Rapid detection method for traditional Chinese medicine compound components

Abstract

The invention relates to the technical field of traditional Chinese medicine component detection and discloses a rapid detection method for traditional Chinese medicine compound components. The method comprises the following steps: performing Raman spectrum scanning on a traditional Chinese medicine compound sample to generate component spectrum distribution information, and performing fitting by adopting a partial least square algorithm to determine a local component anomaly risk area; meanwhile, atomic force microscopic imaging is carried out, a microstructure distribution map is generated, and a potential component mismatch feature region is marked through texture feature extraction and pattern recognition. And generating a detection optimization adjustment region based on the space coordinates of the region, analyzing the gradient change in the molecular arrangement direction of the region to evaluate the component distribution offset state, and analyzing the interface bonding state of the active component and the auxiliary material to evaluate the component interaction coupling strength. And finally, on the basis of an evaluation result, whether the real-time parameter calibration is started in the detection optimization adjustment area is judged through a dual-channel decision-making mechanism. According to the method, correlation analysis of component spectrums and microtopographies is realized, and the accuracy and pertinence of traditional Chinese medicine compound component detection are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of traditional Chinese medicine component detection, in particular to a rapid detection method for traditional Chinese medicine compound components. BACKGROUND

[0002] As an important part of traditional Chinese medicine, traditional Chinese medicine compound has complex and diverse components, including various active ingredients, excipients and potential impurities. The types, contents and interactions of the components directly affect the efficacy and safety. With the advancement of modernization of traditional Chinese medicine, the demand for accurate detection of traditional Chinese medicine compound components is increasingly urgent, but the existing detection techniques still have many limitations. Traditional detection methods for traditional Chinese medicine components, such as high performance liquid chromatography and gas chromatography, can achieve quantitative analysis of some components, but the operation process is complicated, complex sample pretreatment is required, it takes a long time, and it is difficult to simultaneously and rapidly screen multiple components. Although the methods of infrared spectroscopy and ultraviolet spectroscopy in spectral detection technology are relatively simple to operate, they are easily disturbed by the matrix, have limited specific recognition ability for components, and especially when multiple components coexist in the compound, the spectral signals overlap seriously, making it difficult to accurately distinguish the distribution characteristics of different components. Raman spectroscopy technology has been applied in the analysis of traditional Chinese medicine components due to its advantages of no sample pretreatment and non-destructive detection, but when used alone, it can only obtain spectral information of the components, and it is difficult to associate the microscopic morphology characteristics of the components. Atomic force microscopy technology can directly reflect the chemical composition and distribution of the components, but it is difficult to accurately mark the feature area related to component mismatch.

[0003] Existing detection methods mainly focus on single-dimensional analysis, lack of targeted focus on component abnormal areas, and often detect the whole sample, which may ignore the component fluctuations in small local areas. At the same time, the interaction between components is not evaluated, it is difficult to judge whether the combination of active ingredients and excipients is reasonable, and the detection results cannot fully reflect the actual quality state of traditional Chinese medicine compound. In the detection process, there is a lack of dynamic adjustment mechanism, which cannot optimize the detection parameters according to the real-time detection situation, and it is difficult to adapt to the complex and variable characteristics of traditional Chinese medicine compound components. These problems restrict the improvement of the quality control level of traditional Chinese medicine compound. SUMMARY

[0004] The purpose of the present application is to provide a rapid detection method for traditional Chinese medicine compound components to solve the problems in the background art.

[0005] To achieve the above purpose, the present application provides a rapid detection method for traditional Chinese medicine compound components, which comprises: Raman spectrum scanning is performed on the traditional Chinese medicine compound sample to generate component spectral distribution information, and partial component abnormal risk areas are determined by fitting and processing the component spectral distribution information using a partial least squares algorithm; Atomic force microscopy is performed on the traditional Chinese medicine compound sample to generate a micro-morphology distribution map, and potential component mismatch feature areas are marked based on texture feature extraction and pattern recognition; Based on the spatial coordinates of the local component abnormal risk areas and the potential component mismatch feature areas, a detection optimization adjustment area is generated; By analyzing the gradient change of the molecular arrangement direction of the detection optimization adjustment area, it is evaluated whether the component distribution offset state is normal; By analyzing the interface binding state of active ingredients and excipients in the detection optimization adjustment area, it is evaluated whether the component interaction coupling strength is reasonable; Based on the evaluation results of the component distribution offset state and the component interaction coupling strength, a dual-channel decision mechanism is used to determine whether the detection optimization adjustment area starts real-time parameter calibration.

[0006] Preferably, the Raman spectrum scanning of the traditional Chinese medicine compound sample to generate component spectral distribution information comprises: Obtain characteristic spectral band data containing target active ingredients by a Raman spectrometer; Extract the peak intensity and half-peak width parameters in the characteristic spectral band data to generate component spectral distribution information; Use a partial least squares algorithm to reduce dimensionality and fit the component spectral distribution information, and mark local component abnormal risk areas according to principal component weight coefficients.

[0007] Preferably, the atomic force microscopy imaging of the traditional Chinese medicine compound sample to generate a micro-morphology distribution map comprises: Obtain sample surface topological morphology data by an atomic force microscope, and enhance the microstructure contrast by morphological filtering processing; Molecular aggregation features in the topological morphology data are extracted using a multi-scale texture segmentation algorithm; Random forest classifier is used to identify the pattern category of the molecular aggregation features, and potential component mismatch feature areas are marked according to the classification results.

[0008] Preferably, the generation of the detection optimization adjustment area comprises: Extract the coordinate set of the local component abnormal risk areas and the coordinate set of the potential component mismatch feature areas; Calculate the spatial union set of the coordinate sets to generate a minimum circumscribed rectangle area covering the maximum abnormal range as the detection optimization adjustment area.

[0009] Preferably, the evaluation of whether the component distribution offset state is normal comprises: obtaining a molecular orientation angle distribution map in the detection optimization adjustment region; calculating a gradient change rate of the molecular orientation angle distribution map in a spatial dimension; determining whether the component distribution deviation state exceeds a preset tolerance threshold according to a standard deviation and a mean value of the gradient change rate.

[0010] Preferably, the evaluation of whether the component interaction coupling strength is reasonable comprises: analyzing molecular bonding spectrum data of the active ingredient and the excipient in the detection optimization adjustment region; constructing an interface binding feature tensor and performing non-negative matrix factorization to obtain a component interaction coupling coefficient distribution; determining whether the component interaction coupling strength is in a preset reasonable interval according to a positive and negative polarity difference value of the coupling coefficient distribution.

[0011] Preferably, the dual-channel decision mechanism comprises: when the component distribution deviation state is normal and the component interaction coupling strength is reasonable, a standard detection protocol is executed; otherwise, a real-time parameter calibration process is triggered to adjust the spectrum scanning resolution and the microscopic imaging sampling frequency.

[0012] Preferably, the real-time parameter calibration process comprises: calculating a texture complexity index and a component concentration coefficient of variation of the detection optimization adjustment region; when the texture complexity index or the component concentration coefficient of variation exceeds a dynamic threshold, a high-precision review sub-region is peeled off from the current detection region; updating probe scanning path parameters of an atomic force microscope according to molecular conformation features of the high-precision review sub-region.

[0013] Preferably, the method further comprises: establishing a real-time feedback link between spectrum data and microscopic imaging data; when the review result of the high-precision review sub-region deviates from the initial detection result, generating a component distribution correction coefficient; adjusting a feature extraction weight matrix of Raman spectrum based on the component distribution correction coefficient.

[0014] Preferably, the method further comprises: comparing differences between real-time detection results and a standard component spectrum library; updating an abnormal pattern recognition rule library in a multi-layer experience pool according to a category attribution of the difference factors; applying the updated abnormal pattern recognition rule library to an initial spectrum fitting process of the next detection.

[0015] Compared with the prior art, the present application has the following advantages: The present method breaks through the limitations of single detection technology by combining Raman spectrum scanning with atomic force microscopy imaging technology, realizing multi-dimensional analysis of traditional Chinese medicine compound components from macroscopic spectral information to microscopic morphological characteristics. Raman spectrum can accurately capture the chemical composition characteristics of the components, and the local component abnormal risk area can be effectively screened out through partial least squares algorithm fitting. Atomic force microscopy imaging can directly show the microstructure of the sample, and the potential component mismatch feature area can be accurately marked by means of texture feature extraction and pattern recognition. The synergistic effect of the two technologies makes the identification of component abnormalities more comprehensive. Based on the spatial coordinates of the local component abnormal risk area and the potential component mismatch feature area, a detection optimization adjustment area is generated, so that the detection focus is more clear, and the waste of resources and low efficiency caused by indiscriminate detection of the whole sample are avoided. By focusing on a specific area, the detection resources can be concentrated on the part where problems may exist, improving the pertinence and effectiveness of the detection and reducing the interference of irrelevant information, making the subsequent analysis more targeted. Analyzing the gradient change of the molecular arrangement direction in the detection optimization adjustment area can further explore the distribution state of the components at the microscale, and evaluate whether the component distribution deviation is within the normal range from the perspective of molecular arrangement. This micro-level analysis complements the shortcomings of traditional component content detection and provides a new perspective for judging the rationality of component distribution. At the same time, analyzing the interface binding state of active ingredients and excipients in this area can understand the interaction between components and evaluate whether the component interaction coupling strength is reasonable, which helps to fully grasp the mutual influence between components in traditional Chinese medicine compound and better meet the characteristics of multi-component synergistic action of traditional Chinese medicine compound. The introduction of the dual-channel decision mechanism makes the judgment of the detection results more scientific and reasonable. Based on the dual evaluation results of component distribution deviation state and component interaction coupling strength, the decision is made, which avoids the possible deviation of single index judgment and can more comprehensively consider the actual situation of the detection area, so as to accurately judge whether real-time parameter calibration is needed. This dynamic adjustment mechanism makes the detection process more adaptive, which can flexibly optimize the detection parameters according to the actual state of traditional Chinese medicine compound components, better cope with the complex and variable characteristics of traditional Chinese medicine compound components, meet the needs of different samples and different detection scenarios, and make the entire detection process more in line with the actual requirements of traditional Chinese medicine compound quality control. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The working principle diagram of the traditional Chinese medicine compound component rapid detection method described in the present application; Figure 2 The flowchart for Raman spectrum scanning and determination of local component abnormal risk area; Figure 3 The flowchart for evaluating the component distribution deviation state; Figure 4 Flowchart for real-time parameter calibration. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] See also Figure 1 The present invention provides a method for rapid detection of components of a traditional Chinese medicine compound, the method comprising: Raman spectroscopy is performed on traditional Chinese medicine compound samples to generate information representing the spatial distribution of chemical components. This information is processed using a partial least squares algorithm to identify areas of local component abnormality risk. Simultaneously, atomic force microscopy imaging is performed on the same sample to generate a map reflecting the microscopic morphology and structure, and potential component mismatch feature areas are marked based on texture feature extraction and pattern recognition technology; Integrate the spatial coordinates of the risk areas identified by the two aforementioned technologies to generate an inspection optimization and adjustment area that requires focused review; In this region, the spatial gradient changes in the molecular arrangement direction are analyzed to assess whether the component distribution offset state is normal; Analyze the molecular bonding state of the active ingredient and excipient interface in this area and evaluate whether the interaction coupling strength of the ingredients is reasonable; Based on the results of the above two evaluations, a dual-path decision-making mechanism is used to decide whether to initiate a real-time parameter calibration process for the detection optimization adjustment area, thereby optimizing the subsequent detection accuracy.

[0019] Example 1: See Figure 2Raman spectral scanning operation is performed under specific environmental conditions using a Raman spectrometer system equipped with a high-sensitivity CCD detector. The traditional Chinese medicine compound sample is fixed on a three-dimensional precision moving platform. The spectrometer emits a laser beam of a specific wavelength focused on the preset starting point on the sample surface. The laser power and integration time are pre-optimized according to the photosensitivity of the sample. The system controls the platform to move point by point according to the preset scanning path and step size. At each spatial point, the laser excites the sample to generate Raman scattered light, which is received by the detector after being spectrally dispersed by the grating of the spectrometer, and is converted into an electrical signal. The signal is processed by a preamplifier and an analog-to-digital converter to form the original spectral data corresponding to this spatial point. The original spectral data of each point is preprocessed, including removing cosmic ray noise and performing baseline correction to eliminate fluorescence background interference. The preprocessed spectral data constitute the preliminary spectral data set. From the data set, the characteristic spectral band data corresponding to the chemical bond vibration of the target active ingredient is identified and extracted. This is achieved by comparing the known standard active ingredient Raman spectral fingerprint region. Within the extracted characteristic spectral band data, the peak intensity of the identified characteristic peak is calculated, that is, the height of the highest point of the characteristic peak relative to the baseline. At the same time, the half-peak width of the characteristic peak is calculated, that is, the peak width value corresponding to the half height of the characteristic peak. The peak intensity and half-peak width together constitute the component representation information of the spatial point. The component representation information of all scanning points is arranged according to its spatial coordinates to form a two-dimensional or three-dimensional component spectral distribution information map. The map directly shows the spatial distribution of the target active ingredient on the sample surface.

[0020] When processing the component spectral distribution information, the partial least squares algorithm is applied. First, the algorithm standardizes the component spectral distribution information data containing multiple bands. Then, the algorithm performs dimension reduction operation to project the high-dimensional spectral data into a low-dimensional space and extracts several principal components. These principal components can represent the variation information in the original spectral data to the greatest extent. The algorithm calculates the weight coefficient of each principal component, which reflects the contribution of each original spectral variable (band) to the principal component. In the principal component space after dimension reduction, according to the principal component score value of each spatial point and the distribution of its weight coefficient, spatial analysis is performed. The algorithm identifies those spatial points whose principal component score values significantly deviate from the average range statistically obtained from most normal regions. The spatial coordinates of these abnormal points are clustered to form a continuous or discrete collection region in space. Finally, the system marks these collection regions as local component abnormal risk regions and outputs their spatial coordinate information.

[0021] Atomic force microscopy operation uses a commercial atomic force microscope system equipped with a micro-cantilever probe with a specific elastic constant. The probe is driven to perform a grid scan on the sample surface. The interaction force between the probe and the sample surface causes the micro-cantilever to deflect or change in amplitude. The detection system monitors this deflection or amplitude change in real time by a laser beam shining on the cantilever tip reflected onto a position sensitive detector. The detector signal is converted into data reflecting the vertical displacement of the probe, which corresponds to the topographic height information of each point on the sample surface. The topographic height information of all point positions is arranged according to the scanning coordinates to generate the original sample surface topographic data. This original data usually contains non-surface true topographic information caused by instrument noise or thermal drift. In order to obtain the true microstructure information, morphological filtering is performed on the original topographic data. Morphological filtering is based on a combination of opening and closing operations on the image with a set structure element. The opening operation first erodes and then expands, which helps to smooth the object contour, disconnect narrow connections, and eliminate small protrusions. The closing operation first expands and then erodes, which helps to fill small holes in the object, connect adjacent objects, and smooth the contour while basically maintaining the area unchanged. By selecting a suitable size and shape of the structure element to perform a series of morphological operations, the background noise is effectively suppressed, the contrast of the microstructure (such as the boundary of molecular clusters, the surface steps of crystals, and the hole structure) is enhanced, and the micro-topographic distribution map with significantly improved clarity is generated. This map more accurately reflects the true microstructure characteristics of the sample surface.

[0022] When analyzing the micro-morphology distribution map, a multi-scale texture segmentation algorithm is used. Texture segmentation aims to identify regions in the map with different texture characteristics. Multi-scale analysis means that the algorithm examines texture features at multiple spatial scales. A set of different scale parameters is usually set, for example, different sizes of analysis windows or filter kernels are used. At each scale, the algorithm calculates the local texture feature values around each spatial point on the map. These feature values can include texture features calculated based on the gray level co-occurrence matrix (such as energy, entropy, contrast, uniformity), or texture features calculated based on the local binary pattern, or multi-resolution texture features extracted using wavelet transform. By extracting texture features at different scales, the algorithm can capture information about microstructure changes at different spatial ranges, from fine molecular arrangement to larger molecular aggregates. For each spatial point, the texture feature values calculated at each scale are integrated to form a multi-dimensional feature vector. This feature vector describes the texture properties of the point and its neighborhood at multiple scales. The algorithm processes the multi-dimensional feature vector data of all spatial points, identifies the boundaries of regions where the texture features change significantly, and thus realizes the multi-scale segmentation of the map. The segmentation result marks the regions with different texture characteristics. The texture features of these regions are considered to be related to the molecular aggregation state, such as uniform dispersion region, ordered crystalline region, disordered aggregation region or phase separation boundary region. The molecular aggregation feature information output by the segmentation algorithm, including the region boundary coordinates and the corresponding feature vector values, is used for subsequent pattern recognition.

[0023] The task of identifying and labeling potential composition mismatch feature regions is accomplished by a random forest classifier. Random forest is an ensemble learning method composed of multiple decision trees. Before use, the classifier needs to be trained based on samples with known composition matching status. The training dataset contains a large number of sample atomic force microscopy topography map regions and their corresponding texture feature vectors, each region is labeled with its composition matching status category (such as "match", "mild mismatch", "severe mismatch") by manual or other reliable methods. During training, the random forest algorithm randomly selects a subset of samples with replacement from the training dataset, and randomly selects some texture feature variables for each tree. Then, based on these randomly selected samples and features, multiple decision trees are constructed. Each tree independently judges the input features, and the final classification result is determined by voting of all decision trees. After the trained random forest classifier is deployed, it will be applied to the analysis of the current sample. The molecular aggregation feature vectors extracted by the multi-scale texture segmentation algorithm, which represent different regions on the map, are input one by one into the trained random forest classifier. The classifier classifies the regions represented by each input feature vector according to the decision rules learned by it, and predicts which composition matching status category the region belongs to. The classifier outputs the predicted category label for each analysis region. According to the prediction results, the system spatially labels the regions identified as "mild mismatch" or "severe mismatch" categories on the microtopography distribution map, and records their boundary coordinate information. These labeled regions are the potential composition mismatch feature regions. The coordinate information of the region is output for integrated use in subsequent steps.

[0024] The atomic force microscopy imaging process scans the sample surface with an atomic force microscope probe to obtain high-resolution sample surface topography data. The original topography data is processed by morphological filtering (such as opening operation, closing operation) to eliminate noise and enhance the contrast of microstructures (such as molecular clusters, crystal boundaries), and a clear microtopography distribution map is obtained. Subsequently, a multi-scale texture segmentation algorithm (such as an algorithm based on wavelet transform or local binary pattern) is used to analyze the processed topography data and extract texture features reflecting the molecular aggregation state at different scales. The extracted molecular aggregation features are classified and trained using a random forest classifier, which makes judgments based on pre-learned known composition matching patterns, and according to its classification output, regions identified as abnormal aggregation patterns are labeled as potential composition mismatch feature regions.

[0025] Example 2: see Figure 3The operation of generating the detection optimization adjustment region starts with the integration of the spatial location information of potential anomalies identified by both independent detection techniques. From the results outputted by the Raman spectral scanning and the partial least squares algorithm analysis segment, the spatial coordinates of the local compositional anomaly risk regions that have been labeled are extracted. These coordinates usually exist in the form of a two- or three-dimensional point set, recording the precise locations of each point site that is judged to have spectral features significantly deviating from the normal range on the surface or inside of the sample. This coordinate point set is referred to as coordinate set A. At the same time, from the results outputted by the atomic force microscopic imaging and the random forest classifier analysis segment, the spatial coordinates of the potential compositional mismatch feature regions that have been labeled are extracted. These coordinates also constitute a point set, recording the precise locations of each point site that is judged by the classifier to have the possibility of compositional mismatch in the microscopic morphology. This coordinate point set is referred to as coordinate set B. The system processes these two coordinate point sets and calculates the spatial relationship. The core task is to obtain the spatial union of coordinate set A and coordinate set B. The spatial union operation means that all coordinate points in the two point sets are merged together, and after removing the duplicate points, a new and more comprehensive coordinate point set is formed. This new union point set includes both the local compositional anomaly risk points discovered by Raman spectral analysis and the potential compositional mismatch points discovered by atomic force microscopic morphology analysis. It represents all suspicious spatial locations that are jointly indicated by the two techniques and need to be reviewed in detail.

[0026] Based on the merged spatial union point set, the system performs a computational geometry operation to determine a single, continuous rectangular region that can completely cover all coordinate points in the union. The goal of this operation is to generate a minimum circumscribed rectangular region. The minimum circumscribed rectangle means that among all rectangles that can completely contain all points in the union point set, its area is the smallest. The system usually uses the convex hull algorithm as the basic step. The convex hull algorithm first calculates the convex hull profile of the union point set, that is, the smallest convex polygon boundary that contains all points. Once the convex hull profile is determined, the system further calculates the minimum circumscribed rectangle of the convex hull. This can be achieved by the rotating calipers algorithm, which finds the direction in which the projection of the convex hull is the widest by rotating a set of calipers parallel to the edges of the convex hull (representing the sides of the rectangle), thereby determining the position and size of the minimum circumscribed rectangle. The boundary coordinates of the finally calculated minimum circumscribed rectangular region are recorded by the system. This rectangular region is defined as the detection optimization adjustment region. The delineation of this region ensures that all potential anomaly points labeled by spectral or morphological analysis alone or jointly are included, providing a clear spatial range definition for subsequent in-depth and fine analysis within this specific region.

[0027] The process of evaluating whether the component distribution deviation state in the optimization adjustment region is normal needs to obtain the arrangement direction information at the molecular level. The acquisition of the molecular orientation angle distribution map is the key first step. This depends on the lateral force microscope mode of the atomic force microscope or the combination of the polarization Raman spectroscopy technology. In the lateral force microscope mode, when the probe of the atomic force microscope scans the sample surface, it synchronously records the lateral torsion amount generated by the friction or viscous force of the sample surface in addition to the vertical direction topography information. This torsion amount is related to the angle between the scanning direction of the probe and the molecular arrangement direction. By measuring the difference in lateral force when the scanning direction is reversed and combining the mechanical properties of the probe, the average molecular orientation angle at each point on the scanning path can be derived. The polarization Raman spectroscopy technology uses the dependence of Raman scattering intensity on the polarization direction of the incident laser. By rotating the polarization direction of the incident laser or using a polarization analyzer to detect the polarization state of the scattered light, the change of the intensity of a specific Raman peak with the polarization angle can be measured, and the orientation angle of the molecular vibration mode at this point can be inverted. By comprehensively applying these technologies, the system obtains high spatial resolution molecular orientation angle data in the optimization adjustment region, and constructs a molecular orientation angle distribution map reflecting the molecular arrangement direction at each spatial point in the region. This map directly depicts the order of molecular arrangement on the sample surface or near the surface and its spatial variation.

[0028] The gradient change rate of the molecular orientation angle distribution map in the spatial dimension is the core quantitative step of evaluating the distribution state. The gradient describes the speed and direction of the change of a physical quantity in space. For a two-dimensional molecular orientation angle distribution map, the gradient change rate in the X (horizontal) direction and the Y (vertical) direction needs to be calculated. The system usually uses a discrete gradient approximation algorithm, such as the Sobel operator. The Sobel operator contains two 3x3 convolution kernels, one for calculating the X direction gradient and the other for calculating the Y direction gradient. The Sobel kernel in the X direction is convolved with the original angle distribution map to obtain the gradient component of each point in the X direction. The Sobel kernel in the Y direction is convolved with the original angle distribution map to obtain the gradient component of each point in the Y direction. For a point, the size of its gradient change rate can be obtained by calculating the vector module length of its X direction gradient component and Y direction gradient component. The module length value represents the degree of change of the molecular orientation angle at that point. The system traverses all the spatial points in the optimization adjustment region and performs the above gradient calculation, finally obtaining a spatial gradient change rate distribution map covering the entire region. This map quantifies the local variation intensity of the molecular arrangement direction in space.

[0029] Determining whether the component distribution shift is normal relies on a statistical analysis of the calculated gradient change rate distribution within the entire detection optimization adjustment area. The system first calculates global statistics for the gradient change rate distribution map, focusing on its standard deviation and mean. The standard deviation measures the degree to which the gradient change rate values ​​at all points in the area deviate from their mean, reflecting the dispersion or fluctuation of the spatial distribution of the change rate. The mean represents the average level of the gradient change rate within the area. Preset tolerance thresholds define the maximum acceptable gradient change rate fluctuation and average level within the normal range for component distribution shift. These thresholds are typically derived from similar analyses of a large number of standard samples with known uniform and normal component distributions. The system compares the calculated standard deviation of the gradient change rate with the preset "standard deviation tolerance threshold." Simultaneously, the calculated mean of the gradient change rate is compared with the preset "mean tolerance threshold." The judgment rule is as follows: If the calculated standard deviation is less than or equal to the standard deviation tolerance threshold, and the calculated mean is less than or equal to the mean tolerance threshold, the component distribution shift within the detection optimization adjustment area is considered normal. Conversely, if the calculated standard deviation exceeds the standard deviation tolerance threshold, or the calculated mean exceeds the mean tolerance threshold, the component distribution shift state in the region is determined to be abnormal. This determination result serves as one of the important inputs to the subsequent dual-path decision-making mechanism.

[0030] When evaluating the compositional distribution shift within the optimized region, a detailed molecular orientation angle distribution map is first obtained. This map, obtained using the lateral force mode of an atomic force microscope or polarized Raman spectroscopy, reflects the molecular orientation at each point within the region. The gradient change rate of this molecular orientation angle distribution map in the spatial dimensions (X and Y directions) is calculated, typically using the Sobel operator or gradient vector calculation. Statistics for the gradient change rate across the entire region are obtained, including its standard deviation and mean. A preset tolerance threshold defines the allowable spatial fluctuation range of the gradient change rate. The calculated standard deviation and mean of the gradient change rate are compared with the preset tolerance threshold. If the statistical value exceeds the threshold, the compositional distribution shift in the region is considered abnormal; otherwise, it is considered normal.

[0031] Example 3: Evaluate whether the detected strength of the component interaction coupling in the optimization adjustment region is reasonable, starting from the high-precision spectral detection of the molecular interface bonding state in the region. Using a high-spatial-resolution Raman spectrometer, focus the laser beam on each fine scanning point in the detection optimization adjustment region. The laser power and integration time are independently optimized according to the light stability of the excipient and active ingredient to obtain high signal-to-noise ratio spectral signals. At each spatial point, the system collects a complete Raman spectrum. Special attention is paid to specific spectral feature bands that can reflect intermolecular interaction forces, such as characteristic vibration modes related to hydrogen bond formation, van der Waals force action, or charge transfer. These modes are usually manifested as shifts in the position of characteristic peaks, significant changes in the intensity of peaks, or broadening of peak shapes. The system performs accurate baseline subtraction and noise filtering on the collected spectral data. Subsequently, in the preset bonding-sensitive spectral interval, identify and extract molecular bonding spectral data directly related to the interface bonding state of the active ingredient-excipient. This includes accurately measuring the peak shift values of key characteristic peaks, calculating the relative intensity ratios of characteristic peaks, and quantifying the half-peak width changes of characteristic peaks. These parameters together constitute the spectral fingerprint information representing the interface bonding state at that point. Repeat the above operation for all spatial points in the region to obtain a high-resolution molecular bonding spectral data set covering the entire detection optimization adjustment region.

[0032] In order to comprehensively characterize the interface bonding state and its spatial variability in the entire region, it is necessary to organize the multi-dimensional molecular bonding spectral data into a structured mathematical representation. This is achieved by constructing an interface bonding feature tensor. The tensor integrates information in three dimensions: spatial dimension, spectral feature dimension, and time dimension. The spatial dimension corresponds to the coordinate positions of all sampling points in the detection optimization adjustment region. The spectral feature dimension includes multiple key quantitative parameters extracted from the original spectrum, such as the displacement of specific peaks, intensity ratios, half-peak widths, etc. The time dimension here can be understood as the sequence relationship between different parameters or different spatial points. The system fills all spectral feature parameter values extracted at each spatial point into a multi-dimensional array according to their corresponding feature types and spatial coordinates. The multi-dimensional array is the interface bonding feature tensor. Its mathematical representation is:

[0033] Where: represents the constructed interface bonding feature tensor; is the spatial coordinate index, identifying a specific position point in the detection optimization adjustment region; k is the spectral feature index, identifying the specific bonding feature parameter type; x represents the physical longitudinal coordinate; y represents the physical longitudinal coordinate; indicates the spectral feature parameter value; represents the auxiliary time series or sequential index associated with this feature. Before constructing the tensor, it is necessary to normalize different spectral feature parameters to eliminate the dimension difference and ensure the fairness of subsequent analysis.

[0034] The non-negative matrix factorization algorithm is applied to process the interface binding feature tensor. Non-negative matrix factorization is a low-rank decomposition method that can discover hidden features from non-negative data matrices. The system first expands the three-dimensional tensor into a two-dimensional matrix form in a certain dimension to adapt to the input requirements of the standard non-negative matrix factorization algorithm. The goal of decomposition is to approximately decompose the original feature matrix V into the product of two non-negative matrices: the basis matrix W and the coefficient matrix H, i.e., V ≈ WH. Among them, the column vectors of W represent the learned basis patterns, and the row vectors of H represent the corresponding activation coefficients in space. The system uses an iterative optimization algorithm to solve the above decomposition problem, and the objective function usually chooses to minimize the Frobenius norm or Kullback-Leibler divergence of the original matrix and the decomposed matrix. The maximum number of iterations and the convergence threshold are set as stopping conditions in the iteration process. After non-negative matrix factorization, the system extracts key information from the coefficient matrix H. The numerical distribution map of this coefficient matrix is the component interaction coupling coefficient distribution map that reflects the interaction strength between active ingredients and excipients in the detection and optimization adjustment region. The numerical size on this map directly represents the relative strength of the coupling strength at each point in space, and a positive value usually indicates strong coupling, while a negative value indicates weak coupling or repulsive effect.

[0035] After obtaining the component interaction coupling coefficient distribution map, the system needs to quantify the polarity difference features shown in the distribution map, i.e., the relative proportion and intensity difference between positive and negative regions. This is achieved by calculating a scalar indicator called the positive-negative polarity difference value. The system first performs binaryzation on the distribution map: all positive pixels are marked as foreground, and all negative pixels are marked as background. Then, the connected region analysis algorithm is applied to identify and label independent positive region clusters in the foreground and independent negative region clusters in the background. The average intensity value of each positive region cluster and the pixel area it occupies are calculated. Similarly, the average intensity value of each negative region cluster and the pixel area it occupies are calculated. Based on this region cluster information, the system calculates a comprehensive difference measure value. The general form of this value considers the ratio or difference between the sum of the average intensity weighted areas of positive regions and the sum of the average intensity weighted areas of negative regions, and may introduce a balance factor to adjust the sensitivity to the asymmetry of positive and negative regions. The calculated positive-negative polarity difference value is a single numerical value that quantifies the overall coupling strength polarity distribution feature.

[0036] The determination of whether the component interaction coupling strength is reasonable depends on comparing the calculated positive-negative polarity difference value with a preset reasonable interval. The preset reasonable interval is determined according to a large amount of historical data or analysis results of known qualified samples, and defines the minimum and maximum range in which the difference value should fall under normal process conditions. The system presets a lower threshold and an upper threshold. The judgment logic is as follows: if the calculated positive-negative polarity difference value is greater than or equal to the preset lower threshold and at the same time is less than or equal to the preset upper threshold, the value is within the preset reasonable interval, and the system determines that the component interaction coupling strength in the detection optimization adjustment region is reasonable. Conversely, if the calculated positive-negative polarity difference value is less than the preset lower threshold or greater than the preset upper threshold, the value is outside the preset reasonable interval, and the system determines that the component interaction coupling strength in the region is unreasonable. This determination result, together with the determination result of the component distribution deviation state, is input into the double-channel decision mechanism to guide whether the subsequent detection process needs to be calibrated in real time.

[0037] Example 4: see Figure 4 The double-channel decision mechanism operates according to the evaluation conclusions of the component distribution deviation state and the component interaction coupling strength in the detection optimization adjustment region. When both evaluations are determined to be normal states, the system maintains the current configuration and executes the preset standard detection protocol, which includes fixed spectral scanning parameters and microscopic imaging frequency. When either evaluation result is determined to be an abnormal state, the system immediately interrupts the standard process and triggers the real-time parameter calibration process. For example, a certain angelica-huangqi compound sample detects a molecular orientation gradient change rate standard deviation of 8.7 (exceeding the threshold of 7.5) in the region X [120-180 pm], Y [80-130 pm], and at the same time the positive-negative polarity difference value drops to -0.15 (below the lower threshold of -0.05), at which time the system automatically activates the parameter calibration module.

[0038] After the real-time parameter calibration process is started, the feature quantization of the detection optimization adjustment region that triggered the anomaly is first performed. The texture complexity index is obtained by analyzing the atomic force microscopic image of the region. The system divides the image into grid cells, calculates the gray level co-occurrence matrix of each cell, derives texture feature parameters such as energy, entropy, and contrast from it, and finally synthesizes a single complexity score. The component concentration coefficient of variation is calculated based on Raman spectral data, and the system extracts the feature peak intensity values of each sampling point in the region and calculates the variation degree. The dynamic threshold setting module adjusts the threshold level in real time according to historical detection data and sample types, for example, the texture complexity threshold is set to 5.2 and the component concentration coefficient of variation threshold is set to 0.35 for the current batch of samples. The system compares the calculation results with the dynamic threshold, and when either index exceeds the limit, the sub-region stripping mechanism is triggered.

[0039] The sub-region stripping process adopts a spatial clustering algorithm. The system divides the optimized adjustment region into a plurality of analysis units, and calculates the independent index value of each unit. Based on the spatial proximity and index similarity between units, clustering is performed, and units with significantly high index abnormal values and geographical continuity are aggregated into a review sub-region.

[0040] Table 1: Index calculation results of key sub-regions in a calibration process.

[0041] Subregion number Center coordinate (pm) Texture complexity Concentration coefficient of variation Out-of-limit status determination SZ-04 (142,92) 5.8 0.28 Texture out-of-limit SZ-11 (157,118) 4.1 0.41 Concentration out-of-limit SZ-19 (169,105) 6.2 0.39 Double-index out-of-limit SZ-23 (132,97) 3.9 0.22 Normal Referring to Table 1, the system automatically marks SZ-04, SZ-11, and SZ-19 as high-precision review sub-regions. The atomic force microscope then performs fine scanning on these sub-regions, increasing the probe sampling frequency from 2 Hz in the standard mode to 8 Hz, and increasing the scanning density by four times. In the high-precision scanning of the SZ-19 region, the system obtains detailed molecular conformation feature data, including a histogram of molecular principal axis orientation angle distribution, a near-neighbor intermolecular distance statistical spectrum, and a surface adhesion force mapping. Feature analysis shows that there are two abnormal conformation patterns in this region: local molecular arrangement presents a directional tilt of 30°-50° (normal should be 0°-15°), and the intermolecular distance appears a bimodal distribution (normal is a unimodal Gaussian distribution).

[0042] Based on the molecular conformation features of the high-precision review sub-region, the system dynamically updates the probe control parameters of the atomic force microscope. For the molecular arrangement tilt region, the system re-plans the scanning path: in the block with a tilt angle greater than 40°, the scanning step is reduced from 100 nm to 30 nm; in the interface region with abnormal intermolecular distance, the scanning speed is reduced from 20 μm / s to 5 μm / s; for the point with high adhesion force detected, the peak force tapping mode is enabled instead of the contact mode. After parameter updating, the system performs a verification scan in the northwest quadrant of the SZ-19 region, and successfully captures the previously unrecognized molecular dislocation chain structure under the new parameter configuration, confirming the effectiveness of the parameter adjustment. At the same time, the system records the parameter mapping rule generated by this calibration, forms a new experience template, and stores it in the database for optimizing the initial detection parameter preset of the same type of sample. The entire calibration process is completed within 120 seconds after triggering, and then the system continues to perform the detection task of the remaining region under the updated parameter configuration.

[0043] Example 5: Establishing a real-time feedback link between Raman spectral scanning data and atomic force microscopy data is the basic framework for dynamic optimization. The link uses a high-speed data bus to connect the spectrometer control unit and the microscope control unit, and opens a shared data buffer in memory. When the atomic force microscope completes high-precision recheck scanning of a sub-region, the molecular conformation feature data it obtains is written in real time to a specific sector of the buffer. Simultaneously, the spectral detection unit stores the Raman spectral scanning results of the same sub-region in another sector of the buffer. The system sets a time stamp matching mechanism to ensure that two sets of data with completely corresponding spatial coordinates are retrieved and analyzed in parallel. After the data is aligned, the system performs pixel-level comparison: the initial spectral analysis results of each coordinate point in the recheck sub-region are compared with the high-precision microscopy recheck results for multi-dimensional feature vector matching. The feature vector includes active ingredient spatial coordinates, relative concentration estimates, molecular orientation angles, and other parameters. When the recheck result vector of a specific coordinate point and the initial detection result vector have a Euclidean distance that exceeds a pre-set tolerance, the point is marked as a deviation point. The system calculates the spatial distribution density and average deviation amount of deviation points in the entire sub-region.

[0044] Based on the deviation analysis results, the system calculates the ingredient distribution correction coefficient. This coefficient is essentially a spatial distribution function, generated by superimposing the deviation point density distribution map and the deviation amount heat map. The calculation process first constructs a deviation density grid: the recheck sub-region is divided into microgrids, and the number of deviation points in each grid is counted and normalized. At the same time, a deviation amount gradient field is constructed: a scalar field is generated according to the average deviation amount of each point in the grid. The final correction coefficient is obtained by convolving the deviation density grid and the deviation amount scalar field, forming a continuous coefficient distribution surface covering the entire recheck sub-region. Areas with a coefficient value greater than 1 indicate that the initial detection significantly underestimated the actual value at that location, and areas with a coefficient value less than 1 indicate that the initial detection overestimated the actual value. The coefficient distribution surface is stored as a floating-point matrix, with a spatial resolution consistent with the initial spectral scanning grid.

[0045] The application of the ingredient distribution correction coefficient is reflected in adjusting the weight matrix for subsequent Raman spectral feature extraction. The system maintains a spectral feature weight matrix, which defines the relative importance of different spatial locations and different spectral bands in the feature extraction process. When a new correction coefficient is obtained, the system performs Hadamard product operation on the original weight matrix and the correction coefficient matrix. For areas with a correction coefficient greater than 1, the spectral feature weight of the corresponding spatial location is increased; otherwise, it is decreased. For example, a certain Baishao compound obtains a correction coefficient of 1.25 in grid G7, so the elements of the weight matrix corresponding to this area are all multiplied by 1.25; while the coefficient of 0.8 in grid D3 causes the corresponding elements of the weight matrix to be multiplied by 0.8. The updated weight matrix is immediately applied to the spectral analysis process of the subsequent undetected areas of the sample, making the feature extraction algorithm focus more on abnormal areas with high correction coefficients.

[0046] The system continuously compares the real-time detection results with the difference factors of the standard ingredient spectrum library. The standard ingredient spectrum library stores the three-dimensional reference model of the compound preparation in the ideal state: including the benchmark spatial distribution pattern, the standard concentration gradient curve, and the typical molecular bonding fingerprint. The comparison operation is carried out in three dimensions: the similarity of the active ingredient distribution pattern and the reference model is compared in the spatial dimension; the relative content deviation index is calculated in the concentration dimension; and the statistical difference of the characteristic peak shift is analyzed in the bonding dimension. The difference factor classification engine automatically classifies according to the main deviation characteristics: the circular abnormality in the spatial dimension distribution pattern is classified as uneven distribution in the Angelica-Glycyrrhiza compound; the unregistered characteristic peak in the Astragalus sample is classified as ingredient category deviation; and the hydrogen bond peak shift exceeding the limit in the Salvia compound is classified as bonding state deviation.

[0047] According to the category of the difference factors, the system updates the abnormal pattern recognition rule library in the multi-layer experience pool. The multi-layer experience pool adopts a tree index structure, the top layer is classified by difference categories, the middle layer stores specific abnormal pattern templates, and the bottom layer records historical treatment parameters. The update operation contains three levels: for the newly identified abnormal pattern category, create a new branch in the top layer directory; for the new variant pattern in the known category, add the feature vector template to the middle layer template library; for the specific abnormality that repeatedly occurs, optimize the treatment parameter threshold in the bottom layer. The newly discovered branch distribution abnormality in the Angelica compound detection is stored as the 12th template in the middle layer template library, and its feature vector includes gradient change rate threshold, spatial autocorrelation parameter, and other core attributes. The template is associated with the uneven distribution category branch.

[0048] The updated abnormal pattern recognition rule library is called in the initial stage of the next detection. When the new sample enters the system, the initial Raman spectrum scanning is performed to generate the initial spectrum distribution information, and the updated rule library is preloaded to the working memory of the partial least squares algorithm. In the dimensionality reduction fitting process, the algorithm not only extracts principal components according to the conventional spectrum characteristics, but also performs similarity matching between the real-time acquired spectrum data and the abnormal templates in the rule library. If the spectrum characteristics of a certain spatial region match the high-risk template in the rule library with a matching degree exceeding the activation threshold, the principal component weight of this region will be automatically enlarged, significantly improving its sensitivity in the local component abnormality risk area recognition process. The pre-application of the rule library enables the system to implement intensive monitoring of historical high-frequency abnormal patterns in the early detection stage.

[0049] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0050] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for rapid detection of ingredients in a traditional Chinese medicine compound, characterized in that: include: Performing Raman spectroscopy scanning on a traditional Chinese medicine compound sample to generate component spectral distribution information, and using a partial least squares algorithm to fit the component spectral distribution information to determine the local component abnormality risk area; Perform atomic force microscopy imaging on traditional Chinese medicine compound samples to generate microscopic morphology distribution maps, and mark potential component mismatch feature areas based on texture feature extraction and pattern recognition; generating a detection optimization adjustment region based on the spatial coordinates of the local component abnormality risk region and the potential component mismatch feature region; By analyzing the gradient change of the molecular arrangement direction in the detection optimization adjustment area, evaluating whether the component distribution offset state is normal; By analyzing the interface binding state between the active ingredient and the excipient in the detection optimization adjustment area, the rationality of the interaction coupling strength of the ingredients is evaluated; Based on the evaluation results of the component distribution offset state and the component interaction coupling strength, a dual-path decision mechanism is used to determine whether to start real-time parameter calibration in the detection optimization adjustment area.

2. The method for rapid detection of Chinese herbal compound ingredients according to claim 1, characterized in that: The Raman spectroscopy scanning of the traditional Chinese medicine compound sample to generate component spectral distribution information includes: Acquire characteristic spectral band data containing target active ingredients through Raman spectrometer; Extracting peak intensity and half-peak width parameters from the characteristic spectral band data to generate component spectral distribution information; The partial least squares algorithm is used to perform dimension reduction fitting on the component spectral distribution information, and the local component abnormality risk area is marked according to the principal component weight coefficient.

3. The method for rapid detection of Chinese herbal compound ingredients according to claim 2, wherein: The atomic force microscopy imaging of the traditional Chinese medicine compound sample to generate a microscopic morphology distribution map comprises: The surface topology data of the sample is obtained by atomic force microscopy, and the microstructure contrast is enhanced by morphological filtering; A multi-scale texture segmentation algorithm is used to extract molecular aggregation features in the topological data; A random forest classifier is used to identify the pattern categories of the molecular aggregation features, and potential component mismatch feature areas are marked based on the classification results.

4. The method for rapid detection of Chinese herbal compound ingredients according to claim 3, characterized in that: The generating detection optimization adjustment area includes: extracting a coordinate set of the local component abnormality risk region and a coordinate set of the potential component mismatch feature region; The spatial union of the coordinate sets is calculated to generate a minimum circumscribed rectangular area covering the maximum anomaly range as the detection optimization adjustment area.

5. The method for rapid detection of Chinese herbal compound ingredients according to claim 4, characterized in that: The evaluating whether the component distribution deviation state is normal includes: Obtaining a molecular orientation angle distribution map within the detection optimization adjustment area; Calculating the gradient change rate of the molecular orientation angle distribution diagram in the spatial dimension; According to the standard deviation and mean of the gradient change rate, it is determined whether the component distribution offset state exceeds a preset tolerance threshold.

6. The method for rapid detection of Chinese herbal compound ingredients according to claim 5, characterized in that: The evaluation of whether the interaction coupling strength of the components is reasonable includes: Analyzing molecular bonding spectral data of the active ingredient and the excipient within the detection optimization adjustment region; Construct the interface binding eigentensor and perform non-negative matrix factorization to obtain the component interaction coupling coefficient distribution; According to the positive and negative polarity difference values ​​of the coupling coefficient distribution, it is determined whether the component interaction coupling strength is within a preset reasonable range.

7. The method for rapid detection of Chinese herbal compound ingredients according to claim 6, characterized in that: The dual-path decision-making mechanism includes: When the component distribution offset state is normal and the component interaction coupling strength is reasonable, performing a standard detection protocol; Otherwise, the real-time parameter calibration process is triggered to adjust the spectral scanning resolution and microscopic imaging sampling frequency.

8. The method for rapid detection of Chinese herbal compound ingredients according to claim 7, characterized in that: The real-time parameter calibration process includes: Calculating the texture complexity index and component concentration variation coefficient of the detection optimization adjustment area; When the texture complexity index or the component concentration variation coefficient exceeds a dynamic threshold, a high-precision verification sub-region is stripped from the current detection region; The probe scanning path parameters of the atomic force microscope are updated according to the molecular conformation characteristics of the high-precision complex nucleus region.

9. The method for rapid detection of Chinese herbal compound ingredients according to claim 8, characterized in that: Also includes: Establish a real-time feedback link between spectral data and microscopic imaging data; When there is a deviation between the recheck result of the high-precision recheck sub-region and the initial detection result, generating a component distribution correction coefficient; The feature extraction weight matrix of the Raman spectrum is adjusted based on the component distribution correction coefficient.

10. The method for rapid detection of components of a traditional Chinese medicine compound according to claim 9, characterized in that: Also includes: Compare the difference factors between real-time detection results and standard component library; According to the category attribution of the difference factors, updating the abnormal pattern recognition rule base in the multi-layer experience pool; The updated abnormal pattern recognition rule base is applied to the initial spectrum fitting process of the next detection.

Images