Diosmin full-link impurity spectrum analysis method based on near infrared spectrum

Abstract

The application discloses a method for analyzing the full-chain impurity spectrum of diltiazem based on near-infrared spectroscopy, and belongs to the technical field of near-infrared spectroscopy analysis. The method comprises the following steps: acquiring the near-infrared spectroscopy data of diltiazem bulk drug and each sampling node in the production of the full chain of diltiazem preparations, and constructing a full-chain spectroscopy data set; jointly decomposing the full-chain spectroscopy data set in the time domain and the frequency domain to obtain multi-scale spectroscopy components of each sampling node; constructing a spectroscopy transmission mapping relationship between the sampling nodes according to the multi-scale spectroscopy components, and forming a full-chain spectroscopy correlation graph; extracting impurity feature vectors of the sampling nodes based on the full-chain spectroscopy correlation graph, arranging the impurity feature vectors according to the production time sequence to generate an initial impurity spectrum matrix; performing sparse low-rank decomposition on the initial impurity spectrum matrix to separate the trend item and the fluctuation item of the impurity signal, and obtaining a pure impurity spectrum matrix; and generating impurity dynamic change information of each sampling node according to the pure impurity spectrum matrix.

Description

Technical Field

[0001] This invention relates to the field of near-infrared spectroscopy analysis technology, specifically to a method for analyzing the whole chain of impurities in diosmin based on near-infrared spectroscopy. Background Technology

[0002] The production of diosmin involves a complete process chain from active pharmaceutical ingredient synthesis to formulation processing, with complex pathways for impurity introduction and dynamic accumulation characteristics. To control drug quality, continuous analysis of the impurity spectrum along the production chain is necessary. Traditional impurity analysis mainly relies on offline detection methods such as high-performance liquid chromatography (HPLC), which can only perform post-processing analysis on samples at discrete sampling times and cannot reflect the transmission and evolution of impurities between adjacent processes during production. When applying near-infrared spectroscopy to process monitoring, existing technologies typically perform multivariate statistical modeling such as principal component analysis or partial least squares regression on the entire spectrum, using the entire spectrum as input to extract features. These methods ignore the multi-scale structural differences of diosmin impurity signals in the time and frequency domains. Low-frequency baseline drift and high-frequency random interference are intertwined, resulting in extracted impurity features that cannot accurately characterize the transmission patterns between nodes.

[0003] Another drawback of existing methods is that the impurity characteristic matrix obtained from different process nodes simultaneously contains both the common impurity baseline trend across the entire chain and the process fluctuations specific to each node. These two aspects are directly used for impurity spectrum analysis without effective separation. The common baseline trend reflects the evolution of raw material characteristics or slowly accumulated impurities, while the node-specific fluctuations reflect local process anomalies or intermittent disturbances. This overlap makes it difficult to identify the main cause of dynamic changes in impurities and to distinguish between systematic drift and occasional deviations.

[0004] To accurately analyze the dynamic changes of impurity spectra from full-chain near-infrared spectral data, it is necessary to solve two problems: how to extract impurity features reflecting process transfer relationships while suppressing interference from irrelevant spectral components, and how to separate aliased impurity signals into global trend terms and local fluctuation terms. Summary of the Invention

[0005] The technical problem this invention aims to solve is that in existing diosmin whole-chain impurity spectrum analysis, the multi-scale structure of the near-infrared spectrum is not fully utilized, and the trend and fluctuation terms of the impurity signal are mixed, making it impossible to accurately characterize the dynamic transmission law of impurities along the process chain and identify the main causes. This invention provides an analysis method that can achieve time-frequency joint decomposition and spectral correlation graph modeling, and perform sparse low-rank separation of the impurity spectrum matrix. It extracts impurity features reflecting the transmission relationship between nodes from the near-infrared spectrum and decouples the global trend and local fluctuations of the impurity signal, achieving accurate analysis of the diosmin whole-chain impurity spectrum.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a method for full-chain impurity spectral analysis of diosmin based on near-infrared spectroscopy. This method includes: acquiring near-infrared spectral data from each sampling node in the entire production chain of diosmin API and its formulations, forming a full-chain spectral dataset; performing joint time-domain and frequency-domain decomposition on the full-chain spectral dataset to obtain multi-scale spectral components for each sampling node; constructing a spectral transfer mapping relationship between each sampling node based on the multi-scale spectral components, forming a full-chain spectral correlation diagram; extracting impurity feature vectors from each sampling node based on the full-chain spectral correlation diagram, arranging the impurity feature vectors according to the production time sequence to generate an initial impurity spectral matrix; performing sparse low-rank decomposition on the initial impurity spectral matrix to separate the trend term and fluctuation term of the impurity signal, obtaining a pure impurity spectral matrix; generating impurity dynamic change information for each sampling node based on the pure impurity spectral matrix, thus completing the full-chain impurity spectral analysis. This method, through the combination of time-frequency joint decomposition and sparse low-rank decomposition, achieves systematic correlation and accurate extraction of impurity information in the entire production chain, significantly improving the ability to trace impurities and dynamically monitor them.

[0007] As a preferred embodiment of the present invention, the full-chain spectral dataset is jointly decomposed in the time and frequency domains to obtain multi-scale spectral components for each sampling node. Specifically, this includes: performing empirical mode decomposition on each near-infrared spectral data to obtain a set of intrinsic mode functions (EMFs); performing Hilbert spectral analysis on each EMF to obtain instantaneous frequency and instantaneous amplitude; dividing each EMF into low-frequency, mid-frequency, and high-frequency bands based on the instantaneous frequency; reconstructing the EMFs within the same frequency band to obtain low-frequency, mid-frequency, and high-frequency spectral components, which serve as the multi-scale spectral components for that sampling node. Through this adaptive decomposition and frequency partitioning, different scale components reflecting impurity baseline drift, process fluctuations, and random noise can be separated from the spectral data, making the subsequently established correlations more robust.

[0008] As another preferred technical solution of the present invention, a spectral transfer mapping relationship between each sampling node is constructed based on the multi-scale spectral components to form a full-chain spectral correlation graph. Specifically, this includes: using each sampling node as a graph node, and taking the low-frequency spectral components in the multi-scale spectral components as the node attribute of that graph node; performing dynamic time warping on the low-frequency spectral components of two adjacent graph nodes along the production time sequence direction, and calculating the spectral morphological similarity; when the spectral morphological similarity is greater than a preset threshold, establishing a directed edge between the corresponding two graph nodes, the weight of which is determined by the spectral morphological similarity; and combining all graph nodes and directed edges to obtain the full-chain spectral correlation graph. Constructing a directed correlation graph using the morphological similarity of low-frequency components can effectively capture the transmission path and evolution relationship of impurity characteristics in the production process, enhancing the adaptability of the analysis process to process time sequence fluctuations.

[0009] In the above scheme, impurity feature vectors of each sampling node are extracted based on the full-chain spectral correlation graph. These impurity feature vectors are then arranged according to production time sequence to generate an initial impurity spectrum matrix. Preferably, this includes: for each graph node in the full-chain spectral correlation graph, extracting the weights of all adjacent directed edges of that node to form an adjacency weight sequence; performing statistical feature calculations on the adjacency weight sequence to obtain the adjacency distribution feature value of that graph node; concatenating the node attributes of the graph node with the adjacency distribution feature value to form the impurity feature vector of that graph node; and vertically stacking the impurity feature vectors of all graph nodes according to production time sequence to generate the initial impurity spectrum matrix. This integrates graph structure information and spectral attributes into a structured matrix representation, fully preserving the impurity information of each node and its contextual relationship within the correlation graph.

[0010] As a further preferred embodiment of the present invention, the initial impurity spectrum matrix is ​​subjected to sparse low-rank decomposition to separate the trend term and fluctuation term of the impurity signal, thereby obtaining a pure impurity spectrum matrix. Specifically, this includes: decomposing the initial impurity spectrum matrix into a low-rank matrix and a sparse matrix, wherein the low-rank matrix represents the common impurity baseline trend among all sampling nodes, and the sparse matrix represents the impurity fluctuation signal unique to each sampling node; iteratively updating the low-rank matrix and the sparse matrix until a preset iteration termination condition is reached; and using the low-rank matrix obtained at the end of the iteration as the pure impurity spectrum matrix. This method effectively separates the stable baseline trend from local abnormal fluctuations in the impurity signal, and the obtained pure impurity spectrum matrix can more accurately depict the true evolution of impurities along the production process.

[0011] In empirical mode decomposition (EMD), the process of decomposing each near-infrared spectral data point to obtain an intrinsic mode function (EMF) set is as follows: For each near-infrared spectral data point, identify all its local maxima and local minima, fit upper and lower envelopes respectively, and calculate the mean curves of the upper and lower envelopes; subtract the mean curves from the near-infrared spectral data to obtain the component to be determined; when the component to be determined satisfies the two determination conditions of the EMF, store the component to be determined as an EMF in the EMF set, and continue decomposition on the remaining signal; when the component to be determined does not satisfy the two determination conditions, repeat the above steps as a new near-infrared spectral data point until the remaining signal is a monotonic function or the number of extreme points is less than a preset number, at which point the decomposition terminates. The two determination conditions for the EMF are: the difference between the number of extreme points and the number of zero-crossing points does not exceed 1, and the mean of the upper and lower envelopes at any point is zero. This adaptive screening process ensures that the decomposed modal components have physical meaning, providing a reliable basis for the separation of multi-scale impurity information.

[0012] When constructing the spectral transfer mapping relationship, the low-frequency spectral components of two adjacent graph nodes are dynamically time-warped along the production time sequence, and the spectral morphology similarity is calculated. This is preferably achieved through the following method: The low-frequency spectral components of the preceding graph node are taken as the first sequence, and the low-frequency spectral components of the following graph node are taken as the second sequence; a cumulative distance matrix is ​​constructed between the first and second sequences, where each element is obtained by adding the distance between the two sequence points at the corresponding position to the previous minimum cumulative distance; the optimal warping path is obtained by backtracking from the last element of the cumulative distance matrix to the first element; the reciprocal of the cumulative distance at the endpoint of the optimal warping path is taken and normalized to obtain the spectral morphology similarity. This method can tolerate local offsets and scaling in the production time sequence and more accurately reflects the substantial similarity of the spectral profiles.

[0013] Preferably, statistical feature calculations are performed on the adjacency weight sequence to obtain the adjacency distribution feature value of the graph node, including: calculating the arithmetic mean and median of the adjacency weight sequence as a central tendency feature; calculating the standard deviation of the adjacency weight sequence as a dispersion feature; calculating the skewness and kurtosis of the adjacency weight sequence as a distribution morphology feature; and combining the arithmetic mean, the median, the standard deviation, the skewness, and the kurtosis to form the adjacency distribution feature value of the graph node. This multidimensional statistical feature comprehensively characterizes the association strength distribution of nodes in the spectral correlation graph, improving the ability of the feature vector to distinguish impurity transmission patterns.

[0014] In the iterative update of sparse low-rank decomposition, the specific update method is as follows: Fix the current sparse matrix, apply singular value thresholding to the difference between the initial impurity spectrum matrix and the current sparse matrix, and update the low-rank matrix; fix the updated low-rank matrix, apply soft thresholding to the difference between the initial impurity spectrum matrix and the updated low-rank matrix, and update the sparse matrix; calculate the ratio of the difference norm of the low-rank matrix before and after this iteration to the norm of the low-rank matrix. When the ratio is less than a preset tolerance, the iteration termination condition is met, and the iteration stops. This iterative convergence strategy effectively avoids over-decomposition while ensuring decomposition accuracy, making the obtained pure impurity spectrum matrix stable and reliable, and providing an accurate basis for finally generating the dynamic change information of impurities at each sampling node.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: Multi-scale spectral components of each sampling node are obtained through joint decomposition in the time and frequency domains, adaptively separating the original near-infrared spectrum into low-frequency, mid-frequency, and high-frequency spectral components. The low-frequency spectral components carry the main information regarding the slow evolution and baseline drift of diosmin impurities, suppressing high-frequency random noise and local interference, thus enabling effective focusing on robust spectral features related to impurity accumulation when constructing subsequent spectral transfer mapping relationships. Based on this, dynamic time warping is used along the production timeline to calculate the spectral morphology similarity of the low-frequency spectral components of adjacent nodes, and directed edges are established between nodes that meet threshold conditions to form a full-chain spectral association graph. This method transforms the transfer relationship between the full-chain spectra into a topological structure of nodes and directed edges in the graph. The impurity feature vector is composed of node attributes and adjacent distribution feature values, reflecting not only the impurity content information carried by the node's own spectrum but also capturing the degree of transfer variation between adjacent processes. Therefore, the initial impurity spectrum matrix generated by stacking according to the production timeline more accurately represents the evolution trajectory of impurities along the process chain.

[0016] A sparse low-rank decomposition is performed on the initial impurity spectrum matrix, separating the impurity signal into a low-rank matrix and a sparse matrix. The low-rank matrix captures the common impurity baseline trend across all sampling nodes, eliminating the interference of node-specific fluctuations and random noise on this trend, and can be directly used as a pure impurity spectrum matrix for analyzing systematic impurity evolution. The sparse matrix retains the impurity fluctuation signals specific to each node, allowing local impurity changes introduced by process anomalies to be presented separately. The combined effect of singular value thresholding and soft thresholding during the iteration process, under the condition of automatic convergence control by the difference norm to norm ratio, achieves high-precision recovery of the low-rank structure. The separated pure impurity spectrum matrix is ​​free from noise superposition and fluctuation contamination. The impurity dynamic change information generated from this matrix can simultaneously and clearly show the global impurity accumulation path and the independent contribution of occasional fluctuation events, providing clear analytical results for real-time monitoring and retrospective diagnosis of the diosmin production process. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a flowchart of the whole-chain impurity spectrum analysis of diosmin based on near-infrared spectroscopy; Figure 2 It is a flowchart for constructing a full-chain spectral correlation graph based on dynamic time warping; Figure 3 This is a flowchart of the process for generating full-chain impurity spectrum analysis results; Figure 4 It is the result of multi-scale spectral component decomposition of near-infrared spectral data; Figure 5 This is a comparison chart of the sparse low-rank decomposition results of the whole chain impurity spectrum of diosmin. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] See Figure 1 This invention provides a method for analyzing the entire chain of impurities in diosmin based on near-infrared spectroscopy, comprising: acquiring near-infrared spectral data of each sampling node in the entire production chain of diosmin active pharmaceutical ingredient and its formulations, forming a whole-chain spectral dataset; performing joint time-domain and frequency-domain decomposition on the whole-chain spectral dataset to obtain multi-scale spectral components of each sampling node; constructing spectral transfer mapping relationships between each sampling node based on the multi-scale spectral components to form a whole-chain spectral correlation diagram; extracting impurity feature vectors of each sampling node based on the whole-chain spectral correlation diagram, arranging the impurity feature vectors according to the production time sequence to generate an initial impurity spectral matrix; performing sparse low-rank decomposition on the initial impurity spectral matrix to separate the trend term and fluctuation term of the impurity signal, obtaining a pure impurity spectral matrix; generating impurity dynamic change information of each sampling node based on the pure impurity spectral matrix, thus completing the whole-chain impurity spectral analysis.

[0021] Example 1: In practice, near-infrared spectral data from each sampling node in the entire production chain of diosmin API and its formulations are acquired to form a full-chain spectral dataset. Each near-infrared spectral data point in the full-chain spectral dataset is represented as an absorbance sequence with wavelength as the independent variable. Empirical mode decomposition is performed on each near-infrared spectral data point to obtain the set of intrinsic mode functions.

[0022] The empirical mode decomposition (EMD) is implemented as follows: For the near-infrared spectral data, all local maxima and local minima are identified. Using cubic spline interpolation, all local maxima are connected and fitted to an upper envelope, and all local minima are connected and fitted to a lower envelope. The arithmetic mean of the values ​​of the upper and lower envelopes at each wavelength position is calculated to obtain a mean curve. The mean curve is subtracted from the near-infrared spectral data to obtain the component to be determined. It is then determined whether the component to be determined simultaneously satisfies two criteria for the intrinsic mode function: the first criterion is that the difference between the number of extrema and the number of zero-crossings does not exceed 1; the second criterion is that at any wavelength, the mean calculated from the upper and lower envelopes is zero. When the component to be determined simultaneously satisfies both determination conditions, the component to be determined is stored as an intrinsic mode function (IMF) in the IMF set. Then, the obtained IMF is subtracted from the original near-infrared spectral data to obtain the remaining signal. The decomposition process of identifying maxima and minima, fitting the envelope, calculating the mean curve, and determining the component to be determined is then performed on the remaining signal. When the component to be determined does not satisfy at least one of the two determination conditions, the component to be determined is treated as new near-infrared spectral data, and the process of identifying maxima and minima, fitting the envelope, calculating the mean curve, and determining the component to be determined is repeated. The decomposition terminates when the remaining signal exhibits a monotonic function across the entire wavelength range, or when the total number of local maxima and minima of the remaining signal is less than a preset number. The preset number is set to 2, based on the fact that when the number of extreme points is less than 2, the remaining signal has at most one extreme point, making it impossible to extract an IMF with oscillatory characteristics from it.

[0023] After performing empirical mode decomposition on each near-infrared spectral data point in the full-chain spectral dataset, an intrinsic mode function (EMF) set corresponding to each near-infrared spectral data point is obtained. Hilbert spectral analysis is then performed on each EMF in the EMF set to obtain the instantaneous frequency and instantaneous amplitude. The specific process of performing Hilbert spectral analysis on each EMF is as follows: For the first... eigenmode functions Perform Hilbert transform to construct analytic signal ,in Represents the Hilbert transform operator. The imaginary unit, The wavelength is a variable. The first wavelength is obtained from the analytic signal. Instantaneous amplitude of each eigenmode function and instantaneous phase ,in and These represent the operations of taking the real part and the imaginary part, respectively. Each intrinsic mode function at wavelength instantaneous frequency at point Calculated using the following formula: in, Indicates the first Each intrinsic mode function at wavelength instantaneous frequency at that point Indicates the first Each intrinsic mode function at wavelength The instantaneous phase at that point, Indicates instantaneous phase For wavelength The first derivative, π is a constant. After performing the Hilbert spectral analysis on all intrinsic mode functions, the instantaneous frequency sequence and instantaneous amplitude sequence of each intrinsic mode function are obtained.

[0024] Based on the instantaneous frequencies, each intrinsic mode function (IMF) is divided into low-frequency, mid-frequency, and high-frequency bands. The division process involves: calculating the average instantaneous frequency of each IMF, where the average instantaneous frequency is the arithmetic mean of the instantaneous frequencies of the corresponding IMF at all wavelengths. The distribution of the average instantaneous frequencies of all IMFs is statistically analyzed to determine the low-frequency threshold. and high frequency threshold In some implementations, the 33rd percentile of the average instantaneous frequency of all intrinsic mode functions is used as... The 67th percentile of the average instantaneous frequency of all intrinsic mode functions is used as... For each eigenmode function, when the average instantaneous frequency of that eigenmode function is less than... When the intrinsic mode function is in the low-frequency band, it is classified as such; when the average instantaneous frequency of the intrinsic mode function is greater than or equal to... and less than or equal to When the intrinsic mode function is in the mid-frequency band, it is classified as such; when the average instantaneous frequency of the intrinsic mode function is greater than 100 MHz, it is classified as such. At this point, the intrinsic mode function is assigned to the high-frequency band. All intrinsic mode functions assigned to the low-frequency band are linearly superimposed to obtain the low-frequency spectral component; all intrinsic mode functions assigned to the mid-frequency band are linearly superimposed to obtain the mid-frequency spectral component; and all intrinsic mode functions assigned to the high-frequency band are linearly superimposed to obtain the high-frequency spectral component. The low-frequency spectral component, the mid-frequency spectral component, and the high-frequency spectral component together constitute the multi-scale spectral components corresponding to this sampling node.

[0025] See Figure 4 In the graph, the horizontal axis represents wavelength in nanometers (nm), ranging from 400nm to 2500nm; the vertical axis represents absorbance, ranging from approximately 0 to 1.8. The solid black line represents the raw near-infrared spectral data, the dashed blue line represents the low-frequency spectral components, the dotted green line represents the mid-frequency spectral components, and the dotted red line represents the high-frequency spectral components.

[0026] Observing the trend of the curves, the original spectrum exhibits multiple absorption peaks, mainly distributed in the range of approximately 700 nm to 2300 nm, with peak absorbance close to 1.7. The low-frequency spectral components are highly consistent with the original spectrum, and can reconstruct the overall absorption characteristics of the original spectrum well, indicating that the low-frequency components contain the main trend information of the original spectrum. The mid-frequency spectral components have small amplitudes, fluctuating around zero, showing oscillatory characteristics of local details, and their peak amplitudes are much lower than those of the low-frequency components, reflecting the mesoscale fluctuation components in the near-infrared spectrum. The high-frequency spectral components have the smallest amplitudes, basically distributed like random noise, with amplitudes close to zero, reflecting high-frequency noise and small fluctuations in the spectrum.

[0027] The data shown in the figure fully reflects the set of intrinsic mode functions (EMFs) obtained by empirical mode decomposition for each near-infrared spectral data in Example 1. After Hilbert spectral analysis, the EMFs are divided into low-frequency, mid-frequency, and high-frequency bands according to their instantaneous frequencies, and then linearly superimposed to reconstruct multi-scale spectral components. The low-frequency components reflect the main trend of the spectrum and can accurately reproduce the macroscopic absorption characteristics of the original spectrum; the mid-frequency components reflect local detail changes in the spectrum; and the high-frequency components characterize small fluctuations and noise in the spectrum.

[0028] Example 2: In specific implementation, please refer to Figure 2 Each sampling node is used as a graph node, and the low-frequency spectral component in the multi-scale spectral components is used as the node attribute of the corresponding graph node. The low-frequency spectral component is a one-dimensional vector with a length equal to the number of wavelength points in the near-infrared spectral data. Each element in the vector is the absorbance value after reconstruction and superposition of all intrinsic mode functions within the low-frequency band at the corresponding wavelength position.

[0029] Along the production time sequence, dynamic time warping is performed on the low-frequency spectral components of adjacent graph nodes to calculate spectral morphology similarity. The graph node preceding the node along the production time sequence is designated as the predecessor node, and the graph node following the node is designated as the successor node. The low-frequency spectral components of the predecessor node are extracted as the first sequence; the low-frequency spectral components of the successor node are extracted as the second sequence. A cumulative distance matrix is ​​constructed between the first and second sequences. The cumulative distance matrix is ​​constructed as follows: the length of the first sequence is set to... Set the length of the second sequence to be The cumulative distance matrix is ​​denoted as , The number of rows is The number of columns is Define the first sequence in which the first... The point and the second sequence Local distance between points This is the absolute value of the difference in absorbance between two points. Cumulative distance matrix. The Middle line, number Column elements Determined by the following formula: in, Represents the cumulative distance matrix The Middle line, number The cumulative distance value of column elements. Indicates the first sequence The point and the second sequence Local distance between points This indicates the minimum value operation. Represents the cumulative distance matrix The Middle line, number The cumulative distance value of column elements. Represents the cumulative distance matrix The Middle line, number The cumulative distance value of column elements. Represents the cumulative distance matrix The Middle line, number The cumulative distance value of the column elements. At the boundary position, when or hour, The calculation only considers those existing in the cumulative distance matrix. The valid adjacent elements within.

[0030] In the cumulative distance matrix After all the filling is complete, backtracking from the last element of the cumulative distance matrix to the first element yields the optimal normalized path. The last element of the cumulative distance matrix is ​​the first element. line, number Column elements The first element of the cumulative distance matrix is ​​the element in the first row and first column. The backtracking process is as follows: taking the th element in the cumulative distance matrix... line, number The column position is used as the current backtracking position, and the element above the current backtracking position is found in the cumulative distance matrix. Left adjacent element and the element adjacent to the top left The next backtracking position is selected as the position of the neighboring element with the smallest cumulative distance among three neighboring elements. The indices of the first and second sequences corresponding to the current backtracking position are recorded. This next backtracking position is then used as the new current backtracking position, and the steps of comparing neighboring elements and recording backtracking positions continue until the backtracking position reaches the first element of the cumulative distance matrix. Until then, the optimal regular path is formed by combining the indices of all first sequence points and second sequence points recorded sequentially during the backtracking process, arranged according to the backtracking order.

[0031] Extract the cumulative distance to the endpoint of the optimal regular path; this cumulative distance is the last element of the cumulative distance matrix. The cumulative distance value is calculated by taking the reciprocal of the cumulative distance to the endpoint. This reciprocal cumulative distance value is then normalized by dividing it by the maximum of the reciprocal cumulative distance values ​​of all adjacent graph node pairs, yielding the spectral morphological similarity. The spectral morphological similarity value is a real number greater than 0 and less than or equal to 1.

[0032] The spectral morphology similarity is compared with a preset threshold. The preset threshold is set at 0.85, based on the following principle: when the impurity variation of adjacent sampling nodes in the local osmin raw material and its formulation production process is within a reasonably acceptable range, the spectral morphology similarity should not be lower than 0.85. A similarity lower than 0.85 indicates a significant difference in the spectral morphology between the two sampling nodes, suggesting a possible abrupt change in impurity components. In practice, when the spectral morphology similarity between two adjacent graph nodes is greater than 0.85, a directed edge is established between the preceding and succeeding nodes. The direction of the directed edge is from the preceding node to the succeeding node, and the weight of the directed edge is determined by the spectral morphology similarity; that is, the value of the spectral morphology similarity is directly assigned to the directed edge as its weight value. When the spectral morphology similarity between two adjacent graph nodes is less than or equal to 0.85, no directed edge is established.

[0033] In some embodiments, all adjacent sampling node pairs in the entire chain are traversed, and the above-described dynamic time warping and spectral morphology similarity calculation process is performed on each pair of adjacent sampling nodes. The decision to establish a directed edge is based on the comparison result of the spectral morphology similarity with a preset threshold. After all adjacent graph node pairs have been processed, all graph nodes and all established directed edges are combined to form the entire chain spectral association graph. The entire chain spectral association graph uses a binary tuple. It means that among them It is a set of graph nodes, containing the graph nodes corresponding to all sampled nodes. This is a set of directed edges, containing all directed edges established that satisfy the condition that the spectral morphology similarity is greater than a preset threshold. Each graph node in the full-chain spectral association graph has a node attribute, which is the low-frequency spectral component vector of the corresponding sampling node. Each directed edge has a weight value, which is the spectral morphology similarity value of the corresponding adjacent graph node pair.

[0034] Example 3: In practice, a fully constructed full-chain spectral correlation graph is obtained. This graph contains a set of graph nodes and a set of directed edges. Each graph node corresponds to a sampling node, and each directed edge corresponds to a pair of adjacent graph nodes with a weight value. For each graph node in the full-chain spectral correlation graph, the adjacency weight sequence is extracted and statistical features are calculated.

[0035] For a given graph node in the full-chain spectral association graph, traverse all directed edges directly connected to that node, extract the weight value associated with each directed edge, and arrange all extracted weight values ​​in adjacency order according to the direction of the directed edges to form an adjacency weight sequence corresponding to that graph node. In specific implementations, when the graph node has only one directed edge connected to it, the adjacency weight sequence contains only one weight value; when the graph node has multiple directed edges connected to it, the adjacency weight sequence contains multiple weight values. Optionally, for the direction of the directed edge, if the directed edge in the full-chain spectral association graph originates from that graph node and points to other graph nodes, then the weight value of that directed edge is extracted and added to the outgoing edge adjacency weight sequence; if the directed edge in the full-chain spectral association graph originates from other graph nodes and points to that graph node, then the weight value of that directed edge is extracted and added to the edge adjacency weight sequence. The outgoing edge adjacency weight sequence and the incoming edge adjacency weight sequence are merged to form the complete adjacency weight sequence for that graph node.

[0036] Statistical feature calculations are performed on the adjacency weight sequence to obtain the adjacency distribution feature values ​​of the graph nodes. The statistical feature calculation includes three parts: central tendency feature calculation, dispersion feature calculation, and distribution pattern feature calculation.

[0037] The central tendency feature is calculated as follows: the arithmetic mean of all weight values ​​in the adjacent weight sequence is calculated, and this is denoted as the mean feature. Calculate the median value of all weights in the adjacent weight sequence, denoted as the median feature. Mean characteristics The calculation method is as follows: sum all weight values ​​in the adjacent weight sequence and divide by the total number of weight values ​​in the adjacent weight sequence. Median characteristics. The calculation method is as follows: Arrange all weight values ​​in the adjacent weight sequence in ascending order. When the total number of weight values ​​is odd, take the weight value located in the middle position of the arranged sequence as the median feature. When the total number of weight values ​​is even, the arithmetic mean of the weight values ​​located in the middle two positions of the permuted sequence is taken as the median feature. .

[0038] The discreteness feature is calculated as follows: the standard deviation of the adjacent weight sequence is calculated, and denoted as the discrete feature. Discrete features The calculation formula is as follows: in, This represents the standard deviation of the adjacent weight sequence. This represents the total number of weight values ​​in the adjacent weight sequence. Indicates the th in the adjacency weight sequence Each weight value This is the index of the weight value in the adjacent weight sequence. This represents the arithmetic mean of the adjacent weight sequences. Indicates to From 1 to Summing all terms.

[0039] The distribution morphology characteristics are calculated by: calculating the skewness and kurtosis of the adjacent weighted sequences. The skewness value is denoted as... kurtosis value is denoted as Skewness value The calculation method is as follows: first calculate the arithmetic mean of each weight value in the adjacent weight sequence. To calculate the deviation, cube each deviation, sum the cubes of all deviations, and then divide by the total number of weight values. Then divide by the standard deviation The cube of . Kurtosis value The calculation method is as follows: first calculate the arithmetic mean of each weight value in the adjacent weight sequence. To calculate the deviation, raise each deviation to the fourth power, sum all the fourth powers of the deviations, and then divide by the total number of weight values. Then divide by the standard deviation The fourth power of the result is used, and 3 is subtracted from the result to make the kurtosis of the normal distribution 0.

[0040] After obtaining the above five statistical characteristics, the arithmetic mean is... Median value Standard deviation skewness value and kurtosis value The nodes are combined in a fixed order to form a five-dimensional vector, which represents the adjacency distribution characteristic values ​​of the graph nodes. The fixed order is as follows: arithmetic mean. Occupy the first dimension of the vector, median value Occupy the second dimension of the vector, standard deviation Occupying the third-dimensional position of the vector, skewness value Occupying the fourth dimension of the vector, kurtosis value It occupies the fifth dimension of the vector.

[0041] The node attributes and adjacent distribution feature values ​​of the graph node are concatenated to form the impurity feature vector of the node. The node attributes are the low-frequency spectral components in the multi-scale spectral components, which are absorbance vectors with a length equal to the number of wavelengths. The concatenation method is as follows: the low-frequency spectral component vector is used as the first half, and the adjacent distribution feature value vector is used as the second half. The two vectors are concatenated end to end to form a vector with an extended length. This extended length vector is the impurity feature vector corresponding to the graph node.

[0042] In some embodiments, the length of the low-frequency spectral component vector is If the length of the adjacent distribution eigenvalue vector is a fixed value of 5, then the length of the impurity eigenvector obtained after splicing is 5. The first impurity feature vector The first five elements represent the absorbance values ​​of the low-frequency spectral components at various wavelengths, and the last five elements are their arithmetic mean values. Median value Standard deviation skewness value and kurtosis value .

[0043] For each graph node in the full-chain spectral correlation graph, the aforementioned operations of adjacency weight sequence extraction, statistical feature calculation, and vector concatenation are performed to obtain the impurity feature vector corresponding to each graph node. The impurity feature vectors of all graph nodes are then stacked vertically according to the production time sequence to generate an initial impurity spectral matrix. The stacking method is as follows: following the production time sequence from earliest to latest, the impurity feature vector corresponding to the graph node of the earliest production stage is placed in the first row of the initial impurity spectral matrix, the impurity feature vector corresponding to the graph node of the second earliest production stage is placed in the second row, and so on, until the impurity feature vector corresponding to the graph node of the latest production stage is placed in the last row. The number of rows in the initial impurity spectral matrix equals the total number of graph nodes in the full-chain spectral correlation graph, and the number of columns in the initial impurity spectral matrix equals the length of the impurity feature vectors.

[0044] Example 4: In practice, an initial impurity spectrum matrix is ​​obtained. The number of rows in the initial impurity spectrum matrix is ​​equal to the total number of nodes in the whole-chain spectral correlation graph, denoted as . The number of columns in the initial impurity spectrum matrix is ​​equal to the length of the impurity eigenvectors, denoted as . The initial impurity spectrum matrix is ​​denoted as... , for OK A column of real numbers.

[0045] For the initial impurity spectrum matrix Perform sparse low-rank decomposition, and Decompose into low-rank matrices sparse matrix The sum of Low-rank matrix The sparse matrix represents the common impurity baseline trend among all sampling nodes. This represents the unique impurity fluctuation signal at each sampling node. Low-rank matrix. sparse matrix The dimensions are all the same as the initial impurity spectrum matrix. They are the same, both are OK List.

[0046] In some embodiments, decomposition parameters are set, including a soft threshold parameter. and preset tolerance Soft threshold parameter Controlling sparse matrices The sparsity of non-zero elements in the middle and lower reaches of the earth. The value is set to , Indicates taking and The larger value in the range is chosen based on the principle that when separating low-rank components from sparse components, selecting a scaling factor related to the matrix size allows the soft thresholding operation to maintain reasonable sparsity constraints across a wide range of matrix dimensions. Preset tolerance. Used to determine whether the iteration has converged. The value is set to The basis for setting this value is: when the relative change of the low-rank matrix between two adjacent iterations is less than... At this point, the structure of the low-rank matrix is ​​already highly stable, and further iteration has little effect on optimizing the baseline trend of impurities, so it can be considered to have reached a convergent state.

[0047] Initialize low-rank matrix sparse matrix , low-rank matrix Set all elements of the sparse matrix to 0. Set all elements to 0. Set the iteration counter. The initial value is 0.

[0048] The process of iteratively updating the low-rank matrix and the sparse matrix is ​​as follows: at the 1st... In the next iteration, the current sparse matrix is ​​fixed. Calculate the initial impurity spectrum matrix With the current sparse matrix The difference is used to obtain the first difference matrix. For the first difference matrix Perform singular value decomposition to obtain a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix. Take all singular values ​​from the singular value diagonal matrix and subtract a preset singular value shrinkage threshold from each singular value. When the subtraction result is less than 0, the corresponding singular value is set to 0; when the subtraction result is greater than or equal to 0, the subtraction result is retained as the shrunk singular value. Singular value shrinkage threshold. The value is set to ,in Represents the first difference matrix The maximum singular value is determined by setting the shrinkage threshold to 5% of the maximum singular value. This threshold preserves the main spectral trend information while removing smaller singular value components caused by noise. The matrix is ​​reconstructed using the shrunken singular values ​​and their corresponding left and right singular vectors. The reconstructed matrix serves as the updated low-rank matrix. .

[0049] After completing the low-rank matrix update, fix the updated low-rank matrix. Calculate the initial impurity spectrum matrix With the updated low-rank matrix The difference is used to obtain the second difference matrix. For the second difference matrix Each element in the matrix undergoes soft-threshold shrinking to obtain the updated sparse matrix. The calculation method for soft threshold shrinkage is as follows: in, Represents the updated sparse matrix The Middle line, number The numerical values ​​of the column elements. For row index, For column indexes, This is a sign function that returns 1 when the argument is positive, -1 when it is negative, and 0 when it is zero. Represents the second difference matrix The Middle line, number The numerical values ​​of the column elements. This means taking the larger of the two values. express The absolute value, This is the soft threshold parameter, with a value of [value to be filled in]. Iterate through all row indices in turn. and column indexes Complete the sparse matrix Assigning values ​​to all elements.

[0050] In some embodiments, each time a low-rank matrix is ​​completed sparse matrix After the update, the iteration termination condition is checked. The low-rank matrix before the current iteration update is calculated. With the updated low-rank matrix The difference norm, and the low-rank matrix before the update. The norm of the difference norm is the Frobenius norm, calculated as follows: [Calculation of the difference norm]. and The sum of squares of all elements in the difference matrix is ​​then taken as the square root. The norm of a low-rank matrix is ​​also calculated using the Frobenius norm, as follows: [Calculation of Frobenius norm is missing from the original text]. The sum of squares of all elements in the matrix is ​​taken, and then the square root is calculated. The difference norm is divided by the norm of the lower-rank matrix to obtain the relative change ratio. This relative change ratio is then compared with the preset tolerance. The size, when the relative change ratio is less than When the iteration termination condition is met, the iteration process is stopped, and the low-rank matrix at this point is set. The low-rank matrix obtained from the decomposition When the relative change ratio is greater than or equal to At that time, the iteration counter will be... Increment the value by 1, and repeat the steps of fixing the current sparse matrix, updating the low-rank matrix, fixing the updated low-rank matrix, updating the sparse matrix, and calculating the relative change ratio until the relative change ratio is less than 1. Until then. The low-rank matrix obtained at the end of the iteration. This is the pure impurity spectrum matrix. The number of rows and columns of the pure impurity spectrum matrix is ​​the same as that of the initial impurity spectrum matrix. Similarly, the vector in each row of the pure impurity spectrum matrix represents the impurity baseline trend of the corresponding sampling node.

[0051] See Figure 5 In the figure, the horizontal axis represents the sampling node number, ranging from 0 to 600, representing the temporal sequence of each sampling node in the entire Osmin production chain; the vertical axis represents the eigenvalue, reflecting the numerical change of the corresponding sampling node in the impurity spectrum matrix. The figure contains three sequence curves: the initial impurity spectrum sequence (gray solid line), the pure impurity spectrum matrix sequence (trend term, blue solid line), and the sparse matrix sequence (fluctuation term, red dashed line).

[0052] The initial impurity spectrum sequence is represented by a relatively volatile gray curve, reflecting the original form of the impurity feature vector at the sampling node, containing a mixture of baseline trend and fluctuation noise information. The pure impurity spectrum matrix sequence is presented as a smooth blue curve, reflecting the impurity baseline trend characteristics after removing fluctuation noise through sparse low-rank decomposition. This blue trend curve shows a slow upward trend overall, with the eigenvalue gradually increasing from about 0.8 to about 1.5 within the range of sampling node indices 0 to 600, exhibiting slight fluctuations in between, indicating that the impurity baseline trend gradually strengthens and exhibits periodic changes during the production process. The sparse matrix sequence is shown as a red dashed line, fluctuating slightly around the zero value, with the amplitude range limited to about ±0.2, representing the impurity fluctuation information specific to the sampling node, reflecting the dynamic changes in local impurity components.

[0053] Example 5: In specific implementation, please refer to Figure 3 After the sparse low-rank decomposition, obtain the pure impurity spectrum matrix and the sparse matrix. The number of rows in the pure impurity spectrum matrix is ​​the total number of sampling nodes, denoted as . The column number is the length of the impurity feature vector, denoted as . The sparse matrix also has the same number of rows. The number of columns is also 1 Each row in the pure impurity spectrum matrix corresponds to the low-rank impurity baseline vector of a sampling node, and each row in the sparse matrix corresponds to the sparse fluctuation vector of a sampling node.

[0054] Read the low-rank impurity baseline vector corresponding to each sampling node row by row from the pure impurity spectrum matrix. The vector consisting of all elements of a row is denoted as ,in This is the time-series index of the sampling node. The value range is from 1 to an integer, and According to the production sequence, they increase from first to last. The dimension is ,Right now , Represents the th element in the pure impurity spectrum matrix. line, number The elements of the column.

[0055] Along the production timeline, calculate the change between the low-rank impurity baseline vectors of adjacent sampling nodes. For time-series indices of... and Calculate the low-rank impurity baseline vector between two adjacent sampling nodes. and The change between Change The calculation uses the norm form of the vector difference, and the formula is as follows: in, Indicates the first The sampling node and the first The change in the low-rank impurity baseline vector between each sampling node Indicates column index From 1 to Summing all terms together, Represents the th element in the pure impurity spectrum matrix. line, number Column elements, Represents the th element in the pure impurity spectrum matrix. line, number Column elements, This represents the total number of sampling nodes. Let be the length of the impurity feature vector. The change is calculated sequentially for all adjacent sampling node pairs to obtain... These change values ​​are arranged according to the production time sequence to form a sequence of impurity trend changes.

[0056] Simultaneously, the sparse fluctuation vector corresponding to each sampling node is extracted from the sparse matrix. The sparse fluctuation vector corresponding to each sampling node in the sparse matrix is ​​then extracted. The vector consisting of all elements of a row is denoted as , The dimension is , , Represents the th element in a sparse matrix. line, number The elements of the column. Calculate the sparse fluctuation vector with each sampling node as the unit. The amplitude range is obtained to obtain the corresponding value of the first... The impurity fluctuation envelope of each sampling node. The impurity fluctuation envelope consists of a lower bound and an upper bound. The lower bound is determined by the minimum element of the sparse fluctuation vector, and the upper bound is determined by the maximum element of the sparse fluctuation vector. In practice, the sparse fluctuation vector is traversed. All The minimum value among the given elements is taken as the lower bound of the impurity fluctuation envelope. The largest element value is taken as the upper bound of the impurity fluctuation envelope. .Right now , , This indicates the minimum value operation. This indicates the operation of finding the maximum value.

[0057] The impurity trend change sequence and impurity fluctuation envelope are combined according to the sampling nodes to obtain the impurity dynamic range description for each sampling node. For time series index... The impurity dynamic range description of a sampling node consists of three parts: the low-rank impurity baseline vector of that sampling node. The upper and lower bounds of the impurity fluctuation envelope of this sampling node. and And the change of this sampling node relative to the previous sampling node (when When the change does not exist, it is set to 0). The specific structure for describing the impurity dynamic range is as follows: , , and (when (when) or 0 (when) Each sample is packaged into a single data tuple. All data tuples from each sampling node are arranged according to production time sequence, forming the complete chain impurity profile analysis result. This complete chain impurity profile analysis result comprehensively describes the baseline levels, fluctuation ranges, and variation amplitudes of impurities in adjacent stages throughout the entire production chain of diosmin API and its formulations.

[0058] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for analyzing the full-chain impurities of diosmin based on near-infrared spectroscopy, characterized in that, include: Near-infrared spectral data from each sampling node in the entire production chain of diosmin API and its formulations were obtained to form a complete chain spectral dataset; The entire chain spectral dataset is decomposed in both the time and frequency domains to obtain the multi-scale spectral components of each sampling node; Based on the multi-scale spectral components, a spectral transfer mapping relationship between each sampling node is constructed to form a full-chain spectral correlation diagram; Based on the full-chain spectral correlation diagram, the impurity feature vectors of each sampling node are extracted, and the impurity feature vectors are arranged according to the production time sequence to generate an initial impurity spectrum matrix; The initial impurity spectrum matrix is ​​subjected to sparse low-rank decomposition to separate the trend term and fluctuation term of the impurity signal, thereby obtaining the pure impurity spectrum matrix; Based on the pure impurity spectrum matrix, the dynamic change information of impurities at each sampling node is generated to complete the full-chain impurity spectrum analysis; The joint time-domain and frequency-domain decomposition of the full-chain spectral dataset yields the multi-scale spectral components for each sampling node, specifically including: Empirical mode decomposition is performed on each near-infrared spectral data in the full-chain spectral dataset to obtain the set of intrinsic mode functions; Hilbert spectral analysis is performed on each eigenmode function in the set of eigenmode functions to obtain the instantaneous frequency and instantaneous amplitude; Based on the instantaneous frequency, each intrinsic mode function is divided into low-frequency band, mid-frequency band, and high-frequency band. The intrinsic mode functions within the same frequency band are reconstructed to obtain low-frequency spectral components, mid-frequency spectral components, and high-frequency spectral components, which serve as the multi-scale spectral components of the sampling node. The initial impurity spectrum matrix is ​​subjected to sparse low-rank decomposition to separate the trend term and fluctuation term of the impurity signal, resulting in a pure impurity spectrum matrix, specifically including: The initial impurity spectrum matrix is ​​decomposed into a low-rank matrix and a sparse matrix, wherein the low-rank matrix represents the common impurity baseline trend among all sampling nodes, and the sparse matrix represents the impurity fluctuation signal unique to each sampling node. Iteratively update the low-rank matrix and the sparse matrix until a preset iteration termination condition is met; The low-rank matrix obtained at the end of the iteration is used as the pure impurity spectrum matrix.

2. The method for analyzing the whole-chain impurities of diosmin based on near-infrared spectroscopy according to claim 1, characterized in that, Based on the multi-scale spectral components, a spectral transfer mapping relationship is constructed between each sampling node to form a full-chain spectral correlation diagram, specifically including: Each sampling node is used as a graph node, and the low-frequency spectral component in the multi-scale spectral components is used as the node attribute of the graph node. Along the production timeline, the low-frequency spectral components of two adjacent graph nodes are dynamically time-warped to calculate the spectral morphology similarity. When the spectral morphological similarity is greater than a preset threshold, a directed edge is established between the two corresponding graph nodes, and the weight of the directed edge is determined by the spectral morphological similarity. By combining all graph nodes and directed edges, the full-chain spectral correlation graph is obtained.

3. The method for analyzing the whole-chain impurities of diosmin based on near-infrared spectroscopy according to claim 1, characterized in that, Based on the full-chain spectral correlation diagram, impurity feature vectors are extracted from each sampling node. These impurity feature vectors are then arranged according to production time sequence to generate an initial impurity spectrum matrix, specifically including: For each graph node in the full-chain spectral correlation graph, extract the weights of all adjacent directed edges of the graph node to form the adjacency weight sequence of the graph node; Statistical feature calculations are performed on the adjacency weight sequence to obtain the adjacency distribution feature values ​​of the nodes in the graph; The node attributes of the graph node are concatenated with the adjacency distribution feature value to form the impurity feature vector of the graph node; The impurity feature vectors of all graph nodes are stacked vertically according to the production time sequence to generate the initial impurity spectrum matrix.

4. The method for analyzing the whole-chain impurities of diosmin based on near-infrared spectroscopy according to claim 1, characterized in that, Empirical mode decomposition is performed on each near-infrared spectral data in the full-chain spectral dataset to obtain the intrinsic mode function set, which specifically includes: For each near-infrared spectral data point, identify all its local maxima and local minima, fit the upper and lower envelopes respectively, and calculate the mean curves of the upper and lower envelopes; Subtract the mean curve from the near-infrared spectral data to obtain the component to be determined; When the component to be determined satisfies the two determination conditions of the intrinsic mode function, the component to be determined is stored as an intrinsic mode function in the intrinsic mode function set, and the remaining signal is further decomposed. When the component to be determined does not meet the two determination conditions, the above steps are repeated with the component to be determined as new near-infrared spectral data until the remaining signal is a monotonic function or the number of extreme points is less than the preset number, at which point the decomposition is terminated.

5. The method for analyzing the whole-chain impurities of diosmin based on near-infrared spectroscopy according to claim 2, characterized in that, Along the production timeline, dynamic time warping is performed on the low-frequency spectral components of two adjacent graph nodes, and the calculation of spectral morphology similarity specifically includes: The low-frequency spectral components of the previous graph node are used as the first sequence, and the low-frequency spectral components of the next graph node are used as the second sequence. Construct a cumulative distance matrix between the first sequence and the second sequence, wherein each element in the cumulative distance matrix is ​​obtained by adding the distance between the two sequence points at the corresponding position to the previous minimum cumulative distance; Starting from the last element of the cumulative distance matrix and backtracking to the first element, the optimal regularized path is obtained; The spectral morphology similarity is obtained by taking the reciprocal of the cumulative distance to the endpoint of the optimal regular path and normalizing it.

6. The method for analyzing the whole-chain impurities of diosmin based on near-infrared spectroscopy according to claim 3, characterized in that, Statistical feature calculations are performed on the adjacency weight sequence to obtain the adjacency distribution feature values ​​of the graph nodes, specifically including: Calculate the arithmetic mean and median of the adjacent weight sequence as a feature of central tendency; Calculate the standard deviation of the adjacency weight sequence as a feature of its dispersion. The skewness and kurtosis of the adjacency weight sequence are calculated as distribution morphology features; The arithmetic mean, the median, the standard deviation, the skewness, and the kurtosis are combined to form the adjacency distribution characteristic value of the graph node.

7. The method for analyzing the whole-chain impurities of diosmin based on near-infrared spectroscopy according to claim 1, characterized in that, Iteratively updating the low-rank matrix and the sparse matrix until a preset iteration termination condition is met specifically includes: With the current sparse matrix fixed, the difference between the initial impurity spectrum matrix and the current sparse matrix is ​​subjected to singular value thresholding to shrink the low-rank matrix and update the low-rank matrix. The updated low-rank matrix is ​​fixed, and the difference between the initial impurity spectrum matrix and the updated low-rank matrix is ​​subjected to soft thresholding to shrink the sparse matrix. Calculate the ratio of the difference norm of the low-rank matrix before and after this iteration to the norm of the low-rank matrix. When the ratio is less than the preset tolerance, it is determined that the iteration termination condition has been met, and the iteration is stopped.

8. The method for analyzing the whole-chain impurities of diosmin based on near-infrared spectroscopy according to claim 4, characterized in that, The two criteria for determining the intrinsic mode function are: the difference between the number of extreme points and the number of zero-crossing points does not exceed 1, and the mean of the upper envelope and the lower envelope at any point is zero.

Images