A method and system for real-time monitoring of lithium mica flotation process based on portable infrared spectroscopy and machine learning

Abstract

This invention discloses a method and device for real-time monitoring of the lepidolite flotation process based on portable infrared spectroscopy and machine learning, belonging to the field of mineral processing automation and intelligent monitoring. This invention acquires near-infrared image data and near-infrared spectral data of the same area on the surface of the flotation froth; it extracts image feature vectors and spectral feature vectors using convolutional neural networks and one-dimensional convolutional neural networks, respectively; it inputs the two feature vectors into a joint sparse coding layer, using the spectral feature vector as the dominant mode and the image feature vector as the auxiliary mode, constructing an overcomplete dictionary based on the Si-O-Si bond absorption peaks, and using a joint orthogonal matching pursuit algorithm to solve the joint sparse representation coefficients. The image features are reconstructed based on the positions of the non-zero elements in the coefficients; the fused feature vector is input into a regression prediction network, outputting predicted values ​​for lepidolite grade, gangue ratio, or reagent residue concentration. This invention achieves real-time, portable, and low-cost monitoring of the lepidolite flotation process.

Description

Technical Field

[0001] This invention relates to the field of mineral processing automation and intelligent monitoring, and in particular to a method and device for real-time monitoring of the lithium mica flotation process based on portable infrared spectroscopy and machine learning. Background Technology

[0002] Lepidolite is an important lithium resource mineral, and the quality control of its flotation process directly affects the lithium recovery rate and concentrate grade. Currently, the detection of key components in the lepidolite flotation process mainly relies on offline sampling and analysis. Operators periodically take samples from the flotation cell, dry and digest them, and then analyze them using methods such as ICP or XRD to obtain indicators such as lepidolite grade, gangue ratio, and reagent residue concentration. This method has a detection cycle of several hours to several days, failing to reflect component fluctuations during the flotation process in real time. This leads to a significant lag in process parameter adjustments, resulting in reagent waste and decreased recovery rates. To address the detection lag problem, existing technologies attempt to introduce online detection methods, such as online X-ray fluorescence analysis. However, XRF has low sensitivity for lithium with low atomic numbers and the equipment is expensive. While laser-induced breakdown spectroscopy can detect lithium, the equipment is complex and has high maintenance costs, making it difficult to promote in mines. In recent years, machine vision-based flotation foam image analysis methods have become a research hotspot. By extracting the color, texture, size, and dynamic characteristics of the foam, a mapping relationship between these characteristics and process indicators can be established. However, image features only reflect the apparent morphology of foam, and there is an indirect and non-linear mapping relationship between them and mineral grade. Furthermore, they are easily affected by on-site factors such as lighting and bubble bursting, resulting in limited prediction accuracy. Near-infrared spectroscopy can directly reflect the vibrational information of mineral molecules, but it is difficult to obtain the spatial distribution information of foam using spectral detection alone. Therefore, how to organically combine spectral fingerprint information with foam image features to achieve low-cost, real-time, and accurate monitoring of the lepidolite flotation process is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0003] To address the technical problems of severe lag in offline testing and insensitivity to lithium, as well as the low prediction accuracy of single data sources in existing technologies, this invention provides a method and device for real-time monitoring of the lithium mica flotation process based on portable infrared spectroscopy and machine learning.

[0004] In a first aspect, the present invention provides a real-time monitoring method for the flotation process of lithium mica based on portable infrared spectroscopy and machine learning, comprising the following steps: Acquire multimodal data of the flotation foam surface, the multimodal data including: near-infrared image data of the foam surface, and near-infrared spectral data from the same foam region as the near-infrared image data; The near-infrared image data is input into a first feature extraction network, which outputs an image feature vector; the near-infrared spectral data is arranged into a one-dimensional spectral vector by sampling points and then input into a second feature extraction network, which outputs a spectral feature vector. The image feature vector and the spectral feature vector are input into a joint sparse coding layer. The joint sparse coding layer uses the spectral feature vector as the dominant mode and the image feature vector as the auxiliary mode. Based on the feature channels corresponding to the Si-O-Si bond absorption peaks in the spectral feature vector, an overcomplete dictionary is constructed. The joint orthogonal matching pursuit algorithm is used to solve for the joint sparse representation coefficients. The positions of the non-zero elements in the joint sparse representation coefficients correspond to the spatial positions in the image feature vector that share the same sparse mode as the spectral feature vector. The image feature vector is reconstructed according to the joint sparse representation coefficients, and the reconstructed image features are fused with the spectral feature vector to output a fused feature vector. The fused feature vector is input into the regression prediction network, which outputs the predicted content of key components in the lithium mica flotation process. The predicted content is then compared with a preset threshold to generate a feedback signal for adjusting the flotation process parameters.

[0005] Furthermore, the overcomplete dictionary is constructed as follows: The spectral feature channels corresponding to the Si-O-Si bond absorption peaks in the spectral feature vector are used as the basic components of the dictionary atoms. Each dictionary atom is jointly formed by the basic components and the local spatial patterns of the image feature vector. The absorption peak intensity of the Si-O-Si bond is positively correlated with the selection priority of the corresponding dictionary atom.

[0006] Furthermore, the joint sparse coding layer also dynamically adjusts the overcomplete dictionary by utilizing the lattice vibration absorption peak shift caused by thallium in lepidolite; wherein, the absorption peak shift is used to distinguish between lepidolite with high thallium content and lepidolite with low thallium content, and adaptively updates the corresponding dictionary atoms in the overcomplete dictionary according to the shift.

[0007] Furthermore, it also includes: The iteration termination condition of the joint orthogonal matching pursuit algorithm is coupled with the physical properties of the flotation foam: The iteration terminates when the energy of the residual vector is lower than a first preset threshold, which is negatively correlated with the signal-to-noise ratio of the spectral feature vector; wherein the signal-to-noise ratio is calculated in real time based on the ratio of the signal intensity of the band containing the Si-O-Si bond absorption peak in the near-infrared spectral data to the baseline noise. The iteration may terminate when the number of non-zero coefficients reaches a second preset threshold, which is positively correlated with the estimated number of bubbles in the flotation bubble image.

[0008] Furthermore, the fusion of the reconstructed image features and the spectral feature vector employs a gating mechanism: The gating coefficients are calculated based on the norm of the residual vector of the joint orthogonal matching pursuit algorithm and the spectral eigenvector. When the norm of the residual vector increases, the gating coefficient decreases, thereby reducing the weight of the reconstructed image features and increasing the weight of the spectral feature vector.

[0009] Furthermore, the gating fusion mechanism includes: defining a gating coefficient as... , The gating fusion mechanism is implemented through the following formula: , , Represents the spectral feature vector, This represents the reconstructed image features. This represents the fused feature vector. This represents the L2 norm of the residual vector of the joint orthogonal matching pursuit algorithm. Indicates the gating coefficient. Represents the gate weight matrix. Indicates the gating bias term. This represents the Sigmoid activation function. This represents element-wise multiplication. When the water film on the flotation foam surface is thick or the bubbles are dense, causing spectral signal attenuation, the gating coefficient... Decrease the weight of the reconstructed image features and increase the weight of the spectral feature vector; When the bursting of bubbles on the surface of flotation foam causes localized image distortion, the gating coefficient... Increase or decrease the weight of the spectral feature vector, and increase the weight of the attention fusion feature vector.

[0010] Furthermore, the first feature extraction network is a convolutional neural network, used to extract the bubble size distribution, morphological texture and spatial distribution features of mineralized regions of flotation foam; The spatial distribution characteristics of the mineralized region are directly indicated by the position of the non-zero elements in the joint sparse representation coefficients. The second feature extraction network is a one-dimensional convolutional neural network used to extract the molecular vibrational spectral fingerprint features of lepidolite and its associated minerals.

[0011] Furthermore, the fused feature vector is input into a regression prediction network to output predicted values ​​of key components in the lithium mica flotation process, specifically including: The fused feature vector is used as input data and fed into the regression prediction network; The regression prediction network comprises, in sequence: a first fully connected layer, a first activation function layer, a second fully connected layer, a second activation function layer, and an output layer; The first fully connected layer maps the fused feature vector to a first intermediate feature vector, and the first activation function layer performs a nonlinear transformation on the first intermediate feature vector to output a second intermediate feature vector. The second fully connected layer maps the second intermediate feature vector to a third intermediate feature vector, the dimension of which is equal to the number of types of key components to be predicted; The second activation function layer performs a nonlinear transformation on the third intermediate feature vector to output a fourth intermediate feature vector; The output layer uses a linear activation function to directly map each element in the fourth intermediate feature vector to the predicted content value of the corresponding key component, and the predicted content value is a continuous value. The key components include at least one or more of the following: lepidolite grade, gangue ratio, or reagent residual concentration.

[0012] Secondly, the present invention also provides a real-time monitoring device for the lithium mica flotation process based on portable infrared spectroscopy and machine learning, applied to any of the aforementioned real-time monitoring methods for the lithium mica flotation process based on portable infrared spectroscopy and machine learning, comprising: The multimodal data acquisition module includes a portable near-infrared spectrometer and a near-infrared camera. The multimodal data acquisition module is used to simultaneously acquire near-infrared image data and near-infrared spectral data of the flotation foam surface, and output the image data and spectral data to the dual-branch feature extraction module. The dual-branch feature extraction module includes a first feature extraction network and a second feature extraction network. The first feature extraction network is used to extract image feature vectors from the image data, and the second feature extraction network is used to extract spectral feature vectors from the spectral data. The two feature vectors are then output to the joint sparse coding fusion module. The joint sparse coding fusion module is used to perform cross-modal feature fusion with the spectral feature vector as the dominant mode and the image feature vector as the auxiliary mode, and output the fused feature vector to the regression prediction module. The regression prediction module is used to output the predicted content of key components in the lithium mica flotation process based on the fused feature vector, and output the predicted value to the display unit or the flotation control system.

[0013] Furthermore, the joint sparse coding fusion module includes: an overcomplete dictionary storage unit for storing dictionary atoms constructed based on Si-O-Si bond absorption peaks; a joint orthogonal matching pursuit solving unit for solving joint sparse representation coefficients; and a gated fusion unit for dynamically adjusting the fusion weights of image features and spectral features according to the norm of the residual vector.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention represents a leap from "single data source estimation" to "cross-modal physical correlation fusion," significantly improving prediction accuracy and interpretability. By combining a sparse coding layer, prioritizing spectral features and supplementing with image features, and utilizing the specific Si-O-Si bond absorption peaks of lepidolite to construct an overcomplete dictionary, the non-zero element positions of the sparse representation coefficients directly correspond to the spatial regions with the highest mineralization in the image. This fundamentally solves the problem of the lack of a direct physical correspondence between image features and mineral grades. Compared to existing image analysis methods, prediction accuracy is improved, and the model possesses clear physical meaning, making it easier to pass review. Furthermore, by combining a portable near-infrared spectrometer, the detection cycle is shortened from several hours to minutes, overcoming the drawbacks of offline testing delays and the high cost of large online equipment (XRF, LIBS).

[0015] This invention achieves adaptive perception and fusion of the dynamic environment of flotation, significantly improving the robustness and efficiency of the system. Furthermore, it dynamically couples the iteration termination condition of sparse coding with on-site physical characteristics (residual threshold correlated with signal-to-noise ratio, number of non-zero coefficients correlated with number of bubbles), and dynamically links the gating coefficient with the reconstructed residual norm. This enables the system to automatically adjust dictionary atoms, iteration strategy, and bimodal confidence weights based on real-time spectral quality, foam morphology, and mineral composition (such as thallium content) fluctuations. Compared with existing methods using fixed parameters and fixed weights, this invention exhibits higher prediction stability and lower computational resource consumption in complex dynamic environments, achieving truly intelligent adaptive monitoring. The technical effects of this invention achieve rapid, low-cost, and high-precision lithium mica grade detection: the detection cycle is significantly shortened from several hours in traditional offline testing to 1-5 minutes; equipment investment is only one-fifth to one-tenth of that of an XRF online analyzer; and under typical operating conditions, the determination coefficient R² of lithium mica grade prediction reaches over 0.92, with an average absolute error of less than 0.15%. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for real-time monitoring of the lithium mica flotation process based on portable infrared spectroscopy and machine learning, disclosed in one embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the specific execution steps of the joint orthogonal matching pursuit algorithm disclosed in one embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a real-time monitoring device for the lithium mica flotation process based on portable infrared spectroscopy and machine learning, disclosed in one embodiment of the present invention. Detailed Implementation

[0019] 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 a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0021] Furthermore, the use of terms such as "first" and "second" in this invention is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" and "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. When the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed by this invention.

[0022] Example 1

[0023] This embodiment is implemented using an embedded hardware platform, including but not limited to NVIDIA Jetson series, Raspberry Pi, industrial PCs, or edge computing gateways. This platform is deployed in the flotation workshop and interacts with the PLC or DCS of the flotation control system via wired or wireless means. When the predicted content value output by the regression prediction network exceeds a preset process threshold, the system automatically sends an adjustment command to the flotation control system, achieving closed-loop control.

[0024] For ease of understanding, the key terms in this invention are defined as follows: The local spatial pattern of image feature vectors refers to the feature sub-vector corresponding to each spatial block after the image feature vector output by the first feature extraction network is divided into blocks according to their corresponding original image spatial locations. The block division method can be regular grid block division or irregular mineralized region mask generated by image segmentation algorithm.

[0025] Attention fusion feature vector refers to the intermediate feature vector obtained by weighted fusion of reconstructed image features and spectral feature vectors through a gating mechanism.

[0026] See Figures 1-2 As shown, the present invention provides a real-time monitoring method for the flotation process of lithium mica based on portable infrared spectroscopy and machine learning, comprising the following steps: S1. Obtain multimodal data of the flotation foam surface, the multimodal data including: near-infrared image data of the foam surface, and near-infrared spectral data from the same foam region as the near-infrared image data; S2. Input the near-infrared image data into the first feature extraction network and output the image feature vector; after arranging the near-infrared spectral data into a one-dimensional spectral vector through sampling points, input the data into the second feature extraction network and output the spectral feature vector. S3. Input the image feature vector and the spectral feature vector into a joint sparse coding layer. The joint sparse coding layer uses the spectral feature vector as the dominant mode and the image feature vector as the auxiliary mode. Based on the feature channels corresponding to the Si-O-Si bond absorption peaks in the spectral feature vector, an overcomplete dictionary is constructed. The joint orthogonal matching pursuit algorithm is used to solve for the joint sparse representation coefficients. The positions of the non-zero elements in the joint sparse representation coefficients correspond to the spatial positions in the image feature vector that share the same sparse mode as the spectral feature vector. The image feature vector is reconstructed according to the joint sparse representation coefficients, and the reconstructed image features are fused with the spectral feature vector to output a fused feature vector. S4. Input the fused feature vector into the regression prediction network, output the predicted content of key components in the lithium mica flotation process, and generate a feedback signal for adjusting the flotation process parameters by comparing the predicted content with a preset threshold.

[0027] According to step S1, it should be noted that the lepidolite flotation process relies on the differences in the physicochemical properties of the mineral surface. By adding collectors, frothers and other reagents, and under aeration and stirring, the target mineral lepidolite selectively adheres to the bubbles and rises to the surface of the slurry to form a mineralized foam layer, while gangue minerals remain in the slurry. This achieves the beneficiation and enrichment process of separating lepidolite from impurities. The flotation foam surface is a solid-liquid-gas three-phase dynamic interface that carries the mineralized bubbles. Its morphological characteristics (such as bubble size, color, texture and breakage rate) have a highly nonlinear mapping relationship with the process indicators such as the lepidolite grade, reagent residual concentration and mineralization degree in the slurry.

[0028] Acquiring multimodal data of flotation froth surface (including near-infrared spectra reflecting mineral molecular vibration information and images reflecting froth surface morphology) is to address the problem of insufficient physical characterization dimensions of single-modal information: Near-infrared spectroscopy can directly detect characteristic bonds of lepidolite (such as Si-O-Si, lattice hydroxyl groups) and characteristic absorption peaks of gangue minerals, providing bulk composition information based on chemical bonds, but it lacks the ability to perceive the spatial structure and mineralization distribution of froth; while machine vision images can capture the spatial distribution and dynamic behavior of bubble groups, but can only indirectly infer grade and are easily affected by on-site interference such as lighting and bubble overlap.

[0029] In some embodiments, at the outlet of the concentrate foam tank of the lepidolite processing equipment, a portable near-infrared spectrometer (wavelength range 900-2500nm, resolution better than 5nm) is used. The spectrometer probe is aimed at the foam layer surface to simultaneously acquire near-infrared image data and near-infrared spectral data of the same foam area. During acquisition, the probe is kept 2cm away from the foam surface, and each sampling point is sampled three times consecutively and averaged. The acquired image data size is 640×480 pixels, and the spectral data is a one-dimensional vector with 512 sampling points. The image data is input into a pre-trained convolutional neural network (ResNet18 structure), which outputs a 512-dimensional image feature vector; the spectral data is arranged into a one-dimensional spectral vector and then input into a one-dimensional convolutional neural network (containing 3 convolutional blocks with 32, 64, and 128 channels respectively), which outputs a 256-dimensional spectral feature vector. The two feature vectors are input into a joint sparse coding layer, with the spectral feature vector as the dominant mode, based on its... A complete dictionary was constructed for the feature channels corresponding to the Si-O-Si bond absorption peaks within the wavenumber range. A joint orthogonal matching pursuit algorithm was used to solve for the joint sparse representation coefficients, resulting in a set of sparse coefficient vectors, where the positions of non-zero elements correspond to the bubble regions with the highest mineralization in the image. The image feature vector was reconstructed based on the sparse coefficients, and the reconstructed image features were fused with the spectral feature vector element-wise, outputting a 512-dimensional fused feature vector. This fused feature vector was then input into a regression prediction network (a three-layer fully connected network with 256, 128, and 64 hidden neurons, respectively), outputting a predicted value for the lepidolite grade.

[0030] Furthermore, the overcomplete dictionary is constructed as follows: The spectral feature channels corresponding to the Si-O-Si bond absorption peaks in the spectral feature vector are used as the basic components of the dictionary atoms. Each dictionary atom is jointly formed by the basic components and the local spatial patterns of the image feature vector. The absorption peak intensity of the Si-O-Si bond is positively correlated with the selection priority of the corresponding dictionary atom.

[0031] According to step S3, it should be noted that the overcomplete dictionary serves to establish a physically meaningful joint sparse representation basis for the spectral feature vector and the image feature vector. Specifically, the overcomplete dictionary enables the joint orthogonal matching pursuit algorithm to force the sparse representation coefficients of the image feature vector and the sparse representation coefficients of the spectral feature vector to share the same non-zero element position pattern during the solution process. Since the dictionary atoms are constructed based on the feature channels corresponding to the Si-O-Si bond absorption peaks in the spectral feature vector, the positions of the non-zero elements directly correspond to the spatial regions with the highest mineralization in the image feature vector. Simultaneously, the overcomplete dictionary also acts as a noise filter: only those image feature components that can be linearly represented by lepidolite feature atoms can be retained and reconstructed, while interfering features unrelated to lepidolite grade (such as illumination changes and pseudo-textures generated by bubble bursting) are automatically eliminated because they cannot be effectively represented by the atoms in the dictionary.

[0032] It should be noted that the overcomplete dictionary is constructed as follows: The absorption peak of the Si-O-Si bond in the spectral characteristic vector ( The corresponding spectral feature channels serve as the basic components of the dictionary atom. Each dictionary atom is jointly constructed by this basic component and the local spatial pattern of the image feature vector (e.g., the flattened vector of a 5×5 pixel block). Let the i-th dictionary atom be represented as... ,in Based on the fundamental spectral components, This represents the local spatial pattern of the image at the corresponding location. (Intensity of Si-O-Si bond absorption peak) With dictionary atoms The selection priority is positively correlated; that is, the greater the absorption peak intensity, the higher the probability that the atom will be preferentially selected in the orthogonal matching pursuit iteration. For example, when the spectrum of a certain region contains... When the absorption peak intensity is 0.8 (normalized value) in one region and 0.3 in another region, the dictionary atoms in the former region will be preferentially selected to participate in sparse coding, thereby enabling the spectrum to guide image feature extraction.

[0033] Furthermore, the joint sparse coding layer also dynamically adjusts the overcomplete dictionary by utilizing the lattice vibration absorption peak shift caused by thallium in lepidolite; wherein, the absorption peak shift is used to distinguish between lepidolite with high thallium content and lepidolite with low thallium content, and adaptively updates the corresponding dictionary atoms in the overcomplete dictionary according to the shift.

[0034] In some embodiments, to address variations in thallium content in lepidolite, this invention utilizes the lattice vibration absorption peak shift to dynamically adjust the overcomplete dictionary. When K⁺ is replaced by Tl⁺ in the lepidolite crystal, the cell parameters change, resulting in a shift from 400-60... The lattice vibration absorption peaks within the wavenumber range shift. Specifically, when the thallium content increases from 0% to 10%, the characteristic absorption peaks shift towards lower frequencies by approximately 15-20%. The system calculates the current spectrum in real time between 400-600. Absorption peak positions within the range Compared with the preset standard peak position (Compare with low-thallium lithium mica) based on displacement. The magnitude of is used to linearly adjust the fundamental frequency of the atoms corresponding to the lattice vibration modes in the overcomplete dictionary.

[0035] The formula is adjusted as follows: , in, Indicates the number before adjustment One dictionary atom; This refers to the shift of the aforementioned absorption peak; This represents the preset scaling factor, used to control the sensitivity of the displacement to dictionary atom adjustments. The value ranges from 0.01 to 0.1, with a preferred value of 0.05. This factor is obtained through experimental calibration. This represents the adjusted dictionary atoms. If the absolute value of the displacement exceeds a preset threshold (e.g., 10),... If the result is positive, it is identified as high-thallium lepidolite mica, and the complete dictionary is updated using the adjusted dictionary atoms so that the model can adapt to lepidolite mica samples with different thallium contents.

[0036] Furthermore, it also includes: The iteration termination condition of the joint orthogonal matching pursuit algorithm is coupled with the physical properties of the flotation foam: The iteration terminates when the energy of the residual vector is lower than a first preset threshold, which is negatively correlated with the signal-to-noise ratio of the spectral feature vector; wherein the signal-to-noise ratio is calculated in real time based on the ratio of the signal intensity of the band containing the Si-O-Si bond absorption peak in the near-infrared spectral data to the baseline noise. The iteration may terminate when the number of non-zero coefficients reaches a second preset threshold, which is positively correlated with the estimated number of bubbles in the flotation bubble image.

[0037] It should be noted that, in this specific embodiment, the joint orthogonal matching pursuit algorithm is a greedy iterative algorithm for solving the joint sparse coding problem. Its core function is to solve for a joint sparse representation coefficient matrix given spectral feature vectors and image feature vectors, such that the product of the dictionary and the coefficient matrix can approximately reconstruct the original signal, while forcing different columns of the coefficient matrix to share the same non-zero element positions, i.e., the same sparse support set. The input of the joint orthogonal matching pursuit algorithm includes: an overcomplete dictionary, spectral feature vectors, and image feature vectors; the output is a joint sparse representation coefficient matrix, where the first column is the sparse representation coefficient of the spectral coefficients, the second column is the sparse representation coefficient of the image coefficients, and the non-zero element positions of the two columns are exactly the same.

[0038] In this embodiment, the specific execution steps of the joint orthogonal matching pursuit algorithm are as follows: S31. Initialize the residual matrix, with the first column being the spectral residual and the second column being the image residual; initialize the support set to an empty set; initialize the iteration counter to one; set the iteration termination conditions, including a residual energy threshold dynamically determined with a negative correlation to the spectral signal-to-noise ratio, and a maximum number of non-zero coefficients dynamically determined with a positive correlation to the estimated number of bubbles. S32. In the current iteration, for each atom in the dictionary, calculate the sum of the absolute values ​​of the correlation between the atom and each column in the current residual matrix, select the atom index that maximizes the sum, and add the atom to the support set. S33. Based on the dictionary submatrix corresponding to the current support set, the projection of the joint sparse representation coefficients onto the support set is solved by the least squares method to obtain the coefficient values ​​on the support set, while the coefficients outside the support set remain zero. S34. Recalculate the residual matrix based on the current coefficients, i.e., subtract the product of the dictionary and the coefficients from the original signal; S35. Calculate the Frobenius norm of the current residual matrix. If the norm is lower than the residual energy threshold, or the number of iterations has reached the maximum number of non-zero coefficients, terminate the iteration; otherwise, increment the iteration counter and return to step S32. The Frobenius norm of the residual matrix refers to the arithmetic square root of the sum of squares of all elements of the matrix, which is used to measure the reconstruction error of the spectral feature vector and the image feature vector in the current iteration. S36. Output the joint sparse representation coefficient matrix, where the coefficients at non-support positions are zero, and the coefficients at support positions are the corresponding values ​​calculated in step three.

[0039] In some embodiments, the iterative termination condition of the joint orthogonal matching pursuit algorithm is coupled with the physical properties of the flotation foam as follows: Let the energy of the current residual vector be... First preset threshold Signal-to-noise ratio with spectral eigenvectors Negative correlation.

[0040] The threshold calculation formula is: , in, Represents the residual vector The square of the L2 norm, i.e. It is used to measure the magnitude of the current sparse coding reconstruction error. is the dimension of the residual vector; This represents the first preset threshold value, and the iteration terminates when the residual energy is lower than this value. This represents the basic threshold, a preset constant with a value ranging from 0.001 to 0.1, with a preferred value of 0.01. This value is determined through cross-validation. The signal-to-noise ratio (SNR) represents the spectral eigenvectors, based on the wavelength range of the Si-O-Si bond absorption peak (1000-1100 nm). The ratio of signal strength to baseline noise is calculated in real time, using the following formula: ,in The average signal amplitude in the band where the absorption peak is located. For non-absorption characteristic bands (e.g., 800-900) The noise amplitude. When the spectral signal-to-noise ratio is high, Smaller signal-to-noise ratio allows for more thorough algorithm iterations; when the signal-to-noise ratio is low, If the noise level is too high, the algorithm terminates early to avoid overfitting to noise. The iteration terminates when the time is right.

[0041] Meanwhile, let the number of currently selected non-zero coefficients be... The second preset threshold The number of bubbles predicted in the flotation foam image Positive correlation.

[0042] The threshold calculation formula is: , in Use the base coefficient (e.g., 5). The upper bound (e.g., 50). When The iteration terminates when the number of bubbles in the bubble image is large (approximately 200). The value is approximately 35; when the number of bubbles is small (approximately 50), The value is approximately 20, and the coupling mechanism allows the iterative process to adapt to the actual complexity of the bubble.

[0043] Furthermore, the fusion of the reconstructed image features and the spectral feature vector employs a gating mechanism: The gating coefficients are calculated based on the norm of the residual vector of the joint orthogonal matching pursuit algorithm and the spectral eigenvector. When the norm of the residual vector increases, the gating coefficient decreases, thereby reducing the weight of the reconstructed image features and increasing the weight of the spectral feature vector.

[0044] According to step S3, it should be noted that reconstruction refers to the process of using the solved joint sparse representation coefficients and the overcomplete dictionary to synthesize an estimate of the original signal. Specifically, under the joint sparse coding framework, let the overcomplete dictionary be D (of size d×K, where d is the feature dimension, K is the number of atoms, and K>d), and the joint sparse representation coefficient matrix be C (of size K×2, with two columns corresponding to the sparse coefficients of the spectral feature vector and the image feature vector, respectively). Then, the reconstructed spectral feature vector is D·C[:,0], and the reconstructed image feature vector is... Because joint sparse coding mandates that the coefficients in both columns share the same non-zero element positions (i.e., the sparse support sets are the same), the reconstructed image feature vector is actually a linear combination of the same dictionary atoms as the spectral feature vector. This means that the reconstructed image features retain only components with the same sparse pattern as the spectral features, while the components that are removed include: noise unrelated to lepidolite mineralization (such as pseudo-textures caused by uneven illumination), changes in foam morphology unrelated to grade (such as size changes of bubbles due to rupture rather than mineralization), and image features caused by other minerals (such as pyrite and calcite).

[0045] Traditional multimodal fusion methods (such as feature stitching and weighted summation) only merge spectral and image features at the numerical level, without physically filtering and purifying the image features. Image feature vectors contain two types of information: features related to mineral grade (such as changes in luster caused by the degree of lepidolite coverage on mineralized bubble surfaces) and interfering features unrelated to mineral grade (such as changes in illumination angle, natural bubble breakage, camera noise, and slurry surface fluctuations). If the original image features are directly fused with spectral features, the interfering features will introduce errors, reducing the accuracy and robustness of grade prediction. The essence of the reconstruction step is to use spectral features as a "physical benchmark" to filter image features: only those image feature components that share the same sparse pattern as the spectral features (i.e., can also be linearly represented by lepidolite-specific atoms) are retained, while other components are discarded. In existing machine vision-based flotation monitoring methods, there is no direct physical correspondence between image features and mineral grade. On-site disturbances such as changes in lighting and bubble bursting can significantly alter image features (e.g., bubbles become brighter or darker, or their texture becomes coarser or smoother), but these changes do not correspond to the actual changes in grade, leading to misjudgments by the model. For example, when sunlight shines directly on the surface of the flotation cell, the overall foam image becomes brighter, and traditional image models may incorrectly interpret this as an increase in lepidolite grade (because mineralized bubbles are usually more reflective).

[0046] Furthermore, the gating fusion mechanism includes: defining a gating coefficient as... ,

[0047] The gating fusion mechanism is implemented through the following formula: , , Represents the spectral feature vector, This represents the reconstructed image features. This represents the fused feature vector. This represents the L2 norm of the residual vector of the joint orthogonal matching pursuit algorithm. Indicates the gating coefficient. Represents the gate weight matrix. Indicates the gating bias term. This represents the Sigmoid activation function. This represents element-wise multiplication. When the water film on the flotation foam surface is thick or the bubbles are dense, causing spectral signal attenuation, the gating coefficient... Decrease the weight of the reconstructed image features and increase the weight of the spectral feature vector; When the bursting of bubbles on the surface of flotation foam causes localized image distortion, the gating coefficient... Increase or decrease the weight of the spectral feature vector, and increase the weight of the attention fusion feature vector.

[0048] Furthermore, the first feature extraction network is a convolutional neural network, used to extract the bubble size distribution, morphological texture and spatial distribution features of mineralized regions of flotation foam; The spatial distribution characteristics of the mineralized region are directly indicated by the position of the non-zero elements in the joint sparse representation coefficients. The second feature extraction network is a one-dimensional convolutional neural network used to extract the molecular vibrational spectral fingerprint features of lepidolite and its associated minerals.

[0049] It should be noted that the first feature extraction network uses a pre-trained ResNet18 convolutional neural network, which contains 16 convolutional layers and 5 pooling layers. The input is a normalized 640×480 pixel bubble image. The first two layers of the network extract bubble edge features (such as bubble contours), the middle layers extract morphological texture features (such as the distribution of mineralized particles on the bubble surface), and the deep layers extract semantic features of the mineralized region. The spatial distribution features of the mineralized region are directly indicated by the position of the non-zero elements in the joint sparse representation coefficients. For example, when the i-th element in the sparse coefficients is non-zero, the i-th pixel block in the corresponding image is marked as a mineralized region. The second feature extraction network is a one-dimensional convolutional neural network containing three one-dimensional convolutional blocks: the first convolutional block contains 32 convolutional kernels of size 3 to extract local features in the spectrum (such as single peaks); the second convolutional block contains 64 convolutional kernels to extract inter-peak combination features; and the third convolutional block contains 128 convolutional kernels to extract the global pattern of molecular vibrations (such as the main peak position of the Si-O-Si bond). The network outputs a 256-dimensional spectral feature vector, where dimensions 50-80 correspond to the feature encoding of the Si-O-Si bond absorption peak.

[0050] It should be noted that, according to step S2, the image feature vector refers to the low-dimensional, compact vector representation output by the first feature extraction network. Its dimensions are typically 64, 128, or 256. It is a condensed representation of information from the original high-dimensional image (e.g., megapixels) after dimensionality reduction and abstraction, retaining the foam morphology information most relevant to the mineralization degree of lepidolite. The spectral feature vector refers to the low-dimensional, compact vector representation output by the second feature extraction network. Its dimensions are consistent with or similar to the image feature vector (e.g., 128 dimensions). It is a condensed representation of information from the original high-dimensional spectral data (e.g., hundreds of wavelength points) after dimensionality reduction and abstraction, retaining the molecular vibrational information most relevant to the chemical composition of lepidolite.

[0051] Traditional methods for analyzing flotation froth images rely on manually designed feature extraction rules. These include calculating bubble area distribution using image segmentation algorithms, extracting texture statistics (such as contrast, correlation, and uniformity) using gray-level co-occurrence matrices, and analyzing froth color features through color space conversion. These manual features have inherent drawbacks: First, their discriminative power is highly dependent on the designer's prior knowledge; froth morphology varies significantly across different mining areas, making it difficult to generalize a single set of manual features to different operating conditions. Second, manual feature extraction is independent of the prediction task and cannot be optimized for the ultimate goal of "grade prediction," potentially resulting in features with weak correlation to grade. Third, the design and parameter tuning of manual features require substantial manual intervention, hindering end-to-end automated monitoring.

[0052] For near-infrared spectral data, traditional methods typically employ linear dimensionality reduction techniques such as partial least squares regression and principal component regression to extract spectral features. However, the relationship between absorption peaks and gradations in spectral data is often non-linear (e.g., peak position shift, peak width variation, and multiple peak overlap), making it difficult for linear methods to effectively capture these non-linear characteristics.

[0053] This specific implementation employs a first and second feature extraction network based on deep learning. Its core advantage lies in its ability to automatically learn, in an end-to-end manner, the feature representations most relevant to the final prediction target (lithium mica grade) directly from the raw data, without requiring manual feature rule design. The inherent local connectivity and weight-sharing structure of convolutional neural networks make them particularly suitable for processing image data with spatial local correlation; one-dimensional convolutional neural networks are also suitable for processing spectral data with continuous wavelength correlation. The independent dual-branch design allows the two modalities to extract features separately under their respective most suitable network architectures, avoiding information loss caused by forcibly inputting images and spectra into the same network.

[0054] Step S2 addresses existing technical problems through deep learning mechanisms: First, a convolutional neural network serves as the first feature extraction network. Through shallow learning of edge spots, mid-layer learning of bubble contour texture, and deep learning of the hierarchical structure of mineralized luster features, it automatically filters out low-level interference such as illumination changes and extracts high-level semantic features related to grade, solving the problems of susceptibility to interference and poor generalization ability of artificial features. Second, a one-dimensional convolutional neural network or multilayer perceptron serves as the second feature extraction network. It automatically learns the nonlinear mapping relationship in spectral data, capturing the saturation relationship between Si-O-Si bond absorption peaks and grade, as well as the peak position shift caused by thallium, solving the problem that traditional linear spectral analysis methods are difficult to adapt to nonlinear relationships and component fluctuations. Third, the independent dual-branch design allows image and spectral features to be extracted to the same dimension (e.g., 128 dimensions) under their respective optimal architectures, providing high-quality, dimension-aligned input for subsequent joint sparse coding, overcoming the shortcomings of existing technologies that suffer from insufficient feature discrimination due to simple splicing and fusion.

[0055] Furthermore, the fused feature vector is input into a regression prediction network to output predicted values ​​of key components in the lithium mica flotation process, specifically including: The fused feature vector is used as input data and fed into the regression prediction network; The regression prediction network comprises, in sequence: a first fully connected layer, a first activation function layer, a second fully connected layer, a second activation function layer, and an output layer; The first fully connected layer maps the fused feature vector to a first intermediate feature vector, and the first activation function layer performs a nonlinear transformation on the first intermediate feature vector to output a second intermediate feature vector. The second fully connected layer maps the second intermediate feature vector to a third intermediate feature vector, the dimension of which is equal to the number of types of key components to be predicted; The second activation function layer performs a nonlinear transformation on the third intermediate feature vector to output a fourth intermediate feature vector; The output layer uses a linear activation function to directly map each element in the fourth intermediate feature vector to the predicted content value of the corresponding key component, and the predicted content value is a continuous value. The key components include at least one or more of the following: lepidolite grade, gangue ratio, or reagent residual concentration.

[0056] The specific structure and prediction process of the regression prediction network are as follows: Let the fused feature vector be... As input.

[0057] In a specific implementation, the first fully connected layer adopts the formula The input 512-dimensional fused feature vector is linearly transformed to a 256-dimensional hidden space to achieve feature dimensionality reduction and recombination. This is the weight matrix of the first fully connected layer. For bias vectors, This is the output vector; subsequently, the first activation function layer uses the ReLU function, i.e. Nonlinearity is introduced and sparse activation is implemented to enhance the network's expressive power and alleviate the gradient vanishing problem. The second fully connected layer uses the formula... The 256-dimensional features are further mapped to 128 dimensions, where For the weight matrix of the second layer, The bias vector for the second layer enables higher-level feature abstraction; the second activation function layer also uses the ReLU function. Nonlinear transformation is introduced again. The third fully connected layer uses the formula... Mapping 128-dimensional features to a 3-dimensional output space, where For the weight matrix of the third layer, The bias vector for the third layer is denoted as , and the three output dimensions correspond to the predicted values ​​of lepidolite grade, gangue ratio, and reagent residue concentration, respectively. The above layer-by-layer dimensionality reduction (512→256→128→3) design can smoothly compress information and retain key features. All hidden layers uniformly use the ReLU activation function to maintain computational efficiency and training stability, and the output layer adopts a linear activation function (identity mapping) to adapt to continuous regression prediction tasks.

[0058] Taking lepidolite grade prediction as an example, when the output value is 3.2, it means that the Li2O content in the concentrate is about 3.2%. The prediction result is displayed on the monitoring screen in real time. When the grade is lower than the set threshold (such as 2.5%), the system will automatically alarm to prompt the operator to adjust the dosage.

[0059] Example 2 refer to Figure 3 The present invention also provides a real-time monitoring device for the lithium mica flotation process based on portable infrared spectroscopy and machine learning, applicable to any of the aforementioned real-time monitoring methods for the lithium mica flotation process based on portable infrared spectroscopy and machine learning, comprising: The multimodal data acquisition module includes a portable near-infrared spectrometer and a near-infrared camera. The multimodal data acquisition module is used to simultaneously acquire near-infrared image data and near-infrared spectral data of the flotation foam surface, and output the image data and spectral data to the dual-branch feature extraction module. The dual-branch feature extraction module includes a first feature extraction network and a second feature extraction network. The first feature extraction network is used to extract image feature vectors from the image data, and the second feature extraction network is used to extract spectral feature vectors from the spectral data. The two feature vectors are then output to the joint sparse coding fusion module. The joint sparse coding fusion module is used to perform cross-modal feature fusion with the spectral feature vector as the dominant mode and the image feature vector as the auxiliary mode, and output the fused feature vector to the regression prediction module. The regression prediction module is used to output the predicted content of key components in the lithium mica flotation process based on the fused feature vector, and output the predicted value to the display unit or the flotation control system.

[0060] In this embodiment, the portable multimodal data acquisition module integrates a handheld near-infrared spectrometer (wavelength range 900-2500nm) and a miniature near-infrared camera (resolution 1280×720) into the same probe housing. A contact-type limiting ring at the probe tip ensures that the distance between the probe and the foam surface remains consistent during each acquisition. The dual-branch feature extraction module is implemented using an embedded GPU (such as NVIDIA Jetson Xavier) running a convolutional neural network algorithm, including a pre-trained ResNet18 image feature extraction network and a one-dimensional CNN spectral feature extraction network. The joint sparse coding fusion module is implemented on the embedded GPU and includes an overcomplete dictionary storage unit, a joint orthogonal matching pursuit solving unit, and a gated fusion unit. The dictionary size is preset to 1024 atoms. The regression prediction module is implemented using a fully connected network, outputting predicted values ​​for lepidolite grade, gangue ratio, and reagent residue concentration. The device is also equipped with a touchscreen display to show real-time prediction results and process adjustment suggestions. The entire device weighs approximately 1.5kg, has a battery life of approximately 4 hours, and is suitable for mobile monitoring operations in mine sites.

[0061] Furthermore, the joint sparse coding fusion module includes: an overcomplete dictionary storage unit for storing dictionary atoms constructed based on Si-O-Si bond absorption peaks; a joint orthogonal matching pursuit solving unit for solving joint sparse representation coefficients; and a gated fusion unit for dynamically adjusting the fusion weights of image features and spectral features according to the norm of the residual vector.

[0062] Example 3 The present invention also provides an electronic device, including: a processor, a transmitting device, an input device, an output device, and a memory. The processor may be implemented using a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit, or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory may be implemented using a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM), and is used to store computer program code. The computer program code includes computer instructions. When the processor executes the computer instructions, the electronic device executes a method as described in any of the above possible implementation methods.

[0063] Example 4 The present invention also provides a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor of an electronic device, cause the processor to perform a method as described in any of the above possible implementations.

[0064] In the description of this specification, the references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0065] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0066] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A real-time monitoring method for the flotation process of lithium mica based on portable infrared spectroscopy and machine learning, characterized in that, Includes the following steps: Acquire multimodal data of the flotation foam surface, the multimodal data including: near-infrared image data of the foam surface, and near-infrared spectral data from the same foam region as the near-infrared image data; The near-infrared image data is input into a first feature extraction network, which outputs an image feature vector; the near-infrared spectral data is arranged into a one-dimensional spectral vector by sampling points and then input into a second feature extraction network, which outputs a spectral feature vector. The image feature vector and the spectral feature vector are input into a joint sparse coding layer. The joint sparse coding layer uses the spectral feature vector as the dominant mode and the image feature vector as the auxiliary mode. Based on the feature channels corresponding to the Si-O-Si bond absorption peaks in the spectral feature vector, an overcomplete dictionary is constructed. The joint orthogonal matching pursuit algorithm is used to solve for the joint sparse representation coefficients. The positions of the non-zero elements in the joint sparse representation coefficients correspond to the spatial positions in the image feature vector that share the same sparse mode as the spectral feature vector. The image feature vector is reconstructed according to the joint sparse representation coefficients, and the reconstructed image features are fused with the spectral feature vector to output a fused feature vector. The fused feature vector is input into the regression prediction network, which outputs the predicted content of key components in the lithium mica flotation process. The predicted content is then compared with a preset threshold to generate a feedback signal for adjusting the flotation process parameters.

2. The method for real-time monitoring of the lithium mica flotation process based on portable infrared spectroscopy and machine learning according to claim 1, characterized in that, The overcomplete dictionary is constructed as follows: The spectral feature channels corresponding to the Si-O-Si bond absorption peaks in the spectral feature vector are used as the basic components of the dictionary atoms. Each dictionary atom is jointly formed by the basic components and the local spatial patterns of the image feature vector. The absorption peak intensity of the Si-O-Si bond is positively correlated with the selection priority of the corresponding dictionary atom.

3. The real-time monitoring method for lithium mica flotation process based on portable infrared spectroscopy and machine learning according to claim 1, characterized in that, The joint sparse coding layer also utilizes the lattice vibration absorption peak shift caused by thallium in lepidolite to dynamically adjust the overcomplete dictionary; wherein, the absorption peak shift is used to distinguish between lepidolite with high thallium content and lepidolite with low thallium content, and the corresponding dictionary atoms in the overcomplete dictionary are adaptively updated according to the shift.

4. The real-time monitoring method for lithium mica flotation process based on portable infrared spectroscopy and machine learning according to claim 1, characterized in that, Also includes: The iteration termination condition of the joint orthogonal matching pursuit algorithm is coupled with the physical properties of the flotation foam: The iteration terminates when the energy of the residual vector is lower than a first preset threshold, which is negatively correlated with the signal-to-noise ratio of the spectral feature vector; wherein the signal-to-noise ratio is calculated in real time based on the ratio of the signal intensity of the band containing the Si-O-Si bond absorption peak in the near-infrared spectral data to the baseline noise. The iteration may terminate when the number of non-zero coefficients reaches a second preset threshold, which is positively correlated with the estimated number of bubbles in the flotation bubble image.

5. The real-time monitoring method for lithium mica flotation process based on portable infrared spectroscopy and machine learning according to claim 1, characterized in that, The fusion of the reconstructed image features and the spectral feature vector employs a gating mechanism: The gating coefficients are calculated based on the norm of the residual vector of the joint orthogonal matching pursuit algorithm and the spectral eigenvector. When the norm of the residual vector increases, the gating coefficient decreases, thereby reducing the weight of the reconstructed image features and increasing the weight of the spectral feature vector.

6. The real-time monitoring method for lithium mica flotation process based on portable infrared spectroscopy and machine learning according to claim 5, characterized in that, The gating fusion mechanism includes: defining a gating coefficient as... , The gating fusion mechanism is implemented through the following formula: ， ， Represents the spectral feature vector, This represents the reconstructed image features. This represents the fused feature vector. This represents the L2 norm of the residual vector of the joint orthogonal matching pursuit algorithm. Indicates the gating coefficient. Represents the gate weight matrix. Indicates the gating bias term. This represents the Sigmoid activation function. This represents element-wise multiplication. When the water film on the flotation foam surface is thick or the bubbles are dense, causing spectral signal attenuation, the gating coefficient... Decrease the weight of the attention fusion feature vector and increase the weight of the spectral feature vector; When the bursting of bubbles on the surface of flotation foam causes localized image distortion, the gating coefficient... Increase or decrease the weight of the spectral feature vector, and increase the weight of the attention fusion feature vector.

7. The real-time monitoring method for lithium mica flotation process based on portable infrared spectroscopy and machine learning according to claim 1, characterized in that, The first feature extraction network is a convolutional neural network, used to extract the bubble size distribution, morphological texture and spatial distribution features of mineralized areas of flotation foam; The spatial distribution characteristics of the mineralized region are directly indicated by the position of the non-zero elements in the joint sparse representation coefficients. The second feature extraction network is a one-dimensional convolutional neural network used to extract the molecular vibrational spectral fingerprint features of lepidolite and its associated minerals.

8. The method for real-time monitoring of the lithium mica flotation process based on portable infrared spectroscopy and machine learning according to claim 1, characterized in that, The fused feature vector is input into a regression prediction network, which outputs predicted values ​​of the content of key components during the lithium mica flotation process, specifically including: The fused feature vector is used as input data and fed into the regression prediction network; The regression prediction network comprises, in sequence: a first fully connected layer, a first activation function layer, a second fully connected layer, a second activation function layer, and an output layer; The first fully connected layer maps the fused feature vector to a first intermediate feature vector, and the first activation function layer performs a nonlinear transformation on the first intermediate feature vector to output a second intermediate feature vector. The second fully connected layer maps the second intermediate feature vector to a third intermediate feature vector, the dimension of which is equal to the number of types of key components to be predicted; The second activation function layer performs a nonlinear transformation on the third intermediate feature vector to output a fourth intermediate feature vector; The output layer uses a linear activation function to directly map each element in the fourth intermediate feature vector to the predicted content value of the corresponding key component, and the predicted content value is a continuous value. The key components include at least one or more of the following: lepidolite grade, gangue ratio, or reagent residual concentration.

9. A real-time monitoring device for the lithium mica flotation process based on portable infrared spectroscopy and machine learning, applied to the real-time monitoring method for the lithium mica flotation process based on portable infrared spectroscopy and machine learning as described in any one of claims 1-8, characterized in that, include: The multimodal data acquisition module includes a portable near-infrared spectrometer and a near-infrared camera. The multimodal data acquisition module is used to simultaneously acquire near-infrared image data and near-infrared spectral data of the flotation foam surface, and output the image data and spectral data to the dual-branch feature extraction module. The dual-branch feature extraction module includes a first feature extraction network and a second feature extraction network. The first feature extraction network is used to extract image feature vectors from the image data, and the second feature extraction network is used to extract spectral feature vectors from the spectral data. The two feature vectors are then output to the joint sparse coding fusion module. The joint sparse coding fusion module is used to perform cross-modal feature fusion with the spectral feature vector as the dominant mode and the image feature vector as the auxiliary mode, and output the fused feature vector to the regression prediction module. The regression prediction module is used to output the predicted content of key components in the lithium mica flotation process based on the fused feature vector, and output the predicted value to the display unit or the flotation control system.

10. A real-time monitoring device for lithium mica flotation process based on portable infrared spectroscopy and machine learning according to claim 9, characterized in that, The joint sparse coding fusion module includes: an overcomplete dictionary storage unit for storing dictionary atoms constructed based on Si-O-Si bond absorption peaks; a joint orthogonal matching pursuit solving unit for solving joint sparse representation coefficients; and a gated fusion unit for dynamically adjusting the fusion weights of image features and spectral features according to the norm of the residual vector.

Images