CN105259136A - Measuring-point-free temperature correction method of near-infrared correction model - Google Patents

Abstract

The invention relates to a measuring-point-free temperature correction modeling method used in near-infrared spectrum measurement. The measuring-point-free temperature correction modeling method comprises the steps that representative samples are collected, and it is ensured that the range of measurement requirements can be covered by the physical parameters to be measured of the samples; near-infrared spectrums of the samples are acquired at different temperature levels; preprocessing and principal component analysis are conducted on the acquired spectrums respectively according to the temperatures and the physical parameters to be measured; the two different spectrums obtained through preprocessing are combined, and combined spectral data are produced; finally, principal component analysis is conducted on the combined spectrums, and a physical parameter correction model is established through partial least squares. The measuring-point-free temperature correction modeling method uses the temperatures as separated implied factor variables to be involved in the near-infrared modeling process, thus when the near-infrared measurement is performed, measurement of physical properties at different temperatures can be completed by relying on the temperature adaptability of the model itself without direct temperature measurement information and relevant calculation, and the established model can have better universality and temperature adaptability.

Description

No-measuring-point temperature correction method for near-infrared correction model

technical field

The invention relates to a non-measuring-point temperature correction modeling method in near-infrared spectrum measurement, which is suitable for the viscosity of substances easily affected by the ambient temperature, the concentration of alanine in the fermentation process, the quality of food, the quality of agricultural products, the quality of medicines, gasoline and oil products, etc. Rapid detection can also be used for the measurement of human non-invasive blood sugar concentration, soil composition and mineral composition.

Background technique

In recent years, near-infrared spectroscopy has been widely used in petrochemical, food, agriculture, medicine and other industries due to its advantages of rapid detection, non-destructive testing, no chemical pollution, easy operation, and simple sample preparation, and has become the fastest growing One of qualitative and quantitative analysis techniques. The absorption in the near-infrared spectral region mainly comes from the state change caused by molecular vibration or rotation. The vibration of each group is easily affected by external conditions such as temperature. Especially when measuring liquid samples, the increase in temperature will cause stretching vibration. The number of hydroxyl groups decreases and the number of free vibrations increases, resulting in a shift in the vibration spectrum, so that the near-infrared spectrum model established at a specific temperature can only be applied to the quality analysis of samples at this temperature, while the effect of quality analysis on samples at different temperatures Not ideal, this shortcoming greatly limits the application of NIR spectroscopy modeling techniques.

In order to obtain better analysis accuracy, measurement at a constant temperature can effectively reduce the influence of temperature changes, but in practical applications, the temperature cannot be precisely controlled, so some methods to solve the influence of temperature on near-infrared spectroscopy have been proposed one after another, such as spectral preprocessing The method eliminates the influence of temperature in the spectrum; selects the band that is not sensitive to the influence of temperature to establish an analysis model; adopts the method of internal correction to include the temperature change information in the mathematical model; measures the temperature of the sample while collecting the spectrum, and establishes a temperature correction model, etc. Wait. These methods can be used to overcome the interference of the temperature change of the sample to be tested on the quantitative analysis model. However, there is no general rule to judge which method to use in which case, but should be selected according to the specific problem. Therefore, establishing a more general and temperature-adaptable near-infrared detection and correction model under temperature changes is critical to the effective application of near-infrared technology.

Contents of the invention

The method proposed in the present invention obtains sample spectra at different temperature levels, and uses temperature as a separate hidden factor variable to participate in the near-infrared modeling process, so when using near-infrared measurements, it can rely on the adaptability of the model itself to temperature The physical property measurement at different temperatures does not require direct temperature measurement information and related calculations, so that the established model has better versatility and temperature adaptability.

In order to achieve the above object, the present invention adopts the following technical solutions:

The steps of the present invention are divided into two parts. The first part is the experimental design and spectrum collection of modeling data; the second part is the preprocessing of near-infrared spectra and the establishment of calibration models.

The experimental equipment for modeling data includes, (1) a sample cell that can adjust the sample temperature (2) a temperature measurer that can display temperature changes (3) a near-infrared spectrum collection instrument (4) an optical sensor that does not significantly affect the sample temperature probe. (5) A computer recording device connected to the near-infrared spectrum collection instrument.

Experiment of the present invention and data collection steps are as follows:

Experimental step 1: Confirm the maximum and minimum temperature values of the sample. Divide the temperature range into levels. Each temperature level is generally 5 times larger than the resolution of the temperature measuring instrument to achieve effective discrimination accuracy.

Experimental step 2: within the range of the highest and lowest temperature, obtain the original standard analysis data at the specified temperature for all the physical parameters of the samples.

Experimental Step 3: Collect spectral data of samples at different temperature levels. At the same time record the corresponding sample temperature value. This temperature value is involved in the modeling as an implicit factor, and recording the exact value of the temperature is not necessary for this method.

The steps of temperature as a separate latent factor variable modeling method are as follows:

Modeling step 1: Preprocessing the spectrum with the temperature model as the target. Perform the first-order derivative or second-order derivative operation on the original spectrum to generate the first-order derivative spectrum or the second-order derivative spectrum. Here, the determination of the derivative order varies with the characteristics of samples and physical parameters. For example, for high-molecular high-viscosity samples, the second derivative is better. The first derivative is better for low viscosity samples.

Modeling step 2: Perform principal component analysis (PCA) on the derivative spectrum generated above, and remove statistical outliers, so that the principal component mode of the entire derivative spectrum data is within a statistical reliability.

Modeling step 3: Preprocessing the original spectrum with the target physical parameter model as the target. These preprocessing include one or more of the following algorithm superposition operation: first derivative, second derivative, max-min normalization, base baseline correction, scatter correction, constant bias correction, etc. The determination of the preprocessing algorithm here varies with the physical parameters to be measured.

Modeling step 4: Perform principal component analysis (PCA) on the preprocessed spectrum generated above, and remove statistical outliers, so that the principal component mode of the entire preprocessed spectral data is within a statistical reliability.

Modeling Step 5: Merge the above-formed derivative spectrum targeting temperature with the preprocessed spectrum targeting physical parameters to be measured.

Modeling step 6: Perform principal component analysis (PCA) on the merged spectrum generated above, and remove statistical outliers, so that the principal component mode of the entire merged spectral data is within a statistical reliability.

The original analytical value of the measured physical parameters at a specified temperature is used as the predictive variable, and the combined spectral wavenumber is used as the independent variable. Use the partial least squares algorithm (PLS) to establish a physical parameter prediction model:

P＝B ₁ y ₁ +B ₂ y ₂ +…B _n y _n +A ₁ x ₁ +A ₂ x ₂ +…A _n x _n

Here, P is the measured value of the physical variable at the specified temperature, B _i , A _i , i=1, 2,...n are the regression coefficients, y _i , _xi are the preprocessed spectrum and the derivative spectrum at wavenumber i= Values at 1,2,…n.

The present invention takes temperature as a separate implicit factor variable in the near-infrared modeling process, so when using near-infrared measurement, it can rely on the adaptability of the model itself to temperature to complete the physical property measurement at different temperatures without direct temperature measurement Information and related calculations make the established model have a better compensation effect on temperature, so it has better versatility.

Description of drawings

Figure 1 Temperature compensation experimental device without measuring point

Figure 2 The second derivative pretreatment spectrum of a polymer compound

Figure 3 PCA mode diagram based on the second derivative preprocessing spectrum

Figure 4 The first derivative preprocessed spectrum

Fig.5 Principal element mode of the first derivative spectrum

Figure 6 block diagram of implementation steps

Figure 7 Combined spectra at different temperatures

Figure 8 PCA model diagram of combined spectra

Figure 9. Viscosity model resulting from merging spectra.

Figure 10 Wavenumbers used by the viscosity model.

The adaptability of Fig. 11 method result of the present invention to temperature

detailed description

The following takes the viscosity measurement of a polymer compound as an example to illustrate the specific implementation method. This example does not constitute a limitation on the scope of the method of the present invention.

The block diagram of the implementation steps of the non-measuring point temperature correction method of the near-infrared correction model is shown in Figure 6, which specifically includes the following steps:

Step 1: Collect a representative sample, and ensure that the physical parameters of the sample to be measured can cover the range required for the measurement. The total number of samples is 40-60.

Step 2: Use the laboratory equipment shown in Figure 1 to collect the near-infrared spectra of each sample at five different temperature levels of 24°C, 35°C, 50°C, 60°C, and 70°C, and record the experimental conditions such as temperature, etc. .

Step 3: Perform temperature-targeted preprocessing and principal component analysis on the collected spectra to generate derivative spectral data. In the example, the second derivative processing and principal component analysis are performed on the high-molecular high-viscosity sample. The processing effect is shown in Figure 2, and the main element pattern is shown in Figure 3. The second-order derivative preprocessing is to re-extract the temperature-sensitive spectrum based on the first-order derivative, effectively reducing the overlap of temperature and physical parameters in the modeling wavenumber. In the PCA model diagram shown in Figure 3, there are two points that are far away from all other points. This point is a singular point, which is eliminated during modeling, so that the principal component model of the entire preprocessed spectral data is within a statistical confidence level.

Step 4: Perform preprocessing and principal component analysis (PCA) on the original spectrum with the target physical parameter model as the target. Generate preprocessed spectral data. In the example, first derivative preprocessing and principal component analysis are performed on polymer samples. The processed spectrum eliminates the up-and-down drift of the spectrum due to factors such as light source aging, probe vibration, and probe-sample contact, while retaining effective information on the influence of temperature on the spectral peak and shape. The preprocessed spectrum is shown in Figure 4. Fig. 5 is a schematic diagram of the principal component of the preprocessed spectrum. There are two singular points in Figure 5, which should be eliminated and no longer participate in modeling.

Step 5: Combine the derivative spectrum generated above with the preprocessed spectrum to generate combined spectral data. Figure 7 is the combined spectrum at different temperatures. In the combined spectrum in Figure 7, the left half is the first derivative, which provides effective spectral information for physical property modeling; the right half, the second derivative, provides spectral information for temperature compensation.

Step 6: Perform principal component analysis (PCA) on the merged spectrum generated above, and remove statistical outliers, so that the principal component mode of the entire merged spectral data is within a statistical reliability. Fig. 8 is a PCA model diagram of combined spectra. There are three singular points in Figure 8, which should be eliminated and no longer participate in modeling.

Step 7: Use the original analysis value of the physical property parameter to be measured as the predictor variable, and use the combined spectral wavenumber as the independent variable, and use the partial least squares algorithm (PLS) to establish a physical property parameter prediction model:

P＝B ₁ y ₁ +B ₂ y ₂ +…B _n y _n +A ₁ x ₁ +A ₂ x ₂ +…A _n x _n

Here, P is the measured value of the physical variable at the specified temperature, B _i , A _i , i=1, 2,...n are the regression coefficients, y _i , _xi are the preprocessed spectrum and the derivative spectrum at wavenumber i= Values at 1,2,…n.

Figure 9 is the viscosity model generated by merging the spectra, Figure 10 is the wavenumber used by the viscosity model, and Figure 11 is the comparison of the results to the temperature adaptability. The correlation between the predicted value and the measured value of the composite spectrum model in Fig. ⁹ is 0.98, and the model precision R2 is 0.97. In the band range shown in Fig. 10, the spectral band of the first order derivative is 9056-4765 cm ^-1 , and the spectral band of the second order derivative is selected as 6024-4528 cm ^-1 . Fig. 11 is the comparison between the temperature compensation algorithm without measuring points proposed by the present invention and the modeling algorithm fixed at a temperature of 50 degrees. From the figure, it can be seen that the measurement results of the fixed temperature model have greater sensitivity to temperature changes, and this The model established by the inventive method has a better compensation effect on temperature.

Claims (5)

1. a method for modeling without measuring point temperature correction in near-infrared spectrum measurement, is characterized in that the method comprises the steps: Step 1: Obtain the original standard data of physical property parameters of multiple samples at a specified temperature, and collect near-infrared spectra at different temperature levels; Step 2: Perform temperature-targeted preprocessing and principal component analysis on the near-infrared spectrum collected in step 1 to generate derivative spectral data; Target preprocessing and principal component analysis to generate preprocessing spectra; Step 3: combining the derivative spectrum generated in step 2 and the preprocessed spectrum generated in step 4 to generate combined spectral data; Step 4: Perform principal component analysis on the merged spectral data described in step 3, and remove statistical outliers, so that the principal component modes of the entire merged spectral data are within a statistical reliability; Step 5: The original analysis value of the physical property parameter to be measured is used as the predictive variable, and the combined spectral wavenumber is used as the independent variable to establish a physical property parameter prediction model. 2. the non-measuring point temperature correction method of near-infrared correction model according to claim 1, it is characterized in that: the temperature range of collecting sample near-infrared spectrum covers real-time physical property parameter measurement temperature range in the described step 1; The physical parameters cover the range required by the measurement and have sufficient distinguishable intervals. 3. The method for correcting the temperature without measuring points of the near-infrared correction model according to claim 1, characterized in that: the spectral preprocessing aimed at temperature in the step 2 is a second derivative. 4. The non-measuring point temperature correction method of the near-infrared correction model according to claim 1, characterized in that: the spectral preprocessing aimed at the physical parameters in the step 2 includes the superposition operation of one or more of the following algorithms : first derivative, second derivative, max-min normalization, base baseline correction, scatter correction, constant bias correction. 5. the non-measuring point temperature correction method of near-infrared correction model according to claim 1, it is characterized in that: the establishment of physical parameter prediction model in the described step 5 adopts partial least squares algorithm to carry out the linearization of temperature as hidden variable return: P＝B ₁ y ₁ +B ₂ y ₂ +…B _n y _n +A ₁ x ₁ +A ₂ x ₂ +…A _n x _n Here, P is the measured value of the physical variable at the specified temperature, B _i , A _i , i=1, 2,...n are the regression coefficients, y _i , _xi are the preprocessed spectrum and the derivative spectrum at wavenumber i= Values at 1,2,…n.