Method for identifying type and producing area of fructus aurantii in fructus aurantii immaturus

Abstract

The invention belongs to the technical field of identification of traditional Chinese medicinal materials, and particularly relates to a method for identifying types and producing areas of fructus aurantii in fructus aurantii immaturus, which comprises the following steps: collecting ATR-FTIR spectrums of a training sample for multiple times to obtain a plurality of ATR-FTIR spectrums, and averaging to obtain an average ATR-FTIR spectrum; decomposing the average ATR-FTIR spectrum to obtain a scale signal; stretching and translating the scale signal to obtain a wavelet signal; reconstructing the wavelet signal, retaining a large trend signal, obtaining smooth average ATR-FTIR light for training, extracting a characteristic peak of the light, and obtaining a characteristic vector; inputting the feature vectors into an SVM classifier, and training to obtain an optimized SVM classifier; and inputting the feature vector of the smooth average ATR-FTIR spectrum of the to-be-detected sample into the optimized SVM classifier to obtain the category and the production place of the to-be-detected sample. The method can be used for identifying the type and production place of the fructus aurantii immaturus, and the identification accuracy is relatively high.

Description

Technical Field

[0001] The invention belongs to the technical field of identification of traditional Chinese medicines, and particularly relates to a method for identifying the types and origins of Citrus aurantium immaturus and Citrus aurantium fructus. Background Art

[0002] Zhishi and Zhike are commonly used Qi-regulating medicines. They have the same origin. Zhishi is the dried young fruit of the Rutaceae plant Citrus aurantium L. and its cultivated varieties or the sweet orange Citrus sinensis Osbeck. The fallen fruits are collected from May to June each year, cut into two halves from the middle, and dried in the sun or at low temperature. The smaller ones are directly dried in the sun or at low temperature. Zhike is the dried immature fruit of the Rutaceae plant Citrus aurantium L. and its cultivated varieties. They are harvested in July when the peel is still green, cut into two halves from the middle, and dried in the sun or at low temperature.

[0003] The different harvesting times of Citrus aurantium and Citrus aurantium fructus lead to different medicinal effects. Citrus aurantium has the effects of breaking up qi and eliminating accumulation, resolving phlegm and removing lumps; Citrus aurantium fructus has the effects of regulating qi and relieving fullness, and removing stagnation and dissolving fullness. Citrus aurantium has a strong medicinal effect and is mostly used to break up accumulation and remove stagnation; Citrus aurantium fructus has a mild medicinal effect and is mostly used to regulate qi and relieve fullness, and remove fullness and distension. Citrus aurantium and Citrus aurantium fructus need to be used differently in clinical practice, so the "Chinese Pharmacopoeia" includes them as two different medicinal materials.

[0004] The types of Citrus aurantium and Citrus aurantium are determined according to the harvesting time of the fruit. The quality is affected by many factors, such as the growing environment (origin) of the plant and storage conditions. If they are not distinguished, it will directly affect the doctor's medication safety and efficacy. Therefore, it is necessary to identify the types and origins of Citrus aurantium and Citrus aurantium.

[0005] In the 2020 edition of the Chinese Pharmacopoeia, synephrine is determined in Citrus aurantium and naringin and neohesperidin are determined in Citrus aurantium. Although these ingredients are active ingredients, they are not exclusive to the genus Citrus and cannot distinguish related Chinese medicines from different growth stages of the same plant. Selecting only a few identical indicator ingredients cannot achieve effective identification, and it cannot distinguish the origin.

[0006] Study on the identification of Citrus aurantium and its counterfeits based on thin layer chromatography, Yuan Hanwen et al., Journal of Hunan University of Chinese Medicine, Vol. 41, No. 10, October 2021, published the following information: The Chinese Pharmacopoeia conducted a comparative study on the thin layer chromatography of Citrus aurantium and its easily confused varieties Citrus aurantium, Citrus aurantium and Citrus reticulata, and found that although the four herbs could be analyzed, their chemical compositions were highly similar and could not be specifically distinguished. Citrus aurantium and Citrus aurantium have the same origin, but are harvested at different times. The experimental results show that the chemical compositions of the two are highly similar, making it difficult to distinguish them from a chemical perspective.

[0007] CN201710555332.X discloses a method for identifying Citrus aurantium, Citrus aurantium and Citrus aurantium based on chemical classification and UPLCT of MS. It distinguishes the varieties and growth stages of samples by measuring and analyzing the primary and secondary metabolites in the samples, but it cannot effectively distinguish the origins of Citrus aurantium and Citrus aurantium. Summary of the invention

[0008] The technical problem to be solved by the present invention is to provide a method for identifying the category and origin of Citrus aurantium and Citrus aurantium, which can identify the category and origin of Citrus aurantium and Citrus aurantium with high accuracy.

[0009] The embodiment of the present invention provides a method for identifying the type and origin of Citrus aurantium and Citrus aurantium, comprising the following steps:

[0010] ATR-FTIR spectra of training samples are collected multiple times to obtain multiple ATR-FTIR spectra, and the training samples are Fructus Aurantii Immaturus or Fructus Aurantii Immaturus from different origins;

[0011] Averaging multiple ATR-FTIR spectra to obtain an average ATR-FTIR spectrum;

[0012] The average ATR-FTIR spectrum was decomposed to obtain the scale signal;

[0013] The scale signal is scaled and translated to obtain a wavelet signal;

[0014] The wavelet signal is reconstructed, the large trend signal is retained, and the smoothed average ATR-FTIR spectrum for training is obtained;

[0015] Extract characteristic peaks of the training smoothed average ATR-FTIR spectrum to obtain a characteristic vector of the training smoothed average ATR-FTIR spectrum;

[0016] The feature vector of the smoothed average ATR-FTIR spectrum used for training is input into the SVM classifier, and the optimized SVM classifier is obtained through training;

[0017] The feature vector of the smoothed average ATR-FTIR spectrum of the sample to be tested is input into the optimized SVM classifier to obtain the category and origin of the sample to be tested;

[0018] Reconstruct the wavelet signal, retain the large trend signal, and obtain the smoothed average ATR-FTIR spectrum for training, including:

[0019] Design the wavelet signal as an orthogonal signal of the scale signal;

[0020] Obtain the spanned signal space of the wavelet signal;

[0021] The square integrable space of wavelet signal is obtained according to the spanned signal space of wavelet signal;

[0022] According to the scale signal, wavelet signal and square integrable space, the time domain signal is obtained;

[0023] The time domain signal is decomposed and reconstructed to obtain a decomposed signal set, and a large trend signal is selected from the decomposed signal set to obtain a smoothed average ATR-FTIR spectrum for training.

[0024] Optionally, the step of designing the wavelet signal as an orthogonal signal of a scale signal includes:

[0025] The wavelet signal is expressed as:

[0026] ψ j,l (t) = 2 j / 2 ψ(2t-l)

[0027] The wavelet signal ψ j,l (t) is designed as a scale signal The orthogonal signal is expressed as:

[0028]

[0029] Optionally, the training sample preparation method is to remove impurities from the Fructus Aurantii Immaturus and Fructus Aurantii Immaturus, split them, dry them, crush and sieve them to obtain the training sample.

[0030] Optionally, the absorbance range of the ATR-FTIR spectrum is 4000-600 cm -1 The resolution of ATR-FTIR spectrum is 4 cm -1 .

[0031] Optionally, the average ATR-FTIR spectrum is decomposed to obtain scale signals including:

[0032] Get the preset wavelet basis;

[0033] According to the preset wavelet basis, the corresponding scaling function is obtained;

[0034] According to the scaling function, the average ATR-FTIR spectrum is decomposed to obtain the scaling signal.

[0035] Optionally, the scaled signal is scaled and translated to obtain a wavelet signal including:

[0036] According to the recursive relationship of the scaling function, the wavelet function is obtained;

[0037] The scale signal is stretched and retracted by the scale function, and the scale signal is translated by the wavelet function to obtain scale signals at different scales.

[0038] The scale signals at different scales are expressed linearly to obtain a signal space composed of linear expression signals;

[0039] According to the signal space, the spanned signal space is obtained;

[0040] According to the spanned signal space, the recursive signal is obtained;

[0041] The recursive signal is scaled and translated through the wavelet function to obtain the wavelet signal.

[0042] Optionally, the time domain signal is decomposed and reconstructed to obtain a decomposed signal set, and a large trend signal is selected from the decomposed signal set to obtain a smoothed average ATR-FTIR spectrum for training, including:

[0043] The wavelet signal is linearly represented by the scaling function to obtain a linear wavelet signal;

[0044] The linear wavelet signal is deformed by recursive equation through the scaling function and wavelet function to obtain the decomposed signal set;

[0045] The large trend signal is selected from the decomposed signal set to obtain the smoothed average ATR-FTIR spectrum for training.

[0046] Optionally, extracting characteristic peaks of the training smoothed average ATR-FTIR spectrum to obtain a characteristic vector of the training smoothed average ATR-FTIR spectrum includes:

[0047] The first-order derivative of the training smoothed average ATR-FTIR spectrum is taken to obtain all zero-point positions, and the peak values ​​of all zero-point positions are obtained to obtain the characteristic vector of the training smoothed average ATR-FTIR spectrum.

[0048] Optionally, the feature vector of the smoothed average ATR-FTIR spectrum for training is input into the SVM classifier, and the optimized SVM classifier obtained by training includes:

[0049] According to the SVM classifier, the preset kernel function and decision function are obtained;

[0050] Establish the objective function according to the kernel function;

[0051] According to the characteristic vector of the smoothed average ATR-FTIR spectrum used for training and the objective function, the SVM classifier is trained to obtain the first parameter, the second parameter and the third parameter of the objective function;

[0052] According to the first parameter, the second parameter, the third parameter and the decision function, an adjustment decision function is obtained;

[0053] The decision function in the SVM classifier is replaced by the adjusted decision function to obtain the optimized SVM classifier.

[0054] Optionally, the SVM classifier is trained according to the feature vector of the training smoothed average ATR-FTIR spectrum and the objective function, and the first parameter, the second parameter and the third parameter of the objective function are obtained, including:

[0055] According to the characteristic vector of the smoothed average ATR-FTIR spectrum used for training and the objective function, a constraint function of the objective function is obtained;

[0056] Convert the constraint function into a Lagrangian function;

[0057] The gradients of the first and second parameters in the Lagrangian function are set to 0, and the third parameter is calculated;

[0058] Substitute the third parameter into the constraint function and calculate the first parameter and the second parameter through the KKT condition.

[0059] The beneficial effects of the present invention are:

[0060] 1. The ATR-FTIR spectrum of the training sample is averaged to obtain an average ATR-FTIR spectrum, and then the average ATR-FTIR spectrum is decomposed to obtain a scale signal, the scale signal is stretched and translated to obtain a wavelet signal, and a large trend signal is selected from the wavelet signal to obtain a smooth average ATR-FTIR spectrum for training, and the characteristic peaks of the smooth average ATR-FTIR spectrum are extracted to form a feature vector, and then the feature vector is input into the SVM classifier for training to obtain an optimized SVM classifier. At this time, the SVM classifier training is completed, and the feature vector of the smooth average ATR-FTIR spectrum of the sample to be tested is input into the optimized SVM classifier to obtain the category and origin of the sample to be tested. Compared with the chemical classification method, the present application can not only identify the category of Citrus aurantium and Citrus aurantium, but also identify the origin of Citrus aurantium and Citrus aurantium, and the accuracy of identification is higher, which can reach 96.7%, which is much higher than 80.2% of the orthogonal partial least squares-discriminant analysis method.

[0061] 2. When the ATR-FTIR spectrum is processed, the wavelet signal obtained is first designed as an orthogonal signal, and then other processing is performed. Compared with the traditional processing method, it is designed as an orthogonal signal to ensure that the signals do not interfere with each other. Each signal can independently represent the information in the signal space, avoiding the inaccuracy of the extracted signal due to interference. Secondly, the use of orthogonal bases makes the signal easier during wavelet transformation. The signal can be represented as a linear combination of orthogonal bases, so that the components of each base in the signal can be calculated by inner product. Using the processing method of the present invention, the processing effect of the ATR-FTIR spectrum is the best, the smoothness is the highest, which is 96.64%, the similarity is 98.02%, and the average value is 97.83%. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 The present invention is a schematic flow chart of a method for identifying the types and origins of Citrus aurantium and Citrus aurantium.

[0063] Figure 2 This is the ATR-FTIR spectrum of the three types of Citrus aurantium.

[0064] Figure 3 Multiple ATR-FTIR spectra were obtained for multiple measurements.

[0065] Figure 4 The smoothed average ATR-FTIR spectra obtained by different treatment methods. Among them, (A) is the smoothed average ATR-FTIR spectrum obtained by the MSC treatment method, (B) is the smoothed average ATR-FTIR spectrum obtained by the SG treatment method, (C) is the smoothed average ATR-FTIR spectrum obtained by the SNV treatment method, and (D) is the smoothed average ATR-FTIR spectrum obtained by the treatment method of the present invention. DETAILED DESCRIPTION

[0066] A method for identifying the types and origins of Citrus aurantium and Citrus aurantium, as shown in the following Figure 1 As shown, the present invention includes:

[0067] S1. Collecting ATR-FTIR spectra of training samples for multiple times to obtain multiple ATR-FTIR spectra, wherein the training samples are Fructus Aurantii Immaturus or Fructus Aurantii Immaturus from different origins;

[0068] The preparation method of the training sample is to remove impurities from the immature and fructus aurantii, split, dry, crush and sieve to obtain the training sample. The training sample of the present invention is immature and fructus aurantii of known origin and type. The immature and fructus aurantii of different types and origins have different feature vectors. The feature vectors are input into the SVM classifier to obtain an optimized SVM classifier, so as to identify immature and fructus aurantii of unknown weight and origin.

[0069] The training sample of the present invention is powder and does not need to be prepared into a solution, so the sample preparation is simpler, which can greatly save the time of sample preparation and improve the identification speed.

[0070] In the ATR-FTIR spectrum, the absorbance collection range is 4000~600cm -1 The resolution of ATR-FTIR spectrum is 4 cm -1 ATR-FTIR spectroscopy has higher sensitivity and resolution, which can collect better detailed information of the sample and improve the accuracy of identification.

[0071] Specifically, the preparation method of the samples to be differentiated is as follows: the samples used are from Hunan Province, an important production area of ​​Citrus aurantium and Citrus aurantium, and Citrus aurantium and Citrus aurantium samples are collected from three places in Hunan Province (Yiyang, Huaihua, and Changde). Citrus aurantium and Citrus aurantium were collected in June and July, respectively, and a total of 60 samples were collected. The collected fresh Citrus aurantium and Citrus aurantium were free of impurities, cut into two halves from the middle, dried, crushed, placed in a sealed bag, marked with the place of origin and variety, and stored in a desiccator for later use. All samples were in powder form for subsequent analysis.

[0072] Before collecting the sample spectrum, background information was collected and subtracted from the sample spectrum before further analysis. The spectrum of each sample was collected three times, and the spectral data was averaged for further analysis. After each sample was collected, the ATR accessory was cleaned with an alcohol cotton pad until it was not contaminated before collecting the next sample.

[0073] A FTIR spectrometer equipped with an ATR accessory was used to obtain the ATR-FTIR spectra of each Citrus aurantium and Citrus aurantium sample. All spectra were recorded using OMNIC software and saved as absorbance values. The absorbance acquisition range of each spectrum was 4000-600 cm -1 , the spectral resolution is 4cm -1 The scanning number is 32 times to obtain the average analysis result and improve the signal-to-noise ratio. The spectrum consists of absorbance and wave number, with absorbance as the vertical axis and wave number as the horizontal axis.

[0074] For example, the ATR-FTIR spectra of the Fructus Aurantii Immaturus samples from three different places (Yiyang, Huaihua, and Changde) were measured. The ATR-FTIR spectra obtained are as follows: Figure 2 shown.

[0075] S2. Average multiple ATR-FTIR spectra to obtain an average ATR-FTIR spectrum.

[0076] Specifically, multiple scans are averaged to avoid inaccurate average ATR-FTIR spectra due to data errors or large errors in a single scan. n scans are performed to obtain n ATR-FTIR spectra, and the absorbance values ​​at the same wavelength are averaged as the average absorbance value at this wavelength to obtain the average ATR-FTIR spectrum.

[0077] S3. Decompose the average ATR-FTIR spectrum to obtain a scale signal.

[0078] The average ATR-FTIR spectrum is decomposed to obtain scale signals including:

[0079] S31, obtaining a preset wavelet basis;

[0080] S32, obtaining a corresponding scaling function according to a preset wavelet basis;

[0081] S33. Decompose the average ATR-FTIR spectrum according to the scaling function to obtain a scaling signal.

[0082] Specifically, before processing the average ATR-FTIR spectrum, the signal must first be preprocessed, such as denoising and smoothing, to ensure the accuracy of the analysis, and then a suitable wavelet basis is selected, including Haar wavelet, Daubechies wavelet, Symlets, etc. Each wavelet has its own specific scale function and waveform function, and the wavelet transform is used to decompose the signal to obtain the performance at different scales.

[0083] The scale function reflects the low-frequency components of the signal at that scale.

[0084] According to the selected wavelet basis, the corresponding scaling filter and initial function are determined. Usually, the unit pulse function can be selected, and the scaling function is obtained by continuous iterative calculation.

[0085] The scale signal is represented as:

[0086]

[0087] S4. Scale and translate the scale signal to obtain a wavelet signal.

[0088] The scale signal is scaled and translated to obtain the wavelet signal including:

[0089] S41, obtaining a wavelet function according to the recursive relationship of the scaling function;

[0090] For the selected wavelet basis, a low-pass filter and a high-pass filter are selected, and the scaling functions of different scales are calculated starting from the basic scaling function (such as the unit impulse function) using the recursive relationship. The recursive relationship of the scaling function is combined with the high-pass filter to generate the corresponding wavelet function.

[0091] S42, scaling the scale signal using a scaling function, and translating the scale signal using a wavelet function to obtain scale signals at different scales.

[0092] Scaling function After scaling and translation transformation, we can get the scale signal at different scales j Defined as:

[0093]

[0094] S43, expressing scale signals at different scales linearly to obtain a signal space composed of linearly expressed signals;

[0095] Specifically, the scale signal The signal space composed of linearly expressed signals is V j :

[0096]

[0097] S44, obtaining a spanned signal space according to the signal space;

[0098] Specifically, the scaling function can contain more information at high resolution, that is, the signal space spanned by the high-resolution scaling signal contains the signal space spanned by the low-resolution scaling signal, which can be expressed as:

[0099]

[0100] S45, obtaining a recursive signal according to the spanned signal space;

[0101] Depend on Get the recursive signal:

[0102]

[0103] Among them, h 0 [k] is the scaling function coefficient.

[0104] S46, scaling and translating the recursive signal through a wavelet function to obtain a wavelet signal.

[0105] Specifically, according to the square integrable space L 2 (R), by the scaling function The corresponding wavelet function ψ(t) can be defined. The wavelet function is defined in accordance with the scaling function, and the space W spanned by the wavelet function j-1 Vj and V j-1 The difference of .

[0106] Wavelet signal ψ j,l (t) is expressed as:

[0107] ψ j,l (t) = 2 j / 2 ψ(2t-l).

[0108] Reconstruct the wavelet signal, retain the large trend signal, and obtain the smoothed average ATR-FTIR spectrum for training, including:

[0109] Design the wavelet signal as an orthogonal signal of the scale signal;

[0110] Obtain the spanned signal space of the wavelet signal;

[0111] The square integrable space of wavelet signal is obtained according to the spanned signal space of wavelet signal;

[0112] According to the scale signal, wavelet signal and square integrable space, the time domain signal is obtained;

[0113] The time domain signal is decomposed and reconstructed to obtain a decomposed signal set, and a large trend signal is selected from the decomposed signal set to obtain a smoothed average ATR-FTIR spectrum for training.

[0114] Specifically, the wavelet signal ψ j,l (t) The signal space spanned is denoted as W j ,but:

[0115]

[0116] Specifically, the square integrable space of wavelet signal is:

[0117] The time domain signal x(t) is a square integrable space L 2 (R), so the signal x(t) can be represented by the scaled signal and the wavelet signal ψ j,l (t) Expression:

[0118]

[0119] Scale signal Expressing rough information, wavelet signal ψ j,l (t) Express detailed information.

[0120] S5, reconstructing the wavelet signal, retaining the large trend signal, and obtaining the smoothed average ATR-FTIR spectrum for training;

[0121] Decompose and reconstruct the time domain signal to obtain a decomposed signal set, select the large trend signal from the decomposed signal set, and obtain the smoothed average ATR-FTIR spectrum for training, including:

[0122] The wavelet signal is linearly represented by the scaling function to obtain a linear wavelet signal;

[0123] The linear wavelet signal is deformed by recursive equation through the scaling function and wavelet function to obtain the decomposed signal set;

[0124] The large trend signal is selected from the decomposed signal set to obtain the smoothed average ATR-FTIR spectrum for training.

[0125] To obtain and d j,k , the recursive equation deformation of the scaling function and wavelet function:

[0126]

[0127] So:

[0128]

[0129] Right now:

[0130]

[0131] d j,l =∑ n h 1 [2n-l]·d j+1,n .

[0132] The decomposed signal set is:

[0133]

[0134] Considering that the purpose of analysis is to smooth the spectrum, only the signal containing the general trend should be retained, that is, the reconstruction of the signal only retains item.

[0135] The orthogonal wavelet function family includes sym, coif and db, and the decomposition level is limited to 1-12. The signal x(t) is decomposed and reconstructed to retain only Item, coming soon The item is output as a smoothing result for subsequent analysis and processing.

[0136] Multiple ATR-FTIR spectra of Fructus Aurantii Immaturus were obtained by multiple measurements. Figure 3 As shown. Take the average of multiple ATR-FTIR spectra to obtain the average ATR-FTIR spectrum, and use different processing methods to process the average ATR-FTIR spectrum to obtain the smooth average ATR-FTIR spectrum for training. The results are shown in Figure 4 shown.

[0137] The results obtained by different processing methods are different, and the smoothing effect evaluation indicators of different processing methods are shown in Table 1. In Table 1, the smoothing degree and similarity of each region are equal under the same smoothing treatment, indicating that in the process of scanning infrared spectra, factors such as environmental noise have the same impact on each sample, and no abnormal situation occurs in the scanning process.

[0138] Table 1 Smoothing effects of four smoothing methods

[0139]

[0140] From Table 1 and Figure 4It can be seen that the smoothed average ATR-FTIR spectrum obtained by the treatment method of the present invention has the best effect, the highest smoothness, which is 96.64%, the similarity is 98.02%, the average is 97.83%, and there is no band with large fluctuation in the spectrum, and the overall smoothing effect is the best. Under the MSC and SNV treatment methods, the smoothness is 0; the smoothness under the SG treatment is 0.48%, indicating that these three methods cannot well reduce the influence of regional factors, and there are still some absorption peaks with large fluctuation in the smoothed spectrum.

[0141] S6, extracting characteristic peaks of the training smoothed average ATR-FTIR spectrum to obtain a characteristic vector of the training smoothed average ATR-FTIR spectrum; specifically comprising the following steps:

[0142] The first-order derivative of the training smoothed average ATR-FTIR spectrum is taken to obtain all zero-point positions, and the peak values ​​of all zero-point positions are obtained to obtain the characteristic vector of the training smoothed average ATR-FTIR spectrum.

[0143] Specifically, calculate the first-order derivative M1 of the smoothed average ATR-FTIR spectrum and obtain all zero-point positions a of M1 0 , a 0 It is also the peak position of the average smoothed spectrum M. Get each smoothed spectrum at a 0 The peak value x at , x is the characteristic vector of each spectrum.

[0144] Set the class label y 2 , Citrus aurantium corresponds to 1, Citrus aurantium corresponds to 2, and the regional label is y 1 , Changde area corresponds to 1, Huaihua area corresponds to 2, and Yiyang area corresponds to 3. Sample set {x i ,y i}, i = 1, 2, ..., 30, where x i is the feature vector, y i The region label.

[0145] S7, inputting the feature vector of the smoothed average ATR-FTIR spectrum for training into the SVM classifier, and training to obtain an optimized SVM classifier; specifically comprising the following steps:

[0146] S71, according to the SVM classifier, obtaining a preset kernel function and a decision function;

[0147] S72, establishing an objective function according to the kernel function;

[0148] S73, training the SVM classifier according to the feature vector of the training smoothed average ATR-FTIR spectrum and the objective function to obtain a first parameter, a second parameter, and a third parameter of the objective function;

[0149] S74, obtaining an adjustment decision function according to the first parameter, the second parameter, the third parameter and the decision function;

[0150] S75. Replace the decision function in the SVM classifier with the adjusted decision function to obtain an optimized SVM classifier.

[0151] Specifically, the kernel function used in this application is

[0152] The decision function is: f(x) = sgn(w T K(x i ,x)+b.

[0153] The objective function and constraints are sty i (w T K(x i ,x)+b)≥1-ξ i ,ξ i ≥0.

[0154] According to the characteristic vector of the smoothed average ATR-FTIR spectrum used for training and the objective function, the SVM classifier is trained to obtain the first parameter, the second parameter and the third parameter of the objective function including:

[0155] According to the characteristic vector of the smoothed average ATR-FTIR spectrum used for training and the objective function, a constraint function of the objective function is obtained;

[0156] Convert the constraint function into a Lagrangian function;

[0157] The gradients of the first and second parameters in the Lagrangian function are set to 0, and the third parameter is calculated;

[0158] Substitute the third parameter into the constraint function and calculate the first parameter and the second parameter through the KKT condition.

[0159] Specifically, the constraint function of the objective function is converted into a Lagrangian function:

[0160]

[0161] where λ i ,r i ,ξ i ≥0,λ i =Cr i .

[0162] The objective function and constraints are equivalent to:

[0163]

[0164] Specifically, the first parameter is w, the second parameter is b, and the third parameter is λ. Set the gradient of L at w and b to 0, and substitute it into the simplified form to get:

[0165]

[0166] Where C is a box constraint. Solve this quadratic programming problem and use the SMO algorithm to get the optimal solution

[0167] Then according to the KKT condition, we can get:

[0168] Solving the KKT condition, we get:

[0169]

[0170] Specifically, the calculated parameters are substituted into the decision function to obtain the adjusted decision function:

[0171]

[0172] The decision function in the SVM classifier is replaced with the adjustment decision function to obtain an optimized SVM classifier. When the training samples of the categories of Citrus aurantium and Citrus aurantium are input into the adjustment decision function, the adjustment decision function outputs 1 and 2, corresponding to Citrus aurantium and Citrus aurantium, respectively. When the training sample input is the Citrus aurantium and Citrus aurantium origin training sample, the parameters of the adjustment decision function are adjusted to be different from the parameters of the adjustment decision function of the category training sample. At this time, the adjustment decision function outputs the corresponding value according to the input value, for example, Changde area corresponds to 1, Huaihua area corresponds to 2, and Yiyang area corresponds to 3, then the output of the adjustment decision function is 1, 2 and 3.

[0173] The optimization parameters for feature extraction include wavelet function and decomposition layers; the SVM estimation optimal model parameters include corresponding mode (Coding), box constraint (BoxConstraint), scaling parameter (KernelScale), kernel function (KernelFunction), polynomial order (PolynomialOrder), and whether to standardize the data (Standardize). Different wavelet functions and decomposition layers produce different smoothing results and waveforms. Smoothed signals with more details have more peaks and more extracted eigenvalues, that is, higher dimensions. The number of features of samples in the SVM classifier is too large, so it is suitable to use linear kernels and Gaussian kernels. When the number of samples is greater than the number of features, it is suitable to use polynomial kernels.

[0174] S8, inputting the feature vector of the smoothed average ATR-FTIR spectrum of the sample to be tested into the optimized SVM classifier to obtain the category and origin of the sample to be tested;

[0175] Specifically, after the model training is completed, you only need to input the feature vector of the smoothed average ATR-FTIR spectrum of the sample to be tested into the optimized SVM classifier to automatically output the result, that is, the category and origin of the sample to be tested. Depending on the different features input during training, the results output by the SVM classifier are different. For example, if only the characteristics of the region are trained, only the results of the region will be output. If the region and category are trained at the same time, the category and origin of the sample to be tested will be output at the same time.

[0176] The optimized SVM classifier of the present invention can process high-dimensional data and realize classification. The optimized SVM classifier of the present invention is used to classify and identify the origin of Fructus Aurantii Immaturus, and the identification accuracy is as high as 96.7%, which is significantly higher than 80.2% of the orthogonal partial least squares-discriminant analysis method. The calculation method of the identification accuracy is: the number of samples correctly identified / the total number of samples identified.

[0177] A person skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of protection of the present application is limited to these examples. In line with the concept of the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of different aspects of one or more embodiments of the present application as above, which are not provided in detail for the sake of simplicity.

[0178] One or more embodiments of the present application are intended to cover all such substitutions, modifications and variations that fall within the broad scope of the present application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of one or more embodiments of the present application should be included in the protection scope of the present application.

Claims

1. A method for identifying the type and origin of Citrus aurantium and Citrus aurantium, characterized in that: The following steps are included: Collecting ATR-FTIR spectra of training samples for multiple times to obtain multiple ATR-FTIR spectra, wherein the training samples are Fructus Aurantii Immaturus or Fructus Aurantii Immaturus from different origins; Averaging multiple ATR-FTIR spectra to obtain an average ATR-FTIR spectrum; Decomposing the average ATR-FTIR spectrum to obtain a scale signal; Scaling and translating the scale signal to obtain a wavelet signal; Reconstructing the wavelet signal, retaining the large trend signal, and obtaining a smoothed average ATR-FTIR spectrum for training; Extracting characteristic peaks of the smoothed average ATR-FTIR spectrum for training to obtain a characteristic vector of the smoothed average ATR-FTIR spectrum for training; Inputting the feature vector of the training smoothed average ATR-FTIR spectrum into the SVM classifier, and training to obtain an optimized SVM classifier; The feature vector of the smoothed average ATR-FTIR spectrum of the sample to be tested is input into the optimized SVM classifier to obtain the category and origin of the sample to be tested; The reconstructing the wavelet signal, retaining the large trend signal, and obtaining the smoothed average ATR-FTIR spectrum for training comprises: Designing the wavelet signal as an orthogonal signal of the scale signal; Obtain the spanned signal space of the wavelet signal; Obtaining a square integrable space of the wavelet signal according to the spanned signal space of the wavelet signal; Obtaining a time domain signal according to the scale signal, the wavelet signal and the square integrable space; The time domain signal is decomposed and reconstructed to obtain a decomposed signal set, and a large trend signal is selected from the decomposed signal set to obtain a smoothed average ATR-FTIR spectrum for training.

2. The identification method according to claim 1, characterized in that: The step of designing the wavelet signal as an orthogonal signal of a scale signal comprises: The wavelet signal is expressed as: ψ j,l (t)=2 j 2ψ(2t-l) The wavelet signal ψ j,l (t) is designed as a scale signal The orthogonal signal is expressed as:

3. The identification method according to claim 1, characterized in that: The training sample preparation method comprises removing impurities from Citrus aurantium and Citrus aurantium fructus, splitting, drying, crushing and sieving to obtain the training sample.

4. The identification method according to claim 1, characterized in that: In the ATR-FTIR spectrum, the absorbance acquisition range is 4000~600cm -1 The resolution of ATR-FTIR spectrum is 4 cm -1 .

5. The identification method according to claim 1, characterized in that: Decomposing the average ATR-FTIR spectrum to obtain a scale signal includes: Get the preset wavelet basis; According to the preset wavelet basis, a corresponding scaling function is obtained; The average ATR-FTIR spectrum is decomposed according to the scaling function to obtain a scaling signal.

6. The identification method according to claim 5, characterized in that: The scaling and translating of the scaled signal to obtain a wavelet signal comprises: According to the recursive relationship of the scaling function, the wavelet function is obtained; The scale signal is scaled by a scaling function, and the scale signal is translated by a wavelet function to obtain scale signals at different scales; The scale signals at different scales are expressed linearly to obtain a signal space composed of linear expression signals; According to the signal space, a spanned signal space is obtained; Obtaining a recursive signal according to the spanned signal space; The recursive signal is scaled and translated through a wavelet function to obtain a wavelet signal.

7. The identification method according to any one of claims 1 to 6, characterized in that: Decomposing and reconstructing the time domain signal to obtain a decomposed signal set, selecting a large trend signal from the decomposed signal set, and obtaining a smoothed average ATR-FTIR spectrum for training includes: Linearly representing the wavelet signal by the scaling function to obtain a linear wavelet signal; Performing recursive equation deformation on the linear wavelet signal through the scaling function and the wavelet function to obtain a decomposed signal set; A large trend signal is selected from the decomposed signal set to obtain a smoothed average ATR-FTIR spectrum for training.

8. The identification method according to any one of claims 1 to 6, characterized in that: The step of extracting characteristic peaks of the training smoothed average ATR-FTIR spectrum to obtain a characteristic vector of the training smoothed average ATR-FTIR spectrum includes: The first-order derivative of the training smoothed average ATR-FTIR spectrum is taken to obtain all zero-point positions, and the peak values ​​of all zero-point positions are obtained to obtain the characteristic vector of the training smoothed average ATR-FTIR spectrum.

9. The identification method according to any one of claims 1 to 6, characterized in that: The step of inputting the feature vector of the training smoothed average ATR-FTIR spectrum into the SVM classifier and training the optimized SVM classifier comprises: According to the SVM classifier, a preset kernel function and a decision function are obtained; Establishing an objective function according to the kernel function; Training the SVM classifier according to the characteristic vector of the training smoothed average ATR-FTIR spectrum and the objective function to obtain a first parameter, a second parameter and a third parameter of the objective function; Obtaining an adjustment decision function according to the first parameter, the second parameter, the third parameter and the decision function; The decision function in the SVM classifier is replaced by the adjusted decision function to obtain the optimized SVM classifier.

10. The identification method according to claim 9, characterized in that: The SVM classifier is trained according to the feature vector of the training smoothed average ATR-FTIR spectrum and the objective function to obtain the first parameter, the second parameter and the third parameter of the objective function, including: Obtaining a constraint function of the objective function according to the characteristic vector of the training smoothed average ATR-FTIR spectrum and the objective function; Converting the constraint function into a Lagrangian function; The gradient of the first parameter and the second parameter in the Lagrangian function is set to 0, and the third parameter is calculated; The third parameter is substituted into the constraint function, and the first parameter and the second parameter are obtained by KKT condition calculation.

Images