CN105717066B - A kind of near infrared spectrum identification model based on weighted correlation coefficient - Google Patents

Abstract

The present invention establishes a near-infrared spectrum identification model, uses a near-infrared spectrometer to scan the spectra of a series of normal similar products, and calculates the average spectrum as a reference spectrum; then calculates the weighted correlation coefficient between the product spectrum and the reference spectrum, and uses the average Values and standard deviations are used to calculate the identification interval and establish the identification model. Compared with traditional instrument analysis such as chromatography and mass spectrometry, it has the advantages of being green, environmentally friendly, simple and fast, and easy to operate, and the built model has high recognition accuracy, high detection efficiency, and low cost. Authenticity identification provides technical support.

Description

A Near Infrared Spectral Identification Model Based on Weighted Correlation Coefficient

technical field

The invention relates to the technical field of near-infrared spectroscopy, in particular to a near-infrared spectroscopy identification model based on a weighted correlation coefficient, which can be used for authenticity identification and quality stability analysis of cigarette products.

Background technique

Near-infrared spectroscopy has the characteristics of efficient, green, and environmentally friendly analysis, so it has become one of the fastest-growing and eye-catching spectral analysis techniques in recent years. According to the provisions of the American Society for Testing and Materials (ASTM), its wavelength range is 780-2526nm. The absorption of molecules in the NIR region is mainly composed of combined frequency absorption and double frequency absorption of groups such as C-H, 0-H, N-H and C=0. The absorption intensity in this region is low, the bands are complex, and the overlap is serious, so it is impossible to use classical qualitative , Quantitative methods must use methods such as multivariate statistics, curve fitting, and cluster analysis in chemometrics to carry out calibration modeling, and combine appropriate models to achieve rapid multi-component analysis. Near-infrared spectroscopy has the advantages of fast, non-destructive, and real-time detection, and has become a powerful tool for industrial product analysis. However, near-infrared spectroscopy has weak features, large amount of data, visual recognition and traditional matching algorithms are difficult to obtain reliable results. Therefore, it is urgent to develop an effective, fast and highly automated recognition algorithm.

For a long time, the internal quality of cigarette products has been mainly characterized by sensory evaluation, lacking an intuitive and quantitative description method. With the increasingly fierce competition in the product market and the increasing automation of industrial enterprises, the stability and control of product quality are becoming more and more prominent. There is an urgent need for fast, efficient and simple analysis methods for the evaluation and control of cigarette product quality stability.

Aiming at the above problems, the present invention proposes a recognition model based on weighted correlation coefficients. The model is based on a series of near-infrared spectra of normal similar products, and the identification model is established through weighted correlation coefficients. In the application of the model, the near-infrared spectrum that is judged to be a normal product of the same kind can be added to the model, so as to realize the update of the recognition model, making it more adaptable and more accurate.

In the present invention, the similarity of the spectrum is measured by the weighted correlation coefficient, and the weighted correlation coefficient can effectively use the spectral features for the calculation of the similarity, thereby improving the reliability of the result. The near-infrared spectra of the above-mentioned normal similar products are highly similar to the reference spectra, but there are differences; they not only reflect common characteristics, but also reflect individual differences. A range of spectral similarity is thus determined, and the products in this range are similar products, otherwise they are non-similar products or abnormal products.

Contents of the invention

In order to further improve the recognition accuracy of the model, the present invention proposes a method based on weighted correlation coefficient to extract the characteristic spectrum. The recognition model proposed by the invention is composed of reference spectrum and recognition interval.

The present invention uses a near-infrared spectrometer to scan the spectra of a series of normal similar products, then extracts the characteristic data of each spectrum, and uses the average spectrum as a reference spectrum; calculates the weighted correlation coefficient between the product spectrum and the reference spectrum, and calculates the weighted correlation coefficient between the product spectrum and the reference spectrum. Values and standard deviations are used to calculate the identification interval, and the identification model is established. The identification model is composed of the reference spectrum and the identification interval.

The specific modeling steps are as follows:

(1) Firstly, pre-treat m samples of the same kind produced in different batches, scan the sample spectrum by using near-infrared, and use the vector s _i to represent the spectral data of the i-th sample (i=1,2,..., m), each spectrum contains n data points.

(2) Taking the spectral vector _si as a row vector, sort out the data matrix S in the following form,

Calculate the average spectrum by

(3) Calculate all spectra s and The weighted correlation coefficient wcc, where w is the weight

Among them, s _j and represent the spectra s and The jth data point of ; w _j represents the jth data point of the weight vector w.

(4) The weight vector w in step 3 is calculated by the following formula

Among them, the vector d is calculated by the following formula

in, The superscript T indicates matrix transpose.

(5) Calculate the mean and standard deviation of wcc, expressed as and d. Create an identification interval Where k is a proportional coefficient, set according to the calibration set data, so that all wcc are greater than Will

As the identification interval of this type of product.

For an unknown product, first scan its near-infrared spectrum, denoted by s _x , and then calculate the weighted correlation coefficient wcc _x by the formula in step 3. Determine whether wcc _x is in the recognition interval If yes, then the unknown product is considered to be the same as the calibration set product; add s _x to the calibration set, repeat steps 1-4, and update the identification interval. If not, then the unknown product is considered to be different from the calibration set product.

Apply the weighted correlation coefficient calculation formula to calculate the mean value and standard deviation of the weighted correlation coefficient wcc of all similar spectra, where the mean value is expressed as to represent, the standard deviation is represented by d, and the identification interval is established where k is the scaling factor.

According to the identification interval calculated by the weighted correlation coefficient and the data setting of the calibration set, the weighted correlation coefficient wcc of all spectra of the same kind is greater than The identification interval for this type of product is

When applying the model, the weighted correlation coefficient wcc _x is calculated by scanning the spectrum of the sample to be analyzed. If the coefficient falls into the identification interval It can be determined that it is a normal product of the same kind.

The model is based on a series of near-infrared spectra of normal similar products, and the identification model is established through weighted correlation coefficients.

The near-infrared spectrum identification model based on the weighted correlation coefficient: includes the step of crushing the sample into 40-80 meshes before scanning. The samples are shredded tobacco, tobacco stems and/or tobacco dust.

In the application of the model, the near-infrared spectrum that is judged to be a normal product of the same kind can be added to the model, so as to realize the supplement and update of the identification model, make the model more adaptable and the prediction result more accurate.

Compared with the prior art, the present invention has the following significant advantages:

1. A method for establishing a product authenticity and stability identification model based on the weighted correlation coefficient proposed by the present invention. The weighted correlation coefficient can effectively use spectral features for similarity calculations, greatly improving the reliability of the identification model .

2. Through the weighted correlation coefficient, determine a spectral similarity interval, which can not only reflect the common characteristics of each spectrum, but also reflect individual differences. Using the high similarity between the near-infrared spectrum and the reference spectrum of normal similar products, the regional space is established. The sample scanning spectrum falls within this range as the same product, and the ones outside the range are non-similar products or abnormal products, which can effectively avoid the phenomenon of inaccurate judgment. The accuracy of model recognition is improved, and technical support is provided for the quality stability analysis and authenticity identification of cigarette manufacturers' products.

3. Compared with traditional chromatography, mass spectrometry and other instrument analysis, the near-infrared spectroscopy technology adopted does not use chemical reagents in the whole analysis process, which has the advantages of green, environmental protection, simple and fast, and easy operation. Matrix and weighting are used in model establishment Chemometric tools such as correlation coefficient, the built model has high recognition accuracy, high detection efficiency and low cost.

Description of drawings:

Fig. 1 is the modeling flowchart of the present invention;

Fig. 2 is the near-infrared scanning original spectrogram of shredded tobacco;

Fig. 3 is the identification model established by near-infrared spectrum of shredded tobacco of brand A cigarette;

Fig. 4 is A, B trade mark classification recognition model;

Detailed ways:

The specific implementation manners of the present invention will be described in further detail below in conjunction with the accompanying drawings.

The modeling process of the present invention and the embodiment is as follows: firstly, the experimental design is carried out, the sample is collected according to the design, the representative sample collected is pretreated, the spectrum is collected with a near-infrared spectrometer, and the collected spectral parameters are optimized. , Spectral preprocessing methods use Norris derivative smoothing filter, differential processing, multivariate scattering correction, standard normalization and other methods; band selection uses partial least squares method, genetic algorithm, non-informative variable elimination and other methods to optimize the spectral band. After the spectrum is optimized, a qualitative analysis model is established, and the near-infrared recognition model is established according to the steps of extracting spectral characteristic data, establishing a data matrix, calculating the average spectrum, and calculating the weighted correlation coefficient. After the model is established, the spectrum of the sample to be tested is scanned and the model is used for analysis. see picture 1

Embodiment one

In this embodiment, the recognition model is used for the stability recognition of similar products.

1. Experimental equipment

The MPA Fourier near-infrared spectrometer produced by BRUKER (Germany), and the 1095Cyclotec (XF-98B) cyclone precision pulverizer.

2. Sample collection

In order to make the established recognition model widely applicable in predicting cigarette products produced by different machines at different times, this experiment sample selected the same brand (A) cigarette products produced by three different machines from January to December 2015 83 normal samples were used for model building, and 31 unknown samples were selected for external validation of the model.

3. Sample preparation

The shredded cigarette tobacco was stripped off and dried in an oven at 40°C to keep the water content of the sample basically consistent, then fully pulverized with a 1095Cyclotec (XF-98B) cyclone precision grinder, and passed through a 60-mesh sieve.

4. Spectral scanning and data processing

Cigarette shredded tobacco sample spectrogram scanning was carried out by MPA Fourier near-infrared spectrometer (with near-infrared quantitative analysis diffuse reflection gold-plated large integrating sphere and sample rotator sampling accessories) produced by BRUKER (Germany), and the qualitative analysis software Bruker OPUS was used QUANT6.5 processes the spectrum, the specific operation is as follows: put the tobacco powder into the sample cup, the height in the cup is about 3cm, press the weight on the sample for 10s and take it out, wipe the quartz glass at the bottom of the cup with gauze Clean, then place the sample cup on the rotating platform for NIR scanning. The operating parameters are: spectral scanning range 12000-4000 cm ^-1 , spectral resolution 8 cm ^-1 , scanning times 64 times (about 30S). Spectral data is collected in transmission and processed as the first differential of the absorption spectrum. The original scanned image of cigarette shredded tobacco is shown in Figure 2. In the modeling process, in order to eliminate the influence of noise and baseline, the original spectrum after scanning was preprocessed by using the first derivative 9-point smoothing (Savitzky-Golay). After the samples were scanned, the spectral data were processed with statistical software.

5. Recognition model establishment

The steps to build the model are as follows:

5.1 Taking the spectral vector _si as a row vector, sort out the data matrix S in the following form,

Calculate the average spectrum by

By the above formula, calculate

5.2 Calculate all spectra s with The weighted correlation coefficient wcc, where w is the weight

Among them, s _j and represent the spectra s and The jth data point of ; w _j represents the jth data point of the weight vector w.

5.3 The weight vector w in the above steps is calculated by the following formula

Among them, the vector d is calculated by the following formula

in, The superscript T indicates matrix transpose.

5.4 Calculate the mean and standard deviation of wcc, expressed as and d. Create an identification interval Where k is a proportional coefficient, set according to the calibration set data, so that all wcc are greater than Will As the identification interval of this type of product.

By scanning the average spectrum of 83 cut tobacco spectra of brand A cigarettes, use this as the reference spectrum; calculate the weighted correlation coefficients of 83 cut tobacco spectra of this brand and the reference spectrum by the above formula; the mean value and standard deviation of these weighted correlation coefficients are respectively 0.9998 and 0.0002025; through analysis, it is found that all weighted correlation coefficients are greater than 0.9998–3*0.0002025=0.9992, so the identification interval of this type of product is determined as [0.9992,1], and the identification established by the near-infrared spectrum of shredded tobacco of brand A cigarette The model is shown in Figure 3:

After the same batch of samples were pretreated, they were placed in the air for 6 hours, and then the spectrum was scanned by the upper level, and the weighted correlation coefficient between the near-infrared spectrum of the product and the reference spectrum in the model was calculated, and the calculation result was 0.9985. This value is in the model identification interval [0.9992,1], so it is determined that the product is different from the normal product described by the identification model. In fact, the product was placed in the air for a long time before the near-infrared spectrum scanning, and the moisture absorption was serious. Although the essence of the shredded tobacco has not changed, the moisture has changed, and the quality has been affected. No abnormality can be seen from the spectrum, but the spectrum after scanning The weighted correlation coefficient of is much lower than that of similar products, indicating that the model can be used for the quality stability analysis of the same product.

6. Model validation

In order to better verify the recognition ability of the model, this experiment adopts the method of external verification, selects 31 batches of samples that have not participated in the modeling, and uses the built model to identify A cigarettes produced by different batches and different machines. The results See Table 1:

Table 1 Recognition results of "A" brand product feature model

The results showed that all 31 samples from different batches and different machines were successfully identified, with a recognition rate of 100%, indicating that the model built had a high prediction accuracy and could be used for quality stability analysis of cigarette products.

Embodiment two

In this embodiment, the identification model is used to identify the authenticity of the product.

1. Experimental equipment

The MPA Fourier near-infrared spectrometer produced by BRUKER (Germany), and the 1095Cyclotec (XF-98B) cyclone precision pulverizer.

2. Sample collection

The cigarette samples selected in this example are brand A and B, and 5 different machines were selected for the product, and the production time was from January to December 2015. A total of 83 normal samples of brand A were selected for model building, and brand B was selected. The 17 unknown samples were used to identify the authenticity of the model.

3. Sample preparation

The shredded cigarette tobacco was stripped off and dried in an oven at 40°C to keep the water content of the sample basically consistent, then fully pulverized with a 1095Cyclotec (XF-98B) cyclone precision grinder, and passed through an 80-mesh sieve.

Spectrum scanning and data processing and model establishment method of the present embodiment are the same as embodiment one, and the built A, B trade mark classification identification model is shown in Fig. 4:

The near-infrared spectra of 51 non-similar products with brand B were scanned by near-infrared, and the weighted correlation coefficients of the reference spectra in the identification model were calculated and identified. The results are shown in solid points in Figure 2. It can be seen from the figure that all these data points are outside the recognition interval, and the recognition rate is 100%, so it is determined as a non-similar product. This conclusion is completely consistent with the actual situation, thus proving the validity of the recognition model.

4. Model verification

In order to better verify the identification ability of the model, this experiment adopts the method of external verification, selects 29 batches of cigarette samples of different brands that have not participated in the modeling, and uses the built model to compare the counterfeit cigarettes of brand A and brands B and B collected in the market. The cigarette of C is identified, and the specific results are shown in Table 1:

Table 1 Recognition results of "A" brand product feature model

The results show that 29 samples of cigarettes that are not genuine brand A are successfully identified, and the identification rate is 100%, which shows that the model built can be used to identify the authenticity of products.

The above-mentioned embodiments of the present invention are only examples for clearly illustrating the present invention, rather than limiting the implementation of the present invention. For those of ordinary skill in the art, other different forms can also be made on the basis of the above description. The changes or changes of the present invention cannot be exhaustively listed here, and any obvious changes or changes that are derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims (6)

1. A near-infrared spectrum identification model based on a weighted correlation coefficient, using a near-infrared spectrometer to scan the spectra of a series of normal similar products, then extract the characteristic data of each spectrum, and use the average spectrum as a reference spectrum; calculate the product spectrum and Referring to the weighted correlation coefficient of the spectrum, the identification interval is calculated by the average value and standard deviation of the weighted correlation coefficient, and the identification model is established. The identification model is composed of the reference spectrum and the identification interval; The model building includes the following steps: (1) Spectrum scanning: scan the near-infrared spectrum of the sample to be tested, and extract the characteristic spectrum; (2) Establish a data matrix: represent the spectral data of the i-th sample (i=1,2,...,m) with a vector si, and each spectrum contains n data points; use _the spectral vector _si as a row vector , the form of the spectral data matrix S is as follows, (3) Calculate the average spectrum: calculate as follows on the basis of step (2) matrix formula: In the formula, (4) Calculate all spectra s and The weighted correlation coefficient of , expressed in wcc, the formula is as follows: In the formula, s _j and Denote the jth data point of the spectra s and s respectively; w _j represents the jth data point of the weight vector w, w is the weight; (5) The calculation formula of the weight vector w in step (4) is as follows: In the formula, the calculation formula of the vector d is as follows: In the formula, The superscript T indicates matrix transpose. 2. the near-infrared spectrum identification model based on weighted correlation coefficient according to claim 1, is characterized in that: apply weighted correlation coefficient computing formula, calculate the mean value and the standard deviation of the weighted correlation coefficient wcc of all similar spectra, wherein mean value uses to represent, the standard deviation is represented by d, and the identification interval is established where k is the scaling factor. 3. The near-infrared spectrum identification model based on weighted correlation coefficient according to claim 1 or 2, characterized in that: according to the identification interval calculated by weighted correlation coefficient and the calibration set data setting, the weighted correlation coefficient wcc of all similar spectra are greater than The identification interval for this type of product is 4. the near-infrared spectrum identification model based on weighted correlation coefficient according to claim 3 is characterized in that: when model is applied, by scanning the spectrum of sample to be analyzed, calculate weighted correlation coefficient wcc _x , if this coefficient falls into the identification interval It can be determined that it is a normal product of the same kind. 5. The near-infrared spectrum identification model based on weighted correlation coefficient according to claim 4, characterized in that: it includes the step of crushing the sample into 40-80 meshes before scanning. 6. The near-infrared spectrum identification model based on weighted correlation coefficient according to claim 5, characterized in that: the samples to which it belongs are shredded tobacco, tobacco stems and/or tobacco powder.