CN104062256A - Soft measurement method based on near infrared spectroscopy - Google Patents

Abstract

The invention relates to a soft measurement method based on near infrared spectroscopy, and especially relates to a soft measurement method, based on near infrared spectroscopy, for relative density in a compound donkey-hide gelatin pulp medicinal material extract concentration process. According to the invention, a sample set is formed by production process samples and samples prepared in a laboratory; near infrared spectra of the sample set are collected; abnormal samples are rejected; characteristic spectral information of concentrated liquid is extracted by selecting appropriate spectral modeling bands and pre-processing methods; a relative density of the concentrated liquid determined by a pycnometer method is used as a reference value; a quantitative correction model of the relation between the concentrated liquid near infrared spectra and the relative densities is established by a partial least squares regression method; the near infrared spectrum of a sample to be determined is collected; and the relative density of the sample is obtained by using the established model. The invention provides a rapid and precise density soft measurement method which facilitates the improvement of the quality control level of the production process.

Description

A Soft Sensing Method Based on Near Infrared Spectroscopy

technical field

The invention relates to a near-infrared spectrum-based soft measurement method, in particular to a near-infrared spectrum-based soft measurement method for the relative density of a compound donkey-hide gelatin pulp medicine extract concentration process, and belongs to the technical field of traditional Chinese medicine production.

Background technique

Soft sensor technology is a new technology that has emerged in the field of process control and detection in recent years. It mainly uses the idea of indirect measurement to select other variables that are easy to measure for important variables that are difficult to measure or cannot be measured temporarily. Some mathematical relationship to infer or estimate. The application of soft sensor technology to achieve on-line detection of component content is not only economical and reliable, but also has a rapid dynamic response and is easy to achieve real-time control of product quality.

Near Infrared Spectroscopy (Near Infrared Spectroscopy, NIRS) is a spectral region between the visible light and the mid-infrared spectrum with a wavelength range of 780 to 2500nm. This spectral region is mainly the double frequency and combined frequency absorption of hydrogen-containing groups (C-H, N-H, O-H). By scanning the near-infrared spectrum of the sample, the characteristic information of the hydrogen-containing groups of organic molecules in the sample can be obtained. This technique needs to be combined with chemometrics, among which the commonly used chemometric techniques mainly include multiple linear regression, principal component regression and partial least squares regression. The use of near-infrared spectroscopy in quality control of traditional Chinese medicine can reflect the physical and chemical composition information of raw materials, intermediates and products of Chinese medicinal materials as a whole. As a soft measurement technology, it has many advantages such as no or minimal pretreatment of samples, easy operation, and no consumption of chemical reagents. Using this technique, we can establish a soft-sensing model of relative density based on the easily measurable mathematical relationship between near-infrared spectral variables and relative density.

The concentration process is one of the important processes in the production of compound donkey-hide gelatin pulp, and it is also a section that consumes a lot of energy and is relatively difficult in quality control in the actual production process. In the concentration process, the relative density is a key evaluation index for judging the pros and cons of the process, and whether the relative density of the obtained concentrated solution is within the set range will also have an impact on the subsequent process of precipitation and impurity removal. The conventional relative density detection in current production mainly adopts the pycnometer method. This method needs to take out the sample and place it in a constant temperature water bath at 20°C until the sample is cooled to room temperature before accurately weighing the measurement. The operation process is cumbersome and time-consuming, and cannot meet the process analysis. The fast and simple conditions are difficult to be used for real-time online analysis in the production process of traditional Chinese medicine.

Contents of the invention

The object of the present invention is to provide a kind of soft measurement method based on near-infrared spectrum, especially for the soft measurement method of the relative density of the compound donkey-hide gelatin pulp medical material extract concentration process, reflect the change information of the concentrated solution in the concentration process in time, can enhance the The understanding of the concentration process improves the quality control level of the production process and lays the foundation for the online monitoring of the process.

The purpose of the present invention is achieved through the following technical solutions:

A soft sensing method based on near-infrared spectroscopy, comprising the following steps:

1. Collection of samples: The sample set is composed of samples obtained from the actual production process and samples prepared in the laboratory. This method can increase the representativeness of the sample set;

Preferably, the specific steps of the samples collected in the actual production process described in step (1) are: collecting a total of b batches in the actual production enrichment process, and collecting samples at k time points for each batch,

Wherein, b≥5; k≥10.

2. Determination of the relative density of each sample in the sample set: the relative density of the sample measured by the pycnometer method is used as a reference value. The specific operation method is:

Take a dry, clean and precisely weighed pycnometer, fill it with samples, insert the cork, wipe off the excess sample overflowing from the plug hole with filter paper, put it in a constant temperature water bath at 20°C, and put the pycnometer out of the water bath after standing for 10 minutes. Take it out of the pycnometer, weigh it accurately, subtract the weight of the pycnometer itself, and obtain the weight of the sample, dump the sample, wash the pycnometer, fill it with distilled water and measure the weight of the distilled water according to the above method, the relative density of the sample is is the ratio of the weight of the sample to the weight of distilled water.

3. Near-infrared spectrum data collection and data preprocessing: Use a near-infrared spectrometer to collect sample near-infrared spectra, remove abnormal samples and divide sample sets, and then select appropriate modeling spectral bands and preprocessing methods to extract spectral feature information.

Preferably, when the near-infrared spectrometer collects the near-infrared spectrum of the sample, the near-infrared spectrum of the concentrated solution is collected in a transflective mode.

More preferably, when the near-infrared spectrometer collects the near-infrared spectrum of the sample, the built-in background of the instrument is used as a reference, the resolution is 4cm ^-1 , the number of scans is 128, and the wavenumber range of the scanned spectrum is 4000-10000cm ^-1 .

Data preprocessing: Before building a model, it is first necessary to identify and eliminate abnormal samples and divide the sample set to obtain representative calibration set and verification set samples. Among them, the samples used to build the model are calibration set samples, using Validation set samples are used for model verification and evaluation.

The invention adopts the method of Chauvenet test and the combination of leverage value and biochemical residual value to eliminate abnormal samples, and simultaneously takes into account the abnormalities of chemical values and spectral data, which is more helpful for the identification and elimination of abnormal samples.

The Chauvenet test method first calculates the average spectrum of all sample spectra, and then calculates the Mahalanobis distance between each sample spectrum and the average spectrum, arranges the distance values in ascending order, and judges whether the sample spectrum with the largest distance value is is abnormal, if so, continue to judge whether the sample spectrum with the second largest distance value is abnormal, and so on, until a certain sample spectrum is judged to be normal. In the present invention, the software automatically judges the abnormal spectrum according to the criterion. The formula of Chauvenet criterion is as follows:

$<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; ></span>{\mathrm{<span\; class="notranslate">\; Z</span>}}_{\mathrm{<span\; class="notranslate">\; c</span>}}\mathrm{<span\; class="notranslate">\; \&sigma;</span>}$

In the formula, is the average value of the Mahalanobis distance of all samples, Z _c is a constant related to the number of samples, which can be found in the table, and σ is the mean square error.

The formula for calculating the leverage value is:

${\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; )</span>}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}$

In the formula, h _i is the leverage value, n is the number of samples, t _i is the regressor vector of the ith prediction sample, and T is the regressor score matrix of the calibration sample.

The formula for calculating the student residual r _i is:

${\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{\sqrt{\mathrm{<span\; class="notranslate">\; RMSE</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}}$

In the formula, f _i is the residual value of the i-th sample, and RMSE is the mean square error of the calibration set.

In the modeling process, the leverage value measures the influence of a calibration set sample on the model, and the student residual value indicates the quality of the predictive ability. Usually, the sample whose content value is at the average value of the calibration set has a small leverage value. If the leverage value of a certain sample is large, it may be that spectral scanning or other analysis methods introduce errors in the measurement; if the student residual value of a sample If it is higher, it means that the calibration set model has poor predictive ability for this sample. When a sample's leverage value or student residual value is relatively high, the sample is temporarily classified as an abnormal sample.

Before building a quantitative model, it is important to collect a representative sample. A representative sample can not only reduce the workload of modeling, but also directly affect the applicability and accuracy of the built model. Commonly used sample set partition methods include random sampling (Random Sampling, RS) method, content gradient method, Kennard-Stone (KS) method, Duplex method and Sample set Partitioning based on joint x-y distance (SPXY) method, etc., among which SPXY method It is developed on the basis of the Kennard-Stone method. Experiments have proved that the SPXY method can be effectively used for the establishment of near-infrared quantitative models. The step-by-step selection process of the SPXY method is similar to the Kennard-Stone method: the Kennard-Stone method regards all samples as calibration set candidate samples, and first selects the two vector pairs with the farthest Euclidean distance to enter the calibration set, and in subsequent iterations In the process, the candidate sample with the maximum value in the minimum distance is selected into the calibration set, and so on until the preset number of samples is reached. The disadvantage of this method is that only the X variable (spectral data) is considered in the calculation; while the SPXY method is in When calculating the distance between samples, the X variable (spectral data) and the y variable (chemical value) are taken into account at the same time. First, the distances of samples p and q in the X and Y spaces are calculated respectively. The formula is as follows:

${\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\sqrt{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; J</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; ;</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span><span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; ]</span>$

${\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\sqrt{{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; ;</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span><span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; ]</span>$

In the formula, d _x (p, q) and d _y (p, q) are the distances of samples p and q in X and Y spaces, respectively, and j is a variable.

In order to ensure that samples have the same weight in X space and y space, they are divided by their maximum value in the data set respectively, and the formula is as follows:

${\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; xy</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; max</span>}{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; +</span>\frac{{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; max</span>}{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ;</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; q</span>}<span\; class="notranslate">\; \&Element;</span><span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; .</span>$

The advantage of the SPXY method is that it can effectively cover the multi-dimensional vector space, thereby improving the predictive ability of the built model. The present invention uses the SPXY method to divide the sample set.

After determining the calibration set and validation set samples, it is necessary to optimize the modeling band and preprocessing method. Although the partial least squares method allows the processing of full-spectrum information, the modeling band is too wide and may contain a lot of redundant information. Therefore, it is necessary to select the band to eliminate irrelevant interference and improve model performance. The correlation coefficient method is to calculate the correlation between the absorbance vector corresponding to each wavelength in the calibration set spectral array and the concentration vector of the component to be measured in the concentration array. Usually, the band with a larger correlation coefficient has more information content. Wavelengths that contribute significantly to the model can be selected. The near-infrared spectrum often contains some interference caused by factors that have nothing to do with the properties of the sample to be measured, resulting in baseline drift and scattering effects in the near-infrared spectrum. Preprocessing the spectrum can weaken the influence of various non-target factors on the spectrum and improve the spectrum. Figure quality. Commonly used spectral preprocessing methods mainly include multivariate scattering correction, standard canonical transformation, derivative and smoothing and their combinations.

Multiplicative Scatter Correction (MSC): In the process of near-infrared spectroscopy measurement, due to the uneven distribution of sample particles and different particle sizes, the obtained spectra often have large differences. In some cases, the spectra caused by sample scattering The change is even greater than the spectral change caused by the sample component content. Multivariate scattering correction is mainly used to eliminate spectral errors caused by scattering between samples.

Standard Normal Variation (Standrad Normal Variate, SNV): Standard Normal Variation believes that the absorbance value of each wavelength point in each spectrum should satisfy a certain distribution, and correct each spectrum through this assumption. Specifically, each element of the original data is subtracted from the mean value of the element in the column where the element is located, and then divided by the standard deviation of the column, so that the data in a column of data are comparable on the data scale. This method is mostly used to eliminate the influence of the change of measuring optical path on the spectral response.

Derivative: Due to the influence of the instrument, sample background or other factors, the shift or drift of the spectrum often occurs in the near-infrared spectral analysis. If it is not treated, it will also affect the quality of the calibration model and the prediction results of unknown samples. accuracy. The most commonly used solution for baseline correction is to perform first-order derivative or second-order derivative processing on the spectrum. The former mainly solves the offset of the baseline, while the latter can eliminate the offset linearly related to the wavelength.

Smoothing: The original spectral data collected by near-infrared spectroscopy instruments contains both useful information and a lot of random noise. The essence of smoothing is to remove high-frequency components in spectral data and retain useful low-frequency information. Window moving polynomial least squares smoothing method (Savitzky-GolaySmoothing) is a more commonly used smoothing algorithm.

4. Establishment of the soft sensor model: use the multivariate analysis method to build a quantitative calibration model between the near-infrared characteristic spectrum of the sample and its relative density value, use the calibration set samples to establish the soft sensor model, and evaluate the model through the validation set samples.

Preferably, the optimization performance evaluation index of the soft sensor model is: using correlation coefficient r, correction set mean square error RMSEC, cross-validation mean square error RMSECV and verification set mean square error RMSEP as indicators to optimize the modeling parameters; The prediction ability of the measurement model for the sample to be tested is evaluated by the correlation coefficient r of the prediction set and the mean square error RMSEP of the prediction set.

Preferably, the multivariate analysis method is a partial least squares regression method.

5. Application of soft sensor model:

Take the sample to be tested, collect the near-infrared spectrum according to the same spectrum acquisition parameters as the modeled sample, and import the spectrum into the built calibration model to quickly calculate the relative density of the sample to be tested.

After a period of actual application of the above soft sensor model, new samples can be added to expand the scope of application of the model, and the model can be continuously updated and improved. The operation steps are the same as before.

In the present invention, a sample set is composed of production process samples and laboratory prepared samples, the near-infrared spectrum of the sample set is collected, abnormal samples are eliminated, and a suitable sample set division method, spectral modeling band, and preprocessing method are selected to obtain the characteristic spectrum of the sample. Information, using the relative density of the sample measured by the pycnometer method as a reference value, and using the partial least squares regression method to establish a quantitative calibration model for the relationship between the near-infrared spectrum of the sample and its relative density. Collect the near-infrared spectrum of the sample to be tested in the same way, and use the built soft sensor model to quickly calculate its relative density.

The present invention introduces near-infrared spectroscopy technology into the quality control of traditional Chinese medicine intermediates in the production process of traditional Chinese medicine pharmacy. Taking the concentration process of compound donkey-hide gelatin pulp extract in the production of compound donkey-hide gelatin pulp as an example, the soft measurement method based on near-infrared spectroscopy is used to realize the comparative analysis of the concentrated liquid. Rapid determination of density. Compared with the traditional pycnometer method, it avoids tedious and time-consuming operations and saves a lot of time and manpower.

Description of drawings

Accompanying drawing 1 is the near-infrared spectrogram of compound donkey-hide gelatin slurry concentration process concentrate;

Accompanying drawing 2 is the Chauvenet test result graph in abnormal sample elimination;

Accompanying drawing 3 is the distribution diagram of leverage value and student chemical residual in abnormal sample elimination;

Accompanying drawing 4 is the principal component score map of correction set and verification set sample;

Accompanying drawing 5 is the relative spectrogram of sample relative density;

Accompanying drawing 6 is the correlation diagram of the predicted value of concentrated solution relative density model and reference value;

Accompanying drawing 7 is the correlation diagram of the predicted value of the relative density of the concentrated solution in the prediction set and the reference value.

Detailed ways

The present invention will be further described below in conjunction with specific embodiments, and the advantages and characteristics of the present invention will become clearer along with the description. However, the examples are merely exemplary and do not limit the scope of the present invention in any way. Those skilled in the art should understand that the details and forms of the technical solutions of the present invention can be modified or replaced without departing from the spirit and scope of the present invention, but these modifications and replacements all fall within the protection scope of the present invention.

Materials: The extract of compound donkey-hide gelatin pulp is provided by Shandong Dong'e-hide gelatin Co., Ltd.;

Instruments: Fourier transform near-infrared spectrometer was produced by Thermo Fisher Company of the United States.

Example 1: A soft measurement method for the relative density of the compound donkey-hide gelatin pulp medicinal material extract concentration process based on near-infrared spectroscopy

1. Collection of extract samples of compound donkey-hide gelatin pulp:

Collect the concentrated liquid in the circulation pipeline of the concentration tank on the actual production line. The samples taken in the pipeline can represent the actual state of the liquid in the evaporator. A total of 88 samples were collected in 6 batches. The first to fourth batches were respectively There are 15 samples, and the fifth and sixth batches have 14 samples respectively. Take another 3 parts of the concentrated solution, about 200 mL each, and then concentrate the 3 parts under reduced pressure, control the temperature of the constant temperature tank of the rotary evaporator at 70 ° C, and concentrate to a volume of about 100 mL respectively. Then dilute it step by step with ultrapure water, add about 300mL ultrapure water to each concentrated solution, add in several times, shake well after adding water each time, take a sample to measure the relative density until the relative density of the sample is about 1.03. A total of 27 samples were obtained from the three concentrates according to the above-mentioned dilution operation. The 27 self-prepared samples and the samples collected during the production enrichment process were combined to form a sample set, with a total of 115 samples. A total of 14 samples from the sixth batch of the actual production enrichment process were reserved as the prediction set of the present invention to test the predictive ability of the model for unknown samples, and the remaining samples were used to establish the model.

2. Determination of the relative density of the sample: the relative density of the sample set sample except the sixth batch of 14 samples in the actual production and concentration process was measured by the pycnometer method as a reference value, and the distribution range of the relative density of the sample set measured was 1.0224- 1.1352.

3. Sample near-infrared spectrum data collection: Use ANTARIS II Fourier transform near-infrared spectrometer to collect sample near-infrared spectrum. The sampling mode is the transmission and reflection spectrum collection mode. The relevant parameters of the collection are: the built-in background of the instrument is used as a reference, and the resolution is 4cm ^-1 , the number of scans is 128, and the spectrum acquisition wavenumber range is 4000-10000cm ^-1 . The original near-infrared spectrum of the collected compound donkey-hide gelatin pulp medicinal material extract is shown in Figure 1.

4. Establishment of calibration model:

(1) Elimination of abnormal samples:

Abnormal samples were eliminated by using the method of Chauvenet test and the combination of leverage value and studentized residual value. The results of the Chauvenet test are shown in Figure 2 (only 40 samples with higher distance values are listed in the figure, and the distance value is the Mahalanobis distance, which refers to the distance from the spectrum of each sample to the average spectrum). According to the Chauvenet test, seven samples (numbered 35, 36, 80, 81, 99, 100, and 101) were significantly different from the average spectra of all samples in the sample set, so they were excluded as abnormal samples.

The distribution of leverage values and studentized residuals of all modeling samples is shown in Figure 3. It can be seen from the figure that the sample numbered 73 has a large studentized residual value, and the samples numbered 17, 18, 93, and 94 have large leverage values and studentized residual values, so these samples are temporarily listed as abnormal sample.

For the abnormal samples (numbered 73, 17, 18, 93, and 94) eliminated for the leverage value and the studentized residual value, if they are directly eliminated, it is possible to mistake the non-abnormal samples as abnormal samples and remove them. In order to avoid such mistakes, it is necessary to recycle the samples judged as abnormal one by one, and determine whether to keep the samples according to the performance of the model after recycling, which largely avoids the misjudgment of samples as abnormal, thus making the model more steady. By recovering the abnormal samples one by one, building a model, determining the effect of the above abnormal samples on the model, comparing the model results under the conditions of no elimination, all elimination and recovery one by one, and selecting the optimal model to determine the abnormality to be eliminated The sample, the results are shown in Table 1. Since the sample set has not been divided, all the samples of the model are used as the calibration set, and the partial least squares regression method is used to establish the quantitative calibration model between the near-infrared spectrum of the concentrate and its relative density, and the r _c , _rcv , RMSEC and RMSECV is used as a measure of model performance. The results show that the correction ability of the model is reduced to varying degrees after the recovery of each sample, so these samples are designated as abnormal samples and removed from the sample set.

Table 1 Model performance table after recovering and removing samples one by one

Note: The principal component score is the factor that the software automatically judges to affect the performance of the model.

(2) Division of the sample set:

Before building the model, the present invention uses the SPXY method to divide the sample set for model building into a correction set and a verification set, the samples in the correction set are used to build the model, and the samples in the verification set are used to verify the model. The SPXY algorithm function was written in Matlab software. Among the remaining 89 samples after removing abnormal samples, 67 samples were selected as the calibration set by SPXY method, and the other 21 samples were used as the validation set. The relative density ranges of the calibration set and validation set samples are 1.0325-1.0988 and 1.0379-1.0694, respectively. It can be seen that the density range of the calibration set samples covers the validation set samples and is sufficiently representative. Observe the PC1-PC2 scatter diagram of the calibration set sample and the validation set sample as shown in Figure 4. The results show that the distribution of the calibration set sample and the validation set sample is relatively uniform.

(3) Modeling band range optimization:

The transmission analysis generally selects the short-wave band, and the reflection analysis generally selects the long-wave band, but the transflective mode adopted in the present invention combines the optical characteristics of transmission and reflection, so the combination of the short-wave band and the long-wave band can be selected. From Figure 1 of the original spectrogram, it can be seen that the noise in the 4000-4400cm ^-1 band is more obvious, which is the interference caused by fiber absorption, so this band was discarded in the modeling. As shown in Figure 5, there are two obvious band regions (5600-6250cm ^-1 and 7400-10000cm ^-1 bands), and the correlation between the absorbance and the relative density is greater than 0.75. Considering the performance of the model and the calculation speed, the final The combined bands of 5600-6250cm ^-1 and 7400-9000cm ^-1 are used as the modeling bands.

(4) Spectral preprocessing method optimization:

For the original spectrum, preprocessing methods such as multivariate scattering correction (MSC), standard canonical transformation (SNV), first-order derivative, second-order derivative, Savitsky-Golay filter smoothing (SG) and Norris derivative filter smoothing were used respectively, and passed the interactive verification method The impact of different preprocessing methods on the performance of the model was compared, and the results are shown in Table 2. It can be seen from Table 2 that different preprocessing methods failed to improve the performance of the model, and the performance parameters of the model after various pretreatments were worse than those built by the original spectrum, so the original spectrum was used for modeling.

Table 2 The influence of different spectral preprocessing methods on the PLS calibration model

Note: The principal component score is the factor that the software automatically judges to affect the performance of the model. Among them, Raw Spectra: original spectrum; MSC: multivariate scattering correction; SNV: standard canonical transformation; SG: SG filter smoothing; Norris: Norris smoothing; ^1st D: first-order derivative spectrum; ^2nd D: second-order derivative spectrum.

(5) Soft sensor model establishment:

After abnormal samples were identified and 12 abnormal samples were eliminated and the sample set was divided into calibration set and verification set by SPXY method, the original spectra with modeling band ranges of 5600-6250cm ^-1 and 7400-9000cm ^-1 were selected, and the partial least squares The multiplication regression method establishes the correction model between the near-infrared spectrum of the concentrated liquid and its relative density, wherein the partial least squares regression algorithm and the optimization of the modeling band and the pretreatment method are all passed through the TQ analyst software (version 8.5.25, Thermo Fisher, Madson , Wisconsin, USA). The calibration set correlation coefficient of the model is 0.9993, and the RMSEC is 5.85×10 ^-4 ; the cross-validation correlation coefficient is 0.9972, and the RMSECV is 1.18×10 ^-3 ; the validation set correlation coefficient reaches 0.9980, and the RMSEP is 7.24×10 ^-4 , indicating that There is a good correlation between the characteristic spectrum of the pulp extract and its relative density. Figure 6 is a correlation diagram between the relative density near-infrared predicted value and the reference value, which also shows that the established quantitative regression model has a good fitting effect.

5. Unknown samples, that is, rapid determination of the relative density of samples in the prediction set:

The soft-sensing model will be established to predict the prediction set samples, that is, the sixth batch of enrichment process samples that did not participate in the modeling. As shown in Figure 7, the density prediction trend is basically consistent with the actual measurement trend, and the prediction accuracy is high, which meets the industrial The requirements of production process analysis can be effectively used for the rapid detection of relative density in the concentration process.

Claims (10)

1. the flexible measurement method based near infrared spectrum, is characterized in that, this flexible measurement method is realized by following steps:

(1) collection of sample: the sample of the sample gathering with actual production process and laboratory preparation forms sample set;

(2) mensuration of each sample relative density in sample set: the pycnometer method of usining records the relative density of sample as reference value;

(3) sample near infrared spectra collection and data pre-service: use near infrared spectrometer collecting sample near infrared spectrum, carry out the division of exceptional sample rejecting and sample set, then select suitable modeling spectral band and preprocess method, extract spectral signature information;

(4) foundation of soft-sensing model: use near infrared characteristic spectrum that multivariable technique the builds sample quantitative correction model between density value corresponding thereto, use calibration set Sample Establishing soft-sensing model, and collect sample by checking model is evaluated;

(5) application of soft-sensing model: with institute's established model, sample to be tested is carried out to forecast analysis, obtain the relative density of testing sample.

2. measuring method according to claim 1, is characterized in that, the described sample collection concrete steps of step (1) are: collect b batch altogether of actual production concentration process, each batch gathers k sample constantly,

Wherein, b >=5; K >=10.

3. measuring method according to claim 1, is characterized in that, the acquisition mode of the near infrared spectrum that step (3) is described is: use reflective-mode collecting sample near infrared spectrum.

4. measuring method according to claim 1, is characterized in that, the correlation parameter of the near infrared spectra collection that step (3) is described is: the built-in background of the instrument of take is reference, and resolution is 4cm ^-1, scanning times is 128 times, scanning optical spectrum wave-number range is 4000-10000cm ^-1.

5. measuring method according to claim 1, is characterized in that, the division of step (3) sample set: adopt SPXY method that sample set is divided into calibration set and checking collection.

6. measuring method according to claim 1, it is characterized in that, the described pretreated concrete steps of data of step (3) are: the method that adopts Chauvenet method of inspection and lever value to combine with studentization residual values is carried out the rejecting of exceptional sample, adopting Sample set Partitioning based on joint x-y distance(SPXY) sample set is divided into calibration set to method and checking collects, adopt correlation spectrometry to select suitable spectrum range, and in conjunction with effective preprocessing procedures, original spectrum diagram data is processed.

7. measuring method according to claim 1, it is characterized in that, the soft-sensing model that step (4) is described, its Optimal performance evaluation index is: with correlation coefficient r, calibration set mean square deviation RMSEC, cross validation mean square deviation RMSECV and checking, integrate mean square deviation RMSEP as index optimization modeling parameters; Described soft-sensing model is examined with forecast set correlation coefficient r and forecast set mean square deviation RMSEP the predictive ability of sample to be tested.

8. measuring method according to claim 1, is characterized in that, the described multivariable technique of step (4) is partial least-squares regression method.

9. according to the measuring method described in claim 1-8, it is characterized in that, described sample is Chinese Traditional Medicine intermediate.

10. measuring method according to claim 9, is characterized in that, described sample is complex prescription glue mucilage herbal extract.