Raman spectrum reverse analysis method for tablets
By employing Raman spectroscopy mapping techniques, noise reduction and differentiation algorithms, singular value decomposition, and oblique projection algorithms, the analytical challenges of complex mixed systems of tablets were solved, enabling qualitative and quantitative analysis and improving the signal-to-noise ratio and analytical efficiency.
Patent Information
- Application Number
- CN202410813918.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-21
- Publication Date
- 2024-11-08
AI Technical Summary
Existing Raman spectroscopy techniques are difficult to effectively resolve complex mixtures of tablets, and cannot provide detailed molecular structure information and quantitative capabilities. Furthermore, conventional methods require combination with other analytical techniques or increase preliminary research costs.
Raman spectral data were acquired using mapping technology, preprocessed using denoising and differentiation algorithms, and eigenvalue decomposition was performed by utilizing singular value decomposition and higher-order derivative space mapping, taking advantage of the discontinuity between the component space and the noise space. Qualitative and quantitative analysis was then performed using oblique projection algorithm.
It improves the signal-to-noise ratio, reduces nonlinear response, enables qualitative and quantitative analysis of tablets in a single method, reduces dependence on other analytical methods, and improves analytical efficiency and accuracy.
Smart Images

Figure CN118914156A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of Raman spectrum reverse analysis, and in particular to a Raman spectrum reverse analysis method for tablets. Background Art
[0002] Tablet reverse engineering is a systematic project that requires overcoming many difficulties to succeed, and requires sufficient time and resources to tackle the problems. Technically, it is difficult to fully analyze tablets with a single method. It is necessary to use a variety of analytical techniques to characterize tablets layer by layer to obtain comprehensive quality attribute data. Among them, Raman, infrared, nuclear magnetic resonance, mass spectrometry, chromatography, etc. are commonly used methods, and the combination of multiple technologies is also necessary; in addition, it is necessary to use information technology to achieve the correlation and integration of various results, analysis and mining, in order to make an overall expression of the tablet.
[0003] Although reverse engineering must rely on the coordination of various technologies to be completed, in research work, Raman spectroscopy, as a "pioneering method", is often used first and plays a role, providing basic support and important clues for subsequent work, playing a role in paving the way and exploring information, and providing key advance support. It is the "pioneer" and "pathfinder" of the entire research process.
[0004] In most cases, due to the complexity of Raman spectroscopy and samples, the collected Raman signals often cannot directly give the conclusions that users want. This is also a challenge in completing reverse engineering.
[0005] Although Raman spectroscopy can provide rich information, the currently commonly used peak position and peak intensity combination judgment, combined with fingerprint spectrum, provides support for single components or simple systems. However, compared with NMR, MS and other methods, it cannot provide more detailed structural information to judge unknown or uncertain molecular structure information, and the analysis of complex mixed responses is uncertain, and it is often helpless. Similarly, its limited quantitative ability will also encounter great challenges. Summary of the invention
[0006] In order to solve the problems existing in the above-mentioned prior art, the object of the present invention is to provide a Raman spectrum reverse analysis method for tablets.
[0007] To achieve the above object, the technical solution of the present invention is:
[0008] A reverse analysis method for Raman spectra of tablets is specifically as follows: first, mapping technology is used to collect Raman spectra data of tablets, and data redundancy is used to reduce the nonlinearity of response; then, denoising and derivative algorithms are used to preprocess the spectral data to overcome low-frequency baseline lift, improve the signal-to-noise ratio, and obtain a mixed spectrum that is as close to an ideal linear model as possible; then, the discontinuity between component space and noise space is used to perform singular value decomposition on the signal to obtain an eigenvalue sequence arranged in the direction of the maximum spatial projection, and the number of components is determined by analyzing the continuity of the eigenvalue derivatives; further, the original spectrum is mapped to a high-order derivative space to obtain a separation matrix that is closer to the true composition of the source signal, and then the spectra of each component in the mixed system are separated, and the composition consistency is judged using the subspace vector angle evaluation, thereby completing qualitative analysis; finally, an "oblique projection" algorithm is applied to estimate the proportion of each component in the sample to achieve quantitative analysis.
[0009] The method consists of the following steps:
[0010] S1 Qualitative judgment
[0011] S1.1 Sample preparation and spectral signal acquisition
[0012] Take the tablet to be analyzed, grind it, and collect the Raman scanning spectrum S of more than 100 points on the tablet surface by Mapping mode;
[0013] S1.2 Spectral signal preprocessing
[0014] Accumulate the signal of each point of the collected sample spectrum S to obtain SM, then perform noise reduction and differentiation to obtain the preprocessed spectrum h; the derivative order is recommended to be 1-2 order;
[0015] The noise reduction and derivation processing uses a fully adaptive smoother algorithm, that is, SM is imported into the computing platform, and the processing results are output through the command [hcfcf1]=Sz1(SM,DA,N), where: SM is the measurement signal; DA is the derivation order; N is the number of smoothing, taking the default value 6); h is the output result, cf and cf1 are frequency response functions;
[0016] S1.3 Determine the number of independent components
[0017] Using the discontinuous properties of the component space and the noise space, the entire space is decomposed into singular values to obtain the eigenvalues arranged in the direction of the maximum spatial projection; each eigenvalue is multiplied by its corresponding number of dimensions, and each dimensional response is weighted; then the derivative is used to determine the continuity of the sequence, find the turning point of the two subspaces, and determine the number of independent components n of h;
[0018] S1.4 Blind source identification of tablets using ICA method based on spatial mapping
[0019] First, the high-order derivative spectrum of SM is obtained, and ICA processing is performed under the principal component number n to obtain the separation matrix. Then, the source signal of each component is output in the first-order derivative space, and the signal of each component is normalized to obtain Fi (i = 1, 2, ... n);
[0020] S1.5 System actual composition confirmation
[0021] According to the category of the tablet to be analyzed, a series of possible pure substance samples are collected, and the Raman spectrum of each pure sample is collected according to S1.1. Noise reduction-derivation is performed according to the signal processing conditions of step S1.2, and normalization is performed to obtain the unit response spectrum Ci (i=1,2,…n) of each pure component;
[0022] The angles between h and Ci are calculated in sequence. When the angle θ is a small value (θ<0.015), it can be determined that the tablet contains Ci.
[0023] S2 Quantitative Analysis
[0024] S2.1 Separation steps
[0025] 5) Arrange Fi (i = 1, 2, ... n) in columns, take the response spectrum of one single component such as F1 as vector x, and the column vectors of other components in parallel to form the background matrix M;
[0026] 6) Obtain the oblique projection operator of component F1 in the overall response according to the orthogonal projection, orthogonal projection residual and oblique projection formulas;
[0027] P=M(M T M) -1 M T Orthographic projection
[0028] P |M =IP F1 residual after orthogonal projection (net signal)
[0029] P H|x =x(x T P |M x) -1 x T P |M Oblique projection operator
[0030] 7) Using the oblique projection operator h |x =P H|x h separates the contribution of only F1 component in h;
[0031] 8) Similarly, take the response spectrum of any other component as x, and the column vectors of the remaining components are arranged in parallel to form the background matrix M, and calculate the contribution of each component Fi (i=2,…n);
[0032] S2.2 Quantification steps
[0033] 4) The component raw materials determined in step S1.5 are pressed into standard sheets according to step S1.1, and the Raman spectra of each standard sheet are obtained by scanning;
[0034] 5) According to the data processing parameters of step S1.2, each standard film is subjected to noise reduction-derivation processing, and the derivative spectrum of each standard film is obtained as the unit spectrum, and the modulus of the unit spectrum is obtained;
[0035] 6) Calculate the modulus of the contribution spectrum of each separated component, and divide it by the unit modulus of the corresponding standard film to obtain the ratio of each component;
[0036] Normalize the ratio to get the ratio of each component in the system.
[0037] S1.1 In sample preparation and spectral signal acquisition, the specific parameters of tablet grinding are: under a pressure of 12000N-15000N, round tablets with a diameter of 8-13mm and a thickness of 1-3mm are obtained.
[0038] In the step S1.2 of spectral signal preprocessing, the computing platform may be Matlab 2018 or above.
[0039] Compared with the prior art, the Raman spectrum reverse analysis method of a tablet of the present invention has the following beneficial effects:
[0040] Chemometric analysis often hopes to obtain a linear model of an ideal mixed system. However, for the actual measured spectrum of tablets, there are usually many interferences: such as nonlinear decrease caused by the response of increased light intensity, multiplicative random errors caused by changes in scattering intensity due to excitation factors such as particle size and optical path changes, low-frequency responses such as fluorescence and other associated signals that cause baseline elevation, and additive random errors caused by white noise of photoelectric conversion or other uncertain factors. The existence of these interferences significantly increases the difficulty of Raman spectroscopy application, and is also the goal currently being worked on in the fields of analysis and signal processing.
[0041] The present invention collects data through mapping, and the number of points collected is much larger than the number of formula components, which can easily meet the necessary condition of the ideal model, that is, the "full rank" requirement, and can provide a large amount of data redundancy for noise reduction and correction. A large amount of data redundancy can be used to improve the signal-to-noise ratio and reduce the nonlinearity of the response.
[0042] The baseline lift and poor signal-to-noise ratio of Raman spectroscopy are important reasons that affect its effective use. Sharp peaks (high frequency) in Raman spectra are more concerned, and low-frequency baselines need to be ignored. In addition to the subtraction approach, data derivation is another feasible approach. Its advantage is that it does not use subjective additions and subtractions, and will not introduce unexplained errors, but its disadvantage is that high-frequency random interference increases. The present invention uses an algorithm with dual functions of noise reduction and derivation, such as the SH or ZX method. The high-order derivative algorithm with noise adaptability has higher accuracy and better performance, and the output signal-to-noise ratio is also significantly improved.
[0043] If the system measurement signal is completely ideal, that is, it does not contain noise and nonlinear factors, and there is no basic additive background, it is not challenging to judge the number of system components. In fact, these factors cannot be completely eliminated, which constitutes a challenging task. The eigenvalue cumulative contribution method and the interactive verification method used in multivariate statistics and chemometrics are the two most commonly used methods in textbooks. The cumulative contribution method lacks interpretability, and CV needs to determine the corresponding component content value. From the perspective of subspace, the present invention regards noise as a large space with the dimension of the entire sample set, while the determined components in the system are determined spaces with limited dimensions. The entire measured space is the sum of two spaces. Due to different degrees of certainty, the two spaces are discontinuous, and there is a node that can be clearly demarcated. Using the discontinuous properties of the component space and the noise space, the entire space is decomposed into singular values to obtain eigenvalues that are equivalent to being arranged in sequence according to the maximum projection direction of the space. Each eigenvalue is multiplied by its corresponding number of dimensions in turn, each dimensional response is weighted, and then the derivative is used to judge the continuity of the sequence, so that it is relatively easy to find the position where the two subspaces turn.
[0044] The prior knowledge of Raman spectra used for reverse analysis is very limited, and the usual approach is to cooperate with other analytical methods, such as instrument coupling conditions, to meet the suitability. If you do not want to give up the convenience of Raman measurement and do not intend to increase the cost of pre-research, it is difficult to meet the suitability requirements by only understanding the profile of specific varieties through Raman Mapping data, and a new approach needs to be selected. The present invention maps the original spectrum to the high-order derivative space, and uses the advantage of better selectivity of derivative spectra to greatly improve the separation results.
[0045] Spectral pattern recognition method is the most widely used pattern recognition method for judging material properties, and commonly used methods include peak comparison, spatial distance, angle cosine and correlation coefficient. Peaks in mixed systems may experience peak shift, peak shape change and merging, which limits the judgment based on peak comparison; for mixed systems, changes in composition content will cause large changes in peak shape, making "similarity" judgment based on spatial distance, angle cosine and correlation coefficient unfeasible. The present invention evaluates whether two subspaces overlap or are close by the subspace (vector) angle, and does not use the usual characteristic peak parameters for qualitative judgment. It can meet the situation of atypical peaks (overlap or overlap), and is suitable for qualitative analysis of substances in multi-component systems.
[0046] The most direct association for multi-component quantitative problems is PLS (partial least squares) or "multivariate statistical modeling", but statistical modeling requires quantitative batch samples in advance, and its cost is unacceptable for reverse engineering. The present invention proposes to divide the system space into two parts: "focus vector" and "remaining subspace". Through oblique projection calculation, there is no need for statistical modeling based on the content of the sample being measured. Through the spectrum of each component in the multi-component sample, the single complete contribution of the measured component in the mixed spectrum can be reversed. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] Figure 1 Scanning spectrum of tablet surface;
[0048] Figure 2 Spectral signal preprocessing results;
[0049] Figure 3 The result of judging the number of independent components based on the subspace-eigenvalue derivative;
[0050] Figure 4 Comparison of the separation results of the ICA method based on spatial mapping and the normalized first-order Raman spectra of each single component measured
[0051] Figure 5 Raman spectra of pure components in the sample library;
[0052] Figure 6 The contribution of each component is calculated using the oblique projection algorithm. DETAILED DESCRIPTION
[0053] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods: Figure 1-6 As shown:
[0054] Test Example 1: Principle and Idea
[0055] 1. Subspace description of the measurement signal
[0056] In order to obtain a more essential description of the problem, the measured signal is abstracted into the data space.
[0057] The spectrum is measured by discretizing the continuous spectrum into m wavelengths (wave numbers), and then discretely measuring the intensity (I) of each wavelength point. This spectrum can be recorded as:
[0058] Spectrum Single =[I1,I2,I3,…,I m ]
[0059] If the spectrum is contributed by n components, each component has its own content C in the whole composition. i , the composite spectrum actually measured once is composed of the spectra of each component weighted by content:
[0060] Spectrum Complex
[0061] =C1*[I 11 ,…,I 1j ,…,I 1m ]+…+C i *[I i1 ,…,I ij ,…,I im ]+…+C n *[I n1 ,…,I nj ,…,I nm ]
[0062] The pth measurement of different contents of the complex:
[0063] Spectrum Comp,p
[0064] =C 1p *[I 11 ,…,I 1j ,…,I 1m ]+…+C ip *[I i1 ,…,I ij ,…,I im ]+…+C np *[I n1 ,…,I nj ,…,I nm ]
[0065] Then, the spectrum arrays measured multiple times are collected to form an array
[0066] S=[Spectrum comp,1 ,…,Spectrum comp,p ]
[0067] According to the matrix writing:
[0068] S=C*S Single
[0069] This is the linear model of the ideal mixed system most commonly used in chemometrics.
[0070] For the actual measured spectra of tablets, this ideal model should also add several factors to express the interferences.
[0071] S=C*(S Single ) α *(1+δ)+f+ε
[0072] in,
[0073] α is the degree of nonlinear response, the ideal α=1, usually the response decreases nonlinearly with increasing light intensity, that is, α<1;
[0074] (1+δ) is the change in scattering intensity caused by excitation factors such as particle size and optical path length change, which is usually called multiplicative random error;
[0075] f is the low-frequency response caused by the accompanying signals such as fluorescence, which leads to the baseline elevation;
[0076] ε is the additive random error caused by white noise or other uncertain factors from photoelectric conversion.
[0077] Both f and ε are additive. f is not random and mainly affects low frequencies, while ε is random and has a more obvious impact on high frequencies.
[0078] The existence of these interferences significantly increases the difficulty of Raman spectroscopy applications, and is also the goal that the current analysis and signal processing fields are working towards. If a mixed spectrum that is as close to the ideal linear model as possible is measured or obtained through preprocessing, it will greatly facilitate subsequent work.
[0079] 2 Raman spectrum acquisition and signal preprocessing
[0080] The necessary condition for an ideal model is "full rank", that is, the number of spectra measured should be equal to or greater than the number of components contained in the system, and the content of each component in each sample point measured cannot be synchronized. Since most solid tablets are uneven, the number of points collected by mapping data is much larger than the number of formula components, which can easily meet this requirement and provide a large amount of data redundancy for noise reduction and correction.
[0081] 2.1 Using data redundancy to improve signal-to-noise ratio and reduce response nonlinearity
[0082] The number of scattered spectra collected by mapping scanning far exceeds the actual number of components in the system. For example, 1,600 spectra can be obtained according to the 40×40 acquisition density. A large amount of data redundancy can be used to improve the signal-to-noise ratio and reduce the nonlinearity of the response.
[0083] The measured signals are grouped into several groups according to their similarity, which are larger than the estimated number of components in the system. For example, if the estimated number of components in the system is less than 10, the 1600 measured spectra can be grouped into 16 groups, and each group can be added as a cumulative value, which can reduce the subsequent calculation complexity and improve the signal-to-noise ratio by 10 times. For general spectrometers, when the response value is less than half of the full scale, it usually does not show obvious nonlinearity. The light response amplitude can be reduced by using a shorter integration time to avoid the influence of nonlinear factors.
[0084] 2.2 Use noise reduction and derivation algorithms to overcome low-frequency baseline lift and further improve the signal-to-noise ratio
[0085] The baseline rise and poor signal-to-noise ratio of Raman spectroscopy are important factors affecting its effective use.
[0086] The sharp peaks (high frequency) in the Raman spectrum are more concerned, and the low frequency baseline needs to be ignored. In addition to the subtraction method, data derivation is another feasible method. Its advantage is that it does not use subjective additions and subtractions, and will not introduce unexplained errors, but its disadvantage is that the random interference of high frequencies increases, and it is necessary to use an algorithm with dual functions of noise reduction and derivation to complete it.
[0087] Noise-adaptive high-order derivative algorithms are very important for spectral signal processing. The inventors proposed higher-precision SH and ZX methods with better performance and significantly improved output signal-to-noise ratio.
[0088] 3. Determine the number of system components
[0089] If the system measurement signal is completely ideal, that is, without noise and nonlinear factors, and without basic additive background, it is not challenging to determine the number of system components. In fact, these factors cannot be completely eliminated, which constitutes a challenging task.
[0090] The eigenvalue cumulative contribution method and the cross-validation method used in multivariate statistics and chemometrics are the two most commonly used methods in textbooks. The cumulative contribution method lacks interpretability, and CV requires the determination of the corresponding component content value.
[0091] From the perspective of subspace, noise can be regarded as a large space with the dimension of the entire sample set, while the determined components in the system are determined spaces with limited dimensions. The entire measured space is the sum of the two spaces. Due to the different degrees of certainty, the two spaces are discontinuous and there is a node that can be clearly demarcated.
[0092] By using the discontinuous properties of the component space and the noise space, the entire space is decomposed into singular values, which is equivalent to the eigenvalues arranged in the direction of the maximum projection of the space. Each eigenvalue is multiplied by its corresponding number of dimensions, each dimension response is weighted, and then the continuity of the sequence is judged by the derivative, so that the turning point of the two subspaces can be found relatively easily.
[0093] 4 How to separate corresponding components from a certain space
[0094] The prior knowledge of Raman spectra used for reverse analysis is very limited, and usually other analytical methods, such as instrument coupling conditions, are used to meet the requirements of suitability. If you do not want to give up the convenience of Raman measurement and do not intend to increase the cost of pre-research, it is difficult to meet the requirements of suitability by only understanding the general appearance of specific varieties through Raman Mapping data, and a new approach needs to be selected.
[0095] Blind source separation transforms the input mixing matrix into an independent signal source through the independence criteria such as mutual information entropy or kurtosis of the signal source, so as to separate the separation matrix required for the mixed signal.
[0096] According to the ideal linear mixed model mentioned above:
[0097] S=C*S Single
[0098] If the spectra of the substances contained in the system meet statistical independence, the independent spectra can be obtained by using blind source separation as the inverse process.
[0099] S single =WS
[0100] Because the Raman spectrum contains relatively rich information, direct blind separation can often achieve a rough match with the true value. However, if the actual signal source deviates significantly from the premise of statistical independence, it is often difficult to further improve the consistency between the output and the true spectrum, resulting in low practicality.
[0101] To address the problem of insufficient well-posedness in blind separation of molecular spectra, the inventors proposed a method of mapping the original spectrum into a higher-order derivative space, taking advantage of the better selectivity of derivative spectra to significantly improve the separation results.
[0102] The principle is that differential calculation will not cause changes in the spatial addition properties, and the separation matrix that is closer to the actual composition of the source signal can be obtained in the derivative space.
[0103]
[0104] By using the improved separation matrix to separate the original spectrum, a more realistic separation effect can be obtained. The separated spectrum can correspond to most of the actual ingredient information. The types of components that may be covered by the tablet formula are limited. By collecting possible raw materials or combining relevant databases and following the map, we can basically determine which components the system specifically or may contain.
[0105] 5. Composition consistency judgment
[0106] Based on the isolated source clues, we can determine the number of components in the formula and their possible identities. However, the conclusion needs a more reliable basis for judgment.
[0107] The system is composed of a certain signal space formed by the real response of each component. If the mixed spectrum is not the sum of these components existing in the system, it must not overlap with this space. Whether two subspaces overlap or are close can be evaluated by the subspace (vector) angle. The so-called space angle refers to the angle between the closest vectors in the space after the subspace unit modulus is quantized.
[0108] Space angles are between 0 and π / 2 (90°), with the two boundaries representing ideal situations. A value of zero means that the spaces are completely overlapping, or a vector (or low-dimensional space) exists completely within another large space; a value of 90° means that they are completely orthogonal, and one object does not exist at all in the other.
[0109] The actual situation will be affected by noise and nonlinearity. Since the angle value is normalized by intensity, its relationship with noise is approximately equal to the value of the signal-to-noise ratio. For example, when the SNR is 100:1, when the spectra of the same sample are measured at different times, even if the intensity of the spectrum is different, the angle value is about 0.01. If the angle value of the two samples is significantly larger, it can be basically determined that the spectra do not correspond to the same sample.
[0110] If all raw materials are selected, the spectra are measured separately to form the "raw material matrix", and the spatial angle between the "original sample matrix" and the mapping data of the sample to be tested is calculated. Considering the spectral signal-to-noise ratio and distortion, the angle threshold is obtained by statistical estimation, which can be used to estimate and judge the composition of the tablets.
[0111] 6. Estimating the proportion of sample composition
[0112] The most direct association for multi-component quantitative problems is PLS (partial least squares) or "multivariate statistical modeling", but statistical modeling requires quantification of batch samples in advance, and its cost is unacceptable for reverse engineering.
[0113] Under the premise of knowing the composition of the system components, there is another feasible quantitative approach, namely the spatial component projection approach. The system space is divided into two parts: the "focus vector" and the "rest of the subspace". If the mixed spectrum is projected into the "rest of the space", then the difference between the mixed spectrum and the projection value will not contain information about the rest of the components. If the projection operation is orthogonal, this difference will appear as an output value corresponding to the real component of interest, but with a reduced amplitude, which is the so-called "net signal" of the measured component.
[0114] The fundamental reason for the deviation in “net signal” processing is “orthogonal projection”. If “oblique projection” operation is adopted, the calculation output in principle is the actual existence form of the measured component in space.
[0115] The space H contains the x component of interest and the subspace M where other components exist, written as a matrix:
[0116] H=[x,M]
[0117] That is, space H is the union of two non-intersecting proper subsets (vector x and subspace M). If a component h in H has contributions from both x and M, and h is projected onto M using the orthogonal projection operator P,
[0118] P=M(M T M) -1 M T
[0119]
[0120] The remaining part of h outside the space M
[0121]
[0122] The signal is derived solely from x, with no residual contribution from M. In chemometric theory, this is called the "net signal". corresponds only to x, but if x and M are not orthogonal, then the orthogonal projection from M It does not coincide with x either, so we need to use the oblique projection operator P H|x Further Mapped to coincide with x. The expression of the oblique projection operator is:
[0123] P H|x =x(x T P |M x) -1 x T P |M
[0124] Use P H|x Separate h |x:
[0125] h |x =P H|x h
[0126] That is, if h is contributed by x and M, h = αx + BM, then,
[0127] h |x =αx
[0128] The significance of oblique projection calculation is that it does not require statistical modeling based on the content of the sample being measured. Through the spectrum of each component in the multi-component sample, the single complete contribution of the measured component in the mixed spectrum can be reversed.
[0129] When discussing signal preprocessing in the previous article, we did not mention the two thorny problems encountered in statistical modeling prediction: uneven distribution and multiplicative interference of signal scattering. The oblique projection quantification based on cumulative values will not be affected by these two problems. The overall cumulative contribution of all spectra accumulated by mapping naturally eliminates the unevenness of each sampling; the reverse analysis is concerned with the content ratio of the components, which is the normalized relative value. In each measured spectrum, the multiplicative effect of each component is consistent, and the multiplicative effect will not affect the relative value.
[0130] A Raman spectrum reverse analysis method for tablets, comprising the following steps:
[0131] S1 Qualitative judgment
[0132] S1.1 Sample spectral signal acquisition
[0133] After grinding the tablets to be analyzed, they are pressed under a pressure of 12000-15000N to obtain round tablets with a diameter of 8-13mm and a thickness of 1-3mm. The mapping method collects Raman scanning spectra S of more than 100 points on the tablet surface;
[0134] S1.2 Spectral signal preprocessing
[0135] The collected sample spectrum S is accumulated at each point to obtain SM, and then noise reduction and differentiation are performed to obtain the preprocessed spectrum h.
[0136] The noise reduction and derivation processing uses a fully adaptive smoother algorithm, that is, SM is imported into the computing platform (such as Matlab2018 or above), and the processing results are output through the [h cfcf1] = Sz1 (SM, DA, N) command, where: SM is the measurement signal; DA is the derivative order; N is the number of smoothing, taking the default value 6); h is the output result, cf and cf1 are frequency response functions.
[0137] S1.3 Determine the number of independent components
[0138] By using the discontinuous properties of the component space and the noise space, the entire space is decomposed into singular values, which is equivalent to the eigenvalues arranged in the direction of the maximum projection of the space. Each eigenvalue is multiplied by its corresponding number of dimensions, each dimension response is weighted, and then the derivative is used to determine the continuity of the sequence. It is relatively easy to find the turning point of the two subspaces and determine the number of independent components n of h.
[0139] S1.4 Blind source identification of tablets using ICA method based on spatial mapping
[0140] First, the high-order derivative spectrum of S is obtained, and ICA processing is performed under the principal component number n to obtain the separation matrix. Then, the source signal of each component is output in the first-order derivative space, and the signal of each component is normalized to obtain Fi (i=1,2,…n).
[0141] S1.5 System actual composition confirmation
[0142] According to the category of the tablets to be analyzed, possible series of pure substance samples are collected, and they are pressed into standard tablets according to step S1.1, and the Raman spectrum of each standard tablet is scanned and obtained; noise reduction-derivation is performed according to the signal processing conditions of step S1.2, and normalization is performed to obtain the unit response spectrum Ci (i=1,2,…n) of each pure component.
[0143] The angles between h and Ci are calculated in turn. When the angle θ is a small value (θ<0.015), it can be determined that the tablet contains Ci.
[0144] S2 Quantitative Analysis
[0145] S2.1 Separation steps
[0146] 1) Arrange Fi (i = 1, 2, ... n) in columns, take the response spectrum of one single component such as F1 as vector x, and the column vectors of other components in parallel to form the background matrix M;
[0147] 2) According to the orthogonal projection, orthogonal projection residual and oblique projection formulas, the oblique projection operator of component F1 in the overall response is obtained;
[0148] P=M(M T M) -1 M T Orthographic projection
[0149] P |M =IP F1 residual after orthogonal projection (net signal
[0150] P H|x =x(x T P |M x) -1 x T P|M Oblique projection operator
[0151] 3) Use the oblique projection operator to separate the contribution of h that only belongs to the F1 component.
[0152] 4) Similarly, any other component response spectrum is taken as x in turn, and the remaining component column vectors are arranged in parallel to form a background matrix M, and the contribution of each component Fi (i=2,…n) is calculated.
[0153] S2.2 Quantification steps
[0154] 1) The component raw materials determined in step S1.5 are pressed into standard sheets according to step S1.1, and the Raman spectra of each standard sheet are scanned and obtained;
[0155] 2) According to the data processing parameters of step S1.2, each standard film is subjected to noise reduction-derivation processing, and the derivative spectrum of each standard film is obtained as the unit spectrum, and the modulus of the unit spectrum is obtained;
[0156] 3) Calculate the modulus of the contribution spectrum of each separated component, and divide it by the unit modulus of the corresponding standard film to obtain the ratio of each component;
[0157] 4) Normalize the ratio to obtain the ratio of each component in the system.
[0158] Example
[0159] In order to verify the feasibility of the method of the present invention and determine the accuracy and certainty of the analysis results, this example uses self-compressed simulated ibuprofen sustained-release tablets as the analysis target.
[0160] Y1 Qualitative judgment
[0161] Y1.1 Sample spectral signal acquisition
[0162] Ibuprofen (API), sodium lauryl sulfate, magnesium stearate, potato starch, and povidone K30 were weighed according to the mass ratio of 50:15:5:15:15, respectively. After grinding evenly, they were pressed under a pressure of 15000N to obtain round tablets with a diameter of 13mm and a thickness of 2mm.
[0163] The Raman scanning spectrum S of 40×40 points on the tablet surface was collected by mapping method. Figure 1 (a). Raman spectrum acquisition parameters are 532nm, 350mW, probe focal length 7.5mm, single point integration time 2s). Figure 1 (b) and (c) are one of the spectra and the accumulation of all sampling points, respectively.
[0164] Y1.2 Spectral signal preprocessing
[0165] The collected sample spectrum S is accumulated at each point to obtain SM, and then noise reduction and differentiation are performed to obtain the preprocessed spectrum h.
[0166] The noise reduction and derivation processing uses a fully adaptive smoother algorithm, that is, SM is imported into the computing platform (Matlab2023b is selected in this embodiment), and the processing results are output through the [h cfcf1] = Sz1 (SM, DA, N) command, where: SM is the measurement signal; DA is 1; N is the number of smoothing times, taking the default value 6); h is the output result, and cf and cf1 are frequency response functions.
[0167] Figure 2 (a) is the result of smoothing and accumulation of 16 original spectra. After preprocessing, the signal-to-noise ratio of two randomly selected spectra has been significantly improved, such as Figure 2 (b). However, due to the contribution of particle scattering and fluorescence, the spectral lines after preprocessing have large background differences, which will cause troubles in subsequent steps.
[0168] It is a more reasonable approach to use derivatives to weaken the contribution of low-frequency signals. Figure 2 (c) is the first-order derivative of (b). In the derivatives of the two spectral lines, the influence of baseline drift is eliminated, and the correspondence of each spectral peak is clear at a glance. The derivative spectrum that overcomes the interference of high-frequency noise can present more details, while avoiding the influence of baseline drift, and has more qualitative and quantitative advantages.
[0169] Y1.3 Determine the number of independent components
[0170] Due to the “smoothing-derivation”, a high signal-to-noise ratio and a low background signal are obtained. The difference between the signal contribution subspace and the random background subspace can be clearly seen in the eigenvalue derivative sequence, such as Figure 3 , it can be clearly determined that the signal has a turning point after the fifth variable, that is, the number of independent components of the system is equal to 5.
[0171] Y1.4 Using ICA method based on spatial mapping to complete blind source identification of tablets
[0172] First, the high-order derivative spectrum of S is obtained, and ICA processing is performed under the principal component number n to obtain the separation matrix. Then, the source signal of each component is output in the first-order derivative space, and the signal of each component is normalized to obtain Fi (i = 1, 2, ... n). The separation matrix is obtained through the third-order derivative spectrum, and the first-order derivative spectrum of each component is separated. After the maximum value of the spectrum is normalized, it is compared with the measured first-order Raman spectra of each single group of ibuprofen (API), sodium lauryl sulfate, magnesium stearate, potato starch, and povidone, as shown in Figure 2. Figure 4 As shown in the figure, the blue line is the separation result and the red line is the measured data. From the separation result, we can infer the five components that the sample to be tested may contain.
[0173] The actual composition of the Y1.5 system is confirmed
[0174] According to the category of the tablet to be analyzed, a series of possible pure substance samples are collected (in this embodiment, some components are selected, including ibuprofen (API), sodium lauryl sulfate, magnesium stearate, potato starch, povidone, lactose, sweet potato starch, and corn starch), and tablets are pressed and sampled according to step Y1.1, and Raman spectra of each pure sample are collected. Figure 5 .
[0175] The first-order derivative of the pure spectrum of each component is used for noise reduction, and the unit response spectrum library Ci (i = 1, 2, ... n) of each pure component is obtained by normalization. The angle between h and Ci is calculated in sequence. When the angle θ is a small value (such as θ < 0.015), it can be determined that the tablet contains Ci.
[0176] The angles between the first-order derivative spectrum of the overall data (matrix) obtained by mapping and the spectrum library Ci are:
[0177]
[0178] As can be seen from the table, the angle values between the derivative spectrum of the tablet and the derivative spectra of ibuprofen (API), sodium lauryl sulfate, magnesium stearate, potato starch, and povidone are all smaller values of the same order of magnitude, indicating that the tablet contains five substances: ibuprofen (API), sodium lauryl sulfate, magnesium stearate, potato starch, and povidone.
[0179] The angle between the sample and lactose is 0.1357, which is one order of magnitude higher than other components, and it is obvious that lactose is not added to the sample. The angles between the sample and potato, sweet potato, and corn starch are 0.0113, 0.0221, and 0.0275, respectively, indicating that the added starch is most likely potato starch.
[0180] Y2 Estimate the content of tablet composition
[0181] Y2.1 Separation steps
[0182] 1) Arrange Fi (i = 1, 2, ... n) in columns, the single component response spectrum of ibuprofen is the vector x, and the column vectors of the other four components are arranged in parallel to form a background matrix M;
[0183] 2) According to the orthogonal projection, orthogonal projection residual and oblique projection formulas, the oblique projection operator of component F1 in the overall response is obtained;
[0184] P=M(M T M) -1 M T Orthographic projection
[0185] P|M =IP ibuprofen residue after orthogonal projection (net signal)
[0186] P H|x =x(x T P |M x) -1 x T P |M Oblique projection operator
[0187] Using the oblique projection operator h |x =P H|x h, separate the contribution of only the F1 component in h.
[0188] Similarly, the response spectrum of any other component is taken as x, and the column vectors of the remaining components are arranged in parallel to form the background matrix M, and the contribution of each component Fi (i = 2, ... n) is calculated, such as Figure 6 , (a)-(e) are ibuprofen (API), sodium lauryl sulfate, magnesium stearate, potato starch, and povidone, respectively.
[0189] It should be pointed out that the oblique projection does not need to satisfy orthogonality. Unlike ICA, the contribution of each separated component is determined by the input x, and its amplitude is the absolute proportion of the overall composition, so it has a quantitative relationship with the actual composition.
[0190] Y2.2 Quantification steps
[0191] 1) The component raw materials determined in step Y1.5 are pressed into standard sheets according to step Y1.1, and the Raman spectrum of each standard sheet is scanned and obtained;
[0192] 2) According to the parameters of Y1.2, each standard film is subjected to denoising-derivation processing, and the derivative spectrum of each standard film is obtained as the unit spectrum, and the modulus of the unit spectrum is obtained;
[0193] 3) Calculate the modulus of the contribution spectrum of each separated component, and divide it by the unit modulus of the corresponding standard film, and the ratios of each component are: 802.94, 247.27, 80.86, 271.54, 239.14;
[0194] 4) The ratios were normalized to obtain the percentages of the components in the system: 48.9% (ibuprofen), 15.1% (sodium lauryl sulfate), 4.9% (magnesium stearate), 16.5% (potato starch), and 14.6% (povidone K30), which were consistent with the actual ratios for preparing tablets.
[0195] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that are not conceived through creative work should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope defined in the claims.
Claims
1. A method for reverse analysis of Raman spectrum of tablets, characterized in that: Specifically, mapping technology was used to collect Raman spectral data of tablets, and data redundancy was used to reduce the nonlinearity of the response. Then, noise reduction and derivation algorithms were used to preprocess the spectral data to overcome the low-frequency baseline lift, improve the signal-to-noise ratio, and obtain a mixed spectrum that was as close to the ideal linear model as possible. Subsequently, the signal is subjected to singular value decomposition by utilizing the discontinuity between the component space and the noise space, and a sequence of eigenvalues arranged in the direction of the maximum spatial projection is obtained. The number of components is determined by analyzing the continuity of the eigenvalue derivatives. Furthermore, the original spectrum is mapped to the high-order derivative space to obtain a separation matrix that is closer to the true composition of the source signal, and then the spectra of each component in the mixed system are separated. The composition consistency is judged using the subspace vector angle evaluation, thereby completing the qualitative analysis. Finally, the "oblique projection" algorithm is applied to estimate the proportion of each component in the sample to achieve quantitative analysis.
2. The method according to claim 1, characterized in that The method consists of the following steps: S1 Qualitative judgment S1.1 Sample preparation and spectral signal acquisition Take the tablet to be analyzed, grind it, and collect the Raman scanning spectrum S of more than 100 points on the tablet surface by Mapping mode; S1.2 Spectral signal preprocessing Accumulate the signal of each point of the collected sample spectrum S to obtain SM, then perform noise reduction and differentiation to obtain the preprocessed spectrum h; the derivative order is recommended to be 1-2 order; The noise reduction and derivation processing uses a fully adaptive smoother algorithm, that is, SM is imported into the computing platform, and the processing results are output through the command [hcfcf1]=Sz1(SM,DA,N), where: SM is the measurement signal; DA is the derivation order; N is the number of smoothing, taking the default value 6); h is the output result, cf and cf1 are frequency response functions; S1.3 Determine the number of independent components Using the discontinuous properties of the component space and the noise space, the entire space is decomposed into singular values to obtain the eigenvalues arranged in the direction of the maximum spatial projection; each eigenvalue is multiplied by its corresponding number of dimensions, and each dimensional response is weighted; then the derivative is used to determine the continuity of the sequence, find the turning point of the two subspaces, and determine the number of independent components n of h; S1.4 Blind source identification of tablets using ICA method based on spatial mapping First, the high-order derivative spectrum of SM is obtained, and ICA processing is performed under the principal component number n to obtain the separation matrix. Then, the source signal of each component is output in the first-order derivative space, and the signal of each component is normalized to obtain Fi (i = 1, 2, ... n); S1.5 System actual composition confirmation According to the category of the tablet to be analyzed, a series of possible pure substance samples are collected, and the Raman spectrum of each pure sample is collected according to S1.
1. Noise reduction-derivation is performed according to the signal processing conditions of step S1.2, and normalization is performed to obtain the unit response spectrum Ci (i=1,2,…n) of each pure component; The angles between h and Ci are calculated in sequence. When the angle θ is a small value (θ<0.015), it can be determined that the tablet contains Ci. S2 Quantitative Analysis S2.1 Separation steps 1) Arrange Fi (i = 1, 2, ... n) in columns, take the response spectrum of one single component such as F1 as vector x, and the column vectors of other components in parallel to form the background matrix M; 2) According to the orthogonal projection, orthogonal projection residual and oblique projection formulas, the oblique projection operator of component F1 in the overall response is obtained; P=M(M T M) -1 M T Orthographic projection P |M = F1 residual after IP orthogonal projection (net signal) P H|x =x(x T P |M x) -1 x T P |M Oblique projection operator 3) Using the oblique projection operator h |x =P H|x h separates the contribution of only F1 component in h; 4) Similarly, take the response spectrum of any other component as x, and the column vectors of the remaining components are arranged in parallel to form the background matrix M, and calculate the contribution of each component Fi (i=2,…n); S2.2 Quantification steps 1) The component raw materials determined in step S1.5 are pressed into standard sheets according to step S1.1, and the Raman spectra of each standard sheet are scanned and obtained; 2) According to the data processing parameters of step S1.2, each standard film is subjected to noise reduction-derivation processing, and the derivative spectrum of each standard film is obtained as the unit spectrum, and the modulus of the unit spectrum is obtained; 3) Calculate the modulus of the contribution spectrum of each separated component, and divide it by the unit modulus of the corresponding standard film to obtain the ratio of each component; Normalize the ratio to get the ratio of each component in the system.
3. The method according to claim 2, characterized in that: S1.1 In sample preparation and spectral signal acquisition, the specific parameters of tablet grinding are: under a pressure of 12000N-15000N, round tablets with a diameter of 8-13mm and a thickness of 1-3mm are obtained.
4. The method according to claim 2, characterized in that: In the step S1.2 of spectral signal preprocessing, the computing platform is Matlab 2018 or above.
