center.

[0018] Furthermore, the MSC multivariate scattering correction is specifically as follows:

[0019] First, calculate the average reference spectrum of all samples using the following formula. This serves as a unified benchmark:

[0020]

[0021] In the formula, M is the total number of samples; This is the spectrum of the i-th original sample;

[0022] Subsequently, the spectrum of a single original sample was analyzed. A univariate linear regression model is constructed, and the unique scattering coefficient and baseline offset for each sample are solved by least squares fitting using the following formula:

[0023]

[0024] In the formula, This is a multiplicative scattering shift; This is the baseline offset;

[0025] Finally, the following correction formula is used to complete the scattering compensation, eliminate the spectral amplitude differences between samples caused by physical morphology, unify the spectra of all samples to the same baseline scale, and significantly reduce the masking effect of scattering interference on the mineral absorption characteristics:

[0026] MSC .

[0027] Further, in step 2), the MNF minimum noise separation and denoising is specifically as follows:

[0028] Let the original spectral matrix after correction be X ∈ M×N, where M is the total number of samples and N is the total number of spectral bands; firstly, the mixed spectrum is decomposed into two parts: pure useful signal and noise, satisfying the decomposition relationship:

[0029]

[0030] In the formula, S represents the effective signal matrix corresponding to the mixture components, and Γ represents the noise matrix caused by the environment and the acquisition equipment; the covariance matrices of the signal and noise are solved separately.

[0031]

[0032] The discreteness of the data distribution of the quantized signal and noise is then used to construct the generalized eigenvalue solution equation:

[0033]

[0034] In the formula For eigenvalues, For each corresponding eigenvector, the magnitude of the eigenvalue represents the energy ratio of the signal to the noise in the corresponding direction; all eigenvectors are sorted according to their eigenvalues. Sort by largest to smallest, and filter by the top results whose cumulative variance percentage is not less than a set value. The group of feature vectors constitutes the projection matrix ;

[0035] Finally, spectral dimensionality reduction is achieved using the following projection transformation formula:

[0036]

[0037] This transformation completes noise whitening, filters out invalid bands with low signal-to-noise ratios and excessive noise, and significantly compresses data dimensionality while preserving the effective information of mineral spectra to the greatest extent.

[0038] Further, in step 3), before the fused features are fed into the SE channel attention module, batch normalization and ReLU nonlinear activation are performed sequentially, as shown in the following expressions:

[0039]

[0040] in, To fuse features; batch normalization standardizes the feature distribution, alleviating the gradient vanishing problem during deep network training; ReLU nonlinear activation is used to mine complex coupling correlations between spectral bands; the result is the final output feature of the multi-scale convolutional block.

[0041] Furthermore, in step 3), the SE channel attention module is used to filter and denoise the mixed channel features output by the multi-scale convolutional blocks. The specific processing procedure is as follows:

[0042] Compression: For each feature channel output by the multi-scale convolutional block, global average pooling is performed along the spectral wavelength dimension to compress the one-dimensional channel sequence of length L into a single scalar value. This scalar value represents the global spectral information of the corresponding channel, realizing the aggregation of information across the entire band.

[0043] Incentive: The compressed global features are fed into two fully connected layers to learn the channel importance weights: The first fully connected layer performs dimensionality reduction compression with an adjustable compression rate on the number of channels, and uses ReLU activation to mine the dependencies between channels; The second fully connected layer restores the original number of channels, i.e. the number of bands, and maps the output to the [0,1] interval through the Sigmoid activation function to obtain the channel-specific importance weight coefficients.

[0044] Feature weighting: The importance weight coefficient of each channel is multiplied with the original multi-scale feature map channel by channel. High weights are assigned to effective channels containing mineral-sensitive absorption bands to amplify the feature response; low weights are assigned to noisy and irrelevant band channels to suppress redundant interference, thus completing feature screening and noise reduction.

[0045] Further, in step 3), the inversion network as a whole adopts a serially stacked architecture of multiple sets of modular basic units with completely unified structure; within a single basic unit, a linear fully connected layer, a BN batch normalization layer, a LeakyReLU activation layer, and a Dropout regularization layer are sequentially connected in series; among them, the linear fully connected layer completes the feature dimension mapping; the BN batch normalization layer constrains the feature distribution range within the unit, weakens the feature offset caused by spectral baseline drift and scattering noise, and accelerates model convergence; the LeakyReLU activation layer introduces nonlinear fitting capability into the network, avoids neuron death defects, and adapts to the complex nonlinear mapping relationship between the spectrum and mineral components; along the feature forward propagation direction, the feature channel dimension and Dropout drop probability of each basic unit decrease step by step; some neuron outputs are randomly shielded to suppress the overfitting memory of training set noise and irrelevant band features in a hierarchical manner.

[0046] Furthermore, in step 4), the prediction error is constrained by introducing SmoothL1 multi-task regression loss during the training process of the neural network model, and the maximum mean difference loss is used to align the spectral feature distributions of different batches to construct a weighted joint loss function optimization model.

[0047] SmoothL1 Multi-task Regression Loss L task Calculate using the following formula:

[0048]

[0049] In the formula: M is the total number of samples; P is the number of components to be detected; For the first The first sample The true content of each component; For the first The first sample Predicted values ​​for each component;

[0050] Maximum mean difference loss L MMD Calculate using the following formula:

[0051]

[0052] Gaussian kernel function in the formula ; These represent the number of samples in the source domain and the target domain, respectively. For the corresponding domain depth features; For core bandwidth;

[0053] The total loss Li of the prediction neural network model is obtained by weighted fusion of the SmoothL1 multi-task regression loss and the maximum mean difference loss. total As shown in the following formula:

[0054]

[0055] In the formula: This is the maximum mean difference loss weighting coefficient, used to balance the component prediction error with the alignment strength of the feature distribution.

[0056] Furthermore, in step 4), when training the neural network model for prediction, prior knowledge of the spectral physics of sintered ore is embedded into the network loss function to complete the decoupling of overlapping spectral lines of multiple elements; corresponding to the prior loss function... as follows:

[0057]

[0058] in, The known standard characteristic peak position for the k-th mineral is... This refers to the peak response learned by the network. These are the weighting coefficients for each prior constraint;

[0059] The total loss L predicted using the neural network model is then... total As shown in the following formula:

[0060]

[0061] In the formula: This is the weighting coefficient for the mineral spectrum prior constraint loss, which adjusts the constraint strength of the physical prior constraint in the overall training.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] 1. A progressive multi-level composite adaptive spectral preprocessing method is constructed to effectively eliminate nonlinear baseline drift and mixed noise caused by cross-date and cross-operating conditions, and solve the technical problem of poor spectral consistency across batches in traditional static preprocessing algorithms.

[0064] 2. A parallel multi-scale one-dimensional convolutional neural network with SE channel attention mechanism was designed to overcome the shortcomings of traditional unsupervised dimensionality reduction, which easily loses sensitive bands of trace elements and linear models, which have weak resistance to spectral line coupling interference.

[0065] 3. A multimodal fusion network was constructed, and the maximum mean difference loss (MMD) was introduced to build a domain-adaptive transfer learning framework. Combined with a regression head, this effectively suppressed the interference of mineral matrix effects, significantly improving the model's cross-domain generalization ability and the accuracy of simultaneous prediction of multiple components under varying industrial conditions. Simultaneously, prior knowledge of the spectral physics of sintered ore was embedded into the network loss function to decouple overlapping spectral lines of multiple elements from a spectral mechanism perspective, reducing matrix effect interference.

[0066] This invention organically combines a three-level progressive composite correction pipeline (SG+MSC+MNF), a multi-scale SE single-attention lightweight network adapted to one-dimensional spectroscopy, a triple-weighted joint loss combining fusion domain adaptation and mineral mechanism constraints, and a progressively decaying canonical multi-layer inversion network to form an integrated online detection solution for variable industrial sintering production lines. Each component works in concert to overcome the shortcomings of multiple coupled technologies, such as baseline drift across date conditions, insufficient sensitivity for trace component detection, poor generalization across batches, model overfitting, and inference delays that cannot meet the requirements for real-time production line detection. This achieves contactless, continuous, real-time, full-domain, high-precision synchronous online detection of multiple components in sintering production lines.

[0067] Experiments show that, on a multi-batch hyperspectral test set of sintered ore spanning multiple days, the average absolute errors of the five components SiO2, CaO, MgO, Al2O3, and TiO2 all meet the industrial static accuracy requirements. TFe and MnO also basically meet the industrial static accuracy requirements. The overall performance is better than the traditional non-transfer, equal-weight single-output convolution model, verifying its effectiveness and advancement in non-destructive rapid detection scenarios at the sintering site. Attached Figure Description

[0068] Figure 1 This is a flowchart of the quantitative detection process for sintered ore mixture components based on hyperspectral imaging, according to the present invention.

[0069] Figure 2 This is a diagram of the multi-scale convolutional attention spectral feature extraction network framework of the present invention.

[0070] Figure 3 This is a diagram of the multi-layer inversion network framework of the present invention.

[0071] Figure 4 This is a schematic diagram of the data acquisition device of the present invention.

[0072] Figure 5 This is a two-dimensional image of sinter sample #1 taken with the hyperspectral camera of this invention. Detailed Implementation

[0073] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0074] This invention provides a complete standardized end-to-end solution for building a dataset, including data preprocessing, feature extraction, and inversion modeling. This solution enables the entire process from spectral acquisition and spectral denoising to mechanistic constraint feature extraction, multimodal learning modeling, multi-element quantitative output, and visualization analysis. The overall workflow includes the following steps, which can be found in [reference 1]. Figure 1 .

[0075] Step 1: Establish a hyperspectral dataset for sintered ore.

[0076] The acquisition hardware includes a 391.8nm~1003.9nm visible-near-infrared hyperspectral camera, a halogen broadband light source, and a precision optical support. The spectral range fully covers the characteristic absorption ranges of the seven mineral components. A single acquisition simultaneously outputs the two-dimensional morphology of the sinter and continuous spectral data. Sinter entity samples are acquired, and raw data is collected simultaneously: offline analysis of ground truth labels (mass fractions of the seven elements). Wavelength alignment, outlier removal, and structured Excel processing are uniformly completed, ultimately constructing a standardized, labeled dataset containing 20 sets of standard sinter and 200 valid sample points per acquisition, spanning various working conditions. This dataset is divided into training and testing sets for model pre-training and performance comparison verification, respectively. The complete dataset can be read in batches using the accompanying Python code.

[0077] Step 2: Spectral data preprocessing:

[0078] To address the issues of baseline drift, mixed noise, and inconsistent spectral amplitude caused by high-temperature thermal radiation, dust scattering, and material thickness differences in industrial settings, this invention constructs a progressive multi-stage composite correction pipeline. The pipeline sequentially performs Savitzky-Gorye smoothing filter (SG), multivariate scattering correction (MSC), and minimum noise separation transform (MNF) on the original spectrum, outputting a standardized spectrum with high signal-to-noise ratio and consistent baseline across batches.

[0079] (1) SG smoothing filter

[0080] Under normal operating conditions, spectral data contains high-frequency random noise caused by dust. SG smoothing filtering can remove this noise, fully preserving the characteristic absorption peaks of minerals without causing blurring or distortion. Let a single original spectral vector... Window size 2m + 1, polynomial order p, local wavelength coordinates within the window Window Center Point smoothing value:

[0081]

[0082] In the formula, Within the filtering window, the convolution weight coefficient corresponding to the local offset at position t is called the SG convolution weight, which is obtained by the least squares fitting polynomial. Its core function is to smooth high-frequency dust noise, preserve the curvature of mineral characteristic peaks, avoid simple mean filtering from smoothing out absorption valleys, and provide a high-fidelity basic spectrum for subsequent scattering correction and baseline correction. x i+t is the original spectral reflectance (absorbance) value at position t, with the i-th wavelength sampling point as the window center.

[0083] (2) MSC multivariate scattering correction

[0084] The sinter samples exhibit natural differences in particle size, thickness, and surface roughness, resulting in varying degrees of scattering effects when light irradiates the material, introducing multiplicative scattering shifts in the spectrum. Additive baseline offset This directly results in inconsistent spectral baselines and significant amplitude differences among similar minerals, severely interfering with subsequent feature extraction and component inversion. This invention corrects baseline shifts caused by particle scattering using MSC multivariate scattering correction, eliminating interference from particle size and packing thickness. First, the average reference spectrum of all samples is calculated. This serves as a unified benchmark:

[0085]

[0086] In the formula, M represents the total number of samples. This is the spectrum of the i-th original sample. Subsequently, the spectra of a single original sample are... A univariate linear regression model is constructed, and the unique scattering coefficient and baseline offset for each sample are solved by least squares fitting:

[0087]

[0088] In the formula, This is a multiplicative scattering shift; This is the baseline offset;

[0089] Finally, according to the correction formula, scattering compensation is completed to eliminate the spectral amplitude differences between samples caused by physical morphology, unifying the spectra of all samples to the same baseline scale, and significantly reducing the masking effect of scattering interference on mineral absorption characteristics.

[0090] MSC

[0091] (3) Minimum Noise Separation and Denoising in MNF

[0092] Despite the previous processing, the spectrum still contains a large number of low signal-to-noise ratio redundant bands. Stray noise from the industrial environment remains in each channel of the spectrum, directly increasing data dimensionality redundancy and interfering with the extraction of effective mineral signals. Therefore, the Minimum Noise Separation Transform (MNF) is introduced to achieve noise whitening and feature dimensionality reduction.

[0093] Let the original spectral matrix after correction be X ∈ M×N, where M is the total number of samples and N is the total number of spectral bands; firstly, the mixed spectrum is decomposed into two parts: pure useful signal and noise, satisfying the decomposition relationship:

[0094]

[0095] In the formula, S represents the effective signal matrix corresponding to the mixture components, and Γ represents the noise matrix caused by the environment and the acquisition equipment; the covariance matrices of the signal and noise are solved separately.

[0096]

[0097] The discreteness of the data distribution of the quantized signal and noise is then used to construct the generalized eigenvalue solution equation:

[0098]

[0099] In the formula For eigenvalues, For each corresponding eigenvector, the magnitude of the eigenvalue represents the energy ratio of the signal to the noise in the corresponding direction; all eigenvectors are sorted according to their eigenvalues. Sort by largest to smallest, and filter by the top [percentages] with a cumulative variance percentage of at least 95%. The group of feature vectors constitutes the projection matrix ;

[0100] Finally, spectral dimensionality reduction is achieved using the projection transformation formula:

[0101]

[0102] This transformation completes noise whitening, filters out invalid bands with low signal-to-noise ratio and excessive noise, and significantly compresses data dimensionality while preserving the effective information of mineral spectra to the greatest extent, thereby reducing the computational overhead and noise interference in subsequent baseline correction and feature extraction stages.

[0103] Sintered minerals suffer from baseline drift interference caused by factors such as uneven particle size, fluctuating material thickness, and dust scattering. Wavelet transform only performs frequency domain noise reduction and cannot address the spectral amplitude shift and baseline drift issues caused by particle scattering. This invention utilizes SG to adapt to high-frequency random noise in industrial environments, MSC to specifically eliminate baseline shifts caused by particle scattering, and MNF to achieve noise whitening, low signal-to-noise ratio band removal, and data dimensionality reduction. The combination of these three technologies closely matches the multi-coupled interference conditions in sintering environments. Through the above preprocessing, the spectral baseline of samples across dates and batches is unified, solving the problem of poor spectral consistency across operating conditions in traditional preprocessing. It can completely preserve the trace absorption characteristic peaks of minerals, unlike simple filtering which smooths out weak component features. It pre-compresses the spectral dimension, reducing the computational load of the backend neural network, while filtering out stray noise from equipment and the environment, thus improving the signal-to-noise ratio of the input data.

[0104] airPLS adaptive baseline correction eliminates the need for manual calibration of baseline control points and automatically fits to eliminate spectral nonlinear baseline drift caused by high-temperature thermal radiation at the sintering site; when processing multiple batches of data, it reuses the training MSC and MNF reference parameters to eliminate date-related systematic spectral shifts.

[0105] Step 3: Construct a neural network model for prediction:

[0106] The predictive neural network model includes a feature extraction network and an inversion network.

[0107] Feature extraction network (see also) Figure 2 To address the shortcomings of traditional PCA unsupervised dimensionality reduction, such as highly overlapping multi-element spectral lines, weak trace component features, and the loss of sensitive bands, this invention constructs a multi-scale 1D-CNN and channel attention deep feature extraction network to replace the traditional linear dimensionality reduction scheme. The network structure and training logic are fully implemented in the accompanying code. The feature extraction network is constructed by serially stacking multiple sets (four sets in this example) of structurally identical units. Within each set of units, parallel feature extraction is first performed using one-dimensional convolutions of different sizes. After channel concatenation of the output features from each branch, cross-scale feature fusion is achieved through 1×1 one-dimensional convolution. The fused features are then fed into the SE channel attention module for adaptive band weighting, and finally downsampling is performed through max pooling. The original spectral features are sequentially abstracted layer by layer through four sets of progressive units to achieve multi-granularity feature abstraction, ultimately outputting a fixed-dimensional deep feature vector, which is then supplied to the downstream inversion network.

[0108] (1) Multi-scale convolutional blocks

[0109] To address the issue of sintered ore spectra exhibiting three characteristics simultaneously: narrow-band trace absorption valleys, continuous band trends within a given range, and global baseline shift, a parallel multi-branch convolutional structure is designed to replace single-size convolutional kernels. Small convolutional kernels have small receptive fields, capturing local narrow-band absorption details corresponding to trace mineral components and preserving weak, fine-grained features. Medium-sized convolutional kernels have medium receptive fields, extracting correlation and variation features across a continuous spectral range and characterizing inter-band coupling. Large convolutional kernels have large receptive fields, capturing long-distance dependent features such as global baseline drift and wide-band mineral absorption bands. The multi-scale features output from the three branches are concatenated along the channel dimension to obtain a hybrid feature map integrating fine, medium, and coarse-grained information. This is then fed into a convolutional layer to achieve cross-channel information fusion and channel dimensionality reduction. This approach enables the model to autonomously learn the contribution weights of features at different scales while compressing the channel dimension, controlling model parameters and computational overhead, and avoiding channel redundancy caused by multi-branch concatenation.

[0110]

[0111] After concatenating the multi-scale features obtained from three sets of convolutional branches of different sizes (3, 7, and 11), the number of feature channels increases to three times that of a single branch, resulting in channel redundancy. To address this, a 1×1 one-dimensional convolution is introduced to perform cross-scale information fusion and channel dimensionality reduction. A fusion convolution kernel is defined, and the concatenated features are transformed using convolution to obtain features that fuse multi-granularity spectral information.

[0112]

[0113] This convolutional layer adaptively learns the weight allocation of branch features at different scales, achieving deep fusion of local, interval, and global spectral features. On the other hand, it compresses the number of channels, reducing the number of model parameters and computational overhead, and alleviating the dimensionality explosion problem caused by multi-branch splicing.

[0114] To improve network training stability and introduce non-linear expressive power, batch normalization (BN) and ReLU non-linear activation are sequentially performed on the fused features:

[0115]

[0116] Batch normalization standardizes the feature distribution, alleviating the gradient vanishing problem during deep network training; the ReLU activation function introduces a nonlinear transformation to uncover complex coupling relationships between spectral bands. The resulting output features of the multi-scale convolutional block are then fed into subsequent modules for feature selection and noise reduction.

[0117] (2) SE Channel Attention Module

[0118] Multi-scale convolutional blocks output a large number of mixed channel features, including sensitive band channels strongly correlated with components, invalid channels such as instrument noise and dust scattering interference. Squeeze-and-Excitation (SE) channel attention is introduced to achieve adaptive feature recalibration. The complete process consists of three steps: compression, excitation, and feature weighting.

[0119] For each feature channel output by the multi-scale module, global average pooling is performed along the spectral wavelength dimension to compress a one-dimensional channel sequence of length L into a single scalar value. This scalar represents the global spectral information of the corresponding channel, thus achieving full-band information aggregation.

[0120] The compressed global features are fed into two fully connected layers to learn the channel importance weights: the first fully connected layer performs dimensionality reduction compression with an adjustable compression rate on the number of channels, and uses ReLU activation to mine the dependencies between channels; the second fully connected layer restores the original number of channels, i.e. the number of bands, and maps the output to the [0,1] interval through Sigmoid activation to obtain the channel-specific importance weight coefficients.

[0121] The learned one-dimensional weight vector is multiplied with the original multi-scale feature map channel by channel: high weights are assigned to effective channels containing mineral-sensitive absorption bands to amplify the feature response; low weights close to 0 are assigned to channels with noise and irrelevant bands to suppress redundant interference, thus completing feature selection and noise reduction.

[0122] Inversion network (see) Figure 3The input is the depth spectral features output by the front-end convolutional network. The overall architecture employs a multi-group (four groups shown in the diagram) modular basic unit stacked serially, with a completely unified structure. Within each basic unit, a linear fully connected layer, a batch normalization (BN) layer, a LeakyReLU activation layer, and a Dropout regularization layer are sequentially connected. The linear fully connected layer maps the feature dimensions; the BN layer constrains the feature distribution range within the unit, reducing feature shift caused by spectral baseline drift and scattering noise, thus accelerating model convergence; the LeakyReLU activation function introduces nonlinear fitting capability into the network, avoiding neuron death defects and adapting to the complex nonlinear mapping relationship between spectra and mineral components; along the feature forward propagation direction, the feature channel dimension and Dropout probability of each basic unit gradually decrease; some neuron outputs are randomly masked, layer-wise suppressing overfitting of the model to training set noise and irrelevant band features. After stacking, an independent 7-dimensional output head is connected, simultaneously outputting the predicted values ​​of the content of seven mineral components. This invention uses a gradually decreasing Dropout strategy, retaining more basic spectral features in the shallow layer and suppressing overfitting in the deep layer, adapting to the characteristics of large noise fluctuations in industrial samples.

[0123] Step 4: Model Training

[0124] The hyperspectral data corresponding to the training set is input into the prediction neural network model. The prediction neural network model is continuously trained using the hyperspectral data corresponding to the training set to establish a mapping relationship between the hyperspectral data and the content of the components to be detected. During the training process, independent validation sets and test sets are divided simultaneously. The generalization performance of the model is monitored in real time using the validation set. Finally, the overall prediction accuracy of the model is quantitatively evaluated through the test set. The complete network after training convergence is the sintered ore multi-component quantitative component prediction model.

[0125] During training, SmoothL1 multi-task regression loss is introduced to constrain prediction error. Gaussian kernel MMD loss (maximum mean difference loss) is used to align the spectral feature distributions of different batches, and a weighted joint loss function is constructed to optimize the model.

[0126] For simultaneous regression of multiple component indicators, the more robust SmoothL1 loss is used to mitigate interference from anomalous spectral samples; the following formula is the calculation formula for the multi-task regression loss of 7 components:

[0127]

[0128] In the formula, M is the total sample size; For the first The first sample The true content of each component; This represents the predicted value corresponding to the MLP network.

[0129] The content of the seven components in the sintered ore varies greatly (TFe is the main component and MnO / TiO2 is the trace component). MSE is not robust to abnormal spectral samples. SmoothL1 alleviates the gradient explosion of extreme outlier samples and balances the fitting weights of the main and secondary components.

[0130] To eliminate the domain distribution shift present in multi-day spectral acquisitions, a maximum mean difference loss (MMD) is constructed in the deep feature space output by the front-end feature extraction network. This reduces the feature distribution distance between the source domain standard samples and the target domain industrial samples, enabling cross-batch spectral distribution adaptation and reducing the dependence on field-labeled samples. The Gaussian kernel MMD loss expression is as follows:

[0131]

[0132] Gaussian kernel function in the formula ; These represent the number of samples in the source domain and the target domain, respectively. For the corresponding domain depth features; This represents the bandwidth of the Gaussian kernel.

[0133] The high overlap of absorption peaks in multi-element spectra makes it easy for pure data-driven models to confuse the characteristics of different minerals. This invention embeds prior knowledge of the spectral physics of sintered minerals into the network loss function, thereby decoupling overlapping multi-element spectral lines from the spectral mechanism level and reducing matrix effect interference. The corresponding prior loss function... as follows:

[0134]

[0135] in, The known standard characteristic peak position for the k-th mineral is... This refers to the peak response learned by the network. These are the weighting coefficients for each prior constraint.

[0136] We weightedly combine multi-task regression loss, maximum mean difference loss, and prior loss to simultaneously optimize inversion accuracy and cross-batch generalization ability. The total loss is as follows:

[0137]

[0138] In the formula: This is the weighting coefficient for the maximum mean difference loss, used to balance the component prediction error with the alignment strength of the feature distribution; This is the weighting coefficient for the mineral spectrum prior constraint loss, which adjusts the constraint strength of the physical prior constraint in the overall training.

[0139] The Maximum Mean Difference Loss (MMD) is used to reduce the data distribution deviation between laboratory source domain samples and industrial target domain samples in the feature space, requiring only a small number of field-labeled samples for model fine-tuning. The dynamic weight allocation mechanism differentiates the feature contribution weights for high-content principal components (such as TFe and SiO2) and low-content trace components (such as TiO2 and MnO), offsetting the interference of mineral matrix effects on prediction accuracy. It balances the fitting accuracy of high-content principal components and low-content trace components, solving the problem of large prediction errors for trace components in traditional models. Relying on the physical laws of mineral spectra to constrain network training reduces prediction bias caused by spectral coupling and matrix effects; the average absolute errors for SiO2, CaO, and TiO2 components are far below industry accuracy limits.

[0140] The complete training process includes: inputting the spectral data of the training set into the model, passing through a multi-scale feature extraction module, a channel attention module, a global pooling and feature enhancement layer, and a regression prediction head, and outputting the predicted chemical composition content; subsequently, calculating the weighted total loss between the predicted components and the actual test labels. The system performs backpropagation to update the entire network parameters with the goal of minimizing the loss. The Adam optimizer is used to update the model parameters based on the loss value, with an initial learning rate of 0.001 and weight decay applied. An adaptive learning rate decay strategy is added to continuously monitor the validation set loss. If the validation loss does not decrease within a set number of training rounds, the current learning rate is halved. In the early stages, a large learning rate is used to quickly converge the global parameters, while in the later stages, a small learning rate is used to finely fit the weak mineral feature absorption peaks. An early stopping optimal weight preservation strategy is added: both learning rate decay and early stopping are triggered by the validation set loss not decreasing for multiple consecutive rounds. When a threshold is met, the training process is terminated, and only the network weights with the minimum global validation set loss are preserved, avoiding overfitting in the later stages of training.

[0141] Step 5: The component prediction model obtained from the training can then be used for actual sinter component detection.

[0142] The entire algorithm of this invention is developed modularly based on the PyTorch framework. The feature extraction network can independently export weight files to achieve offline batch feature extraction, and supports offline loading of models to complete batch inference and result export of multiple batches of sintered ore samples.

[0143] In this embodiment, the hyperspectral data acquisition of the present invention uses a commercial visible-near-infrared hyperspectral camera and an adjustable halogen broadband light source, equipped with a dustproof, light-shielding and heat-insulating shell, without the need for imported laser optical components, and the hardware construction cost is far lower than that of LIBS online detection equipment.

[0144] The entire process of reasoning in this invention, from inputting the original hyperspectral image to predicting the content of seven components, calculating accuracy indicators, and marking abnormal areas, is much shorter than offline detection. It can achieve contactless, continuous, online, real-time component detection in sintering production lines.

[0145] To better illustrate the technical solution and effects of this application, an embodiment is given below.

[0146] This invention is based on a mobile processor equipped with an NVIDIA GeForce RTX 4090 Laptop GPU (16GB GDDR6 dedicated video memory). The software environment consists of a Windows operating system, the PyTorch deep learning framework, and supporting scientific computing libraries such as NumPy and SciPy. All models are trained and tested in this environment, and the parallel computing capabilities of the GPU are utilized for acceleration. Actual sintered ore samples are used in experiments to verify the algorithm's performance.

[0147] Construction of hyperspectral dataset

[0148] First, a dedicated hardware platform for hyperspectral acquisition of sintered ore was built. A hyperspectral camera with a wavelength range of 391.8 nm to 1003.9 nm was used to perform line scanning of the sintered ore samples, generating a three-dimensional hyperspectral data cube. Two-dimensional spatial images and one-dimensional continuous spectral information of the sintered ore samples were exported. Wavelength alignment was performed on the original hyperspectral data, outlier samples caused by camera shake or sensor malfunctions were removed, and resampling was performed. All data and labels were then organized into a structured Excel spreadsheet. Finally, a standardized sintered ore spectral dataset containing multiple batches and spanning multiple dates was constructed. Table 1 shows the ground truth dataset of the samples in this embodiment. Figure 4 This is a schematic diagram of a data acquisition device. Figure 5 This is a two-dimensional image of sinter sample #1 taken with a hyperspectral camera.

[0149] Table 1 Sample True Value Dataset

[0150] 1# 1601 55.17 5.42 10.88 2.76 2.42 0.17 0.34 2# 1602 55.22 5.61 10.74 2.77 2.45 0.13 0.36 3# 1603 54.45 5.64 11.58 2.52 2.72 0.12 0.34 4# 1600 53.81 5.97 12.09 2.63 2.84 0.16 0.5 5# 1664 54.82 5.65 11.31 2.22 2.61 0.19 0.3 6# 1665 54.41 5.48 11.63 2.76 2.68 0.17 0.38 7# 1667 55.98 5.76 12.03 2.8 2.78 0.14 0.44 8# 1685 53.55 5.85 11.25 2.59 2.64 0.16 0.31 9# 1633 54.71 5.54 11.32 2.63 2.54 0.22 0.61 10# 1652 55.06 5.48 11.23 2.52 2.61 0.21 0.47 11# 1635 53.74 6.07 12.07 2.61 2.74 0.16 0.41 12# 1470 54.88 5.51 11.27 2.75 2.54 0.17 0.44 13# 1469 54.4 5.51 11.4 2.68 2.67 0.14 0.43 14# 1641 54.72 5.75 11.34 2.88 2.57 0.18 0.45 15# 1640 54.79 5.74 11.09 2.64 2.56 0.19 0.58 16# 1634 55.1 5.66 10.95 2.75 2.49 0.16 0.38 17# 1643 54.5 5.59 11.31 2.76 2.58 0.17 0.33 18# 1642 55.11 5.56 10.93 2.83 2.43 0.18 0.41 19# 1653 54.74 5.74 11.32 2.66 2.62 0.14 0.33 20# 1652 54.75 5.81 11.11 2.9 2.55 0.2 0.43

[0151] Data preprocessing

[0152] Data preprocessing is used to perform progressive multi-level composite correction on the input spectrum to eliminate various noises and interferences introduced by the industrial environment, filter out high-frequency random noise caused by dust scattering, and at the same time fully preserve the characteristic absorption peaks of minerals.

[0153] First, perform SG smoothing filtering: the program sets the window size to... (Right now =5), polynomial order =3. For each wavelength point within the window. Its smoothed value is obtained by least-squares fitting of the convolution weights. calculate:

[0154]

[0155] Then, MSC multivariate scattering correction is performed: First, the average spectrum of all samples in the training set is calculated as a unified benchmark. Then, univariate linear regression is performed on each spectrum. Finally, correction is performed to unify the spectra of all samples to the same baseline scale, thereby achieving noise whitening and effective information dimensionality reduction.

[0156] Finally, minimum noise separation and denoising are performed using MNF: the corrected spectral matrix X ∈ The M×N problem (where M is the number of samples and N is the number of bands) is decomposed into signal and noise components, and the generalized eigenvalue problem is solved. The eigenvalues ​​are sorted from largest to smallest, and the top eigenvalues ​​with a cumulative variance contribution rate ≥ 95% are selected. The eigenvectors form a projection matrix, which ultimately transforms the original... 3D spectral projection to =64-dimensional low-dimensional feature space This significantly compresses data dimensions and suppresses noise.

[0157] For subsequent batches of data, the program can reuse the trained MSC reference spectrum and MNF projection matrix to ensure consistency of spectral data across dates and eliminate systematic biases.

[0158] Feature extraction

[0159] A mechanistic-constrained multi-scale one-dimensional convolutional neural network (1D-CNN) with channel attention mechanism is constructed to convert preprocessed spectra into high-resolution deep features. To simultaneously capture the narrow absorption valleys, interval continuity trends, and global contours of trace elements, a parallel three-branch convolutional structure is designed: Branch 1, with a kernel size of 3 and a small receptive field, is used to extract local subtle absorption features; Branch 2, with a kernel size of 7 and a medium receptive field, is used to extract interval correlation variation features; Branch 3, with a kernel size of 11 and a large receptive field, is used to extract long-range global dependency features. The multi-scale features output from the three branches are concatenated along the channel dimension to obtain a hybrid feature map. Subsequently, a 1×1 convolution is used for cross-channel information fusion and channel dimensionality reduction, compressing the number of channels to the original level. The map is then sequentially processed through batch normalization (BN) and ReLU activation functions to obtain features that fuse multi-granularity information. .

[0160] To emphasize the importance of adaptive recalibration of feature channels, the program... Perform a Squeeze operation: global average pooling is performed along the spectral dimension, compressing each channel into a scalar. Then, perform an Excitation operation: this scalar vector is input into a two-layer cascaded fully connected network (with a compression ratio of 4 in the middle layer), and a channel weight vector is generated using the Sigmoid activation function. Finally, the original features With weight vector Multiply by each channel to obtain the weighted features. This enables effective amplification of mineral-sensitive bands and suppression of interference bands.

[0161] During network training, this invention uses the characteristic peak positions and spectral broadening patterns of minerals such as iron oxides (Fe-O, characteristic peak positions 500nm~700nm) and silicates (Si-O, characteristic peak positions 900nm~1000nm) as prior knowledge, embedding them into the network loss function in the form of a regularization term:

[0162]

[0163] in The known standard characteristic peak position for the k-th mineral is... This refers to the peak response learned by the network. These are the weight coefficients for each prior constraint. This constraint guides the network to focus on physically reasonable spectral line positions during feature extraction, reducing feature confusion caused by overlapping spectral lines of different elements and achieving effective decoupling of overlapping multi-element spectral lines. Finally, global average pooling compresses the SE-weighted feature map along the spectral dimension into a fixed-length 256-dimensional depth feature vector for use by downstream multi-task regression networks. The feature extraction network can independently save weight files and supports offline batch extraction of sample features.

[0164] This invention employs a multi-scale one-dimensional convolutional neural network with parallel computation using convolutional kernels of various sizes. Shallow convolutions extract subtle absorption peak features of minerals in local spectra, while deep convolutions extract global spectral contour features. The channel attention mechanism autonomously learns the sensitivity weights of each band to the seven components, amplifying the characteristic bands of minerals and suppressing the interference bands of dust and thermal radiation. The physical priors of mineral spectra are used to construct constraint loss based on the characteristic peak positions of iron oxides and silicates and the spectral broadening patterns, thereby reducing feature confusion caused by the coupling and overlap of different element spectra.

[0165] Inverted Network

[0166] The inversion network is a multi-layer fully connected stacked structure that receives 256-dimensional deep features from the feature extraction network. Layered linear transformation, batch normalization, LeakyReLU activation, and progressively decaying Dropout regularization are applied to extract nonlinear correlation features layer by layer. A 7-dimensional output layer is placed at the end of the network to simultaneously output standardized predicted values ​​for seven sintered ore components. During the training phase, a weighted total loss is used as the loss function to calculate the overall error between the predicted and ground truth values ​​for all components. The Adam optimizer is used for backpropagation to update all network parameters, coupled with an adaptive learning rate decay strategy to automatically reduce the learning rate when the loss does not decrease.

[0167] The training and validation processes are executed iteratively, and the training loss and validation loss are recorded in real time. The optimal network weights with the lowest validation set loss are saved. After training, the complete inference model, the standardized parameters of the dataset, and the preprocessing baseline parameters of SG / MSC / MNF are uniformly packaged and stored to obtain a complete component prediction model that can be directly used for batch quantitative inversion of the spectra to be measured.

[0168] The test set data was input into the component prediction model for testing, and the test results are shown in Table 2.

[0169] Table 2 Comparison of Inversion Prediction Experiment Accuracy

[0170] TFe 0.0020 1.5490 0.3930 ≤0.35 SiO2 0.0010 0.4720 0.1355 ≤0.27 Al2O3 0.0010 1.0620 0.3144 ≤0.39 CaO 0.0010 0.5777 0.1114 ≤0.22 MgO 0.0002 0.2889 0.0900 ≤0.39 TiO2 0.0001 0.0666 0.0200 ≤0.04 MnO (Mn) 0.0003 0.1870 0.0559 ≤0.03

[0171] As shown in Table 2, this invention achieves ideal accuracy across most component inversion evaluation metrics, fully validating its effectiveness and robustness in simultaneous quantitative detection under scenarios involving spectral shifts in multiple batches of sinter and interference from differences in component content. Specifically, in a multi-day mixed sinter sample set, this method shows particularly significant improvement in the average absolute error control of the four components: SiO2, CaO, MgO, and TiO2. For example, the average absolute error of the TiO2 component is only 0.0200%, far below the industrial static accuracy limit of 0.04%; the average absolute error of the CaO component is as low as 0.1114%, significantly lower than traditional models, demonstrating the accurate extraction capability of this model's multi-scale convolution and SE channel attention for weak mineral absorption spectral features. This confirms that multi-stage preprocessing can effectively filter out spectral noise caused by dust and instrument baseline, while MMD domain adaptation can better eliminate spectral distribution shifts between different production batches, significantly improving the prediction accuracy of stable components.

[0172] The method of this invention also demonstrates strong generalization ability on complex sample subsets with greater fluctuations in raw material ratios and sintering conditions. In particular, it achieves leading error control performance for SiO2, Al2O3, and MgO components. For example, the average absolute error for the Al2O3 component is 0.3144%, meeting the industrial precision control standard of 0.39%, proving that the model, through adaptive balancing of the loss weights of the seven components using a dynamic weighted multi-task regression head, can effectively overcome the technical challenge of gradient-dominated training for high-content principal components and insufficient fitting for trace low-content components. It is worth noting that for TFe and MnO components, the average absolute error of this model is slightly higher than the industry static precision requirement. This may be due to the extremely wide range of TFe content in sintered ore, whose spectral absorption characteristics highly overlap with those of silicon and aluminum oxides, making them difficult to distinguish; while MnO is a trace component with an extremely low effective spectral signal-to-noise ratio, and its characteristics are easily masked by cross-batch illumination and dust disturbance; at the same time, the differences in spectral domain distribution caused by multiple days of production cannot be completely offset by MMD loss to eliminate extreme sample bias, which to some extent reveals the limitations of modeling single hyperspectral modal data in extreme working conditions and trace component detection scenarios.

[0173] This invention proposes a hyperspectral inversion framework for sintered ore that integrates multi-level spectral preprocessing, MMD domain adaptive transfer learning, and dynamically weighted multi-task regression. Through innovations such as multi-level hierarchical noise reduction and correction, multi-scale CNN channel attention networks, SE attention feature extraction, and multi-kernel MMD cross-batch domain alignment, it effectively solves two major bottleneck problems in existing models: insufficient inversion accuracy and poor cross-batch generalization. These problems stem from spectral noise scattering interference, cross-condition spectral domain shift, and neglect of trace components by equal-weighted training. This invention significantly improves the model's generalization ability and robustness to multi-day, multi-ratio sintered ore spectral samples. By achieving balanced fitting of spectral features of primary and secondary mineral components through domain distribution alignment and adaptive loss weighting mechanisms, it enables simultaneous and accurate quantitative inversion of multiple components in mixed industrial samples under various conditions. The inference time for a single sample throughout the entire process is significantly shorter than offline detection, meeting the continuous online real-time detection requirements of sintering production lines.

[0174] Finally, it should be noted that the above examples of the present invention are merely illustrative and not intended to limit the implementation of the invention. Although the applicant has described the present invention in detail with reference to preferred embodiments, those skilled in the art can make other variations and modifications based on the above description. It is impossible to exhaustively list all possible implementations here. All obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for online detection of multiple components in sintered ore mixtures based on hyperspectral imaging, characterized in that: The steps are as follows: 1) Establishment of hyperspectral dataset: Obtain different sintered ore mixture samples. All sintered ore mixture samples contain all the components to be detected and the content of each component is known. The content of any one component to be detected is different in different batches. Use a hyperspectral camera to collect hyperspectral data of all sintered ore mixture samples. 2) Data preprocessing: All hyperspectral data were preprocessed separately, including SG smoothing filtering, MSC multivariate scattering correction and MNF minimum noise separation and denoising performed sequentially. The preprocessed hyperspectral data were divided into training and testing sets. 3) Constructing a predictive neural network model: The predictive neural network model includes a feature extraction network and an inversion network; the feature extraction network is composed of multiple sets of structurally identical unit modules connected in series; each set of unit modules includes a multi-scale convolutional block and an SE channel attention module. The multi-scale convolutional block first uses one-dimensional convolutional branches of different sizes to perform parallel feature extraction. The multi-scale features output by each branch are concatenated in the channel dimension. The concatenated mixed features are then fused across scales through a 1×1 one-dimensional convolution. The fused features are fed into the SE channel attention module for adaptive feature weighting, and then downsampled by max pooling; the original spectral features are sequentially abstracted layer by layer through multiple sets of unit modules to complete multi-granularity feature abstraction, and finally output a fixed-dimensional deep feature vector and provide it to the inversion network; 4) Model training: Input the hyperspectral data corresponding to the training set into the prediction neural network model, continuously train and adjust the parameters of the neural network model using the hyperspectral data corresponding to the training set, establish the mapping relationship between the hyperspectral data and the content of the component to be detected, and verify the model performance through the test set. The trained prediction neural network model is the component prediction model. 5) Use a hyperspectral camera to collect hyperspectral data of the sinter mixture sample to be tested and preprocess it according to step 2). Then input it into the component prediction model trained in step 4). The component prediction model outputs the content of the component to be detected in the sinter mixture.

2. The method for online detection of multiple components in sintered ore mixture based on hyperspectral imaging according to claim 1, characterized in that: In step 2), the SG smoothing filter is specifically as follows: assuming a single original spectral vector... Window size 2m + 1, polynomial order p, local wavelength coordinates within the window ; Window Center Point smoothing value : In the formula, x represents the convolution weight coefficient at position t within the filtering window, where the local offset is t. i+t is the original spectral reflectance value at position t, with the i-th wavelength sampling point as the window center.

3. The method for online detection of multiple components in sintered ore mixture based on hyperspectral imaging according to claim 2, characterized in that: The specific details of the MSC multivariate scattering correction are as follows. First, calculate the average reference spectrum of all samples using the following formula. This serves as a unified benchmark: In the formula, M is the total number of samples; This is the spectrum of the i-th original sample; Subsequently, the spectrum of a single original sample was analyzed. A univariate linear regression model is constructed, and the unique scattering coefficient and baseline offset for each sample are solved by least squares fitting using the following formula: In the formula, This is a multiplicative scattering shift; This is the baseline offset; Finally, the following correction formula is used to complete the scattering compensation, eliminate the spectral amplitude differences between samples caused by physical morphology, unify the spectra of all samples to the same baseline scale, and significantly reduce the masking effect of scattering interference on the mineral absorption characteristics: MSC 。 4. The method for online detection of multiple components in sintered ore mixture based on hyperspectral imaging according to claim 2, characterized in that: In step 2), the minimum noise separation and denoising of MNF is specifically as follows: Let the original spectral matrix after correction be X ∈ M×N, where M is the total number of samples and N is the total number of spectral bands; firstly, the mixed spectrum is decomposed into two parts: pure useful signal and noise, satisfying the decomposition relationship: In the formula, S represents the effective signal matrix corresponding to the mixture components, and Γ represents the noise matrix caused by the environment and the acquisition equipment; the covariance matrices of the signal and noise are solved separately. The discreteness of the data distribution of the quantized signal and noise is then used to construct the generalized eigenvalue solution equation: In the formula For eigenvalues, For the corresponding eigenvector, the magnitude of the eigenvalue represents the energy ratio of the signal to the noise in the corresponding direction; All eigenvectors are sorted according to their eigenvalues. Sort by largest to smallest, and filter by the top results whose cumulative variance percentage is not less than a set value. The group of feature vectors constitutes the projection matrix ; Finally, spectral dimensionality reduction is achieved using the following projection transformation formula: This transformation completes noise whitening, filters out invalid bands with low signal-to-noise ratios and excessive noise, and significantly compresses data dimensionality while preserving the effective information of mineral spectra to the greatest extent.

5. The method for online detection of multiple components in sintered ore mixture based on hyperspectral imaging according to claim 1, characterized in that: In step 3), the multi-scale convolutional block uses three one-dimensional convolutional branches of different sizes to extract features in parallel. The kernel sizes of the three different sizes are 3, 7 and 11.

6. The method for online detection of multiple components in sintered ore mixture based on hyperspectral imaging according to claim 1, characterized in that: In step 3), before the fused features are fed into the SE channel attention module, batch normalization and ReLU nonlinear activation are performed sequentially, as shown in the following expressions: in, To fuse features; batch normalization standardizes the feature distribution, alleviating the gradient vanishing problem during deep network training; ReLU nonlinear activation is used to mine complex coupling correlations between spectral bands; the result is the final output feature of the multi-scale convolutional block.

7. The method for online detection of multiple components in sintered ore mixture based on hyperspectral imaging according to claim 1, characterized in that: In step 3), the SE channel attention module is used to filter and denoise the mixed channel features output by the multi-scale convolutional blocks. The specific processing procedure is as follows: Compression: For each feature channel output by the multi-scale convolutional block, global average pooling is performed along the spectral wavelength dimension to compress the one-dimensional channel sequence of length L into a single scalar value. This scalar value represents the global spectral information of the corresponding channel, realizing the aggregation of information across the entire band. Incentive: The compressed global features are fed into two fully connected layers to learn the channel importance weights: the first fully connected layer performs dimensionality reduction compression with an adjustable compression rate on the number of channels, and is combined with ReLU activation to mine the dependencies between channels; The second fully connected layer restores the original number of channels, i.e., the number of bands. The output is mapped to the [0,1] interval by the Sigmoid activation function to obtain the specific importance weight coefficient of each channel. Feature weighting: The importance weight coefficient of each channel is multiplied by the original multi-scale feature map channel by channel, and the effective channels containing mineral-sensitive absorption bands are assigned high weights to amplify the feature response; Low weights are assigned to noisy and irrelevant band channels to suppress redundant interference, thereby completing feature selection and noise reduction.

8. The method for online detection of multiple components in sintered ore mixture based on hyperspectral imaging according to claim 1, characterized in that: In step 3), the inversion network as a whole adopts a serially stacked architecture of multiple sets of modular basic units with completely unified structure. Within a single basic unit, a linear fully connected layer, a BN batch normalization layer, a LeakyReLU activation layer, and a Dropout regularization layer are connected in series. Among them, the linear fully connected layer completes the feature dimension mapping; the BN batch normalization layer constrains the feature distribution range within the unit, weakens the feature offset caused by spectral baseline drift and scattering noise, and accelerates model convergence; the LeakyReLU activation layer introduces nonlinear fitting capability into the network, avoids neuron death defects, and adapts to the complex nonlinear mapping relationship between the spectrum and mineral components; along the feature forward propagation direction, the feature channel dimension and Dropout drop probability of each basic unit decrease step by step; some neuron outputs are randomly shielded to suppress the overfitting memory of training set noise and irrelevant band features in a hierarchical manner.

9. The method for online detection of multiple components in sintered ore mixture based on hyperspectral imaging according to claim 1, characterized in that: In step 4), the prediction neural network model training process introduces SmoothL1 multi-task regression loss to constrain the prediction error, and the maximum mean difference loss is used to align the spectral feature distributions of different batches to construct a weighted joint loss function optimization model. SmoothL1 Multi-task Regression Loss L task Calculate using the following formula: In the formula: M is the total number of samples; P is the number of components to be detected; For the first The first sample The true content of each component; For the first The first sample Predicted values ​​for each component; Maximum mean difference loss L MMD Calculate using the following formula: Gaussian kernel function in the formula ; These represent the number of samples in the source domain and the target domain, respectively. For the corresponding domain depth features; For core bandwidth; The total loss Li of the prediction neural network model is obtained by weighted fusion of the SmoothL1 multi-task regression loss and the maximum mean difference loss. total As shown in the following formula: In the formula: This is the maximum mean difference loss weighting coefficient, used to balance the component prediction error with the alignment strength of the feature distribution.

10. The method for online detection of multiple components in sintered ore mixture based on hyperspectral imaging according to claim 9, characterized in that: In step 4), when training the neural network model for prediction, prior knowledge of the spectral physics of sintered ore is embedded into the network loss function to complete the decoupling of overlapping spectral lines of multiple elements; the corresponding prior loss function... as follows: in, The known standard characteristic peak position for the k-th mineral is... This refers to the peak response learned by the network. These are the weighting coefficients for each prior constraint; The total loss L predicted using the neural network model is then... total As shown in the following formula: In the formula: This is the weighting coefficient for the mineral spectrum prior constraint loss, which adjusts the constraint strength of the physical prior constraint in the overall training.

Images