CN103398971A - Chemometrics method for determining cetane number of diesel oil - Google Patents

Abstract

The invention discloses a chemometrics method for determining a cetane number of diesel oil. The chemometrics method comprises the following steps of firstly, carrying out standardized treatment on determined original near infrared spectrum data of the diesel oil and screening variables (wavelengths) in the data by adopting an uninformative variable elimination UVE method; then, establishing a partial least squares PLS model between a known diesel oil spectrum matrix containing different variable numbers and the corresponding cetane number; predicating the cetane number of an unknown sample by adopting a multi-model consensus method; calculating a total predicated value by carrying out weighting method calculation on a predicated value of a single model; finally, calculating the predicated values by an averaging method through repeating a predication process for a plurality of times so as to obtain the final predicated result. The method disclosed by the invention is applied to the determination for the cetane number of an actual diesel oil sample to obtain a good effect; the performance for predicating the cetane number of the diesel oil by combining the PLS linear model with a near-infrared spectrometry is improved in the aspects of accuracy and stability; the cetane number of the diesel oil can be detected accurately, rapidly and simply.

Description

Chemometrics method for determining cetane number of diesel oil

Technical Field

The invention relates to a method for measuring the cetane number of diesel oil by combining chemometrics with a near infrared spectrum technology.

Background

The structural groups that absorb in the near infrared spectral region (NIRs) are predominantly X-H, and the various compounds that make up the petroleum product have this structure. Therefore, the NIRs are almost used in each link of petrochemical engineering detection by the characteristics of simplicity, rapidness, no damage, good reproducibility and the like, and play an important role. As a complex hydrocarbon mixture, the diesel oil has a very complex near infrared spectrum structure and a serious spectrogram overlapping phenomenon. Therefore, the characteristic information extraction of the near infrared spectrum data of the diesel oil is carried out, and a relation model between indexes (cetane number, density, condensation point and the like) and spectrum data is established, so that the method is a key step for ensuring the quality of the oil product.

The Cetane Number (CN) of diesel oil is a main index for measuring the combustion performance of diesel oil. The traditional cetane number measuring method adopts an ASTM D613 experimental method internationally, and GB/T386 established in China also refers to the ASTM D613 method. The method needs a large amount of oil samples, takes long time and has high measuring cost. At present, the near infrared spectrum technology is combined with a chemometric method for measuring the cetane number of diesel oil, and the method is reported in documents. The cetane number of diesel oil and a spectrum signal have a nonlinear relation, and the modeling is better than linear algorithms such as Multiple Linear Regression (MLR) and Partial Least Squares (PLS) by adopting a nonlinear algorithm Support Vector Machine (SVM), an Artificial Neural Network (ANN) and the like.

The invention adopts the non-information variable elimination (UVE) method combined with the multi-model consensus PLS method to establish a novel quantitative model, improves the performance of predicting the cetane number of diesel oil by a PLS linear model, and obtains good effect.

Disclosure of Invention

The invention aims to provide a method for rapidly and accurately detecting the cetane number of diesel oil, and provides a new correction model for the quantitative regression analysis of near infrared spectrum and the determination of the cetane number of the diesel oil.

The principle of the method is that firstly, variables contained in the diesel near infrared spectrum data after standardized processing are screened by a non-information variable elimination (UVE) method, and then a plurality of PLS models containing different variable numbers are used for jointly predicting the cetane number of an unknown diesel sample, so that higher prediction accuracy and stability than those of a single PLS model are obtained.

The method comprises the following specific steps:

1. the original near infrared spectrum of the diesel oil is subjected to standardization treatment, and a data set is randomly divided into a training set, a checking set and a prediction set.

2. The number of factors for the PLS model was established according to the Monte Carlo cross validation in combination with the F-test method. The Root Mean Square Error (RMSEP) of the test set predicted values is used as an optimization criterion for all parameters in the model.

$\mathrm{RMSEP}={\left[\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(y_i-{\hat{y}}_i)}^2\right]}^{1/2},---\left(1\right)$

Wherein,

denotes the predicted value of the i-th sample, y_iRepresenting the true value of the ith sample, and n is the number of predicted samples.

3. According to the size of a stability value of a variable (wavelength) in a UVE algorithm, different variable screening thresholds are established, wavelength optimization screening is carried out, and therefore a plurality of PLS models with different variable numbers are established.

4. Randomly selecting a certain number of models from the models established in the step 3, and optimizing the number of the selected models through predicting a test set sample; and predicting all training set samples by using the selected fixed number of models, calculating the prediction error of each model and the weight of the corresponding model, and finally calculating the prediction result of each model by a weighting method to obtain a total prediction value.

5. The prediction set samples are predicted using the same prediction method as in step 4. And then, setting the times of repeated prediction, and combining the predicted values of the prediction set in the multiple prediction processes by an averaging method to obtain the final prediction result of the prediction set. And calculating the mean and standard deviation (sigma) of the Root Mean Square Error (RMSEP) of the prediction results of the prediction set.

The chemometric method improves the performance of predicting the cetane number of the diesel oil by combining a PLS linear model with near infrared spectrum from the aspects of accuracy and stability, and can accurately, quickly and simply detect the cetane number in the diesel oil.

Drawings

FIG. 1 is a flow chart of the method of the present invention

FIG. 2 is a near infrared spectrum of diesel oil in the example

Detailed Description

Example (b):

NIR spectra and corresponding cetane number profiles for 245 diesel samples were downloaded from http:// software. The near infrared spectrum scanning range is 750-1550nm, the scanning interval is 2nm, and the scanning interval comprises 401 wavelength points (variables). The method comprises the following specific steps:

1. and carrying out standardization processing on the original NIR spectral data to obtain a standardized data matrix. The normalized matrix has a mean value of 0 and a variance of 1 for each variable in the variable space.

2. The original data set was randomly divided into 155 samples as a training set, 45 samples as a test set, and the remaining 45 samples as a prediction set.

3. A Partial Least Squares (PLS) correlation model was established between NIR spectra of diesel training set samples and corresponding cetane number data according to Monte Carlo (Monte Carlo) cross validation in combination with the F-test method. The number of factors for determining the PLS predictive model method is 13, taking Root Mean Square Error (RMSEP) of the predicted values of the test set samples as an optimization criterion.

4. And obtaining the stability value of each variable by a non-information variable UVE (ultraviolet ray) elimination method, and setting K times of the maximum value in the absolute values of the stability values of the added noise as a threshold value for variable screening. The K value is in the interval from 0 to 0.8, and the value is taken at intervals of 0.02. In the normalized data obtained in step 1, variables having stability values with absolute values greater than or equal to the size of the screening threshold are retained. Thus 40 PLS models were obtained which retained a different number of variables.

5. And (4) randomly selecting 25 models (the number of the models is determined after optimization) from the 40 models established in the step (4), and respectively modeling by using a PLS method. Calculating all training set samples x for each model prediction_iRelative error δ between predicted value and actual value of (i ═ 1.. multidot.n)_i，t

${\mathrm{\δ}}_{i,t}=\frac{e_{i,t}}{y_i}---\left(2\right)$

Wherein,

t (t ═ 1.., 25) is the model number.

Then delta will be_i，tThe normalization processing is carried out, and the normalization processing is carried out,

${\mathrm{\δ}}_{i,t}=\frac{{\mathrm{\δ}}_{i,t}}{\max \left({\mathrm{\δ}}_{i,t}\right)};---\left(3\right)$

then delta will be_i，tThe average value is taken to obtain the prediction error beta of the model_tThereby calculating each model C_tIs/are as follows

${\mathrm{\α}}_t=1/21g\frac{{1-\mathrm{\β}}_t}{{\mathrm{\β}}_t};---\left(4\right)$

Then calculate each model C_tThe weight of (a) is determined,

$w_t={\mathrm{\α}}_t/\underset{t=1}{\overset{25}{\mathrm{\Σ}}}{\mathrm{\α}}_t;---\left(5\right)$

calculating the results of 25 PLS model prediction set samples by a weighting method to obtain a total prediction value,

$C\left(x_i\right)=\underset{t=1}{\overset{25}{\mathrm{\Σ}}}{w}_t{C}_t\left(x_i\right).---\left(6\right)$

6. and (5) repeating the operation 100 times to predict the cetane number of the prediction set sample. The mean of the root mean square error RMSEP and the relative standard deviation (σ) of RMSEP of the predicted results are calculated.

7. The method of the present invention having the above steps is named multi-model BPLS method. Each model C in the step 5 is processed_tThe weights of (a) are set to be equal, namely, a total predicted value is obtained through a simple averaging method, other steps are not changed, and the method is named as a multi-model CPLS method. The operation is repeated 100 times, and the cetane number of the prediction set sample is predicted. The results of the two methods are compared and are shown in table 1. The result shows that the chemometric method provided by the invention has good effect on the accuracy and stability of prediction, and provides a novel and practical correction model for the diesel cetane number determination and the quantitative regression analysis of the near infrared spectrum.

TABLE 1 comparison of results of different methods for predicting cetane number of diesel oil samples

Claims (5)

1. A chemometric method for determining the cetane number of diesel oil, which is characterized by comprising the following steps:

(1) measuring to obtain near infrared spectrum data of diesel oil, and carrying out standardization treatment on the data;

(2) screening variables contained in the data obtained in the step (1) by adopting a non-information variable UVE elimination method;

(3) establishing a Partial Least Squares (PLS) correlation model between diesel spectrum matrixes containing different variable quantities and corresponding diesel cetane numbers;

(4) and (4) randomly extracting a certain number of models established in the step (3) and jointly predicting the cetane number in the unknown diesel oil sample. Then, calculating the total predicted value of the prediction results of the models by a weighting method;

(5) and (4) repeatedly operating the step (4) for multiple times, and combining the prediction results in the prediction processes through an averaging method to finally give a prediction result of the cetane number of the unknown diesel oil sample.

2. The chemometric method for determining the cetane number of diesel oil as claimed in claim 1, wherein, in the step (1), the wavelength range of the characteristic region in the near infrared spectrum is 750-1550 nm.

3. The chemometric method for determining the cetane number of diesel oil according to claim 1, characterized in that in step (2), K times the maximum value of the absolute values of the stability values of the additive noise (K is in the interval from 0 to 0.8 and is 0.02 per interval) in the UVE method is taken as the threshold value of the variable screening, and the variables with the absolute values of the stability values greater than or equal to the threshold value are retained.

4. The chemometric method for determining the cetane number of diesel oil according to claim 1, wherein in the step (4), 25 models are randomly selected from the 40 PLS models established in the step (3) in a ratio of 5/8.

5. The chemometric method for measuring the cetane number of diesel oil according to claim 1, wherein in the step (4), the prediction error of each model and the weight of the model are calculated, and the total prediction result is calculated from the prediction results of these models by a weighting method:

according to each model C_tIs predicted by the prediction error beta_tCalculating

Wherein t (t ═ 1.., 25) is a model number;

each model C_tHas a weight of $w_t={\mathrm{\α}}_t/\underset{t=1}{\overset{25}{\mathrm{\Σ}}}{\mathrm{\α}}_t---\left(2\right)$

The final prediction result is $C\left(x_i\right)=\underset{t=1}{\overset{25}{\mathrm{\Σ}}}{w}_t{C}_t\left(x_i\right)---\left(3\right)$

Wherein x is_i(i 1.., n) is a prediction sample.

Images