Intelligent modeling method for near infrared spectrum quantitative analysis

Abstract

The invention discloses an intelligent modeling method for near infrared spectrum quantitative analysis, and belongs to the technical field of chemometrics modeling. The method comprises the following steps: acquiring a training set of a near infrared spectrum of a tested sample, and performing sample preprocessing and spectrum preprocessing on the training set to obtain a preprocessed data set; dividing the preprocessed data set into a plurality of groups of corresponding training subsets and verification subsets, for each group of training subsets in each group of preprocessed data set, based on an extreme learning machine, establishing a plurality of extreme learning machine submodels, predicting the corresponding verification subsets, obtaining an alternative model library, integrating a plurality of optimal submodels, and obtaining a candidate model library; and obtaining a final model. The near infrared spectrum quantitative analysis intelligent modeling method is high in adaptability, easy to use and high in prediction accuracy, the generalization ability and stability of the model are effectively improved, and dependence on professional knowledge is reduced.

Description

Technical Field

[0001] The present application belongs to the technical field of chemometric modeling in the field of analytical chemistry, and in particular relates to an intelligent modeling method for near-infrared spectroscopy quantitative analysis. Background Art

[0002] In the field of near-infrared spectroscopy quantitative analysis modeling, multivariate correction technology plays a key role, including multiple linear regression (MLR), principal component regression (PCR), partial least squares regression (PLS), artificial neural network (ANN), support vector regression (SVR) and extreme learning machine (ELM) methods. Among them, PLS, as a classic linear correction technology, is highly favored due to its wide application; and extreme learning machine ELM, as an efficient feedforward neural network technology, has been applied in multiple near-infrared quantitative analysis scenarios with its fast learning speed and simplified implementation process.

[0003] The training process of the extreme learning machine (ELM) is mainly divided into two stages: first, the hidden layer weights are randomly generated and nonlinearly mapped through the activation function; second, the output layer weight matrix is ​​solved using the regularization method. Since the hidden layer weights of the extreme learning machine (ELM) are randomly initialized and the training process only involves analytical linear equations, the training efficiency of the extreme learning machine (ELM) is extremely high. In addition, compared with traditional neural networks, the extreme learning machine (ELM) is easier to avoid overfitting because of the randomness of its hidden layer weights and the omission of additional optimization steps, thereby enhancing the generalization ability of the model.

[0004] Ensemble learning, as an important branch of machine learning, improves the stability and predictive ability of the model by combining multiple different learners. In recent years, near-infrared quantitative modeling methods based on ensemble learning strategies have become increasingly popular. For example, patents such as CN201310537968.3, CN201710569312.8, CN201810857499.6, CN201810857673.7, CN105117525A, CN106650926A, and CN107389592A all involve different ensemble methods for near-infrared spectroscopy modeling. These technologies mainly integrate one or two key parameters of data samples, input attributes, and algorithm parameters, but other parameters still need to be tuned. In addition, member models are mostly based on the linear regression method of PLS, for example, CN113607683A is an automatic modeling method for near-infrared spectroscopy quantitative analysis. Although these techniques perform well on most datasets, the generalization performance of the model may decrease when faced with nonlinear data. Therefore, developing new modeling methods to improve adaptability and generalization ability to nonlinear data is a technical challenge in this field. Summary of the invention

[0005] In order to solve the above technical problems, the present application provides an intelligent modeling method for near-infrared spectral quantitative analysis, which realizes an intelligent modeling method for near-infrared spectral quantitative analysis with strong adaptability, ease of use, and high prediction accuracy, reduces the impact of data nonlinearity on the near-infrared quantitative model, effectively improves the generalization ability and stability of the model, reduces dependence on professional knowledge, and promotes the popularization and application of technology.

[0006] The technical solution is as follows:

[0007] An intelligent modeling method for near infrared spectroscopy quantitative analysis, comprising:

[0008] S1: obtaining a training set of near-infrared spectra of the tested samples, and performing sample preprocessing and spectrum preprocessing on the training set to obtain a preprocessed data set;

[0009] S2: Divide the preprocessed data set into several groups of corresponding training subsets and validation subsets;

[0010] S3: For each training subset in each group of preprocessed data sets, several extreme learning machine sub-models are established based on the extreme learning machine, the corresponding validation subsets are predicted, and the sub-models are sorted from small to large according to RMSEP, and several better sub-models are selected to enter the candidate model library;

[0011] S4: Repeat step S3 for each training subset in each preprocessed data set to obtain the final candidate model library, select several best sub-models for integration, and obtain the final model.

[0012] A further technical solution is to preprocess the samples of the training set, use the partial least squares method to establish a corresponding model to eliminate extreme abnormal samples and obtain a clean data set.

[0013] A further technical solution is to use a Monte Carlo sampling algorithm to divide the preprocessed data set into multiple training subsets and validation subsets.

[0014] A further technical solution is to use the extreme learning machine ELM as an individual learner, randomly select values ​​of the activation function, the number of hidden layer nodes and the regularization coefficient through random search technology, and establish several ELM sub-models.

[0015] According to a further technical solution, the spectrum preprocessing adopts spectrum preprocessing algorithms and combinations thereof.

[0016] A further technical solution is that the spectral preprocessing algorithm is one of the following algorithms: mean centering, smoothing + mean centering, first-order derivative + mean centering, second-order derivative + mean centering, multivariate scattering correction + mean centering, vector normalization + mean centering, standard normal variable transformation + mean centering, elimination of constant offset + mean centering, subtracting a straight line + mean centering, detrending + mean centering, maximum and minimum normalization + mean centering, standard normal variable transformation + detrending + mean centering, first-order derivative + multivariate scattering correction + mean centering, first-order derivative + standard normal variable transformation + mean centering, first-order derivative + subtracting a straight line + mean centering, first-order derivative + vector normalization + mean centering, and their combinations.

[0017] Further technical solutions, the calculation formula of the root mean square error of prediction (RMSEP) is:

[0018]

[0019] in: —The predicted value of the i-th sample (i=1,2…n); y i —reference value of the i-th sample (i=1,2…n); d i —Prediction error, whose value is

[0020] The technical solution includes at least the following technical effects:

[0021] The application provides an intelligent modeling method for near-infrared spectral quantitative analysis, which is based on an integrated learning strategy, uses an extreme learning machine as an individual learner, combines partial least squares regression to eliminate extreme abnormal samples, and uses Monte Carlo sampling and random search techniques to integrate models from five aspects, including samples, preprocessing methods, activation functions, number of hidden layer nodes, and regularization coefficients. The member models are more diverse. A highly adaptable, easy-to-use, and highly accurate intelligent modeling method for near-infrared spectral quantitative analysis is implemented, which reduces the impact of data nonlinearity on near-infrared quantitative models, effectively improves the generalization ability and stability of the model, reduces dependence on professional knowledge, and promotes the popularization and application of technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0023] Figure 1 A flow chart of an intelligent modeling method for near-infrared spectroscopy quantitative analysis provided in a preferred embodiment of the present application;

[0024] Figure 2A comparison diagram of the moisture model of meat samples in a preferred embodiment;

[0025] Figure 3 A comparison diagram of meat sample fat models in a preferred embodiment;

[0026] Figure 4 A comparison diagram of a meat sample protein model according to a preferred embodiment;

[0027] Figure 5 A comparison diagram of a free alkali model of a sodium chlorate sample in a preferred embodiment;

[0028] Figure 6 This is a comparison diagram of the nicotine model of tobacco powder samples in a preferred embodiment. DETAILED DESCRIPTION

[0029] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0030] Glossary:

[0031] Training Data: training data;

[0032] Clean Data: clean data;

[0033] Pretreated Data: preprocessed data;

[0034] PLS: Partial Least Squares:

[0035] MonteCarlo sampling: Monte Carlo sampling;

[0036] ELM: Extreme Learning Machine;

[0037] Sigmoid activation function: logistic function or S-type function;

[0038] ReLU activation function: linear rectification function or rectified linear unit;

[0039] Linear activation function: often called linear function or pure linear function.

[0040] RMSEP: Root mean square error of prediction.

[0041] like Figure 1As shown in the figure, in view of the shortcomings of traditional modeling methods and modeling software, such as numerous hyperparameters and high requirements for users' chemometric knowledge background, an intelligent modeling method for near-infrared spectroscopy quantitative analysis is proposed. Based on the ensemble learning strategy, the partial least squares method is used to eliminate extreme abnormal samples, and different spectral preprocessing methods are used to obtain multiple groups of different Pretreated Data preprocessed data. The Monte Carlo sampling algorithm is used to divide them into several training subsets and validation subsets respectively. For each group of training subsets, the extreme learning machine (ELM) is used as an individual learner. The activation function, the number of hidden layer nodes and the regularization coefficient are randomly selected from the fixed parameter list by random search multiple times, so as to establish several ELM sub-models with different performances, and then the corresponding validation subsets are predicted. According to the RMSEP, several good sub-models are selected to enter the candidate model library. The same operation is performed on all training subsets in all Pretreated Data preprocessed data to obtain a final candidate model library, which is sorted from small to large according to the RMSEP, and several best sub-models are selected from them for integration to obtain the final model.

[0042] The specific process is as follows:

[0043] S101: Collect a certain number of samples to be tested, collect near-infrared spectra of the samples to be tested, and use standard methods to determine the reference values ​​of their components or properties, which are recorded as Training Data;

[0044] S102: Use partial least squares (PLS) to build a corresponding model. The algorithm will intelligently remove extreme abnormal samples in the data set based on the results of cross-validation to form Clean Data.

[0045] S103: 16 different spectral preprocessing methods are used to preprocess the Clean Data, including mean centering, smoothing + mean centering, first-order derivative + mean centering, second-order derivative + mean centering, multivariate scattering correction + mean centering, vector normalization + mean centering, standard normal variable transformation + mean centering, constant offset elimination + mean centering, subtracting a straight line + mean centering, detrending + mean centering, maximum and minimum normalization + mean centering, standard normal variable transformation + detrending + mean centering, first-order derivative + multivariate scattering correction + mean centering, first-order derivative + standard normal variable transformation + mean centering, first-order derivative + subtracting a straight line + mean centering, first-order derivative + vector normalization + mean centering. 16 common spectral preprocessing methods are used to process the Clean Data to obtain 16 groups of Pretreated Data. The algorithm will intelligently select the window width for smoothing and derivation according to the number of data points in the spectrum.

[0046] S104: For each group of Pretreated Data, divide it into m groups of training subsets and validation subsets respectively using the Monte Carlo sampling algorithm. The algorithm will intelligently select the number of sampling times according to the number of samples;

[0047] S105: For each set of training subsets, an extreme learning machine (ELM) is used as an individual learner, and the activation function (sigmoid, relu, linear), the number of hidden layer nodes (50-1000) and the regularization coefficient (2 to the power of i, i = -10, -9, -8...23, 24) are randomly selected from a fixed parameter list through a random search technique for multiple times (200-1000, for example, 500 times) to form a certain random parameter combination, thereby establishing a number (for example, 500) of ELM sub-models with different performances, and then predicting the corresponding validation subsets, sorting them from small to large according to RMSEP, and selecting j (50-100) better sub-models to enter the candidate model library. The preferred ELM sub-models may have different preprocessing methods, corresponding samples, activation functions, number of hidden layer nodes, and regularization coefficients, so as to achieve the effect of model integration.

[0048] S106: Repeat step S105 for each training subset in each group of Pretreated Data to obtain a large number of candidate model libraries, and then select the best from the best, sort them from small to large according to RMSEP, and select k (100-200) best sub-models for integration to obtain the final model.

[0049] Assuming that a test set (or validation set) contains n samples, the calculation formula for the root mean square error of prediction (RMSEP) is:

[0050]

[0051] in: —The predicted value of the i-th sample (i=1,2…n); y i —reference value of the i-th sample (i=1,2…n); d i —Prediction error, whose value is

[0052] Embodiment 1:

[0053] This example uses near infrared spectroscopy data analysis. The data in this example comes from a public data set. The near infrared absorption spectra of 215 meat samples collected by transmission are in the wavelength range of 850-1048nm and the wavelength interval is 2nm. Each sample contains water, fat, and protein content. It is divided into 129 training set samples and 86 prediction set samples. The specific steps are as follows:

[0054] 1. Use partial least squares method to build the model, and intelligently remove extreme samples with large errors. The serial numbers are: 15, 26, 38, 105, forming Clean Data;

[0055] 2. Use 16 different preprocessing algorithms to perform spectral preprocessing on Clean Data to obtain Pretreated Data;

[0056] 3. For each group of Pretreated Data, the Monte Carlo sampling algorithm is used to divide it into 50 training subsets and validation subsets. The training subset is used to train the extreme learning machine sub-model, and the validation set subset is used to evaluate the prediction performance of the aforementioned sub-model;

[0057] 4. For each set of training subsets, the extreme learning machine (ELM) is used as the individual learner, and the random search technology is used to randomly combine the activation function (sigmoid, relu, linear), the number of hidden layer nodes (50-1000) and the regularization coefficient (2 to the power of i, i = -10, -9, -8...23, 24) from the fixed parameter list for 500 times to form 500 parameter combinations, and then 500 ELM sub-models with different performances are established. Then the corresponding validation subsets are predicted, sorted from small to large according to RMSEP, and 50 better sub-models are selected to enter the candidate model library.

[0058] 5. Repeat step 4 for each training subset in each group of Pretreated Data to obtain a large number of candidate model libraries. Then select the best from the best, sort them from small to large according to RMSEP, and select the 100 best sub-models for integration to obtain the final model.

[0059] 6. When predicting unknown samples, simply average the prediction results of 100 models to get the final prediction result.

[0060] Table 1: Comparison results of meat sample models

[0061]

[0062] From the data in the above table and Figure 2-Figure 4 As shown, Figure 2 Figure a shows the prediction results of the meat sample moisture model using the PLS-based integrated model, which is shown in blue dots. Figure 2 Figure b in the figure is the prediction result of the moisture model of the meat sample using the method of this embodiment, which is shown by a red triangle in the figure; Figure 3 Figure a shows the prediction results of the meat sample fat model using the PLS-based integrated model, which is shown in blue dots. Figure 3Figure b in the figure is the prediction result of the meat sample fat model using the method of this embodiment, which is shown by a red triangle in the figure; Figure 4 Figure a shows the prediction results of the meat sample protein model using the PLS-based integrated model, which is shown in blue dots. Figure 4 Figure b is the prediction result of the meat sample protein model using the method of this embodiment, which is shown by a red triangle in the figure.

[0063] The prediction results of an intelligent modeling method for near-infrared spectroscopy quantitative analysis adopted in this embodiment are as follows:

[0064] The determination coefficient R2 of the meat sample moisture model is 0.9952, the root mean square error RMSEP is 0.6831, the mean absolute error MAE is 0.4917, and the mean relative error MRE is 0.8572%.

[0065] The determination coefficient R2 of the meat sample fat model is 0.9976, the root mean square error RMSEP is 0.6262, the mean absolute error MAE is 0.4157, and the mean relative error MRE is 3.4153%.

[0066] The determination coefficient R2 of the meat sample protein model is 0.9720, the root mean square error RMSEP is 0.5118, the mean absolute error MAE is 0.3938, and the mean relative error MRE is 2.3192%.

[0067] Comparing the model prediction effects between the PLS-based integrated model and the method adopted in this embodiment, the determination coefficient R2, prediction root mean square error RMSEP, mean absolute error MAE, and mean relative error MRE of the prediction results obtained by the method of this embodiment are all better than those of the PLS-based integrated model, and can effectively solve the nonlinear problems existing in the data set.

[0068] Embodiment 2:

[0069] The data in this embodiment comes from multiple sodium hypochlorite production companies, including 1000 near-infrared spectra and corresponding free alkali content data collected by multiple MEMS near-infrared spectrometers IAS-L1 during the production process. The near-infrared spectrum ranges from 900 to 1700nm, the spectrum interval is 1nm, and the free alkali content is 0.26% to 14.79%. A systematic sampling method is used to divide 800 training sets, and the remaining 200 samples are used as prediction sets. The specific steps are as follows:

[0070] 1. Use partial least squares method to build the model, and intelligently remove extreme samples with large errors. The serial numbers are: 6, 17, 58, 165, 199, 398, 557, 766, 789, forming Clean Data;

[0071] 2. Use 16 different preprocessing algorithms to perform spectral preprocessing on Clean Data to obtain Pretreated Data;

[0072] 3. For each group of Pretreated Data, the Monte Carlo sampling algorithm is used to divide it into 30 training subsets and validation subsets. The training subset is used to train the extreme learning machine sub-model, and the validation set subset is used to evaluate the prediction performance of the aforementioned sub-model;

[0073] 4. For each set of training subsets, the extreme learning machine (ELM) is used as the individual learner, and the random search technology is used to randomly combine the activation function (sigmoid, relu, linear), the number of hidden layer nodes (50-1000) and the regularization coefficient (2 to the power of i, i = -10, -9, -8...23, 24) from a fixed parameter list for 800 times to form 800 parameter combinations, and then 800 ELM sub-models with different performances are established. Then the corresponding validation subsets are predicted, sorted from small to large according to RMSEP, and 100 better sub-models are selected to enter the candidate model library.

[0074] 5. Repeat step 4 for each training subset in each group of Pretreated Data to obtain a large number of candidate model libraries. Then select the best from the best, sort them from small to large according to RMSEP, and select the 100 best sub-models for integration to obtain the final model.

[0075] 6. When predicting unknown samples, simply average the prediction results of 100 models to get the final prediction result.

[0076] Table 2: Comparison of free base models of sodium chlorate samples

[0077] EnsemblePLS This method <![CDATA[Coefficient of determination R 2 > 0.9984 0.9987 Root mean square error of prediction RMSEP 0.1517 0.1360 Mean absolute error (MAE) 0.1014 0.0813 Mean relative error (MRE) 4.1635％ 3.3068％

[0078] From the data in the above table and Figure 5 As shown, Figure 5 Figure a shows the prediction results of the free alkali model of sodium chlorate sample using the PLS-based integrated model, which is shown in blue dots. Figure 5 Figure b in the figure is the prediction result of the free base model of the sodium chlorate sample using the method of this embodiment, which is shown by a red triangle in the figure.

[0079] The prediction results of an intelligent modeling method for near-infrared spectroscopy quantitative analysis adopted in this embodiment are as follows:

[0080] The determination coefficient R2 of the free alkali model of sodium chlorate sample is 0.9987, the root mean square error RMSEP is 0.1360, the mean absolute error MAE is 0.0813, and the mean relative error MRE is 3.3068%.

[0081] Comparing the model prediction effects of the integrated model based on PLS and the method adopted in this embodiment, the determination coefficient R2, prediction root mean square error RMSEP, mean absolute error MAE, and mean relative error MRE of the free alkali model prediction results of the sodium chlorate sample obtained by the method of this embodiment are better than those of the integrated model based on PLS, and can effectively solve the nonlinear problem existing in the data set.

[0082] Embodiment 3:

[0083] The data in this embodiment comes from two tobacco production enterprises, including the near-infrared spectra and corresponding nicotine content of 9508 tobacco powder samples scanned by two different models of Fourier transform near-infrared spectrometers. The near-infrared spectrum ranges from 10000 to 4000 cm-1, the spectral data interval is 3.8570 cm-1, and the nicotine content is 0.40% to 5.29%. A systematic sampling method is used to divide 10% of them as a prediction set (950), and the remaining 8558 samples are used as a training set. The specific steps are as follows:

[0084] 1. Use partial least squares method to build the model, and intelligently remove extreme samples with large errors. The serial numbers are: 16, 37, 65, 113, 159, 197, 206, 288, 306, 396, 451, 603, 995, 1401, 1453, 2504, 3538, 4442, 5057, 6766, 6861, 6991, 7189, 7521, 7938, 8024, 8511, forming Clean Data;

[0085] 2. Use 16 different preprocessing algorithms to perform spectral preprocessing on Clean Data to obtain Pretreated Data;

[0086] 3. For each group of Pretreated Data, the Monte Carlo sampling algorithm is used to divide it into 20 training subsets and validation subsets. The training subset is used to train the extreme learning machine sub-model, and the validation set subset is used to evaluate the prediction performance of the aforementioned sub-model;

[0087] 4. For each set of training subsets, the extreme learning machine (ELM) is used as the individual learner, and the random search technology is used to randomly combine the activation function (sigmoid, relu, linear), the number of hidden layer nodes (50-1000) and the regularization coefficient (2 to the power of i, i = -10, -9, -8...23, 24) from the fixed parameter list 400 times to form 400 parameter combinations, and then 400 ELM sub-models with different performances are established. Then the corresponding validation subsets are predicted, sorted from small to large according to RMSEP, and 80 better sub-models are selected to enter the candidate model library.

[0088] 5. Repeat step 4 for each training subset in each group of Pretreated Data to obtain a large number of candidate model libraries. Then select the best from the best, sort them from small to large according to RMSEP, and select the 50 best sub-models for integration to obtain the final model.

[0089] 6. When predicting unknown samples, simply average the prediction results of 50 models to get the final prediction result.

[0090] Table 3: Comparison of nicotine models of tobacco powder samples

[0091]

[0092]

[0093] From the data in the above table and Figure 6 As shown, Figure 6 Figure a shows the prediction results of the nicotine model of tobacco powder samples using the PLS-based integrated model, which is shown in blue dots. Figure 6 Figure b in the figure is the prediction result of the nicotine model of the tobacco powder sample using the method of this embodiment, which is shown by a red triangle in the figure.

[0094] The prediction results of an intelligent modeling method for near-infrared spectroscopy quantitative analysis adopted in this embodiment are as follows:

[0095] The determination coefficient R2 of the nicotine model for tobacco powder samples was 0.9893, the root mean square error RMSEP was 0.0897, the mean absolute error MAE was 0.0707, and the mean relative error MRE was 3.3541%.

[0096] Comparing the model prediction effects between the PLS-based integrated model and the method adopted in this embodiment, the determination coefficient R2, prediction root mean square error RMSEP, mean absolute error MAE, and mean relative error MRE of the prediction results of the nicotine model of tobacco powder samples obtained by the method of this embodiment are better than those of the PLS-based integrated model, and can effectively solve the nonlinear problems existing in the data set.

[0097] Through the above three embodiments, it can be seen that the method is adaptive to various types of near-infrared spectrometers, various analysis scenarios (offline and online) and data sets of various sizes, and has achieved good results. The determination coefficient R2, prediction root mean square error RMSEP, mean absolute error MAE, and mean relative error MRE of the prediction results obtained by this method are better than those of the PLS-based integrated model, and can effectively solve the nonlinear problems existing in the data set.

[0098] The present application provides an intelligent modeling method for near-infrared spectroscopy quantitative analysis, which is based on an integrated learning strategy, uses an extreme learning machine as an individual learner, combines partial least squares regression to eliminate extremely abnormal samples, and uses Monte Carlo sampling and random search technology to integrate models from five aspects, including samples, preprocessing methods, activation functions, number of hidden layer nodes, and regularization coefficients, so that member models are more diverse.

[0099] The intelligent modeling method for near-infrared spectroscopy quantitative analysis provided in the present application is easy to engineer, and can easily develop intelligent quantitative modeling software based on various operating systems, including local modeling and cloud modeling; users do not need to have a chemometrics knowledge background, and there are no hyperparameter items, so it is easy and convenient to get started; the prediction results of unknown samples are accurate, the generalization ability is strong, and it is suitable for quantitative analysis of various complex near-infrared data; it can promote the popularization and promotion of near-infrared spectroscopy technology to a certain extent.

[0100] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention herein. The present application is intended to cover any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary techniques in the art that the present application does not apply for. The specification and embodiments are intended to be exemplary only, and the true scope and spirit of the present application are indicated by the appended claims.

[0101] It should be understood that the above detailed description of the embodiments of the present application is only a preferred embodiment of the present application and cannot be considered to limit the scope of implementation of the present application. All equivalent changes and improvements made within the scope of the present application should still fall within the scope of the patent coverage of the present application.

Claims

1. An intelligent modeling method for near infrared spectroscopy quantitative analysis, characterized in that: include: S1: obtaining a training set of near-infrared spectra of the tested samples, and performing sample preprocessing and spectrum preprocessing on the training set to obtain a preprocessed data set; S2: Divide the preprocessed data set into several groups of corresponding training subsets and validation subsets; S3: For each training subset in each group of preprocessed data sets, several extreme learning machine sub-models are established based on the extreme learning machine, the corresponding validation subsets are predicted, and the sub-models are sorted from small to large according to RMSEP, and several better sub-models are selected to enter the candidate model library; S4: Repeat step S3 for each training subset in each preprocessed data set to obtain the final candidate model library, select several best sub-models for integration, and obtain the final model.

2. The modeling method according to claim 1, characterized in that: The training set is subjected to sample preprocessing, and a corresponding model is established using the partial least squares method to remove extreme abnormal samples and obtain a clean data set.

3. The modeling method according to claim 1, characterized in that: The Monte Carlo sampling algorithm is used to divide the preprocessed dataset into multiple training subsets and validation subsets.

4. The modeling method according to claim 1, characterized in that: The extreme learning machine (ELM) is used as the individual learner. The activation function, the number of hidden layer nodes and the regularization coefficient are randomly selected through random search technology to establish several ELM sub-models.

5. The modeling method according to claim 1, characterized in that: The spectrum preprocessing adopts spectrum preprocessing algorithms and combinations thereof.

6. The modeling method according to claim 5, characterized in that: The spectral preprocessing algorithm is one of the following algorithms: mean centering, smoothing + mean centering, first-order derivative + mean centering, second-order derivative + mean centering, multivariate scattering correction + mean centering, vector normalization + mean centering, standard normal variable transformation + mean centering, elimination of constant offset + mean centering, subtracting a straight line + mean centering, detrending + mean centering, maximum and minimum normalization + mean centering, standard normal variable transformation + detrending + mean centering, first-order derivative + multivariate scattering correction + mean centering, first-order derivative + standard normal variable transformation + mean centering, first-order derivative + subtracting a straight line + mean centering, first-order derivative + vector normalization + mean centering, and combinations thereof.

7. The modeling method according to claim 1, characterized in that: The root mean square error of prediction (RMSEP) is calculated as: in: —The predicted value of the i-th sample (i=1,2…n); y i —reference value of the i-th sample (i=1,2…n); d i —Prediction error, whose value is

Images