Lithium ore multi-component content prediction and detection optimization system based on machine learning

Abstract

The invention relates to the technical field of neural networks and machine learning, in particular to a lithium ore multi-component content prediction and detection optimization system based on machine learning. The system comprises a multispectral data fusion and preprocessing part, a correction and interference elimination part and a machine learning model prediction part. The multispectral data fusion and preprocessing part is used for performing fusion and normalization processing on the first light intensity signal and the second light intensity signal to obtain a fusion signal; the correction and interference elimination part is used for performing spectral line overlapping interference elimination processing on the fusion signal to obtain an interference elimination signal, and performing matrix effect correction on the interference elimination signal to obtain a correction signal; the machine learning model prediction part is used for performing feature extraction on the correction signal by using a preset machine learning model to obtain features; based on the feature response function, an initial content of each element is predicted. According to the invention, high-precision and intelligent detection of multiple components of the lithium ore is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of neural networks and machine learning, and in particular relates to a multi-component content prediction and detection optimization system in lithium ore based on machine learning. Background Art

[0002] As one of the main sources of lithium resources, lithium ore has complex multi-component characteristics, which determines that accurate and rapid component detection is of great significance for the efficient utilization of lithium resources. In the analysis of lithium ore components, traditional chemical analysis methods, such as gravimetric analysis and titration, although with high accuracy, can no longer meet the needs of modern industry for rapid detection due to complex operations, long time consumption and large sample consumption. Therefore, spectral analysis technology has gradually become the main method for lithium ore component detection. At present, inductively coupled plasma spectroscopy (ICP-OES) and X-ray fluorescence spectroscopy (XRF) are the two most widely used spectral detection technologies, but in actual use, these technologies still have some difficult-to-overcome technical bottlenecks.

[0003] ICP-OES uses high-temperature plasma to excite atoms or ions in the sample, causing them to emit spectra of characteristic wavelengths, and uses light intensity signals to characterize the element content. This method is known for its high sensitivity and ability to simultaneously analyze multiple elements, and is particularly suitable for the detection of low-concentration elements (such as lithium). The advantages of ICP-OES also lie in its high linear range and good reproducibility. However, ICP-OES also has obvious technical bottlenecks: the spectral emission wavelengths of different elements may be similar, resulting in signal overlap, especially in complex samples, where interference seriously affects the accuracy of quantitative analysis. The composition of the sample matrix can significantly change the excitation efficiency of the plasma, resulting in a nonlinear response of the element signal, thereby reducing the reliability of the measurement results. Traditional ICP-OES analysis relies on linear regression models, which are difficult to adapt to nonlinear interference and complex matrix effects in multi-component systems.

[0004] X-ray fluorescence spectrometry (XRF) XRF bombards the sample with high-energy X-rays to excite its internal atoms to produce characteristic fluorescence signals, which are used to characterize the type and content of elements. The XRF method has the advantages of non-destructiveness, easy operation, and fast detection speed, especially in the quantitative analysis of high-concentration elements. However, XRF also has the following technical limitations: Due to the weak intensity of the fluorescence signal, XRF is difficult to achieve high-precision detection of low-concentration elements such as lithium. The thickness, surface roughness and density distribution of the sample have a significant impact on the intensity and distribution of the fluorescence signal, causing the test results to deviate from the true value. Similar to ICP-OES, the fluorescence signal spectra of different elements may overlap within the wavelength range, thereby introducing interference. Summary of the invention

[0005] The main purpose of the present invention is to provide a multi-component content prediction and detection optimization system in lithium ore based on machine learning, which realizes high-precision and intelligent detection of multi-component lithium ore. Its technical advantages are not only reflected in the accuracy of spectral signal processing, but also in the adaptability to multi-component complex systems and the high efficiency of concentration prediction.

[0006] The technical solution of the present invention is achieved in this way: A prediction and detection optimization system for multi-component content in lithium ore based on machine learning, the system comprising: a multi-spectral data fusion and preprocessing part, a correction and interference elimination part and a machine learning model prediction part; the multi-spectral data fusion and preprocessing part is used to measure the emission spectrum of the lithium ore multi-element sample by inductively coupled plasma spectroscopy to obtain the first light intensity signal of each element, use X-ray fluorescence spectroscopy to measure the fluorescence signal of the lithium ore multi-element sample to obtain the second light intensity signal of each element, fuse and normalize the first light intensity signal and the second light intensity signal to obtain a fused signal; wherein, the inductively coupled plasma spectroscopy is used to measure the emission spectrum of the lithium ore multi-element sample by inductively coupled plasma spectroscopy to obtain the first light intensity signal of each element, and the fluorescence signal of the lithium ore multi-element sample is measured by X-ray fluorescence spectroscopy to obtain the second light intensity signal of each element. When the emission spectrum of a lithium ore multi-element sample is measured by a spectroscopic method and the fluorescence signal of the lithium ore multi-element sample is measured by an X-ray fluorescence spectroscopic method, the same observation distance, the same resolution, the same wavelength and the same energy level are used; the correction and interference elimination part is used to eliminate the interference of spectral line overlap on the fusion signal to obtain an interference elimination signal, and to correct the interference elimination signal for matrix effect to obtain a correction signal; the machine learning model prediction part is used to extract features from the correction signal using a preset machine learning model to obtain features; a feature response function is constructed based on the features, and based on the feature response function, the initial content of each element is predicted.

[0007] Furthermore, it is assumed that the emission spectrum of the lithium ore multi-element sample is measured using inductively coupled plasma spectroscopy to obtain the first Elements in wavelength In the The first light intensity signal of the spectral line is ; Assume that the emission spectrum of lithium ore multi-element sample is measured by inductively coupled plasma spectroscopy, and the first Elements at wavelength In the The first light intensity signal of the spectral line is ; The first light intensity signal and the second light intensity signal are fused and normalized by the following formula to obtain a fused signal : ; in, Indicates Elements in wavelength The fusion signal at The following table is the index for integers; For the The total number of spectral lines of an element; is the base of natural logarithms; For the The characteristic energy of an element; For the The element in The spectral line energy of each spectral line; is the Boltzmann constant, is temperature; For the The element in The absorption coefficient of each spectral line; For the The atomic number of the element; For the The density of the element; For the The mass absorption coefficient of the element; is the average thickness of multi-element samples of lithium ore; For the The atomic weight of an element; is the surface roughness of multi-element samples of lithium ore; For the The excited state temperature of an element; For the The temperature coefficient of the fluorescence yield of the element; The wavelength of the inductively coupled plasma spectrometry The response function at ; The wavelength of X-ray fluorescence spectroscopy The response function at .

[0008] Furthermore, the correction and interference elimination part eliminates the spectral line overlap interference of the fused signal by the following formula to obtain an interference elimination signal: ; in, ; Indicates The fusion signal of the elements is The first element other than All elements are matrix elements; is the number of elements in the multi-element sample of lithium ore; Indicates Elements in wavelength Interference cancellation signal at For the The element The spectral line interference coefficient of the element is the set value; For the The element The central wavelength of the interference spectrum of the element; For the The element The half width of the interference line of the element; Indicates Elements in wavelength The interference cancellation signal at the

[0009] Furthermore, the correction and interference elimination part performs matrix effect correction on the interference elimination signal by the following formula to obtain a correction signal: ; in, Indicates Elements in wavelength The correction signal at For the Elements in wavelength Interference cancellation signal at For the The element The matrix influence coefficient of the element is the set value; For the The characteristic energy of an element; is Avogadro's constant; is the thermodynamic temperature of the plasma in ICP spectroscopy; For the The mass of an element.

[0010] Furthermore, in the machine learning model prediction part, the preset machine learning model is a multi-level feature extraction network; the correction signal is subjected to feature extraction using the preset machine learning model through the following formula to obtain features: ; in, For the Spectral contribution coefficient of the layer; is the feature activation function; The number of layers of the multi-level feature extraction network; For the The feature extraction function of the layer is as follows: ; in, is the lower limit of wavelength; is the upper limit of wavelength; For the The weight function of the layer; is the feature extraction kernel function; is the wavelength integration variable; is a nonlinear transfer function; is the spectral resolution.

[0011] Furthermore, the feature extraction kernel function The formula is as follows: ; in, is the central wavelength; For the Spectral broadening parameters of the layer; is the plasma emission intensity of inductively coupled plasma spectroscopy; is the X-ray emission intensity of X-ray fluorescence spectrometry; is the optical path length.

[0012] Further, Layer weight function Use the following formula to express it: ; in, is Planck's constant; is the transition frequency; is the upper energy level degeneracy; is the lower energy level degeneracy; is the probability of spontaneous radiation transition; is the upper energy level energy; is the Boltzmann constant.

[0013] Furthermore, the feature response function is constructed according to the feature through the following formula: ; in, For the Characteristic response function of an element.

[0014] Furthermore, the initial content of each element is predicted based on the characteristic response function using the following formula: ; in, For the The initial content of the element; For the The mass of the element; For the The mass of an element.

[0015] The multi-component content prediction and detection optimization system in lithium ore based on machine learning of the present invention has the following beneficial effects: The present invention effectively compensates for the limitations of single spectral technology by fusing multi-spectral data. ICP-OES excels in its high sensitivity to low-concentration elements, while XRF has advantages in detecting high-concentration elements. The two complement each other in signal characteristics and scope of application. The present invention integrates the advantages of the two spectral technologies through multi-spectral fusion, significantly improving the system's ability to detect multiple elements in lithium ore. The construction of the fused signal fully considers the characteristic differences between the two spectral signals, eliminates the inconsistency of the signal scale through normalization and weighted fusion, and makes the fused signal more unified and complete. This method not only improves the characterization ability of the target element signal, but also expands the scope of application of the system, laying a solid foundation for subsequent feature extraction and content prediction.

[0016] The present invention significantly improves the accuracy and stability of spectral signals by eliminating spectral line overlap interference and correcting matrix effects. In the complex system of lithium ore, spectral line overlap and matrix effect are the two main factors leading to signal distortion. The present invention accurately models the overlapping effect of matrix elements on the spectral lines of target elements through a dynamic interference correction formula, thereby separating the independent signal of the target element.

[0017] In addition, the matrix effect correction formula based on the exponential model comprehensively considers the parameters such as the spectral intensity, characteristic energy and plasma temperature of the matrix elements, and realizes the dynamic adjustment of the complex matrix effect of the sample. Compared with the existing empirical correction method, the correction method of the present invention has a stronger physical basis and adaptability, thereby effectively reducing the interference of the complex matrix on the target element signal, so that the system can maintain high-precision detection under various complex sample conditions.

[0018] The present invention introduces a multi-level feature extraction network, which significantly improves the expression ability of spectral features and the accuracy of concentration prediction. The feature extraction network extracts deep features from the correction signal through a hierarchical design to capture the complex relationship between the target element and the matrix element. The spectral signal is locally convolved by a kernel function, and the contribution of the hierarchical features is dynamically adjusted in combination with a weight function. The present invention can refine the spectral features at different scales and levels. This deep learning method can not only capture the local characteristics of the spectral signal, but also comprehensively consider the global features, providing a more comprehensive and accurate feature input for the concentration prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A schematic diagram of the system structure of a system for predicting and detecting multi-component content in lithium ore based on machine learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In the following description, specific details such as specific system structures, interfaces, and technologies are provided for illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present invention. However, it should be clear to those skilled in the art that the embodiments of the present invention may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the embodiments of the present invention.

[0021] Example 1, reference Figure 1 :A prediction and detection optimization system for multi-component content in lithium ore based on machine learning, the system comprising: a multi-spectral data fusion and preprocessing part, a correction and interference elimination part and a machine learning model prediction part; the multi-spectral data fusion and preprocessing part is used to measure the emission spectrum of the lithium ore multi-element sample using inductively coupled plasma spectroscopy to obtain a first light intensity signal of each element, use X-ray fluorescence spectroscopy to measure the fluorescence signal of the lithium ore multi-element sample to obtain a second light intensity signal of each element, fuse and normalize the first light intensity signal and the second light intensity signal to obtain a fused signal; wherein, the inductively coupled plasma spectroscopy is used to measure the emission spectrum of the lithium ore multi-element sample using inductively coupled plasma spectroscopy to obtain a first light intensity signal of each element, use X-ray fluorescence spectroscopy to measure the fluorescence signal of the lithium ore multi-element sample to obtain a second light intensity signal of each element, and fuse and normalize the first light intensity signal and the second light intensity signal to obtain a fused signal. When the emission spectrum of a lithium ore multi-element sample is measured by a spectroscopic method and the fluorescence signal of the lithium ore multi-element sample is measured by an X-ray fluorescence spectroscopic method, the same observation distance, the same resolution, the same wavelength and the same energy level are used; the correction and interference elimination part is used to eliminate the interference of spectral line overlap on the fusion signal to obtain an interference elimination signal, and to correct the interference elimination signal for matrix effect to obtain a correction signal; the machine learning model prediction part is used to extract features from the correction signal using a preset machine learning model to obtain features; a feature response function is constructed based on the features, and based on the feature response function, the initial content of each element is predicted.

[0022] Specifically, inductively coupled plasma spectroscopy (ICP-OES) uses high-temperature plasma to excite atoms or ions of elements in the sample to a high-energy state. When these particles fall back to a low-energy state, they emit a spectrum with a specific wavelength. The intensity of this emission spectrum is positively correlated with the concentration of the element and can provide characteristic light intensity signals of multiple elements in lithium ore. However, the accuracy of the signal of ICP-OES depends on the matrix environment of the sample. Especially in complex samples, matrix effects and spectral line overlap interference will reduce the reliability of the signal. In addition, the detection sensitivity of ICP-OES for high-concentration elements is limited, which is a limitation of its physical principle. On the other hand, X-ray fluorescence spectroscopy (XRF) bombards the sample with high-energy X-rays to induce the transition of the inner-shell electrons of atoms in the sample. When the high-energy electrons return to a low-energy state, they release fluorescent radiation, and the energy and intensity of this fluorescence correspond to the characteristics and concentration of the element, respectively. XRF is less sensitive to matrix effects and has significant advantages in detecting high-concentration elements. However, XRF is relatively insensitive to low-concentration elements, and the resolution of the signal may be affected by background noise. Therefore, ICP-OES and XRF are complementary in physical measurement: ICP-OES is suitable for the detection of low-concentration elements, while XRF is more suitable for the quantitative analysis of high-concentration elements. The core principle of multispectral data fusion is to utilize the complementarity of the two spectral signals and integrate them into a new composite signal to retain the characteristics of both ICP-OES and XRF. During the measurement process, the system ensures the comparability of the two spectral data by strictly controlling key parameters such as spectral resolution, observation distance, energy level and wavelength. The consistency of the observation distance ensures that the excitation light and fluorescence of the sample are recorded at the same measurement scale, avoiding signal attenuation caused by distance changes; the matching of wavelength and energy level enables the spectral signals of the two methods to be compared and fused in the same feature space. Normalization is a key step in multispectral fusion. Its principle is to adjust the amplitude of each spectral signal and convert it into a dimensionless form to eliminate the deviation caused by instrument response or environmental changes. The mathematical basis of this processing method is to remap the signal value to a uniform interval so that it has a similar scale distribution while keeping its physical meaning unchanged in the signal characteristics. In this way, the emission spectrum intensity from ICP-OES and the fluorescence intensity from XRF can be regarded as different dimensions of the same set of data features, providing a basis for subsequent fusion operations. Signal fusion is implemented using multivariate statistical methods, and its physical basis is to map the data dimensions of the two signals to a unified space with lower redundancy by recombining the feature space. For example, the principal component analysis (PCA) method calculates the covariance matrix between the signals and extracts the direction with the largest variance in the two signals, thereby reducing the redundancy of the data while retaining the most important feature information.This process is similar to projecting the original high-dimensional spectral signal into a low-dimensional feature subspace, so that the two signals are more closely combined in information expression. The fused signal is not just a simple linear superposition, but a deep reorganization of information. In the fused signal, the spectral advantages of ICP-OES and the spectral advantages of XRF are amplified simultaneously. For example, some elements have different response characteristics in the two spectra. ICP-OES may show stronger sensitivity at specific wavelengths, while XRF may provide more accurate background correction capabilities. The fused signal organically combines these characteristics to form a spectral feature with wider applicability within the element concentration range. Overall, the principle of multi-spectral data fusion and preprocessing is based on the basic characteristics of physical measurement methods. Through data consistency correction, normalization processing and feature space mapping, efficient signal reorganization is achieved, providing a high-quality input data format for the prediction of multi-element content in lithium ore.

[0023] In the process of measuring multi-spectral signals, due to the complexity of elements in lithium ore samples, the spectral characteristics of some elements often overlap in a similar wavelength range, and this phenomenon is called spectral line overlap interference. The formation mechanism of spectral line overlap comes from the similarity of the electronic transition characteristics of different elements. When the resolution of the spectrometer is not enough to completely separate these characteristic wavelengths, the superposition contribution of multiple elements will appear in the detection signal, resulting in the light intensity signal of a single element cannot be directly and accurately separated. The present invention uses a mathematical algorithm to deconvolve the spectral lines in the fusion signal to separate the characteristic spectral lines of each element. The core principle of deconvolution is to model the superposition signal as a weighted combination of multiple single component signals based on the known characteristics of the spectrum of each element, and solve these weighting coefficients through an optimization algorithm. Such a processing method can accurately separate the characteristic signals of each element and eliminate the influence of spectral line overlap. In the process of measuring multi-spectral signals, due to the complexity of elements in lithium ore samples, the spectral characteristics of some elements often overlap in a similar wavelength range, and this phenomenon is called spectral line overlap interference. The formation mechanism of spectral line overlap comes from the similarity of the electronic transition characteristics of different elements. When the resolution of the spectrometer is not enough to completely separate these characteristic wavelengths, the superposition contribution of multiple elements will appear in the detection signal, resulting in the inability to directly and accurately separate the light intensity signal of a single element. The present invention uses a mathematical algorithm to deconvolve the spectral lines in the fusion signal to separate the characteristic spectral lines of each element. The core principle of deconvolution is to model the superposition signal as a weighted combination of multiple single component signals based on the known characteristics of the spectrum of each element, and solve these weighting coefficients through an optimization algorithm. Such a processing method can accurately separate the characteristic signals of each element and eliminate the influence of spectral line overlap. In the correction process, noise removal of the fusion signal is also a key link. Spectral signals often contain random noise, which comes from electronic noise of the instrument, environmental background interference, and sample inhomogeneity. These noise components usually exist in high-frequency form, which will significantly reduce the signal-to-noise ratio of the signal. The present invention uses signal filtering and noise reduction technology to smooth the signal through mathematical methods, thereby extracting the low-frequency effective components in the signal. The selection of filtering algorithm depends on the characteristics of spectral signals. For example, wavelet transform can decompose the signal into multi-scale components, extract low-frequency signals as the main features, and remove high-frequency noise. Such processing not only improves the stability of the signal, but also enhances the significance of the characteristic components.

[0024] As a carrier of multi-component information of samples, spectral signals contain the mixed characteristics of target element concentration, background noise and interfering components. Therefore, in the prediction part of machine learning model, it is necessary to extract features of correction signals first. The core of this process is to identify the features that are most sensitive to changes in target element content from multidimensional spectral data. Machine learning models automatically learn these features from training samples through data-driven methods, without relying on manually set rules. The mathematical basis of feature extraction lies in the distribution characteristics of signals and the changing rules of element concentrations in sample space. For example, principal component analysis (PCA) can extract the main direction of change in data by calculating signal variance, while deep learning models can automatically generate high-order features through hierarchical nonlinear mapping of neural networks. These extracted features can be regarded as low-dimensional expressions of spectral signals, which not only retain the main information of the signal but also eliminate irrelevant redundant features. After extracting features, the model describes the relationship between spectral features and element concentrations by establishing feature response functions. The construction of this function is the key to the entire prediction part. Its principle is to enable the model to output the predicted concentration of each element based on the input signal through regression analysis or classifier training. Traditional linear regression models are difficult to adapt to the complex nonlinear characteristics of spectral data, while the present invention adopts more flexible nonlinear models, such as support vector regression (SVR), random forest regression (RFR) or deep neural network (DNN). These models can handle high-dimensional, nonlinear and noisy data, and show strong fitting and generalization capabilities in the mapping process of spectral signals and concentrations. For example, support vector regression maps data to high-dimensional space through kernel functions, and searches for the optimal hyperplane between spectral features and concentrations in high-dimensional space; deep neural networks capture complex relationships in data through multi-layer perceptron structures and generate highly abstract feature response functions. The training process of machine learning models relies on a large amount of sample data, and its core goal is to minimize the error between the predicted value and the actual value of the model by optimizing the loss function. In this process, the model parameters are continuously adjusted through the back-propagation algorithm to improve the accuracy of the prediction. The choice of loss function directly affects the performance of the model. For nonlinear and noisy data such as spectral signals, mean square error (MSE) is often used as the objective function. In addition, to avoid the decline in prediction performance caused by overfitting of the model, the present invention introduces regularization technology, such as L1 regularization, which can sparse the parameters of the model, and L2 regularization, which can reduce the fluctuation of the parameters, thereby improving the robustness of the model. After training, the performance of the model on unknown samples is used for verification and testing to ensure that it has sufficient generalization ability. During the prediction process, the machine learning model processes the corrected spectral signal in real time to generate a concentration prediction value of the target element.The inference speed of the model depends on its complexity and the optimization level of the implementation framework. For example, deep learning models can achieve efficient batch prediction when GPU acceleration is used, while lightweight models such as random forests show high real-time performance under limited computing resources. The final result of the prediction part is an estimate of the initial content of each element in the lithium ore. These estimates not only provide accurate data for ore composition analysis, but also serve as an important basis for composition control in industrial production.

[0025] Example 2: Suppose that the emission spectrum of a multi-element lithium ore sample is measured using inductively coupled plasma spectroscopy to obtain Elements in wavelength In the The first light intensity signal of the spectral line is ; Assume that the emission spectrum of lithium ore multi-element sample is measured by inductively coupled plasma spectroscopy, and the first Elements in wavelength In the The first light intensity signal of the spectral line is ; The first light intensity signal and the second light intensity signal are fused and normalized by the following formula to obtain a fused signal : ; in, Indicates Elements in wavelength The fusion signal at The following table is the index for integers; For the The total number of spectral lines of an element; is the base of natural logarithms; For the The characteristic energy of an element; For the The element in The spectral line energy of each spectral line; is the Boltzmann constant, is temperature; For the The element in The absorption coefficient of each spectral line; For the The atomic number of the element; For the The density of the element; For the The mass absorption coefficient of the element; is the average thickness of multi-element samples of lithium ore; For the The atomic weight of an element; is the surface roughness of multi-element samples of lithium ore; For the The excited state temperature of an element; For the The temperature coefficient of the fluorescence yield of the element; The wavelength of the inductively coupled plasma spectrometry The response function at ; The wavelength of X-ray fluorescence spectroscopy The response function at .

[0026] Specifically, the formula is based on the characteristics of ICP-OES and XRF signals, and generates a fusion signal through a weighted combination of signal intensity, response function, and physical parameters. The core feature of ICP-OES is its high sensitivity to low-concentration elements and the emission spectrum produced by atoms excited by high-temperature plasma. The intensity of these spectral signals is directly related to the concentration of the element and its specific energy state. The first part of the formula introduces a weight factor, which uses the Boltzmann distribution to describe the proportion of different energy states of the target element in lithium ore. In the formula is the key to the Boltzmann distribution, indicating that at a given temperature This description can effectively characterize the temperature dependence of the ICP spectral signal, and the absorption coefficient term The spectral signal is further corrected to reflect the absorption loss of photons when they propagate in the sample. At the same time, the formula also takes into account the atomic properties of the elements themselves, such as This reflects the relationship between the emission spectrum intensity and the square of the atomic number, which is a basic physical law determined by the element's ionization energy and transition probability. In addition, the formula takes the temperature factor It is listed separately to describe the overall effect of excited state temperature on signal intensity. In particular, when the sample is under different thermal excitation conditions, dynamic correction of the signal is necessary. Finally, by introducing , that is, the spectrometer at a specific wavelength The response function at further refines the measurement characteristics of the signal intensity in the optical system, making it possible to quantify the ICP signal under actual measurement conditions.

[0027] For the XRF signal, the formula uses another set of physical parameters and characteristic descriptions to express the nature of the fluorescence signal. The detection mechanism of XRF relies on high-energy X-rays to excite atoms in the sample, causing them to produce specific fluorescence radiation. This fluorescence signal is not only related to the element concentration, but also affected by the physical properties of the sample such as density, thickness and surface roughness. The formula is and describes the attenuation effect of density and sample thickness on XRF signals, where is the mass absorption coefficient, which reflects the absorption capacity of the sample material to X-rays, and is the average thickness of the sample, indicating the energy loss of the fluorescence signal during the propagation process. These factors comprehensively reflect the matrix effect of the sample in XRF detection. In addition, the formula is calculated by the fluorescence yield Temperature coefficient To adjust the signal intensity, the temperature is adjusted to correct the effect of fluorescence efficiency. This adjustment is important for enhancing the physical consistency of the signal and reducing external interference factors. In order to unify the two spectral signals, the formula uses normalization processing to separate the ICP-OES signal and the XRF signal with their specific physical constants (such as atomic weight). , surface roughness The goal of normalization is to eliminate the signal scale differences caused by the different physical properties of the two spectral methods, so that they can express the spectral characteristics on a unified scale. In this way, the ICP-OES signal and the XRF signal can be combined with the same weight to generate a fusion signal. The two parts of the formula are combined in the form of a sum, which represents a weighted calculation of the comprehensive contribution of each element on all possible spectral lines, and the spectral response function and The core idea of ​​this fusion formula is to maximize the physical advantages of ICP-OES and XRF, while solving the characteristic differences between spectral methods by adjusting parameters and allocating weights. The final generated fusion signal It has a high signal-to-noise ratio and high sensitivity, and can cover the multi-dimensional spectral characteristics of the target elements, providing input data with rich information for subsequent machine learning models. The construction of the formula not only relies on the physical basis of the spectral technology itself, but also comprehensively considers the material properties and experimental conditions of the lithium ore samples, and realizes the optimized expression of the spectral signal through mathematical methods, thus providing important support for the accurate detection of multiple components in lithium ore.

[0028] Characteristic Energy It is an inherent physical property of the target element, which is usually closely related to the electronic transition energy level of the element. It defines the energy difference required for an atom or ion to transition from one energy level to another under specific external energy excitation. In spectral measurement, this characteristic energy corresponds to the characteristic wavelength or frequency of the target element, which directly affects the distribution of the spectral signal. The first step is to refer to a standard spectral database, such as the atomic spectral database of the National Institute of Standards and Technology (NIST), which provides transition information for a large number of elements, including characteristic energies and related transition probabilities. This standardized database is highly accurate and can provide a reliable basis for the analysis of the present invention. Spectral line energy The acquisition of is more specific, it is the energy value corresponding to the emission or absorption of photons by the target element under experimental conditions. Unlike characteristic energy, spectral line energy may be affected by the external environment (such as plasma temperature, electron density), so its acquisition requires a combination of experimental measurement and theoretical model. In the present invention, the measurement of spectral line energy is completed by a high-resolution spectrometer, and the spectral intensity distribution of the target element at different wavelengths is recorded by inductively coupled plasma spectroscopy (ICP-OES) or X-ray fluorescence spectroscopy (XRF). By analyzing the peak position information in the spectrum, the wavelength corresponding to the spectral line can be accurately determined. , and then use the energy-wavelength relationship Calculate the energy of the spectrum line .in, is Planck's constant, The speed of light ensures the consistency of energy calculation with physical laws. In order to further calibrate the measured spectral line energy, the experiment also needs to correct the spectral resolution, noise level and instrument drift of the system. These factors may cause the actual measured spectral line position to deviate from the theoretical value. The present invention uses the instrument response function and line width correction parameters The experimental measurements are calibrated to ensure the accuracy of the spectral line energy. At the same time, the use of known standard samples (such as single element standard solutions) for benchmark correction can further reduce the systematic error. At the theoretical level, quantum mechanical calculations provide a supplement for the verification of characteristic energies and spectral line energies. Based on the electronic structure of the target element, quantum chemistry software can accurately calculate the transition energy by solving the Schrödinger equation. This result can not only be mutually verified with the experimental value, but also provide supplementary information for transition paths not covered in the spectral database. For example, for some rare impurity elements in lithium ore, theoretical calculations can quickly provide their characteristic energies and potential spectral line positions.

[0029] Plasma temperature It is one of the most important parameters in inductively coupled plasma spectroscopy (ICP-OES). In the plasma, the sample is excited by high temperature, and the atoms or ions release characteristic photons when they transition from high energy state to low energy state. These photons form the emission spectrum of the target element. It directly determines the excitation and ionization efficiency of atoms in the plasma. According to the Boltzmann distribution, temperature is closely related to the distribution of particles in different energy states. The higher the temperature, the more particles in the high energy state there are, and the greater the probability of characteristic energy level transitions. Therefore, temperature not only affects the intensity distribution of the spectral line, but also controls the emission characteristics within a specific energy range. The temperature in the formula It usually appears in the Boltzmann distribution and the Saha equation, which are used to describe energy level distribution and ionization equilibrium, respectively. For example, the exponential term of the Boltzmann distribution is reflects the modulation effect of temperature on the probability of high energy level occupation, where is the upper energy level, is the Boltzmann constant. This relationship shows that higher temperatures can significantly increase the occupation number of high-energy states, thereby enhancing the signal intensity of high-energy transitions. In X-ray fluorescence spectroscopy (XRF), although temperature has little direct effect on the excitation process, it may indirectly affect the detection of fluorescence signals through changes in the physical state and density of the sample. In particular, when the sample thickness or surface state is affected by temperature changes, the intensity and distribution of the fluorescence signal may fluctuate. By introducing the temperature parameter into the XRF signal correction formula, these effects can be effectively reduced.

[0030] The absorption coefficient of a spectral line is usually expressed as , which is used to describe the photon absorption intensity of the target element at a specific spectral line (defined by wavelength or frequency). The physical definition of this coefficient is the ratio of the sample to the light intensity absorbed per unit path length, and its value is related to the concentration of the target element in the sample, the transition probability of a specific energy level, and the initial intensity of the photon. Specifically, when a photon passes through a sample, its intensity With path length The change of can be expressed by an exponential decay relationship: . Spectral absorption coefficient It is the quantitative parameter of this attenuation rate. In order to quantify the spectral line absorption coefficient, it is usually necessary to combine experimental measurements and theoretical models. Experimentally, the absorption spectrum of the sample at different wavelengths is recorded by a high-resolution spectrometer to identify the absorption peak and width of the characteristic spectral line of the target element. By comparing the change in light intensity before and after absorption, the absorption intensity can be calculated, and then combined with the path length of the sample to obtain the spectral line absorption coefficient. In addition, theoretical models (such as transition probability calculations in quantum mechanics) provide predictions for specific transition absorption coefficients. These theoretical results are usually included in standard spectral databases, such as the NIST database, which can be used to calibrate with experimental values ​​or supplement missing data.

[0031] Mass absorption coefficient is another key parameter that describes the absorption capacity of the target element or the sample as a whole for photons within a certain mass range. It is defined as the absorption rate of photons per unit mass of the sample, usually expressed in It is a value closely related to the composition and spectral characteristics of the material, and is affected by the density, atomic structure and energy (or wavelength) of the target element. The calculation formula for the mass absorption coefficient is: ,in is the linear absorption coefficient, which indicates the absorption intensity per unit length, is the density of the sample. The quantification of the mass absorption coefficient also combines experimental and theoretical methods. Experimentally, it can be directly measured by the transmission method: using a sample of known thickness and density, the intensity of the incident light and the transmitted light is determined, and the absorbance is calculated based on the change in light intensity to obtain the mass absorption coefficient. In theory, the mass absorption coefficient can be decomposed and calculated through the element's photoelectric effect, Compton scattering, and the transition absorption mechanism of photons of a specific wavelength. These calculation results are also stored in the spectral database as a verification benchmark for experimental measurements. In the application of the present invention, the spectral line absorption coefficient and the mass absorption coefficient are not only important parameters for describing the spectral signal, but also directly affect the results of spectral correction and concentration prediction. For example, in the interference correction formula, the spectral line absorption coefficient It is used to correct the spectral overlap effect of target elements and matrix elements, and the mass absorption coefficient Used to calculate sample thickness The influence on signal propagation. These parameters provide a physical basis for signal correction by accurately describing the spectral attenuation process.

[0032] In the present invention, atomic weight, surface roughness and excited state temperature are important parameters for lithium ore component detection and spectral signal processing. These parameters each reflect the basic physical properties of the element, the surface state of the sample, and the excitation conditions of the target element in the plasma. Their acquisition method relies on the combination of standard databases, experimental measurements and theoretical calculations, and provides key support for the correction, feature extraction and concentration prediction of spectral analysis through accurate quantification.

[0033] Atomic weight is a basic property of an element, defined as the mass ratio of an atom relative to the carbon-12 nucleus, and its value can be directly consulted from standardized chemical databases. For example, the International Union of Pure and Applied Chemistry (IUPAC) provides standard values ​​for the atomic weights of all known elements. These values ​​are usually based on the weighted average of the abundance of different isotopes in nature, and are therefore highly representative of the characteristics of actual samples. In the present invention, atomic weight It is used for normalization in concentration prediction formulas, especially when calculating the mass ratio of the target element to the matrix element, to correct the spectral intensity deviation caused by the mass difference. When there is a non-natural isotopic distribution of the target element in the sample (such as artificially synthesized samples), the atomic weight can be measured by mass spectrometry (such as inductively coupled plasma mass spectrometry, ICP-MS). Mass spectrometry can accurately resolve the abundance of each isotope in the sample and generate the actual atomic weight of the sample based on weighted calculations.

[0034] Surface roughness is an important parameter to describe the physical state of the sample surface. Its value directly affects the scattering and absorption characteristics of the spectral signal. In spectral measurement, surface roughness determines the distribution of incident light and scattered light, so it is a key factor in interference correction. Surface roughness is usually expressed in root mean square roughness. It indicates that its acquisition mainly depends on the measurement of surface morphology. Commonly used measurement techniques include atomic force microscopy (AFM) and laser confocal microscopy (LCM), which can analyze the morphological characteristics of the sample surface with nanometer-level precision. In the present invention, surface roughness is used for signal correction of X-ray fluorescence spectroscopy (XRF). The greater the roughness, the more obvious the scattering and absorption of photons, which may lead to a weakening of the fluorescence signal intensity. By introducing the measured roughness parameters into the correction formula, the signal deviation caused by the surface state can be effectively compensated. In addition, under experimental conditions, the roughness parameters can also be obtained by measuring and calculating the signal of the standard sample. For example, by using a reference sample with known roughness, recording the change characteristics of its spectral signal, and deducing the influence of the sample roughness on the signal.

[0035] Excited state temperature It is a thermodynamic parameter that describes the excited state of the target element in the plasma. It affects the probability of the distribution of atoms or ions between different energy states, thereby determining the intensity of the spectral signal. The acquisition of the excited state temperature is usually based on the Boltzmann distribution and the Saha equation, and is calculated by the spectral line intensity ratio method. In actual measurement, two or more characteristic spectral lines of the target element can be selected and their spectral intensities can be recorded separately. , and derive the temperature using the following relationship: ; in is the wavelength of the spectral line, is the degeneracy of the energy levels, is the transition probability, is the energy level, is the Boltzmann constant. By analyzing the intensity ratio of multiple sets of spectral lines, the excited state temperature of the plasma can be obtained. In specific experiments, the width of the spectral line (such as Doppler broadening) can also be used to estimate the temperature. Doppler broadening is directly related to the thermal motion of particles, and its formula is: ; in is the central wavelength of the spectral line, is the mass of the particle. By measuring the broadening of the spectral line, the plasma temperature can be inferred. The excited state temperature is widely used in the present invention for spectral signal correction and feature extraction. It directly affects the weight distribution of energy level distribution and adjusts the nonlinear relationship of the response function in concentration prediction. By combining the above-mentioned multiple methods to measure and calibrate the temperature, the present invention can dynamically adjust the temperature parameters under different experimental conditions, thereby enhancing the accuracy and adaptability of spectral analysis. Response function Mainly used to characterize the optical detection characteristics in ICP-OES methods. Its expression is based on the optical characteristic function of the spectrometer, usually with wavelength The distribution of is directly related to the sensitivity of the ICP spectrometer. Experimentally, the detection sensitivity of the ICP spectrometer shows a non-uniform distribution at different wavelengths, which is mainly limited by factors such as the grating diffraction efficiency, the quantum efficiency of the photodetector, and the transmittance of the optical system. In order to accurately describe these characteristics, the response function It can be approximated by a Gaussian distribution combined with instrument characteristic parameters, as shown below: ; in, is the central wavelength of the spectrometer in the target band, represents the standard deviation of the signal broadening of the spectrometer. This Gaussian response function can accurately capture the focusing characteristics of the spectrometer on the target wavelength and describe the signal resolution capability of the instrument optical system. Experimentally, by scanning the standard sample and measuring its spectral line intensity, The specific shape of the signal can be corrected according to the actual sample signal.

[0036] For the XRF method, the response function It describes the spectral distribution characteristics of the fluorescence signal of the target element after X-ray excitation. The response characteristics of the XRF instrument are not only affected by the quantum efficiency of the detector, but also closely related to the thickness, surface roughness and light source characteristics of the sample. It can be expressed as follows: ; Among them, the denominator Describes the inverse relationship between X-ray fluorescence intensity and wavelength. This is because the energy of X-rays is inversely proportional to the wavelength, and high-energy photons have a stronger ability to penetrate the sample. and They represent the response center and broadening parameters of the XRF instrument to a specific wavelength signal. Through experimental calibration, the XRF signal response of different wavelengths can be fitted to obtain specific Expression form.

[0037] Embodiment 3: The correction and interference elimination part performs a process of eliminating the spectral line overlap interference on the fused signal by the following formula to obtain an interference elimination signal: ; in, ; Indicates The fusion signal of the elements is The first element other than All elements are matrix elements; is the number of elements in the multi-element sample of lithium ore; Indicates Elements in wavelength Interference cancellation signal at For the The element The spectral line interference coefficient of the element is the set value; For the The element The central wavelength of the interference spectrum of the element; For the The element The half width of the interference line of the element; Indicates Elements in wavelength The interference cancellation signal at the

[0038] Specifically, as a multi-component complex system, the spectral signal of lithium ore is often interfered by the overlapping spectral lines of other elements. This interference arises from the fact that the emission or fluorescence spectra of different elements are adjacent or overlapping within the wavelength range. When multiple elements coexist, their spectral signals will be superimposed on the same detector, resulting in the intensity at a certain wavelength not only reflecting the characteristic signal of the target element, but also including the contribution of other matrix elements. In order to separate the real signal of the target element, it is necessary to mathematically eliminate the interference of the fused signal. The formula of Example 3 simulates the interference characteristics of the matrix element on the target element, and uses a parametric method to dynamically subtract the interference component, thereby restoring the independent signal of the target element. The main body of the formula consists of two parts. First, The fusion signal represents the target element, which is the spectral response obtained by fusing the two spectral methods of ICP-OES and XRF. Other matrix elements are gradually removed by applying a series of interference subtraction factors, i.e., the product term in the formula, to the signal. For the target element The structure of the product term It is the core of the whole formula. Its physical meaning is to simulate the The mathematical model of interference removal is based on the Lorentz distribution, which reflects the influence of the central wavelength, half width and intensity of the spectral line on the interference contribution. In the formula, Represents the matrix element For the target element The central wavelength of the interference spectrum. Since the intensity of the spectrum reaches its maximum at the central wavelength, the interference effect is most significant at this point and gradually weakens on both sides away from the central wavelength. The decay rate of interference is determined by the half-width of the spectrum. Control, which is defined as the wavelength spacing when the spectral line intensity drops to half of the maximum value. The wider the spectral line, the larger the distribution range of the interference in the wavelength space. Express this spatial attenuation relationship.

[0039] The intensity of the interfering line is given by the coefficient Fusion signals with matrix elements Decide. is a set spectral line interference coefficient used to quantify matrix elements For the target element This coefficient depends on experimental determination or theoretical calculation and can be calibrated according to the element pair, experimental conditions and the response characteristics of the spectrometer. This means that the matrix element has a wavelength The fusion signal at , provides the actual contribution of the matrix element line. The product term in the formula iterates all The matrix elements are introduced one by one, and the signal of the target element is gradually optimized by subtracting the current interference component of each item. The product form of the formula reflects the cumulative effect of interference elimination, that is, the effect of each interference factor is independent, but their impact on the final signal of the target element is global. Therefore, after all matrix elements are subtracted, the signal of the target element is is restored to a value as close as possible to its true value. is a signal after interference removal, and its wavelength is Contains only the target element The accuracy of this interference removal process depends on the model parameters (such as , and ) and matrix element fusion signals The quality of the target element can be restored through accurate calculation of the formula, providing a reliable data basis for subsequent content prediction. The formula not only reflects the physical properties of the spectral signal, but also effectively solves the complex problem of spectral line overlap in a multi-component system through mathematical methods, making the present invention more accurate and reliable in multi-component analysis of lithium ore.

[0040] The physical meaning of the spectral interference coefficient is to quantify the matrix elements The characteristic spectral lines of the target element in the wavelength space This interference is usually caused by two main mechanisms: one is spectral overlap interference, that is, the spectrum of the matrix element and the spectrum of the target element are adjacent or overlapped in the wavelength space, resulting in the matrix signal being mixed into the target signal; the other is matrix effect interference, where the matrix element indirectly affects the signal intensity of the target element by changing the plasma excitation conditions (such as temperature and electron density).

[0041] In the experiment, the spectral line interference coefficient The initial setting of can be calibrated by using the spectral measurement data of the standard sample. First, a single matrix element is selected and the target element The spectral signals of the two elements under the same experimental conditions are measured by comparing the spectral characteristics of the two elements. The possible spectral line overlap area and intensity relationship are identified. Matrix elements For the target element The interference degree can be calculated by the following relationship: ; in and Respectively represent the matrix element and target element at wavelength Through this ratio, we can preliminarily determine The numerical range of .

[0042] In practical applications, the setting of the spectral line interference coefficient also needs to be corrected in combination with the theoretical model. Based on the transition characteristics of the target element and the matrix element (such as transition energy level, transition probability and spectral line broadening), the interference intensity of the matrix element in the target spectral line band can be predicted by quantum mechanics calculation or spectral simulation model. For example, by performing convolution calculation on the spectral line shapes of the two elements, their superposition effect in wavelength space can be simulated, thereby further optimizing Settings.

[0043] In addition, the dynamic adjustment of the spectral interference coefficient is the key to ensure its adaptability. In a multi-component complex system, the component distribution and concentration changes of the actual sample may cause fluctuations in the interference intensity. To this end, the present invention designs a dynamic update mechanism based on experimental data fitting. Specifically, by introducing the spectral signal matrix of the sample, The functional relationship between the experimental measurement data and the experimental data can be dynamically optimized by fitting the experimental data using machine learning models (such as multivariate linear regression or neural networks) to maintain high accuracy under different sample conditions.

[0044] For example, in actual sample measurement, the following objective function can be minimized to optimize Setting value: ; in is the target element signal after interference elimination, and are the original fusion signals of the target element and the matrix element, respectively. By minimizing the objective function, the optimal interference coefficient value can be obtained.

[0045] Central wavelength of interference line It is a matrix element The wavelength of the characteristic spectrum line of the matrix element, which marks the maximum position of the spectrum line intensity. In the experiment, this parameter can be determined by directly measuring the spectrum signal of the matrix element with a high-resolution spectrometer. The specific operation is to select a standard sample of a single matrix element, record its emission or fluorescence spectrum under the same experimental conditions, and determine the central wavelength by identifying the peak position in the spectrum line. Wavelength calibration of the spectrometer is an integral step in this process, and is usually accomplished by using a standard calibration light source (such as an argon or mercury lamp) with known spectral line positions to accurately correct the wavelength axis to ensure that the measured peak positions are consistent with the true values.

[0046] In theory, the central wavelength of the interfering spectral line can be obtained from a reference spectral database, such as the atomic spectral database of the National Institute of Standards and Technology (NIST). These databases contain characteristic spectral lines of various elements and their transition information, including wavelength, intensity, and transition energy level. By matching the wavelength range of the target element and the matrix element, possible spectral line overlap areas can be quickly found and the central wavelength can be accurately obtained. Theoretical value.

[0047] Half Width of Interference Lines It describes the distribution range of spectral line intensity in wavelength space, usually defined as the wavelength interval when the spectral line intensity drops to half of the maximum value. The measurement of half-width needs to take into account the broadening mechanism of the spectral line, including physical effects such as natural broadening, Doppler broadening and collision broadening. After recording the characteristic spectral lines of the matrix elements by a high-resolution spectrometer, the half-width parameters can be extracted by fitting the shape of the spectral lines (usually Gaussian or Lorentzian distribution). For example, for Doppler broadening, its half-width is related to the plasma temperature. and particle mass Correlation can be calculated by the following relationship: ; in is the Boltzmann constant. Experimentally, by measuring the actual width of the characteristic spectral line and combining it with known temperature conditions and particle mass, the actual half-width can be corrected. .

[0048] In actual samples, the broadening of the interfering spectral lines may be affected by additional factors, such as the surface roughness, thickness and matrix composition of the sample. These factors can lead to asymmetry or additional broadening of the spectral line shape. In this case, the present invention corrects these external influences by introducing standard samples and reference signals. For example, by measuring standard samples with known spectral line widths and comparing the measured values ​​with the theoretical values, the influence of the experimental system on the spectral line broadening can be deduced, thereby the half-width of the sample can be corrected. Make corrections.

[0049] Embodiment 4: The correction and interference elimination part performs matrix effect correction on the interference elimination signal by the following formula to obtain a correction signal: ; in, Indicates Elements in wavelength The correction signal at For the Elements in wavelength Interference cancellation signal at For the The element The matrix influence coefficient of the element is the set value; For the The characteristic energy of an element; is Avogadro's constant; is the thermodynamic temperature of the plasma in ICP spectroscopy; For the The mass of an element.

[0050] Specifically, the matrix effect in lithium ore mainly comes from the modulation of the physical and chemical properties of different elements on the excitation and ionization efficiency of the target element. This effect may be manifested through changes in temperature distribution in the plasma, changes in ionization energy, or differences in electron density. The formula is based on this principle and uses interference to eliminate signals. A dynamic correction model is established by combining the matrix-dependent correction factors. Indicated in wavelength The characteristic signal of the target element is the result of eliminating the spectral line overlap in the previous step, and the correction factor is calculated through the exponential function Introduced, dynamically adjust the target signal to compensate for the influence of matrix effect. The core of the correction factor is to quantify the influence of matrix elements on target elements. In the formula, the matrix effect is described by a series of parameters, including the matrix effect coefficient , characteristic energy of matrix elements , interference elimination signal These parameters together construct a mathematical model for matrix effect correction, so that the correction factor can fully consider the concentration, characteristic energy of matrix elements and their modulation effect on target elements. For example, the matrix effect coefficient It is obtained through experimental calibration and is used to measure the matrix elements For the target element The strength of the interaction under different experimental conditions. This interaction may manifest as an enhancement (positive modulation) or inhibition (negative modulation) of the excitation efficiency of the target element, thereby affecting its spectral signal intensity.

[0051] Exponential function form in formula It can not only describe the nonlinear characteristics of the matrix effect, but also effectively amplify or weaken the influence range of the correction factor. The weighted sum term includes the interference elimination signal of the matrix element and its characteristic energy The ratio of Normalization is performed to unify physical quantities of different dimensions. Interference elimination signal is the actual spectral intensity of the matrix element, reflecting the concentration distribution of the matrix element in the sample, while the characteristic energy It represents the energy characteristics of the matrix element in the spectral range and is used to quantify its modulation ability on the target element signal. The temperature and mass terms introduced in the correction factor Further linking the thermodynamic state of the plasma to the physical properties of the target element. It is a key parameter in ICP spectroscopy, which determines the effectiveness of element excitation and the distribution of ionization efficiency. Higher plasma temperature can generally improve the excitation efficiency, thereby enhancing the spectral signal of the target element, but its effect on different elements varies significantly, and this difference is determined by the mass of the target element. Adjustment. Larger atomic mass generally means higher ionization energy and therefore requires higher temperatures for efficient excitation. The formula is given by These influencing factors are comprehensively expressed in the form of , so that the correction factor can dynamically adapt to the physical properties of different elements. The corrected signal It is the real spectral signal of the target element under the influence of the matrix effect. Through the dynamic adjustment of the correction factor, the formula eliminates the interference of the matrix elements, making the signal purer and physically consistent. This not only provides a reliable data basis for subsequent content predictions, but also significantly improves the prediction accuracy and robustness of the model. In the lithium ore component analysis system based on machine learning, accurate signal correction is a prerequisite for the machine learning algorithm to effectively extract features and perform modeling. The design of this formula starts from physical principles, combined with experimental parameters and mathematical expressions, so that the correction signal can accurately reflect the true state of the target element in the sample, laying the foundation for the entire system to achieve high-precision content prediction.

[0052] Matrix influence coefficient The physical meaning of is to describe the interference of matrix elements on the spectral signal of the target element through physical or chemical effects. Specifically, this interference may come from changes in plasma excitation efficiency, spectral line overlap effects, sample surface states, and multiple scattering effects in the optical path. For example, a high concentration of matrix elements may significantly reduce the plasma temperature, thereby indirectly weakening the excitation intensity of the target element; or the matrix elements may affect the transmission path and absorption characteristics of the target signal by changing the physical density of the sample. In experimental calibration, the matrix influence coefficient The initial value of is obtained by measuring the standard sample. The specific method is to select a sample with a known composition ratio, in which the target element and matrix elements The concentration of .set up is the signal intensity of the target element without matrix influence, and In the presence of matrix elements The matrix influence coefficient can be determined by the following relationship: ; in It is a matrix element The signal intensity in a mixed sample. This formula reflects the degree of suppression of the target element signal by the matrix element signal.

[0053] The theoretical model is The setting of the matrix element provides important support. By analyzing the physical and chemical properties of the matrix elements and the target elements, the interaction between the two can be predicted. For example, in inductively coupled plasma spectroscopy (ICP-OES), the matrix effect is closely related to the thermodynamic parameters of the plasma (such as temperature and electron density). According to the Saha equation and Boltzmann distribution, the modulation effect of different concentrations of matrix elements on the plasma state can be calculated, and then the change amplitude of the target element signal can be derived. These theoretical results can be used as the initial value of the matrix influence coefficient, or calibrated with the experimental value.

[0054] Data fitting and machine learning techniques further optimize the setting of matrix influence coefficients. Due to the large variety of matrix elements and wide concentration range in lithium ore samples, a single theoretical model is difficult to fully cover all possible interactions. The present invention uses large sample experimental data and establishes the relationship between the matrix influence coefficient and sample composition through a data fitting model (such as multiple regression or neural network). In the fitting process, the target element signal is Matrix element signal The change pattern of the input model is considered, and other experimental conditions (such as temperature, surface state, etc.) are considered to optimize the The specific value of .

[0055] For example, the following error function can be optimized by minimizing : ; in is the predicted signal calculated by the matrix influence model, is the actual signal measured experimentally. Through iterative optimization, it can be dynamically adjusted , so that it can adapt to the complex characteristics of different samples.

[0056] Embodiment 5: In the prediction part of the machine learning model, the preset machine learning model is a multi-level feature extraction network; the correction signal is subjected to feature extraction using the preset machine learning model through the following formula to obtain features: ; in, For the Spectral contribution coefficient of the layer; is the feature activation function; The number of layers of the multi-level feature extraction network; For the The feature extraction function of the layer is as follows: ; in, is the lower limit of wavelength; is the upper limit of wavelength; For the The weight function of the layer; is the feature extraction kernel function; is the wavelength integration variable; is a nonlinear transfer function; is the spectral resolution.

[0057] Specifically, the feature extraction function The function of is to extract hierarchical feature information from the correction signal through a series of convolution, weighting and nonlinear transformation operations. In the formula, the integral term It is the key to feature extraction, which converts the correction signal within the wavelength range Perform weighted aggregation to generate The integration range is determined by the wavelength lower limit. and upper limit The decision means extracting features within a specific spectral range, so that the spectral characteristics of the target element can be captured for different wavelength regions. The weight function included in the integral term Reflects the The sensitivity distribution of layer feature extraction gives higher weights to certain wavelength regions to highlight the important features of the target elements. During the integration process, the correction signal Also through an exponential factor The physical basis of this term is the Boltzmann distribution, which describes the excitation characteristics of the target element under specific wavelength and temperature conditions. With characteristic energy And spectral resolution Combined, the formula can adjust the signal according to the actual spectrometer's resolution, while the temperature The introduction of further enhances the adaptability of the model to different experimental conditions.

[0058] Feature extraction kernel function is another key part of the formula, which captures the spatial characteristics of the spectral signal in the form of local convolution. The design of can be Gaussian, Lorentz or other kernel forms suitable for spectral signals, depending on the spectral line distribution characteristics of the target element. The introduction of this kernel function enables feature extraction to focus more on the key areas in the spectral signal while ignoring irrelevant or redundant information. Nonlinear conversion function It plays an important role in the feature extraction process. It makes the extracted features more expressive by mapping the linear integral results to a nonlinear space. Commonly used nonlinear functions include ReLU (rectified linear unit) and Sigmoid function, which can capture the complex nonlinear relationship in the spectral signal and provide rich input for the feature fusion of subsequent levels. Based on the feature extraction, the formula is The features extracted by the layer are weighted and combined to generate a global feature representation This combination process is based on the spectral contribution coefficient The features at different levels are integrated with weights. The value of is usually obtained through network training optimization and is used to measure the importance of each layer of features to the target task. The combined features are activated by the activation function Further nonlinearization is performed to enhance the discriminative power of global features. The activation function can be selected as the classic Sigmoid function or the more complex Swish function, depending on the design requirements of the feature extraction network. The final output of the formula is It is a multi-level, multi-scale feature representation that can fully reflect the spectral characteristics of the target element and the overall composition distribution of the sample. Through this multi-level feature extraction network, the formula gradually converts the original signal in the spectral data into high-dimensional features, retaining the physical information in the spectral signal to the greatest extent, while providing operational input features for the machine learning model. This feature extraction method not only improves the predictive ability of the model, but also enhances the system's adaptability to complex sample conditions, making the detection and prediction of the multi-component content of lithium ore more accurate and efficient.

[0059] Example 6: Feature extraction kernel function The formula is as follows: ; in, is the central wavelength; For the Spectral broadening parameters of the layer; is the plasma emission intensity of inductively coupled plasma spectroscopy; is the X-ray emission intensity of X-ray fluorescence spectrometry; is the optical path length.

[0060] Specifically, the core structure of the feature extraction kernel function is based on Gaussian distribution, which takes the central wavelength of the target element as The exponential term of the Gaussian distribution is It describes the rapid attenuation trend of the spectral signal intensity as the wavelength deviates from the central wavelength. The greater the wavelength deviation, the lower the signal intensity. This design accurately simulates the physical properties of the spectral signal, that is, the characteristic spectral line of the target element usually appears as a sharp peak concentrated near a specific wavelength. This characteristic is determined by the resolution of the spectrometer and the atomic energy level transition characteristics of the target element. Therefore, in the Gaussian distribution, the parameter It reflects the degree of spectral line broadening. The value of is determined by the extraction requirements of a specific level: the smaller Used to extract local detail features, larger This flexible adjustment capability enables the kernel function to adapt to the characteristic requirements of different target elements and sample conditions. In addition to describing the spatial distribution characteristics of the spectrum, the kernel function also introduces This weighting factor dynamically adjusts the contribution ratio of the two spectral methods (ICP-OES and XRF). and XRF fluorescence emission intensity They reflect the response signals of the target elements under high temperature plasma and fluorescence excitation conditions respectively. , the kernel function can flexibly balance the weights of the two signal sources to ensure robustness under different experimental conditions. For example, for low-concentration elements, ICP-OES usually has higher sensitivity, while for signals under high-concentration elements or complex matrix conditions, the contribution of XRF may be more significant. The introduction of this weight factor enables the kernel function to dynamically adapt to the performance characteristics of the target element in different spectral methods, thereby improving the generalization ability of the feature extraction network. At the same time, the kernel function also uses the exponential decay term The attenuation effect of the spectral signal during propagation is simulated. Here, is the mass absorption coefficient of the target element, is the optical path length, and the two together determine the energy loss of the signal intensity when it propagates in the sample and the medium. This exponential form is derived from the Beer-Lambert law in optical propagation, which describes the law that the light intensity decays exponentially with the propagation distance. This design incorporates the influence of experimental conditions (such as sample thickness and optical path length) into the kernel function modeling, thereby more realistically reflecting the spectral signal characteristics under the actual measurement environment. Overall, the kernel function The design achieves a high degree of unity between physical meaning and mathematical expression. The Gaussian distribution term focuses on the characteristic spectral lines of the target element and captures the spatial characteristics of the spectrum through local convolution; The weight factor dynamically balances between the two spectral methods, improving the adaptability and robustness of the kernel function; the exponential decay term further takes into account the influence of experimental conditions, making the kernel function closer to the actual measurement scenario. In the feature extraction network, the kernel function generates a more targeted feature representation for each layer through the convolution operation of the input signal. These features not only retain the physical information in the spectral signal, but also eliminate redundancy and noise, providing high-quality input for further modeling of the machine learning model. In the prediction of multi-component content of lithium ore, the kernel function The design of is of great significance. It converts the original signal into a highly recognizable feature representation by localizing the spectral signal of the target element.

[0061] Embodiment 7: Layer weight function Use the following formula to express it: ; in, is Planck's constant; is the transition frequency; is the upper energy level degeneracy; is the lower energy level degeneracy; is the probability of spontaneous radiation transition; is the upper energy level energy; is the Boltzmann constant.

[0062] Specifically, in the formula, the weight function The first item This reflects the decreasing effect of the feature extraction network layer. As the weight increases, the weight value gradually decreases. This design intention is to reduce the direct dependence on the initial input signal in the higher layers of feature extraction, and instead make more use of the abstract features extracted by the previous layers. This decreasing design is in line with the concept of hierarchical modeling of deep network features, which allows the network to focus on detailed signals at low levels and global features at high levels. The next part of the formula It expresses the intensity contribution of the spectral signal in the spatial distribution. Here, is Avogadro's constant, which quantifies the number of atoms per unit volume and thus relates to signal strength. Planck's constant and transition frequency Combined with, it defines the energy of each transition photon and describes the contribution of a specific transition process to the spectral signal. is the diffusion area of ​​radiation in space, describing the spherical diffusion characteristics of the signal as the propagation distance increases. The design of this part directly relates the spectral intensity to the geometric distribution and energy mechanism of the signal source, so that the weight function can accurately model the signal intensity of different transition mechanisms. In the formula, Describes the statistical properties and radiative probability of a particular transition. Upper level degeneracy and the lower level degeneracy It reflects the energy level distribution characteristics of the target element and defines the proportion of states at different energy levels. It is an important parameter that describes a specific transition channel. The larger its value, the easier it is for the transition process to occur. Through this term, the formula can capture the distribution law of the spectral characteristics of the target element on different transition paths, thereby highlighting the transition signals with significant contributions in the weight allocation of the feature extraction network.

[0063] Finally, in the formula is a classic Boltzmann distribution term used to describe the relative occupancy of particles at a specific energy level under thermodynamic equilibrium conditions. Indicates the energy height of the energy level, Boltzmann constant and plasma temperature Together they define the thermal energy scale of the system. This term reflects the effect of thermodynamic temperature on the distribution of excited states of the target element: at higher temperatures, particles are more likely to occupy high energy levels, thereby enhancing the transition signal corresponding to that energy level; while at low temperatures, the probability of high energy level transitions is significantly reduced. This description enables the weight function to dynamically respond to changes in experimental conditions, especially in the practical application of plasma spectroscopy, where changes in temperature have a particularly significant effect on signal intensity. Through this formula, the weight function It can accurately capture the weight characteristics of spectral signals in different energy level transitions and spatial distributions. This modeling method based on physical principles not only gives the feature extraction network a clear physical meaning, but also effectively improves the robustness of the network when processing multi-component spectral signals of lithium ore. The various parameters in the formula jointly construct a dynamic connection between signal contribution and experimental conditions, allowing the feature extraction network to adaptively adjust its response to the input signal, thereby achieving higher prediction accuracy and feature expression capabilities under complex sample conditions. This physical-statistical property of the weight function combined with the hierarchical feature extraction mechanism of the deep network provides a highly reliable and physically consistent modeling method for multi-component analysis of lithium ore, further improving the analysis accuracy and adaptability of the system.

[0064] Embodiment 8: Construct a characteristic response function according to the characteristic by the following formula: ; in, For the The characteristic response function of the element.

[0065] Specifically, the design of the formula is based on the principle of feature normalization and signal comparison. Feature response function The core of the formula is to use the ratio of the characteristic intensity of the target element to the characteristic intensity of other matrix elements to describe the characteristic significance of the target element. Through this normalized ratio, the formula can effectively eliminate the global deviation caused by sample matrix effect, instrument response inhomogeneity or experimental condition fluctuations in spectral measurement, thereby highlighting the relative importance of the target element. The numerator of the formula is Represents the target element In wavelength The characteristic intensity at comes directly from the output of the feature extraction network. After multi-level processing of the network, the characteristic intensity has integrated the information of the target element in different spectral methods and captured the key spectral characteristics of the target element through specific physical weights and kernel functions. As the characteristic representative of the target element, It reflects the spectral signal significance of the target element under the current measurement conditions. Except for the target element This term is calculated by multiplying the characteristic intensities of all matrix elements except The combination of the characteristics of other elements reflects the overall spectral contribution of the matrix elements. In lithium ore samples, matrix elements usually exist in multiple forms and directly or indirectly interfere with the spectral signals of the target elements. By calculating the product of all matrix element characteristics, the formula can capture their overall impact intensity and correct these interferences during the normalization process.

[0066] Ratio structure is the key part of the formula, which compares the characteristic intensity of the target element with the comprehensive intensity of the matrix element, thereby quantifying the relative contribution of the target element in the entire sample composition space. This design effectively solves the problem of confusion between the target signal and the matrix signal caused by the matrix effect in the spectral data, making the characteristic expression of the target element more independent and clear. The formula also introduces The ratio is used to dynamically adjust the characteristic response function. This ratio reflects the plasma emission intensity of inductively coupled plasma spectroscopy (ICP-OES) and the fluorescence emission intensity of X-ray fluorescence spectroscopy (XRF). Since the sensitivity of different spectral methods to the target element signal may vary significantly, The introduction of provides a dynamic balance mechanism for the formula, so that the characteristic response function can adaptively adjust the weights of different spectral signals. For example, when ICP-OES shows high sensitivity to certain low-concentration elements, this ratio can enhance the contribution of the ICP-OES signal; and for the robustness of XRF under complex matrix conditions, the formula reasonably reduces the impact of ICP-OES through this ratio, thereby ensuring that the characteristic response function has consistent applicability to multiple spectral methods. The characteristic response function constructed by this formula , which can not only reflect the characteristic intensity of the target element, but also quantify the relative contribution of the target element in the multi-component spectral signal. This normalized ratio form is of great significance in the multi-component analysis of lithium ores. Since the elements in lithium ores usually have complex spectral characteristics and significant matrix effects, it is difficult for traditional methods to accurately separate the characteristic signals of the target elements. Through the characteristic response function, the system can highlight the signal of the target element in a physically meaningful way, while minimizing matrix interference and measurement deviation. Characteristic response function The output is a numerical indicator directly related to the concentration of the target element, which provides a key input for the training and prediction of the machine learning model. The construction process of this indicator is based on the physical properties of spectral data and the feature extraction ability of machine learning, which converts complex multi-spectral information into simple and meaningful response signals, significantly improving the accuracy and stability of multi-component detection of lithium ore.

[0067] Example 9: The initial content of each element is predicted based on the characteristic response function using the following formula: ; in, For the The initial content of the element; For the The mass of the element; For the The mass of an element.

[0068] Specifically, the formula is constructed by first considering the characteristic response function of the target element As a direct expression of the input signal, It is a comprehensive description of the spectral signal intensity of the target element, reflecting the significance of the target element after feature extraction and response normalization. The characteristic response function itself removes the systematic deviation of the spectral instrument response and experimental conditions through normalization, making it a reliable input parameter for concentration prediction. The concentration of the target element cannot be directly derived because the presence of matrix elements in the sample may significantly interfere with the target element signal. Therefore, the formula introduces a matrix effect correction through the denominator to ensure that the concentration calculation not only depends on the signal intensity of the target element, but also can dynamically correct the influence of the matrix elements. The core of the weighted sum term in the denominator is the description of the interference of the matrix elements on the target elements. For complex multi-component samples, matrix elements affect the spectral signals of the target elements through various mechanisms, such as changing plasma conditions, introducing spectral line overlap interference, or changing the excitation efficiency of the target elements through chemical matrix effects. These influences are described in the formula through the response function of the matrix element and its interaction coefficient with the target element Express. The introduction of quantifies the specific interference degree of matrix elements on target elements. Its value is determined based on experimental calibration or theoretical modeling to reflect the differential contribution of different matrix elements under actual sample conditions.

[0069] To further correct for spectral line interference, the formula introduces a correction term based on wavelength characteristics. Spectral line interference is a common problem in spectral analysis, especially in complex multi-element samples such as lithium ore, where spectral lines of different elements may overlap, resulting in the target element signal containing contributions from other elements. , the formula can dynamically evaluate and correct the impact of this interference. Here, is the interference coefficient, which indicates the specific interference intensity of the matrix element on the target element; and The center position and width of the interference spectral line are defined. This design enables the formula to accurately capture the spectral line relationship between the target and matrix elements in wavelength space, thereby eliminating the influence of interference on the target element concentration estimation. The signal is corrected from another dimension. The mass of different elements directly affects their excitation behavior in the plasma and the intensity distribution of the spectral signal. Lighter elements generally have higher excitation efficiency and transition probability, while heavier elements may produce weaker signals due to their higher ionization energy. By introducing the mass ratio, the formula can compensate for the signal deviation caused by the difference in physical properties, so that the predicted results are closer to the actual concentration of the element. Finally, the formula calculates the initial concentration through the ratio of the characteristic response function of the target element to the matrix correction factor in the denominator. This ratio form is the key to the formula for concentration prediction, because it combines the intensity information of the target element signal with the matrix correction, which maximizes the retention of the characteristics of the target element while eliminating matrix interference. This design not only has a clear physical meaning, but can also dynamically adapt to different samples and experimental conditions, thereby achieving high-precision concentration estimation in multi-component complex systems.

[0070] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A multi-component content prediction and detection optimization system in lithium ore based on machine learning, characterized in that: The system comprises: a multi-spectral data fusion and preprocessing part, a correction and interference elimination part and a machine learning model prediction part; the multi-spectral data fusion and preprocessing part is used to measure the emission spectrum of the lithium ore multi-element sample by inductively coupled plasma spectroscopy to obtain the first light intensity signal of each element, measure the fluorescence signal of the lithium ore multi-element sample by X-ray fluorescence spectroscopy to obtain the second light intensity signal of each element, and fuse and normalize the first light intensity signal and the second light intensity signal to obtain a fused signal; wherein, when the emission spectrum of the lithium ore multi-element sample is measured by inductively coupled plasma spectroscopy and the fluorescence signal of the lithium ore multi-element sample is measured by X-ray fluorescence spectroscopy, the same observation distance, the same resolution, the same wavelength and the same energy level are used; the correction and interference elimination part is used to eliminate the interference of spectral line overlap on the fused signal to obtain an interference elimination signal, and perform matrix effect correction on the interference elimination signal to obtain a correction signal; the machine learning model prediction part is used to extract features from the correction signal by using a preset machine learning model to obtain features; a feature response function is constructed based on the features, and the initial content of each element is predicted based on the feature response function.

2. The multi-component content prediction and detection optimization system in lithium ore based on machine learning as claimed in claim 1, characterized in that: Suppose that the emission spectrum of a multi-element lithium ore sample is measured using inductively coupled plasma spectroscopy, and the Elements at wavelength In the The first light intensity signal of the spectral line is ; Assume that the emission spectrum of lithium ore multi-element sample is measured by inductively coupled plasma spectroscopy, and the first Elements at wavelength In the The first light intensity signal of the spectral line is ; The first light intensity signal and the second light intensity signal are fused and normalized by the following formula to obtain a fused signal : ； in, Indicates Elements at wavelength The fusion signal at The following table is the index for integers; For the The total number of spectral lines of an element; is the base of natural logarithms; For the The characteristic energy of an element; For the The element in The spectral line energy of each spectral line; is the Boltzmann constant, is temperature; For the The element in The absorption coefficient of each spectral line; For the The atomic number of the element; For the The density of the element; For the The mass absorption coefficient of the element; is the average thickness of multi-element samples of lithium ore; For the The atomic weight of an element; is the surface roughness of multi-element samples of lithium ore; For the The excited state temperature of an element; For the The temperature coefficient of the fluorescence yield of the element; For inductively coupled plasma spectroscopy at wavelength The response function at ; The wavelength of X-ray fluorescence spectroscopy The response function at .

3. The multi-component content prediction and detection optimization system in lithium ore based on machine learning as claimed in claim 2, characterized in that: The correction and interference elimination part eliminates the spectral line overlap interference of the fused signal by the following formula to obtain an interference elimination signal: ； in, ; Indicates The fusion signal of the elements is The first element other than All elements are matrix elements; is the number of elements in the multi-element sample of lithium ore; Indicates Elements at wavelength Interference cancellation signal at For the The element The spectral line interference coefficient of the element is the set value; For the The element The central wavelength of the interference spectrum of the element; For the The element The half width of the interference line of the element.

4. The multi-component content prediction and detection optimization system in lithium ore based on machine learning as claimed in claim 3, characterized in that: The correction and interference elimination part performs matrix effect correction on the interference elimination signal by the following formula to obtain a correction signal: ； in, Indicates Elements at wavelength The correction signal at For the Elements at wavelength Interference cancellation signal at For the The element The matrix influence coefficient of the element is the set value; For the The characteristic energy of an element; is Avogadro's constant; is the thermodynamic temperature of the plasma in ICP spectroscopy; For the The mass of an element.

5. The multi-component content prediction and detection optimization system in lithium ore based on machine learning as claimed in claim 4, characterized in that: In the machine learning model prediction part, the preset machine learning model is a multi-level feature extraction network; the correction signal is subjected to feature extraction using the preset machine learning model through the following formula to obtain features: ； in, For the Spectral contribution coefficient of the layer; is the feature activation function; The number of layers of the multi-level feature extraction network; For the The feature extraction function of the layer is as follows: ； in, is the lower limit of wavelength; is the upper limit of wavelength; For the The weight function of the layer; is the feature extraction kernel function; is the wavelength integration variable; is a nonlinear transfer function; is the spectral resolution.

6. The multi-component content prediction and detection optimization system in lithium ore based on machine learning as claimed in claim 5, characterized in that: Feature extraction kernel function The formula is as follows: ； in, is the central wavelength; For the Spectral broadening parameters of the layer; is the plasma emission intensity of inductively coupled plasma spectroscopy; is the X-ray emission intensity of X-ray fluorescence spectrometry; is the optical path length.

7. The multi-component content prediction and detection optimization system in lithium ore based on machine learning as claimed in claim 6, characterized in that: No. Layer weight function Use the following formula to express it: ； in, is Planck's constant; is the transition frequency; is the upper energy level degeneracy; is the lower energy level degeneracy; is the probability of spontaneous radiation transition; is the upper energy level energy; is the Boltzmann constant.

8. The multi-component content prediction and detection optimization system in lithium ore based on machine learning as claimed in claim 7, characterized in that: The feature response function is constructed based on the feature using the following formula: ； in, For the The characteristic response function of the element.

9. The multi-component content prediction and detection optimization system in lithium ore based on machine learning as claimed in claim 8, characterized in that: The initial content of each element is predicted based on the characteristic response function using the following formula: ； in, For the The initial content of the element; For the The mass of the element; For the The mass of an element.

Images