A method for intelligent analysis of carbon powder purity using deep learning

Abstract

The present application provides a method for intelligently analyzing toner purity using deep learning, comprising the following steps: obtaining scattering spectrum data of a toner sample, calculating the scattered light intensity of the toner sample at different wavelengths, and obtaining original scattering spectrum curve data of the toner sample; preprocessing the original scattering spectrum curve data of the toner sample, using a data smoothing and denoising algorithm to eliminate high-frequency noise and abnormal points in the curve data, and obtaining a smoothed scattering spectrum curve; extracting characteristic parameters from the smoothed scattering spectrum curve, using a peak detection algorithm to determine the main peak position, peak width, and peak intensity of the scattering spectrum curve, and constructing a scattering spectrum feature vector of the toner sample; and using a principal component analysis method to qualitatively and quantitatively analyze the type and content of the adsorbed impurities based on the position and intensity information of the scattering peaks of adsorbed impurities obtained by decomposition, and obtaining the main impurity components and the relative content of the main impurities that affect the purity of the toner.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for realizing intelligent analysis of carbon powder purity by utilizing deep learning. Background Art

[0002] A key technical challenge in intelligent analysis of toner purity lies in the effect of impurity particle adsorption on the surface morphology of toner agglomerates. When impurity particles adsorb on the toner surface, they cause changes in the surface roughness of the toner agglomerates, which in turn alters the peak shape of the scattering spectrum. These subtle changes in the spectral curve contain rich surface morphological information, but accurately extracting this characteristic information from the spectral data is a pressing technical challenge. Traditional spectral analysis methods struggle to capture these subtle spectral changes, but deep learning models offer a new approach to addressing this problem. However, further research is needed to design a reasonable deep learning model structure to automatically extract and optimize the representation of key characteristic parameters of the spectral curve and to clarify the intrinsic relationship between these characteristic parameters and toner purity. Furthermore, further research is needed to explore the mechanism by which the characteristic parameters extracted by the model affect the accuracy of purity determination and to reveal their physicochemical nature. This is crucial for further optimizing model performance and improving the reliability of toner purity analysis. Summary of the Invention

[0003] The present invention provides a method for realizing intelligent analysis of carbon powder purity using deep learning, which mainly includes:

[0004] Acquiring scattered light spectrum data of the toner sample, calculating the scattered light intensity of the toner sample at different wavelengths, and obtaining original scattered light spectrum curve data of the toner sample;

[0005] The original scattering spectrum curve data of the carbon powder sample is preprocessed by using data smoothing and denoising algorithms to eliminate high-frequency noise and abnormal points in the curve data and obtain a smooth scattering spectrum curve;

[0006] Characteristic parameters are extracted from the smoothed scattering spectrum curve. The peak detection algorithm is used to determine the main peak position, peak width and peak intensity of the scattering spectrum curve, and the scattering spectrum characteristic vector of the carbon powder sample is constructed.

[0007] Based on the scattered spectrum feature vector of the toner sample, the support vector machine algorithm is used to determine the purity of the toner sample. If the scattered spectrum feature vector falls within the feature space of the preset high-purity toner sample, the toner sample is judged to have reached the target purity; otherwise, it is judged to have not reached the target purity.

[0008] For carbon powder samples that do not reach the target purity, the abnormal change characteristics of the corresponding scattering spectrum curve are analyzed, and the scattering spectrum curve is decomposed using the peak fitting method to obtain the scattering peak caused by adsorbed impurities and the characteristic scattering peak of the carbon powder itself;

[0009] According to the position and intensity information of the scattering peaks of adsorbed impurities obtained by decomposition, the principal component analysis method is used to conduct qualitative and quantitative analysis on the type and content of adsorbed impurities, and the main impurity components and relative contents of the main impurities that affect the purity of carbon powder are obtained;

[0010] Based on the purity determination results and impurity analysis results of the toner samples, a decision tree algorithm is used to establish a purity evaluation model for toner purity. According to the toner scattering spectrum characteristics and impurity analysis results, the toner purity is quantitatively scored to provide guidance for toner quality optimization.

[0011] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0012] The present invention discloses a method for intelligently analyzing carbon powder purity using deep learning. The method obtains scattering spectrum data from a carbon powder sample, preprocesses and extracts features from the original spectral curve, and constructs a scattering spectrum feature vector. A support vector machine algorithm is used to determine whether the carbon powder purity meets the standard, and impurity analysis is performed on samples that do not meet the standard. Spectral peak fitting and principal component analysis algorithms are used to qualitatively and quantitatively analyze the main impurity components that affect purity. Finally, an intelligent carbon powder purity assessment model is established to quantitatively score the purity. The present invention achieves rapid and accurate determination of carbon powder purity and can analyze key factors affecting purity, providing effective guidance for optimizing carbon powder quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a flow chart of a method for intelligently analyzing carbon powder purity using deep learning according to the present invention.

[0014] Figure 2 This is a schematic diagram of a method for intelligently analyzing carbon powder purity using deep learning according to the present invention.

[0015] Figure 3 This is another schematic diagram of the method for intelligently analyzing carbon powder purity using deep learning according to the present invention. DETAILED DESCRIPTION

[0016] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] like Figure 1-3 In this embodiment, a method for realizing intelligent analysis of toner purity using deep learning may specifically include:

[0018] S101 , obtaining scattered light spectrum data of a toner sample, calculating the scattered light intensity of the toner sample at different wavelengths, and obtaining original scattered light spectrum curve data of the toner sample.

[0019] Light signals are collected from the surface of a carbon powder sample at preset collection intervals using a scattering spectrometer, and a first set of scattering spectrum data within a wavelength range is acquired using a photodetector to obtain a first scattering spectrum sequence. A smoothing method is used to process data at each wavelength point in the first scattering spectrum sequence. If the fluctuation amplitude of the wavelength point signal exceeds a calibration threshold of the scattering spectrometer, the data is repeatedly collected at that wavelength point to obtain a second scattering spectrum sequence. A spectral absorption curve is constructed based on the second scattering spectrum sequence, and the spectral absorption curve is fitted using a three-layer forward neural network, in which the input layer of the neural network is the wavelength value and the output layer is the scattered light intensity value at the corresponding wavelength, to obtain a third scattering spectrum sequence. A frequency domain analysis is performed on the third scattering spectrum sequence using Fourier transform, and high-frequency noise and low-frequency baseline drift are filtered out using a bandpass filter to obtain scattering spectrum curve data of the carbon powder sample.

[0020] Exemplarily, a spectral scan is performed on the surface of the carbon powder sample at preset spectral acquisition intervals using a scattering spectrometer. Light signals reflected from the carbon powder sample are collected using a photodetector. A first set of scattering spectrum data is collected from the photodetector within at least three wavelength bands. The first set of scattering spectrum data is sorted by wavelength to produce a first scattering spectrum sequence. A five-point smoothing algorithm is used to smooth each wavelength point in the first scattering spectrum sequence. If the signal fluctuation amplitude at that wavelength point exceeds the factory calibration threshold of the scattering spectrometer after processing, data is repeatedly collected at that wavelength point until the fluctuation amplitude falls below the calibration threshold, thereby producing a second scattering spectrum sequence. Data quality is assessed for the second scattering spectrum sequence by calculating the signal-to-noise ratio to determine the data quality at each wavelength point. If the signal-to-noise ratio falls below the factory calibration parameters of the scattering spectrometer, the data at that wavelength point is discarded, thereby producing a third scattering spectrum sequence. A spectral absorption curve is constructed based on the data from the third scattering spectrum sequence. The spectral absorption curve is then fitted using a three-layer feed-forward neural network. The input layer of the neural network is the wavelength value, the hidden layer uses a hyperbolic tangent activation function, and the output layer is the scattered light intensity value at the corresponding wavelength, thereby producing a fourth scattering spectrum sequence. Cubic spline interpolation is performed between adjacent wavelength points in the fourth scattering spectrum sequence. The number of interpolation points is adaptively determined based on the light intensity difference between adjacent wavelength points. If the light intensity difference exceeds the resolution threshold of the scattering spectrometer, an interpolation point is added to this interval to obtain the scattering spectrum curve data of the toner sample. The scattering spectrum curve data of the toner sample is analyzed in the frequency domain using Fourier transform. High-frequency noise and low-frequency baseline drift are filtered out using a bandpass filter to obtain the final scattering spectrum curve data of the toner sample. During the scattering spectrum data acquisition process for the toner sample, the photodetector typically captures wavelengths in the range of 400-700 nanometers, with an acquisition interval set at 2 nanometers. This results in a total of 150 wavelength points within the visible light range. In the first scattering spectrum sequence, the light intensity value at each wavelength point is expressed as relative reflectance, ranging from 0 to 100. When performing five-point smoothing on the scattered spectrum data, the data of the current wavelength point and the two wavelength points before and after it are selected for weighted averaging, with weight coefficients of 0.1, 0.2, 0.4, 0.2, and 0.1, respectively. This method can effectively reduce the impact of random noise. When the fluctuation amplitude of the processed signal exceeds the calibration threshold of 0.5%, the data of the wavelength point is recollected. The signal-to-noise ratio of the scattered spectrum data is calculated using the ratio of the peak signal to the standard deviation. For the scattered spectrum data at 500 nanometers, the peak signal is 85.6 and the standard deviation is 0.42. The calculated signal-to-noise ratio is approximately 203.8, which is higher than the calibration parameter of 180, indicating that the data quality of this wavelength point is good. At 650 nanometers, due to the weak energy of the light source, the peak signal drops to 45.3, the standard deviation is 0.38, and the signal-to-noise ratio is only 119.2, which is lower than the calibration parameter. The data of this wavelength point needs to be eliminated.In a three-layer feedforward neural network, the input layer has 1 neuron, corresponding to the wavelength value, the hidden layer has 16 neurons, and uses a hyperbolic tangent activation function. The output layer has 1 neuron, corresponding to the scattered light intensity value. Taking a wavelength of 500 nanometers as an example, the input layer value is 0.5, which is passed to the hidden layer after calculation of the weight matrix and bias term, resulting in an output layer light intensity value of 85.6. During cubic spline interpolation, if the light intensity difference between adjacent wavelength points exceeds the resolution threshold of 0.2%, an interpolation point is added to that interval. For example, in the wavelength range of 498-500 nanometers, the light intensity values ​​of the two end points are 84.2 and 85.6, respectively, with a difference of 1.4, exceeding the threshold. Therefore, four interpolation points are added to this interval: at 498.4, 498.8, 499.2, and 499.6 nanometers. After Fourier transformation, bandpass filtering was performed in the frequency domain with a cutoff frequency set between 0.05 and 0.8. This effectively removed high-frequency noise components above 0.8 and baseline drift components below 0.05. The filtered scattering spectrum adjusted the light intensity at 500 nm from 85.6 to 85.4, resulting in a smoother curve.

[0021] S102 , preprocessing the original scattering spectrum curve data of the carbon powder sample, using a data smoothing and denoising algorithm to eliminate high-frequency noise and abnormal points in the curve data, and obtaining a smooth scattering spectrum curve.

[0022] A median filter is used to perform sliding window processing on the spectral data sequence, and outliers are marked by calculating the local interval mean and standard deviation to obtain a first spectral data sequence; based on the outliers marked in the first spectral data sequence, the outliers are replaced by a local weighted averaging method, and a second spectral data sequence is obtained from the replaced data sequence; wavelet multiscale decomposition is performed on the second spectral data sequence, and soft threshold processing is performed on the high-frequency noise coefficient using an adaptive threshold, and a third spectral data sequence is obtained from the reconstructed wavelet coefficients; for the third spectral data sequence, the Savitzky-Golay smoothing method is used for data fitting, and polynomial coefficients are estimated for the local interval using the recursive least squares method, and a smoothed scattering spectrum curve is obtained from the fitting curve.

[0023] Exemplarily, data statistics are calculated based on the raw scattering spectrum curve data. A sliding window process is performed on the spectral data sequence using a median filter with a window length of five data points. Outliers are identified from the spectral data sequence, and data points outside the range of three standard deviations centered on the local interval mean are marked as outliers, resulting in a first spectral data sequence. For the marked outliers in the first spectral data sequence, the outliers are replaced by a local weighted average of three adjacent points on each side. The weighting coefficient is set inversely proportional to the distance between the data point and the outlier. If the adjacent points include an outlier, the calculation interval is expanded until a normal data point is obtained, resulting in a second spectral data sequence. A four-layer multiscale decomposition is performed on the second spectral data sequence using a wavelet transform. The db4 wavelet basis function is used. A soft thresholding process is performed on the high-frequency noise coefficient using an adaptive threshold based on local variance estimation, and a third spectral data sequence is reconstructed from the wavelet coefficients. Data quality assessment is performed on the third spectral data sequence by calculating three indicators: signal-to-noise ratio, spectral resolution, and baseline drift. If the indicator values ​​meet preset threshold requirements, a fourth spectral data sequence is output. The Savitzky-Golay smoothing algorithm of the third-order polynomial is used to fit the fourth spectral data sequence. The sliding window length is set to seven data points. The polynomial coefficients of each local interval are estimated by recursive least squares method, and a smoothed scattering spectrum curve is obtained from the fitting curve. In the preprocessing of scattering spectrum data, the raw data usually contains various interferences and noises. Taking the scattered light intensity data of the carbon powder sample at a wavelength of 500nm as an example, the raw data sequence is

[0024] The data points [85.2, 84.9, 92.3, 85.1, 84.8] are significantly out of alignment with the other data points. Using a five-point median filter, these five data points are sorted to obtain [84.8, 84.9, 85.1, 85.2, 92.3]. The median value 85.1 is selected as the center point of the filtered result. The mean of this local interval is 86.46, with a standard deviation of 3.24. Point 92.3 falls outside the range of [76.74, 96.18] ±3 times the standard deviation of the mean and is therefore marked as an outlier. For the marked outlier point 92.3, a weighted average of its three neighboring points on each side is used to replace it. The weight coefficients for the neighboring points are set inversely proportional to their distance, specifically 1 / 1, 1 / 2, and 1 / 3, which, after normalization, are 0.55, 0.27, and 0.18, respectively. Taking the three data points [84.8, 84.9, 85.1] to the left of the outlier as an example, the weighted average calculation results in 84.9, which replaces the original outlier value. In wavelet transform processing, the db4 wavelet basis function is used to perform a four-layer decomposition of the data sequence. Taking a data segment of length 16 as an example, the first-layer decomposition yields 8 low-frequency coefficients and 8 high-frequency coefficients. When adaptively thresholding the high-frequency coefficients, the threshold T = σ√(2logN) = 1.28 is calculated based on the local variance σ = 0.45, where N is the data length. For high-frequency coefficients with amplitudes less than 1.28, soft thresholding is performed, setting them to 0 to achieve denoising. The data quality assessment phase comprehensively examines three key indicators: signal-to-noise ratio, spectral resolution, and baseline drift. Taking the spectral range near 500nm as an example, the calculated signal-to-noise ratio is 42.5dB, which is higher than the preset threshold of 40dB; the spectral resolution is 0.15nm, which is better than the preset threshold of 0.2nm; the baseline drift amplitude is 0.8%, which is lower than the preset threshold of 1%, indicating that the data processing quality is good. The Savitzky-Golay smoothing algorithm uses a third-order polynomial to fit the seven-point window data.

[0025] For example, using the polynomial [84.8, 84.9, 85.1, 85.0, 84.9, 85.2, 85.0], we use recursive least squares to solve the polynomial coefficients [a, b, c, d] = [0.02, -0.15, 0.43, 84.9], obtaining a smoothed center value of 85.0. We then slide the window point by point, performing similar processing on each data point, ultimately obtaining a smoothed scattering spectrum curve.

[0026] S103. Extract characteristic parameters from the smoothed scattering spectrum curve, use a peak detection algorithm to determine the main peak position, peak width, and peak intensity of the scattering spectrum curve, and construct a scattering spectrum characteristic vector of the carbon powder sample.

[0027] A five-point difference method is used to calculate the first-order derivative and the second-order derivative of the spectral curve, and a set of peak candidate points is obtained according to the position where the positive and negative signs of the first-order derivative change and the second-order derivative is negative; a three-point fitting method is used to calculate the local curvature value for the peak candidate point set, and a first spectral feature sequence is obtained from the local curvature value; based on the intensity ratio of the peak position to the left and right adjacent valley points in the first spectral feature sequence, it is judged whether the peak-to-valley ratio exceeds the calibration threshold range of the spectrometer; if the peak-to-valley ratio exceeds the calibration threshold range, the peak position is eliminated to obtain a second spectral feature sequence; for the second spectral feature sequence, a cubic spline interpolation method is used to determine the wavelength value of the peak position, and the peak width parameter is obtained by calculating the left and right wavelength difference at half the peak height through linear interpolation, and a third spectral feature sequence is obtained from the wavelength difference and the peak intensity.

[0028] Exemplarily, the first and second derivatives of the five-point difference are calculated based on the smoothed scattering spectrum curve. By determining the position where the sign of the first-order derivative changes and the second-order derivative is negative, a set of peak candidate points is obtained from the spectrum curve. The local curvature value is calculated using the three-point fitting method for the candidate points, and a first spectral feature sequence is obtained from the candidate point set. The peak positions in the first spectral feature sequence are screened, and a discrimination method based on the peak-to-valley ratio is used. By calculating the intensity ratio of the peak position to the left and right adjacent valley points, if the peak-to-valley ratio exceeds the spectrometer calibration threshold range, the peak position is eliminated to obtain a second spectral feature sequence. For each peak position in the second spectral feature sequence, the cubic spline interpolation method is used to determine the precise wavelength value of the peak position. The peak width parameter is obtained by linear interpolation to calculate the left and right wavelength difference at half the peak height. The third spectral feature sequence is obtained from the wavelength difference and peak intensity. The three characteristic parameters of peak position, peak intensity, and peak width in the third spectral feature sequence are normalized to their maximum and minimum values, and a standardized characteristic parameter sequence is obtained through normalization. A support vector machine based on a Gaussian kernel function was used to reduce the dimensionality of the standardized feature parameter sequence. The kernel function parameters were determined through cross-validation, and a fixed-dimensional feature vector was obtained after dimensionality reduction. The feature extraction quality was verified based on the feature vector. The reliability of the feature extraction results was assessed by calculating the feature reconstruction error and information retention rate. The feature vector of the carbon powder sample scattering spectrum was obtained from the verification results. When extracting the feature of the carbon powder scattering spectrum curve, the first-order derivative was calculated using five-point difference. For the light intensity value f(x) at wavelength point x, the first-order derivative was calculated as f'(x) = [f(x+2)-8f(x+1)+8f(x-1)-f(x-2)] / 12. For example, the light intensity values ​​at five adjacent points at a wavelength of 500 nm are [84.8, 85.1, 85.3, 84.9, 84.5], respectively. Substituting this into the formula yields a first-order derivative value of -0.15. The second-order derivative is calculated using a similar method and the value is -0.42, indicating that the point is a local maximum point and is used as a peak candidate point. The local curvature is calculated using the three-point fitting method, selecting the peak candidate point and its left and right adjacent points to form the parabola equation y=ax 2 +bx+c. Taking the peak candidate point at a wavelength of 500nm as an example, three points [498nm,85.1], [500nm,85.3], and [502nm,84.9] with a wavelength interval of 2nm are selected. By solving the linear equation system, the parabola coefficient a=-0.05 and the local curvature κ=|2a| / [1+(f'(x)) 2]^(3 / 2)=0.092. In the peak-to-valley ratio discrimination, the calibration threshold range of the spectrometer is usually set to 1.2-5.0. For the peak point at 500nm, its left valley point is located at 495nm with an intensity of 82.4, and the right valley point is located at 506nm with an intensity of 82.1. The peak-to-valley ratio is calculated to be (85.3-82.4) / (82.4-82.1)=3.1, which is within the calibration threshold range, and the peak point is retained. For the determined peak point, cubic spline interpolation is used for precise positioning. 10 interpolation points are set in the wavelength range [499nm, 501nm] to obtain a more refined light intensity data sequence. Half of the peak height is determined by linear interpolation, that is, the left and right wavelength values ​​corresponding to 84.2 are 499.4nm and 500.7nm respectively, and the peak width is calculated to be 1.3nm. The characteristic parameter normalization uses the maximum and minimum method. Taking peak intensity as an example, if the original data range is [82.1, 85.3], the normalization formula is y = (x-82.1) / (85.3-82.1). The peak intensity of 85.3 at 500nm is normalized to 1.0. The Gaussian kernel function of the support vector machine is K(x,y) = exp(-||xy|| 2 / 2σ 2 ), where σ is the kernel function parameter. Five-fold cross-validation determined the optimal σ value to be 0.8. The normalized feature parameters were then subjected to dimensionality reduction, reducing the original 9-dimensional features (3 peak points × 3 class parameters) to a 4-dimensional feature vector. Feature extraction quality was evaluated by calculating the reconstruction error and information retention rate. The reconstruction error, defined as the root mean square error between the original and reconstructed features, was calculated to be 0.075, and the information retention rate reached 94.3%, indicating that the feature extraction results were reliable.

[0029] S104. Based on the scattered spectrum feature vector of the toner sample, a support vector machine algorithm is used to determine the purity of the toner sample. If the scattered spectrum feature vector falls within the feature space of a preset high-purity toner sample, the purity of the toner sample is determined to have reached the target purity; otherwise, the purity of the toner sample is determined to have not reached the target purity.

[0030] According to the characteristic vector of the scattering spectrum of the carbon powder sample, the principal component analysis method is used to perform an orthogonal transformation on the characteristic vector, and the first characteristic space coordinate value is obtained from the projection matrix; for the first characteristic space coordinate value, the characteristic space boundary is constructed by a support vector machine based on the Gaussian kernel function, and the second characteristic space coordinate value is obtained from the kernel function mapping; according to the second characteristic space coordinate value, the sample intra-class scatter matrix is ​​calculated, and the distance from the sample to the characteristic space boundary is calculated using the Mahalanobis distance, and the sample purity judgment threshold is obtained from the distance distribution histogram; for the characteristic vector of the carbon powder sample to be judged, it is projected to the second characteristic space through the kernel function mapping. If the Mahalanobis distance value is less than the lower limit of the confidence interval, it is judged that the sample does not reach the target purity.

[0031] Exemplarily, based on the eigenvectors of the scattering spectrum of a toner sample, principal component analysis is used to perform an orthogonal transformation on the eigenvectors. Eigenvectors corresponding to eigenvalues ​​whose cumulative contribution rates exceed a preset threshold are selected to construct a feature space projection matrix, and coordinate values ​​in the first feature space are obtained from the projection matrix. For the first feature space coordinate values, a support vector machine based on a Gaussian kernel function is used to construct the boundary of the feature space of the high-purity toner sample. The kernel function parameters are optimized through cross-validation, and coordinate values ​​in the second feature space are obtained from the kernel function mapping. Based on the second feature space coordinate values, the sample intra-class scatter matrix is ​​calculated. The distance from the sample to the feature space boundary is calculated using the Mahalanobis distance. The confidence interval within the preset parameter is determined from the distance distribution histogram to obtain the sample purity determination threshold. The second feature space coordinate values ​​are then subjected to feature distribution verification, random sample points are generated using the Monte Carlo method, and a stability index of the feature space boundary is calculated. The feature space confidence score is obtained from the stability index. For the toner sample eigenvector to be determined, the coordinates are transformed using the feature space projection matrix, projected into the second feature space using kernel function mapping, and the Mahalanobis distance from the sample point to the feature space boundary is calculated. The sample purity is judged based on the Mahalanobis distance value and the confidence score of the feature space. If the distance value is less than the lower limit of the confidence interval or the confidence score is lower than the preset threshold, the sample is judged to have not reached the target purity. Otherwise, the sample is judged to have reached the target purity. In the process of determining the purity of the toner sample, the principal component analysis of the eigenvector is a key step. The original eigenvector contains 9 dimensions, corresponding to the wavelength value, intensity value and peak width value of the 3 peak positions. By calculating the eigenvalues

[0032] [2.45,1.82,1.36,0.85,0.42,0.28,0.15,0.09,0.06] and its cumulative contribution rate

[0033] [32.6%, 56.8%, 74.9%, 86.2%, 91.8%, 95.5%, 97.5%, 98.7%, 100%], select the eigenvectors corresponding to the first 6 eigenvalues ​​with cumulative contribution rates exceeding 95% to construct the projection matrix. The parameter γ in the Gaussian kernel function K(x, y) = exp(-γ||xy||2) is determined by five-fold cross validation.

[0034] Among the values ​​[0.1, 0.2, 0.5, 1.0, 2.0], the highest classification accuracy is achieved when γ = 0.5, reaching 96.8%. For the sample points (0.82, 0.65, 0.43, 0.38, 0.25, 0.12) in the 6-dimensional feature space, the coordinates in the second feature space are obtained by kernel mapping (0.78, 0.61, 0.40). The intra-class scatter matrix S is obtained by calculating the covariance matrix within a high-purity sample group. Taking the 3-dimensional feature space as an example, the diagonal elements of the scatter matrix are [0.024, 0.018, 0.015], representing the variance of each dimension. For the sample point (0.78, 0.61, 0.40) to be determined, the class center coordinates are (0.80, 0.62, 0.41), and the calculated Mahalanobis distance is 1.85. 1000 random sample points were generated using the Monte Carlo method to evaluate the stability of the feature space boundary. The Mahalanobis distance distribution of random samples follows a chi-square distribution, with a critical value of 7.81 at a 95% confidence level. The feature space boundary stability index, defined as the proportion of random points within the boundary, was calculated to be 0.942, corresponding to a feature space confidence score of 94.2%. The Mahalanobis distance value of the sample to be determined was 1.85, significantly lower than the critical value of 7.81 corresponding to a 95% confidence level. At the same time, the feature space confidence score of 94.2% was higher than the preset threshold of 90%, indicating that the sample met the target purity. The reliability of this determination is supported by two factors: first, the position of the sample points in the feature space is relatively stable, with the Mahalanobis distance being much lower than the determination threshold; second, the feature space is constructed with high confidence, with accurate and reliable boundary demarcation, forming a complete chain of feature extraction, space construction, and purity determination. Each link is connected by specific numerical indicators, ensuring the traceability and reliability of the determination results.

[0035] S105. For carbon powder samples that do not reach the target purity, the abnormal change characteristics of the corresponding scattering spectrum curve are analyzed, and the scattering spectrum curve is decomposed using a spectrum peak fitting method to obtain the scattering peak caused by adsorbed impurities and the characteristic scattering peak of the carbon powder itself.

[0036] A four-layer wavelet transform is used to perform multi-scale decomposition on the scattering spectrum curve of the carbon powder sample that does not reach the target purity. The spectral baseline is obtained by reconstructing the low-frequency coefficients, and the baseline is subtracted from the spectral curve to obtain a first scattering spectrum sequence. Based on the first scattering spectrum sequence, the scattering peak is fitted using a Gaussian-Lorentzian mixed function, and the function parameters are iteratively optimized using the least squares method until the residual is less than a preset threshold to obtain a second scattering spectrum sequence. For the second scattering spectrum sequence, the area ratio of the Gaussian component to the Lorentzian component in the mixed function is calculated, and a scattering peak identification criterion is established based on the ratio distribution to obtain a third scattering spectrum sequence. For the scattering components in the third scattering spectrum sequence, the peak position distribution is clustered using the Gudan kernel density estimation method, and the carbon powder intrinsic characteristic scattering peak and impurity scattering peak are obtained from the cluster center.

[0037] For example, based on the scattering spectrum curve of the carbon powder sample that does not reach the target purity, a four-layer wavelet transform is used to perform multi-scale decomposition of the spectrum curve, and the spectrum baseline is obtained by reconstructing the low-frequency coefficients. The baseline is subtracted from the spectrum curve to obtain the first scattering spectrum sequence. For the first scattering spectrum sequence, a Gaussian-Lorentzian mixed function is used to fit the scattering peak. The mixed function form is f(x) = A1exp[-(x-μ) 2 / 2σ 2 ]+A2 / [1+(x-x0) 2 / γ 2], the function parameters are iteratively optimized by the least squares method, and the iteration is stopped when the residual is less than the preset threshold, thereby obtaining the second scattering spectrum sequence. According to the fitting parameters of the second scattering spectrum sequence, the area ratio of the Gaussian component to the Lorentzian component in the mixing function is calculated, and the identification criterion of the scattering peak is established through the ratio distribution, and the third scattering spectrum sequence is obtained from the identification criterion. The fitting quality of the third scattering spectrum sequence is evaluated, and the reliability of the fitting result is judged by calculating the goodness of fit and the residual distribution characteristics, and the fourth scattering spectrum sequence is obtained from the reliability judgment. According to the scattering peak parameters in the fourth scattering spectrum sequence, the peak shape is orthogonally decomposed by principal component regression, and the impurity scattering component and the carbon powder intrinsic scattering component are separated by variance maximization rotation. For the separated scattering components, the peak position distribution is clustered by the Gudan kernel density estimation method. The cluster radius is determined based on the standard deviation of the peak position, and the carbon powder intrinsic characteristic scattering peak and the impurity scattering peak are obtained from the cluster center. In the analysis of the carbon powder scattering spectrum curve, the baseline extraction of the four-layer wavelet transform adopts the Debye Chebyshev wavelet basis function. Taking the spectral curve in the wavelength range of 400-700 nm as an example, the first-level decomposition yields scale coefficients [2.45, 1.82, 1.36] and detail coefficients [0.15, 0.09, 0.06]. Reconstructing the low-frequency coefficients yields a spectral baseline value of 85.3. Subtracting the baseline reduces the amplitude of the spectral curve to 3.2. During the Gaussian-Lorentzian mixture function fitting process, for the scattering peak at 500 nm, the initial parameters are set as Gaussian component parameters A1 = 2.8, μ = 500, σ = 1.2, and Lorentzian component parameters A2 = 1.5, x0 = 500, and γ = 0.8. After least-squares iterative optimization, the iterations are terminated when the residual is less than 0.01, resulting in the optimized parameters: A1 = 2.6, μ = 500.2, σ = 1.1, A2 = 1.4, x0 = 500.1, and γ = 0.7. In the calculation of the component area ratio of the mixing function, the Gaussian component area is A1σ√(2π)=7.2, the Lorentzian component area is A2πγ=3.1, and the area ratio is 2.32. According to a large number of sample statistics, the characteristic scattering peak area ratio of pure carbon powder is distributed between 2.0-2.5, while the area ratio of impurity scattering peaks is usually less than 1.5 or greater than 3.0. The determination coefficient R is used to evaluate the fitting quality. 2 For the scattering peak at 500nm, R 2 =0.986, residual standard deviation =0.008, both better than the preset indicators, R 2>0.95, and the residual standard deviation <0.01, indicating that the fitting results are reliable. During principal component regression decomposition, a covariance matrix was constructed and eigenvalue decomposition was performed, resulting in eigenvalues ​​[2.8, 1.5, 0.4] and corresponding eigenvectors. A rotation angle θ = 32° was determined using the maximum variance criterion to separate the impurity scattering component from the intrinsic scattering component. The separated impurity scattering peaks were located at 465nm and 545nm, while the intrinsic scattering peaks were located at 500nm and 600nm. Gudan kernel density estimation used a Gaussian kernel function with a bandwidth of h = 2nm. When clustering the separated scattering peaks, the calculated standard deviation of the peak positions was 1.8nm, which led to a cluster radius of 5.4nm. The clustering results showed that the positional coefficient of variation of the intrinsic scattering peaks was 0.4%, while the positional coefficient of variation of the impurity scattering peaks reached 1.2%, indicating a more dispersed distribution of the impurity scattering peaks.

[0038] Identify the peak intensity, peak width and peak symmetry of the scattering spectrum curve to analyze the corresponding abnormal change characteristics, calculate the first-order derivative and second-order derivative of the scattering peak to determine the inflection point and overlapping position of the peak, determine the boundary range and baseline position of the single scattering peak based on the abnormal change characteristics, accurately separate the overlapping scattering peaks, and extract the characteristic parameters of the impurity adsorption peak and the carbon powder intrinsic peak.

[0039] According to the light intensity value of the scattering spectrum curve, the first-order derivative and the second-order derivative are calculated using the five-point difference method, and the peak apex is determined by judging the position where the sign of the first-order derivative changes from positive to negative and the second-order derivative is negative, and a first scattering peak sequence is obtained from the peak apex; for the first scattering peak sequence, a third-order polynomial is used to fit the baseline, and the fitting coefficient is iteratively optimized by the least squares method until the residual does not exceed the residual threshold, and a second scattering peak sequence is obtained from the fitting result; based on the second scattering peak sequence, the overlapping coefficients of adjacent peaks are calculated, and if the overlapping coefficients exceed the overlapping coefficient threshold, the adjacent peaks are determined to be overlapping peaks, and a third scattering peak sequence is obtained from the overlapping state; for the overlapping peaks in the third scattering peak sequence, the maximum entropy deconvolution algorithm is used to separate them, and the iteration is stopped when the entropy function increment is less than the preset threshold, and a fourth scattering peak sequence is obtained from the separation result.

[0040] For example, according to the light intensity value of the scattering spectrum curve, the first-order derivative d1=(f2-8f1+8f -1 -f -2 ) / 12h and the second-order derivative d2=(f2-2f0+f -2 ) / 4h 2, by judging the position where the sign of the first-order derivative changes from positive to negative and the second-order derivative is negative, the peak vertex is determined, and the first scattering peak sequence is obtained from the peak vertex. For the first scattering peak sequence, a third-order polynomial is used to fit the baseline, and the fitting coefficient is iteratively optimized by the least squares method. If the residual exceeds the preset threshold, the polynomial order is increased to the fifth order in sequence, and the second scattering peak sequence is obtained from the fitting result. According to the second scattering peak sequence, the overlap coefficient of adjacent peaks is calculated, and the overlapping state of the peak is determined by judging whether the overlap coefficient exceeds the preset threshold. The third scattering peak sequence is obtained from the overlapping state. For the overlapping peaks in the third scattering peak sequence, the maximum entropy deconvolution algorithm is used for separation, and the entropy function increment is set to be less than the preset threshold as the iterative termination condition, and the fourth scattering peak sequence is obtained from the separation result. The peak intensity normalized value and peak width half-height value are calculated for the fourth scattering peak sequence, and the peak shape symmetry is characterized by the third-order origin moment and central moment. The fifth scattering peak sequence is obtained from the peak shape parameters. Based on the characteristic parameters of the fifth scattering peak sequence, the type of scattering peak is determined by calculating the peak symmetry deviation. If the symmetry deviation exceeds the preset threshold range, it is determined to be an impurity adsorption peak; otherwise, it is determined to be a carbon powder intrinsic peak. In the processing of the scattering spectrum curve, for the scattering peak near a wavelength of 500nm, when the five-point difference method is used to calculate the derivative, the light intensity values ​​of the five points with an interval of h = 2nm are selected as [84.8, 85.1, 85.3, 84.9, 84.5], respectively. Substituting into the first-order derivative formula, d1 = (84.5-8×84.9+8×85.1-85.3) / 24 = -0.15, and the second-order derivative d2 = (84.5-2×85.3+84.5) / 8 = -0.42. Since the first-order derivative changes from positive to negative and the second-order derivative is negative at this point, it is determined that this is the peak apex. The baseline fitting uses the third-order polynomial y = ax 3 +bx 2 +cx+d for initial fitting. For the data points in the wavelength range of 490-510nm, the coefficients [a, b, c, d] = [0.002, -0.015, 0.43, 82.5] are obtained by least squares method. The root mean square value of the calculated fitting residual is 0.28, which exceeds the preset threshold of 0.2. It is necessary to increase the polynomial order to the fourth order to continue optimization. In the overlap judgment, for two adjacent scattering peaks, the overlap coefficient is defined as R = H / (H1+H2), where H is the height of the lowest point between the two peaks, and H1 and H2 are the heights of the two peaks. Taking the two overlapping peaks at wavelengths of 500nm and 515nm as an example, H = 83.2, H1 = 85.3, H2 = 84.8, and the calculated overlap coefficient R = 0.49 exceeds the preset threshold of 0.4 and is determined to be an overlapping peak. When the maximum entropy deconvolution algorithm is used to separate overlapping peaks, the entropy function is defined as S = -∑p i lnp i , where p iis the normalized light intensity value. After setting the initial distribution, iterative calculation is performed, and the iteration is stopped when the entropy function increment is less than 0.001. For the above overlapping peaks, two separated Gaussian peaks are obtained after 15 iterations, and the peak positions are located at 500.2nm and 514.8nm respectively. The peak shape symmetry analysis is characterized by the third-order moment. For the separated scattering peaks, the origin moment M1 = 500.2 is calculated, which is the peak position, M2 = 1.24, which is the peak width, and M3 = 0.086, which is the asymmetry. At the same time, the central moment μ3 = M3-3M1M2+2M1 is calculated. 3 = 0.028. The closer the central moment is to zero, the more symmetrical the peak shape. Based on extensive sample statistics, the absolute value of the third-order central moment of the carbon powder intrinsic peak is typically less than 0.05, while the absolute value of the third-order central moment of the impurity adsorption peak is often greater than 0.1. Based on this, the scattering peak at 500.2nm is determined to be the carbon powder intrinsic peak, while the scattering peak at 514.8nm is the impurity adsorption peak. The above analysis process achieves quantitative characterization of complex spectral curves through multiple aspects such as derivative analysis, baseline fitting, overlap determination, peak shape separation, and symmetry calculation.

[0041] S106. Based on the position and intensity information of the scattering peaks of the adsorbed impurities obtained by decomposition, the principal component analysis method is used to perform qualitative and quantitative analysis on the type and content of the adsorbed impurities, thereby obtaining the main impurity components and relative contents of the main impurities that affect the purity of the carbon powder.

[0042] The maximum and minimum values ​​of the wavelength positions and intensity values ​​of the impurity scattering peaks are normalized, and the first impurity characteristic matrix is ​​obtained by eliminating abnormal values ​​that exceed the preset standard deviation range; the Pearson correlation coefficient is calculated for the first impurity characteristic matrix, and the second impurity characteristic matrix is ​​obtained by performing singular value decomposition on the correlation coefficient matrix and selecting characteristic values ​​whose cumulative contribution rates exceed the contribution rate threshold; the second impurity characteristic matrix is ​​projected using the maximum variance method, and the third impurity characteristic matrix is ​​obtained by calculating the projection error and judging whether the projection error exceeds the preset threshold; the probability density is calculated by kernel density estimation for the third impurity characteristic matrix, and the impurity component type is obtained by calculating the Mahalanobis distance between the sample point and the cluster center and judging whether the Mahalanobis distance exceeds the distance threshold.

[0043] For example, the data is normalized using the maximum-minimum normalization method based on the wavelength positions and intensities of the impurity scattering peaks. Outliers exceeding three standard deviations are removed to obtain a first impurity characteristic matrix from the normalized data. For the first impurity characteristic matrix, the Pearson correlation coefficient is used to calculate the correlation between characteristic wavelengths. The correlation coefficient matrix is ​​diagonalized using singular value decomposition, and eigenvalues ​​whose cumulative contribution exceeds a preset threshold are selected. A second impurity characteristic matrix is ​​obtained from the eigenvectors. Based on the second impurity characteristic matrix, the data is projected using the maximum variance method. The projection error is calculated using cross-validation. If the error exceeds a preset threshold, the number of eigenvalues ​​is increased. A third impurity characteristic matrix is ​​obtained from the projection results. For the third impurity characteristic matrix, probability density is calculated using kernel density estimation with a fixed bandwidth. Clustering effectiveness is evaluated using the silhouette coefficient. A fourth impurity characteristic matrix is ​​obtained from the clustering results. Based on the cluster centers of the fourth impurity characteristic matrix, the Mahalanobis distance is used to calculate the distance between the sample points and the cluster centers. The impurity type is assigned using a distance threshold, and the impurity component type is determined from the assignment results. For the identified impurity types, a content prediction equation was established using multivariate linear regression. Residual analysis was used to verify the model's reliability, and the relative content of the main impurities was derived from the regression equation. In impurity component analysis, the raw scattering peak data consists of two characteristic dimensions: wavelength position and intensity. For a particular set of impurity scattering peaks, the wavelength positions were distributed between [465, 545, 625] nm, corresponding to intensity values ​​of [32.4, 28.7, 25.5]. By normalizing the data to the interval [0, 1] using maximum and minimum values, the standardized intensity values ​​were mapped to [1.0, 0.46, 0]. The calculated intensity mean was 28.87, with a standard deviation of 3.48. The outlier determination range was determined to be [18.43, 39.31], indicating that this data set was within the normal range. For the wavelength and intensity features, the correlation coefficient matrix R = [[1.0, -0.92], [-0.92, 1.0]]. The eigenvalues ​​[1.92, 0.08] were obtained by singular value decomposition, with cumulative contribution rates of 96% and 100% respectively. The first eigenvalue with a contribution rate exceeding 95% was selected, and the corresponding eigenvector was [0.707, -0.707]. In the maximum variance projection, the two-dimensional data was projected into one-dimensional space through the eigenvector. The above data points were projected to obtain projection values ​​[-0.495, 0.318, 0.857]. The projection error was calculated using five-fold cross validation, and the root mean square value of the reconstruction error was 0.068, which was less than the preset threshold of 0.1, indicating that the projection effect was good. The kernel density estimation uses the Gaussian kernel function K(x) = (1 / √(2π))exp(-x 2 / 2), the bandwidth parameter h is determined to be 0.25 by the Silverman criterion. The probability density function is calculated, and three obvious peaks are shown in the projection space, corresponding to different types of impurity components. The clustering effect is evaluated by the silhouette coefficient. When the number of clusters is 3, the silhouette coefficient is the largest, reaching 0.68. For the three types of impurities obtained by clustering, the Mahalanobis distance is used to determine the attribution of samples. Taking the first type of impurity as an example, the cluster center is located at -0.482 in the projection space, and the covariance is 0.024. For the sample point with a projection value of -0.495, the Mahalanobis distance is calculated to be 0.84, which is less than the threshold value of 2.0, and it is determined to belong to this type of impurity. The least squares method is used in multivariate linear regression to establish the impurity content prediction equation y=β1x1+β2x2+β3x3. The parameters β1=0.42, β2=0.35, β3=0.23, and R 2 =0.93. Residual analysis showed a normal distribution, with 95% of the residuals falling within the ±2σ range, validating the reliability of the regression model. Substituting the sample scattering peak intensity data, the relative contents of the three impurities were calculated to be 42%, 35%, and 23%, respectively.

[0044] The chemical bond structure characteristics of the impurity are judged according to the wavelength position of the scattering peak, the crystal structure characteristics of the impurity are analyzed according to the shape and symmetry of the scattering peak, the relative concentration of the impurity is calculated using the relative intensity ratio of the scattering peak, and the standard curve of the impurity content is established using the external standard method. The scattering peak intensity of the unknown impurity is compared with the standard curve to obtain the content level of the unknown impurity. The impurity component fingerprint is constructed based on the position and intensity characteristics of the scattering peak, and the main impurity types and contents in the carbon powder are analyzed.

[0045] The spectral wavenumber is obtained according to the wavelength position of the impurity scattering peak, and the spectral wavenumber is compared with the bond vibration characteristic wavenumber database through a conversion formula to obtain a first impurity characteristic sequence; the first impurity characteristic sequence is subjected to Fourier series expansion, and the odd-even ratio of the Fourier series expansion coefficient characterizes the peak shape symmetry to obtain a second impurity characteristic sequence; the second impurity characteristic sequence is used to establish a third-order polynomial equation, and the third-order polynomial equation is iteratively optimized with the least squares method to obtain a third impurity characteristic sequence; a random forest model is constructed for the third impurity characteristic sequence, and the random forest model selects a feature subset through the Gini coefficient, and the cosine similarity between the feature subset and the standard spectrum is calculated to obtain the impurity type discrimination result.

[0046] For example, based on the wavelength position and offset of the impurity scattering peak, the spectral wavenumber conversion formula ν = 1 / λ is used to calculate the chemical bond vibration frequency, and the frequency is assigned to a spectral band through the bond vibration characteristic wavenumber database, and the first impurity characteristic sequence is obtained from the characteristic wavenumber. For the first impurity characteristic sequence, the scattering peak shape is expanded using an eight-term Fourier series, and the peak shape symmetry is characterized by calculating the odd-even ratio of the expansion coefficient, and the second impurity characteristic sequence is obtained from the symmetry parameter. According to the second impurity characteristic sequence, the third-order polynomial equation y = ax is used. 3 +bx 2 A curve was constructed using the least squares method to iteratively optimize the fitting parameters, yielding a third impurity signature sequence. The dependent variable, y, was the scattering peak intensity; the independent variable, x, was the impurity concentration; the coefficients of the cubic term, b and c were the quadratic term, and d were the constant term. The third impurity signature sequence was validated against the standard curve using a leave-one-out cross-validation algorithm. The reliability of the standard curve was statistically evaluated using the prediction error. The validation results yielded a fourth impurity signature sequence. Based on the fourth impurity signature sequence, a random forest algorithm was used to extract impurity type features. The optimal feature subset was selected using the Gini coefficient, and a fifth impurity signature sequence was derived from this feature subset. A fingerprint was constructed for the fifth impurity signature sequence. The cosine similarity was used to calculate the degree of match between the unknown sample and the standard signature. The impurity type was determined by setting a similarity threshold. The matching results revealed the type and content of the carbon powder impurities. In impurity scattering spectral analysis, the vibration frequency of chemical bonds has an inverse relationship with the wavelength. Taking the scattering peak at a wavelength of 465 nm as an example, the vibration frequency is calculated to be 21505 cm by the wave number conversion formula ν = 1 / λ. -1 , corresponding to the stretching vibration characteristic frequency range of C=C double bond 21000-22000cm -1 Similarly, the scattering peaks at 545nm and 625nm correspond to 18348cm -1 and 16000cm -1 , matching the vibration frequencies of CO and CH bonds. The peak symmetry analysis was performed using the Fourier series expansion f(x)=a0+Σ(a n cos(nx)+b n sin(nx)), a0 is the DC component (constant term), a n is the cosine term coefficient, n is the harmonic order, b nis the sinusoidal term coefficient. Taking the scattering peak at 545nm as an example, the eight-term series expansion is used to obtain the coefficients [a0, a1, b1, ..., a4, b4] = [28.7, 12.4, -0.8, 5.2, -0.3, 1.8, -0.2, 0.6, -0.1]. The calculated odd-even coefficient ratio (a1 + a3) / (a2 + a4) = 2.85 is greater than the preset threshold of 1.5, indicating that the peak shape has obvious asymmetry. The standard curve is established using a third-order polynomial fit. For the characteristic scattering peak of the C = C bond, 5 standard samples with concentration gradients of [0.5%, 1.0%, 2.0%, 4.0%, 8.0%] are selected, and the measured peak intensity values ​​are [15.2, 28.7, 54.3, 102.8, 198.5] respectively. The equation y = 0.024x is obtained by least squares fitting. 3 -0.185x 2 +3.246x+2.157, goodness of fit R 2 =0.998. In the leave-one-out cross-validation, each standard point is used as a validation sample in turn. Taking the 2.0% concentration point as an example, the standard curve established using the other four points predicts a concentration value of 2.15%, with a relative error of 7.5%. The prediction errors of all standard points were statistically analyzed, and the average relative error was 8.2%, which verified the reliability of the standard curve. 200 decision trees were selected for random forest feature extraction, and the Gini coefficient was used for feature selection of each node. For the impurity feature sequence [21505, 2.85, 0.024, 8.2], the importance scores of each feature were calculated to be [0.42, 0.28, 0.18, 0.12], respectively. The wavenumber feature has the highest discriminant contribution. Fingerprint matching is calculated using cosine similarity, and the similarity is defined as

[0047] cos(θ)=(a·b) / (|a|·|b|). Taking an unknown sample as an example, its eigenvector is

[0048] [21498, 2.82, 0.025, 8.4], with a calculated similarity of 0.986 with the standard spectrum, exceeding the preset threshold of 0.95, and identified as an organic impurity containing a C=C bond. The standard curve predicted the impurity content to be 2.8%.

[0049] S107. Based on the purity determination results and impurity analysis results of the toner samples, a decision tree algorithm is used to establish a purity assessment model for the toner purity. Based on the toner scattering spectrum characteristics and impurity analysis results, the toner purity is quantitatively scored to provide guidance for toner quality optimization.

[0050] According to the carbon powder scattering spectrum feature vector and impurity analysis data, the maximum and minimum value normalization method is used to map the carbon powder scattering spectrum feature vector and impurity analysis data to the interval [0, 1], and the first feature matrix is ​​obtained by eliminating abnormal values ​​beyond the preset standard deviation range; for the first feature matrix, the Pearson correlation coefficient is used to calculate the correlation between features, and the second feature matrix is ​​obtained by setting the absolute value of the correlation coefficient greater than the preset threshold to filter features; for the second feature matrix, the variance contribution rate is used to calculate the feature importance score, and the third feature matrix is ​​obtained by filtering features by the cumulative contribution rate exceeding the preset threshold; for the third feature matrix, a decision tree is used to establish a purity assessment model, and the optimal splitting feature is selected through the Gini coefficient to obtain the fourth feature matrix, and the upper limit of the decision tree depth is set to the preset number of layers.

[0051] For example, normalization is performed based on the carbon powder scattering spectrum feature vectors and impurity analysis data. The data is mapped to the interval [0, 1] using the maximum-minimum normalization method. Outliers exceeding three standard deviations are removed to obtain a first feature matrix from the normalized data. For the first feature matrix, the Pearson correlation coefficient is used to calculate the correlation between features. A strong correlation threshold is set at an absolute value of the correlation coefficient greater than 0.8. From this correlation analysis, a second feature matrix is ​​obtained. Based on the second feature matrix, features are ranked by importance using the variance contribution ratio. Features are screened by cumulative contribution ratios exceeding a preset threshold. The screening results yield a third feature matrix. For the third feature matrix, a decision tree is used to model purity assessment rules, with an upper limit of 5 levels set for the tree depth. The Gini coefficient is used to select the optimal splitting features. The fourth feature matrix is ​​obtained from the tree structure. Based on the fourth feature matrix, a weighted summation method is used to calculate the comprehensive score. Weight coefficients are set according to the feature importance scores. The scores are mapped to the interval [0-100] using a piecewise function. The mapping results yield a fifth feature matrix. For the fifth characteristic matrix, a five-fold cross validation is used to evaluate the model performance. By calculating the average accuracy and mean square error, if the indicators do not meet the preset requirements, the feature selection step is returned to re-optimize, and the toner purity score is obtained from the verification results. In the toner purity evaluation, the original features include two categories: scattering spectrum features and impurity analysis results. Taking a certain sample as an example, the scattering spectrum features include wavelength position [465,545,625] nm, peak intensity [32.4,28.7,25.5] and peak width [15.2,14.8,16.3]. The impurity analysis results include the number of impurity types 3 and the total content 8.6%. By normalizing the maximum and minimum values, the eigenvalues ​​are mapped to the [0,1] interval to obtain the standardized eigenvector

[0052] [0.82, 0.65, 0.43, 1.0, 0.46, 0.0, 0.0, 0.24, 0.67, 0.75, 0.86]. The correlation coefficients of the standardized feature matrix were calculated, and the correlation coefficients between peak position and intensity were -0.92, 0.35, and 0.84, respectively. Strongly correlated features were removed by screening based on a threshold with an absolute value of the correlation coefficient greater than 0.8. Feature importance was assessed using the variance contribution rate, calculating the proportion of each feature's variance to the total variance. The contribution rates of the retained features were calculated, yielding the following: peak position 0.42, peak width 0.28, impurity type 0.18, and impurity content 0.12. The feature subset with a cumulative contribution rate exceeding 0.9 included peak position and peak width. Decision tree modeling used the Gini coefficient criterion for feature splitting. Taking the peak position feature as an example, the Gini coefficients at different splitting points were calculated, with the minimum value of 0.28 being achieved at 0.6. The first-level splitting rule for the tree is that if the peak position is ≤ 0.6, the sample is classified as high purity; otherwise, it is classified as low purity. A complete decision tree structure was constructed within the preset five-level depth limit. The score was calculated using the weighted summation formula score = Σwixi, with weights set according to feature importance to [0.42, 0.28, 0.18, 0.12]. Substituting the sample feature values ​​into the original score of 0.724, which was mapped to the [0-100] interval using a piecewise function, yielding a final score of 72.4. The score interval is 90-100 points for high-quality products, 70-89 points for acceptable products, and scores below 70 for unacceptable products. In the five-fold cross-validation, the dataset was divided into five parts, with four of these parts used as training sets and one as validation sets. The validation results for one fold yielded an accuracy of 0.92 and a mean squared error of 0.086. The average accuracy of the 50% fold reached 0.90, and the mean square error was 0.094, meeting the preset verification index requirements of accuracy > 0.85 and mean square error < 0.1.

[0053] The above embodiment is only one of the preferred implementation methods of the present invention and should not be used to limit the scope of protection of the present invention. Any changes or modifications that have no substantive meaning made to the main design concept and spirit of the present invention, as long as the technical problems they solve are still consistent with the present invention, should be included in the scope of protection of the present invention.

Claims

1. A method for intelligent analysis of carbon powder purity using deep learning, characterized in that: The method comprises: Acquire scattering spectrum data of the toner sample, calculate the scattered light intensity of the toner sample at different wavelengths, and obtain original scattering spectrum curve data of the toner sample; The original scattering spectrum curve data of the carbon powder sample is preprocessed by using data smoothing and denoising algorithms to eliminate high-frequency noise and abnormal points in the curve data and obtain a smooth scattering spectrum curve; Characteristic parameters are extracted from the smoothed scattering spectrum curve. The peak detection algorithm is used to determine the main peak position, peak width and peak intensity of the scattering spectrum curve, and the scattering spectrum characteristic vector of the carbon powder sample is constructed. Based on the scattered spectrum feature vector of the toner sample, the support vector machine algorithm is used to determine the purity of the toner sample. If the scattered spectrum feature vector falls within the feature space of the preset high-purity toner sample, the toner sample is judged to have reached the target purity; otherwise, it is judged to have not reached the target purity. For carbon powder samples that do not reach the target purity, the abnormal change characteristics of the corresponding scattering spectrum curve are analyzed, and the scattering spectrum curve is decomposed using the peak fitting method to obtain the scattering peak caused by adsorbed impurities and the characteristic scattering peak of the carbon powder itself; According to the position and intensity information of the scattering peaks of adsorbed impurities obtained by decomposition, the principal component analysis method is used to conduct qualitative and quantitative analysis on the type and content of adsorbed impurities, and the main impurity components and relative contents of the main impurities that affect the purity of carbon powder are obtained; Based on the purity determination results and impurity analysis results of the toner samples, a decision tree algorithm is used to establish a purity evaluation model for toner purity. According to the toner scattering spectrum characteristics and impurity analysis results, the toner purity is quantitatively scored to provide guidance for toner quality optimization.

2. The method according to claim 1, characterized in that The step of obtaining scattered light spectrum data of the toner sample, calculating the scattered light intensity of the toner sample at different wavelengths, and obtaining original scattered light spectrum curve data of the toner sample comprises: The light signal is collected on the surface of the carbon powder sample according to a preset collection interval of the scattering spectrometer, and a first set of scattering spectrum data within a wavelength band is obtained by a photoelectric detector to obtain a first scattering spectrum sequence; For each wavelength point in the first scattering spectrum sequence, a smoothing method is used to process the data. If the fluctuation amplitude of the wavelength point signal exceeds the calibration threshold of the scattering spectrometer, the wavelength point is repeatedly collected to obtain a second scattering spectrum sequence; Constructing a spectral absorption curve based on the second scattering spectrum sequence, and fitting the spectral absorption curve using a three-layer forward neural network, wherein the input layer of the neural network is the wavelength value and the output layer is the scattered light intensity value corresponding to the wavelength, to obtain a third scattering spectrum sequence; For the third scattering spectrum sequence, Fourier transform is used to perform frequency domain analysis, and high-frequency noise and low-frequency baseline drift are filtered out by a bandpass filter to obtain scattering spectrum curve data of the carbon powder sample.

3. The method according to claim 1, characterized in that The raw scattering spectrum curve data of the carbon powder sample is preprocessed, and a data smoothing and denoising algorithm is used to eliminate high-frequency noise and abnormal points in the curve data to obtain a smooth scattering spectrum curve, including: A median filter is used to perform sliding window processing on the spectral data sequence, and abnormal points are marked by calculating the local interval mean and standard deviation to obtain a first spectral data sequence; According to the abnormal points marked in the first spectral data sequence, the abnormal points are replaced by a local weighted average method, and a second spectral data sequence is obtained from the replaced data sequence; performing wavelet multi-scale decomposition on the second spectral data sequence, performing soft threshold processing on the high-frequency noise coefficient using an adaptive threshold, and obtaining a third spectral data sequence from the reconstructed wavelet coefficients; For the third spectral data sequence, the Savitzky-Golay smoothing method is used for data fitting, the polynomial coefficients of the local interval are estimated by the recursive least squares method, and a smoothed scattering spectrum curve is obtained from the fitting curve.

4. The method according to claim 1, wherein The method extracts characteristic parameters from the smoothed scattering spectrum curve, uses a peak detection algorithm to determine the main peak position, peak width and peak intensity of the scattering spectrum curve, and constructs a scattering spectrum characteristic vector of the carbon powder sample, including: A five-point difference method is used to calculate the first-order derivative and the second-order derivative of the spectral curve, and a set of peak candidate points is obtained according to the positions where the first-order derivative changes in sign and the second-order derivative is negative; Calculating local curvature values ​​using a three-point fitting method for the peak candidate point set, and obtaining a first spectral feature sequence from the local curvature values; Determining whether the peak-to-valley ratio exceeds a spectrometer calibration threshold range based on an intensity ratio between a peak position and its left and right adjacent valley points in the first spectral feature sequence; and removing the peak position if the peak-to-valley ratio exceeds the calibration threshold range to obtain a second spectral feature sequence; For the second spectral characteristic sequence, the wavelength value of the peak position is determined by cubic spline interpolation method, the peak width parameter is obtained by linear interpolation calculation of the left and right wavelength difference at half the peak height, and the third spectral characteristic sequence is obtained from the wavelength difference and peak intensity.

5. The method according to claim 1, wherein The method comprises: determining the purity of the toner sample using a support vector machine algorithm based on the scattered spectrum feature vector of the toner sample; if the scattered spectrum feature vector falls within the feature space of a preset high-purity toner sample, determining that the purity of the toner sample reaches the target purity; otherwise, determining that the purity of the toner sample does not reach the target purity, including: According to the characteristic vector of the scattering spectrum of the carbon powder sample, the principal component analysis method is used to perform an orthogonal transformation on the characteristic vector, and a first characteristic space coordinate value is obtained from the projection matrix; For the first feature space coordinate value, constructing a feature space boundary by a support vector machine based on a Gaussian kernel function, and obtaining a second feature space coordinate value from the kernel function mapping; Calculating the sample intra-class scatter matrix based on the coordinate values ​​of the second feature space, calculating the distance from the sample to the feature space boundary using the Mahalanobis distance, and obtaining the sample purity determination threshold from the distance distribution histogram; The feature vector of the toner sample to be judged is projected into the second feature space through kernel function mapping. If the Mahalanobis distance value is less than the lower limit of the confidence interval, it is determined that the sample does not reach the target purity.

6. The method according to claim 1, characterized in that For the carbon powder sample that does not reach the target purity, the abnormal change characteristics of the corresponding scattering spectrum curve are analyzed, and the scattering spectrum curve is decomposed using the peak fitting method to obtain the scattering peak caused by adsorbed impurities and the characteristic scattering peak of the carbon powder itself, including: A four-layer wavelet transform is used to perform multi-scale decomposition on the scattering spectrum curve of the carbon powder sample that does not reach the target purity, and a spectral baseline is obtained by reconstructing the low-frequency coefficients. The baseline is subtracted from the spectral curve to obtain a first scattering spectrum sequence; According to the first scattering spectrum sequence, a Gauss-Lorentz mixed function is used to fit the scattering peak, and the function parameters are iteratively optimized by the least squares method until the residual is less than a preset threshold, thereby obtaining a second scattering spectrum sequence; For the second scattering spectrum sequence, calculating the area ratio of the Gaussian component to the Lorentzian component in the mixing function, establishing a scattering peak identification criterion based on the ratio distribution, and obtaining a third scattering spectrum sequence; For the scattering components in the third scattering spectrum sequence, the peak position distribution is clustered using the Gudan kernel density estimation method, and the carbon powder intrinsic characteristic scattering peak and the impurity scattering peak are obtained from the cluster center; It also includes: identifying the peak intensity, peak width and peak symmetry of the scattering spectrum curve to analyze the corresponding abnormal change characteristics, calculating the first-order derivative and second-order derivative of the scattering peak to determine the inflection point and overlapping position of the peak, determining the boundary range and baseline position of a single scattering peak based on the abnormal change characteristics, accurately separating overlapping scattering peaks, and extracting the characteristic parameters of the impurity adsorption peak and the carbon powder intrinsic peak.

7. The method according to claim 6, characterized in that The method includes identifying the peak intensity, peak width, and peak symmetry of the scattering spectrum curve to analyze the corresponding abnormal change characteristics, calculating the first-order derivative and second-order derivative of the scattering peak to determine the inflection point and overlapping position of the peak, determining the boundary range and baseline position of the single scattering peak based on the abnormal change characteristics, accurately separating the overlapping scattering peaks, and extracting the characteristic parameters of the impurity adsorption peak and the carbon powder intrinsic peak, including: Calculating the first-order derivative and the second-order derivative using a five-point difference method based on the light intensity value of the scattering spectrum curve, determining the peak vertex by judging the position where the sign of the first-order derivative changes from positive to negative and the second-order derivative is negative, and obtaining a first scattering peak sequence from the peak vertex; For the first scattering peak sequence, a third-order polynomial is used to fit the baseline, and the fitting coefficient is iteratively optimized by the least squares method until the residual does not exceed the residual threshold, and a second scattering peak sequence is obtained from the fitting result; Calculating the overlap coefficient of adjacent peaks according to the second scattering peak sequence, and if the overlap coefficient exceeds an overlap coefficient threshold, determining that the adjacent peaks are overlapping peaks, and obtaining a third scattering peak sequence based on the overlapping state; The overlapping peaks in the third scattering peak sequence are separated by a maximum entropy deconvolution algorithm, and the iteration is stopped when the entropy function increment is less than a preset threshold, and a fourth scattering peak sequence is obtained from the separation results.

8. The method according to claim 1, characterized in that According to the position and intensity information of the scattering peak of the adsorbed impurities obtained by decomposition, the principal component analysis method is used to perform qualitative and quantitative analysis on the type and content of the adsorbed impurities, and the main impurity components affecting the purity of the carbon powder and the relative content of the main impurities are obtained, including: Performing maximum and minimum value normalization processing based on the wavelength position and intensity value of the impurity scattering peak, and obtaining a first impurity feature matrix by eliminating abnormal values ​​that exceed the preset standard deviation range; Calculating the Pearson correlation coefficient for the first impurity feature matrix, and obtaining a second impurity feature matrix by performing singular value decomposition on the correlation coefficient matrix and selecting eigenvalues ​​whose cumulative contribution rates exceed a contribution rate threshold; Performing a projection transformation on the second impurity characteristic matrix using a maximum variance method, and obtaining a third impurity characteristic matrix by calculating a projection error and determining whether the projection error exceeds a preset threshold; Performing kernel density estimation on the third impurity feature matrix to calculate probability density, and obtaining the impurity component type by calculating the Mahalanobis distance between the sample point and the cluster center and determining whether the Mahalanobis distance exceeds a distance threshold; It also includes: judging the chemical bond structure characteristics of impurities based on the wavelength position of the scattering peak, analyzing the crystal structure characteristics of impurities based on the shape and symmetry of the scattering peak, calculating the relative concentration of impurities using the relative intensity ratio of the scattering peak, establishing a standard curve for impurity content using the external standard method, comparing the scattering peak intensity of unknown impurities with the standard curve to obtain the content level of the unknown impurity, constructing an impurity component fingerprint based on the position and intensity characteristics of the scattering peak, and analyzing to obtain the main impurity types and contents in the carbon powder.

9. The method according to claim 8, characterized in that The method comprises the following steps: determining the chemical bond structure characteristics of the impurity according to the wavelength position of the scattering peak, analyzing the crystal structure characteristics of the impurity according to the shape and symmetry of the scattering peak, calculating the relative concentration of the impurity using the relative intensity ratio of the scattering peak, establishing a standard curve for the impurity content by using the external standard method, comparing the scattering peak intensity of the unknown impurity with the standard curve to obtain the content level of the unknown impurity, constructing an impurity component fingerprint based on the position and intensity characteristics of the scattering peak, and analyzing and obtaining the main impurity types and contents in the carbon powder, including: Obtaining spectral wavenumbers according to the wavelength position of the impurity scattering peak, and comparing the spectral wavenumbers with a bond vibration characteristic wavenumber database through a conversion formula to obtain a first impurity characteristic sequence; Performing a Fourier series expansion on the first impurity characteristic sequence, wherein the odd-even ratio of the Fourier series expansion coefficients represents the peak symmetry to obtain a second impurity characteristic sequence; Establishing a third-order polynomial equation using the second impurity characteristic sequence, and iteratively optimizing fitting parameters of the third-order polynomial equation by a least squares method to obtain a third impurity characteristic sequence; A random forest model is constructed for the third impurity feature sequence. The random forest model selects a feature subset through the Gini coefficient. The cosine similarity between the feature subset and the standard graph is calculated to obtain an impurity type discrimination result.

10. The method according to claim 1, characterized in that The purity determination results and impurity analysis results of the toner samples are used to establish a toner purity assessment model using a decision tree algorithm. The toner purity is quantitatively scored based on the toner scattering spectrum characteristics and impurity analysis results, providing guidance for toner quality optimization, including: According to the carbon powder scattering spectrum feature vector and the impurity analysis data, the carbon powder scattering spectrum feature vector and the impurity analysis data are mapped to the interval [0, 1] using the maximum and minimum value normalization method, and the first feature matrix is ​​obtained by eliminating abnormal values ​​that exceed the preset standard deviation range; For the first feature matrix, the Pearson correlation coefficient is used to calculate the correlation between features, and the features are screened by setting the absolute value of the correlation coefficient to be greater than a preset threshold to obtain the second feature matrix; For the second feature matrix, the feature importance score is calculated using the variance contribution rate, and the features whose cumulative contribution rate exceeds the preset threshold are selected to obtain the third feature matrix; For the third feature matrix, a decision tree is used to establish a purity assessment model, and the fourth feature matrix is ​​obtained by selecting the optimal split feature through the Gini coefficient. The upper limit of the decision tree depth is set to a preset number of layers.

Images