CN112750507B - Method for simultaneously detecting nitrate and nitrite contents in water based on hybrid machine learning model - Google Patents

Abstract

The invention belongs to the field of spectrum signal analysis, and particularly relates to a method for simultaneously detecting nitrate and nitrite contents in water based on a hybrid machine learning model. Comprising the following steps: acquiring a series of machine spectrum data of mixed solution samples of nitrate and nitrite with different nitrogen contents; classifying samples according to the optimal critical concentration to obtain four types of samples; establishing a relation model of the nitrogen content of the nitrate and the nitrite corresponding to each of the four types of samples and the corresponding spectrum data; screening characteristic wavelengths with high sensitivity and correlation, and establishing a regression sub-model; and acquiring spectrum data of the sample to be detected, determining the type of the sample to be detected according to the relation model, and adopting a regression sub-model corresponding to the type of the sample to be detected to carry out analysis and prediction so as to obtain the concentration of nitrate and nitrite. The method of the invention realizes accurate and rapid detection of nitrate nitrogen and nitrite nitrogen, and can ensure detection sensitivity under low concentration.

Description

Simultaneous detection of nitrate and nitrite in water based on hybrid machine learning model quantitative method

Technical field

The invention belongs to the field of spectral signal analysis, and specifically relates to a method for simultaneously detecting nitrate and nitrite content in water based on a hybrid machine learning model.

Background technique

There are currently a variety of nitrogen-containing compound detection technologies on the market, which differ greatly in detection principles, calculation methods, operating techniques, and application fields. The relatively mature multi-component concentration instrument analysis methods at home and abroad mainly include: electrochemistry, capillary electrophoresis, ion chromatography, biosensing and spectrophotometry. Electrochemical measurement technology is not perfect in monitoring the concentration of trace analytes, and in actual samples, the electrode surface is easily contaminated, which can easily lead to unstable detection results. The method based on capillary electrophoresis is relatively reliable, but requires large-scale instruments and is complex to operate, making it difficult to achieve on-site automated monitoring. Chromatography can analyze the concentrations of multiple ion components simultaneously and is highly safe, but the equipment requires frequent maintenance, which is time-consuming and expensive. Biosensor approaches need to address issues of robustness, selectivity, and standardization of operation. Spectroscopic technologies such as ultraviolet-visible, near-infrared, and fluorescence are non-destructive, versatile, and flexible detection methods that have all the characteristics required for online monitoring. They are currently a relatively economical, feasible, fast and simple method. Based on the light absorption characteristics of nitrate and nitrite, the fast and simple UV-visible spectrophotometry was selected as the basic detection method.

Sequential analysis is commonly used in traditional spectrophotometric methods for detecting nitrate and nitrite: first, use the Griess reagent method to analyze the nitrite in the sample, and then reduce another identical sample (usually using a copper/cadmium column) to ensure After all nitrates have been converted to nitrites and the nitrite analysis is repeated, the nitrate concentration can be calculated from the difference. This method is an indirect analysis of nitrate, which is time-consuming and highly dependent on the accuracy of nitrite detection. Secondly, the Griess method involves toxic chemical reagents, which are harmful to the body and pollute the environment. Some researchers have proposed that the UV absorption spectra of nitrate and nitrite can be used to directly measure them. Since the UV absorption spectra of nitrate and nitrite have similar shapes in the first half, and the absorption peak wavelengths are very close and almost overlap. , in actual operation, it is difficult to separate the contributions of nitrite and nitrate from the collected spectra, and the existing direct spectroscopy methods still use traditional chemometric methods to process spectral data, facing the problem of narrow application range and detection The problem of low accuracy. In recent years, the combination of UV spectroscopy and machine learning methods has been successfully applied to the rapid detection of a variety of compounds. However, there are still few studies on the separation of nitrate and nitrite. Preliminary experiments have shown that when ordinary machine learning models are used for mixed solutions of nitrate and nitrite within a certain concentration range, they are not sensitive enough to predict components at low concentrations. There is an urgent need to find a method for detecting analytes when their concentration changes greatly. Machine learning methods that still maintain the same level of accuracy.

Contents of the invention

Based on this, the present invention aims at the above-mentioned problems and provides a method for simultaneously detecting nitrate and nitrite content in water based on a hybrid machine learning model. This method combines classification and regression algorithms to ensure that nitrate content is detected within the entire model range. The detection accuracy of nitrate and nitrite is balanced, easy to operate, low cost, and can achieve accurate and rapid detection of nitrate and nitrite in simple environments at the same time.

The present invention provides a method for simultaneously detecting nitrate and nitrite content in water based on a hybrid machine learning model, which specifically includes:

S1: Prepare a series of nitrate and nitrite mixed solution samples with different nitrogen contents, and measure the spectral data of the samples;

S2: Construct a two-dimensional plane based on the nitrogen content of nitrate and nitrite in the sample, and obtain the optimal critical concentration. Divide the two-dimensional plane into four sub-regions, and the samples in each sub-region are classified into one category. Samples, four types of samples are obtained;

S3: Establish a relationship model between the nitrogen content of nitrate and nitrite corresponding to each type of the four types of samples and the corresponding spectral data to achieve automatic classification of samples;

S4: Use samples within the sub-region and on the classification boundary as modeling samples, screen characteristic wavelengths with high sensitivity and correlation, and establish a regression sub-model;

S5: Obtain the spectral data of the sample to be tested, determine the category of the sample to be tested according to the relationship model, and use the regression sub-model corresponding to the category of the sample to be tested to perform analysis and prediction, and obtain the content of nitrate and nitrite in the sample to be tested. concentration.

Further, the step S3 is specifically:

The nitrogen content of nitrate and nitrite corresponding to each type of the four types of samples and the corresponding spectral data are trained to obtain a support vector machine classification model, a random forest classification model and a logistic regression model.

Further, obtaining a support vector machine classification model specifically includes:

The objective function of the support vector classification model is:

st y _i (ω ^T x _i +b)≥1-ξ _i , ξ _i ≥0, i=1, 2,...,l

The x _i is the sample vector, x _j is the sample classification label, ω is a vector whose dimension is equal to the characteristic dimension of the sample, b is a real number, n is the total number of samples, C is the penalty factor, and ξ _i represents the slack variable ;

The Gaussian kernel function is selected as the kernel function of the support vector machine, and its function expression is as follows:

In the formula, x _{i and} x _j represent the characteristic vector of the sample in the low-dimensional space, and σ is the bandwidth of the Gaussian kernel, that is, the kernel parameter.

Further, obtaining a random forest classification model specifically includes:

Sampling the sample to obtain a self-service sample and constructing a CART tree, extracting several features from each node of the CART tree, calculating the Gini index of each feature, and obtaining classification features with classification capabilities; the sample The calculation method of Gini index D is The C _k is the number of the Kth category;

Classify according to the classification features to obtain a tree structure with completely split nodes.

Further, obtaining the logistic regression model specifically includes:

The logistic regression model is:

in the formula is the weight, x is the input sample data, and y is the probability that the sample is the positive class of the classifier;

The loss function of the model is:

In the formula, is the weight, N is the number of samples,/> is the probability that the sample is a positive class, yn is the sample category label, 0 or 1.

Further, in step S4, the stable variable replacement method is used to select the characteristic wavelength, and the optimal variable subset is established as follows:

Monte Carlo sampling is used to obtain sub-data sets of the sample space and variable space. The stability of each variable is calculated in the sub-data set of the sample space to obtain elite variables with high stability. The stability S _j calculation formula is: In the formula, b _ij is the regression coefficient of the j-th variable of the i-th sample,/> is the average regression coefficient of the j-th variable, M is the total number of samples;

Perform variable substitution analysis in the sub-data set of the variable space, calculate the degree of substitution, and obtain important variables with high degree of substitution. The calculation formula of the degree of substitution PD _j is: PD _j = PCE _j -SCE _j , where PCE _j is used or not. The mean value of the root mean square error of the model established with multiple wavelength subsets containing the j variable, and SCE _j is the mean value of the root mean square error of the model established with the remaining multiple wavelength subsets containing the j variable;

The elite variables and important variables are combined, and a cross-validation method is used to obtain the optimal variable subset.

Further, the final model structure in step S4 is:

in, x _i is the sample vector, σ is the bandwidth of the Gaussian kernel, that is, the kernel parameter, [b α ₁ α ₂ … α _n ] is a constant, which can be obtained by solving the least squares support vector machine objective function using the Lagrangian method.

Further, in step S5, determining the category of the sample to be tested based on the relationship model is specifically:

The support vector machine classification model, the random forest classification model and the logistic regression model were respectively used for classification to obtain three categories, and the majority category among the three categories was selected as the category of the sample to be tested.

Further, the conditions for measuring spectral data in steps S1 and S5 are:

The spectral scanning range is 190-400nm, and the spectral scanning interval is 1nm.

Further, the optimal critical concentration in step S2 is 0.4 mg NL ^-1 .

Beneficial effects:

The present invention pre-configures a series of mixed solutions of nitrate and nitrite, measures its spectral data, and uses the above data to establish a hybrid machine learning model through classification and regression algorithms. Through the above learning model, only measurement is required. The spectral data of the sample to be tested can accurately and quickly detect the nitrate and nitrite content in the sample to be tested, which can ensure that the detection accuracy of nitrate and nitrite is balanced within the entire modeling range, and improve the detection of low concentrations. Prediction accuracy of components, easy operation and low cost.

It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and do not limit the present disclosure.

Description of the drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without exerting creative efforts.

Figure 1 is a flow chart of a method for simultaneously detecting nitrate and nitrite content in water based on a hybrid machine learning model provided by an embodiment of the present invention;

Figure 2 is a schematic diagram of sample classification provided by an embodiment of the present invention;

Figure 3 is an algorithm framework diagram for content analysis of a sample to be tested provided by an embodiment of the present invention;

Figure 4 is a comparison chart of the effects of single model and hybrid model in predicting nitrate concentration provided by the embodiment of the present invention;

Figure 5 is a comparison chart of the effects of single model and hybrid model in predicting nitrite concentration provided by the embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention and are not intended to limit the present invention.

The present invention is based on the research discovery that combining ultraviolet spectroscopy with machine learning methods can be used for rapid detection of nitrate and nitrite at the same time. However, when the ordinary machine learning model is used for mixed solutions of nitrate and nitrite within a certain concentration range, it is difficult to detect nitrate and nitrite at the same time. The sensitivity of predicting components at low concentrations is insufficient, and there is an urgent need to find a machine learning method that can maintain the same level of detection accuracy when the analyte concentration changes greatly.

As shown in Figure 1, in one embodiment, a flow chart of a method for simultaneously detecting nitrate and nitrite content in water based on a hybrid machine learning model is proposed, which specifically includes the following steps:

Step S101: Prepare a series of nitrate and nitrite mixed solution samples with different nitrogen contents, and measure the spectral data of the samples.

In the embodiment of the present invention, first prepare nitrate nitrogen and nitrite nitrogen standard stock solutions: weigh dried 0.7221g potassium nitrate or 0.4928g sodium nitrite, dissolve it in an appropriate amount of fresh deionized water, and move it into a 1000ml volumetric flask. , dilute to the mark with deionized water, mix well and set aside. Before use, dilute to a standard solution of 10 mg NL ^-1 . All reagents were of analytical grade (Sinopharm Chemical Reagent Co., Ltd., China). The nitrite nitrogen concentrations are respectively prepared as 0.1, 0.2, 0.3, 0.4, 0.8, 1.2, 1.6, 2.0, 2.5, 3.0mg NL ^-1 and the nitrate nitrogen concentrations are 0.1, 0.2, 0.3, 0.4, 0.8, 1.2, 1.6 , 2.0, 2.5, 3.0 mg NL ^-1 mixed solutions, a total of 100 sets of mixed samples. Use deionized water as a reference solution for background subtraction, and measure the spectral data at each wavelength point in the wavelength range of 190-400 nm at an interval of 1 nm.

Step S102, use the nitrogen content of nitrate and nitrite in the sample to form a two-dimensional plane, obtain the optimal critical concentration, and divide the two-dimensional plane into four sub-regions, and the sample in each sub-region is one Class samples, four classes of samples are obtained.

As shown in Figure 2, the embodiment of the present invention provides a schematic diagram of sample classification. The concentration plan of nitrate and nitrite is divided into four sub-regions for modeling respectively. Due to insufficient sensitivity in predicting analytes at low concentrations, it is used for dividing The critical concentration of the sub-region is selected at a lower position, and the critical concentrations are 0.3, 0.4 and 0.8mg NL ^-1 respectively for modeling analysis. The results are shown in Table 1. When the critical concentration is 0.4mg NL ^-1 , The overall model has high classification accuracy and low average relative error; the nitrate and nitrite contents in each sub-region have different characteristics: the nitrate and nitrite contents in region 1 are both low; The nitrate content in 2 is much higher than that of nitrite; the nitrate concentration in zone 3 is much lower than that of nitrite; and the nitrate and nitrite content in zone 4 are both high. Compared with a single full model, each sub-model is more adaptable to the sample characteristics of each sub-region and has higher prediction accuracy.

Table 1 Comparison of model performance under different critical concentrations (mg NL ^-1 )

Step S103: Establish a relationship model between the nitrogen content of nitrate and nitrite corresponding to each type of the four types of samples and the corresponding spectral data to achieve automatic classification of samples;

In the embodiment of the present invention, the nitrogen content of nitrate and nitrite corresponding to each type of the four types of samples and the corresponding spectral data are trained to obtain a support vector machine classification model, a random forest classification model and a logistic regression model. .

In the embodiment of the present invention, the MATLAB toolbox of LIBSVM-farutoUltimateVersion is used to train the support vector machine classification model, and its objective function is:

st y _i (ω ^T x _i +b)≥1-ξ _i , ξ _i ≥0, i=1, 2,...,l (1)

In the formula, x _i is the sample vector, y _i is the sample classification label, ω is a vector whose dimension is equal to the characteristic dimension of the sample, b is a real number, l is the total number of samples, C is the penalty factor, and ξ _i represents the slack variable. ;

The Gaussian kernel function is selected as the kernel function of the support vector machine, and its function expression is as follows:

In the formula, x _{i and} x _j represent the characteristic vector of the sample in the low-dimensional space, and σ is the bandwidth of the Gaussian kernel;

In the process of using SVM modeling, the absorbance data is first normalized and pre-processed, and the data is mapped to the range of 0 to 1 to speed up the convergence speed of the training network, and then principal component analysis (PCA) is used to reduce the data of the input layer. Dimension, the particle swarm algorithm (PSO) was used to tune the two hyperparameters, the penalty factor C and the kernel parameter σ.

The SVC function in the LIBSVM-farutoUltimateVersion toolbox integrates to achieve the above functions. The function is as follows: [predict_label, accuracy, bestc, bestg] = SVC (train_label, train_data, test_label, test_data, Method_option), where Method_option is a structure. The settings are: Method_option.scale=1, Method_option.pca=0, Method_option.type=2. The SVM classification model can be established to obtain the predicted sample category predict_label, and the optimal penalty factor C and kernel parameter g can be output at the same time.

In the embodiment of the present invention, the MATLAB toolbox of RF_MexStandalone-v0.02 is used to train the random forest classification model. First, from the original training samples, the bootstrap method is used to randomly extract k new bootstrap sample sets with replacement, and thereby Construct k CART trees, and the unsampled samples each time constitute k out-of-bag data; assuming there are n features, randomly extract m features at each node of each tree, and calculate the Gini index, select a feature with the most classification ability for node splitting. For a given sample D, assuming that there are K categories, the number of K-th categories is CK, the calculation formula of the Gini index of sample D is as follows:

If the selected attribute is A, then the Gini index calculation formula of the split data set D is as follows:

In the formula, k means that the sample D is divided into K parts, and the data set D is split into K data sets D _j ;

The tree structure is formed by completely splitting the nodes, and each CART tree is allowed to grow to the maximum extent. Finally, each generated tree is allowed to vote on the sample category, and the final classification result of the unknown sample is determined according to the principle of the minority obeying the majority.

In the embodiment of the present invention, a program is written in MATLAB to implement logistic regression, and the output of the linear regression model is used as the input of the sigmoid function to obtain the mathematical expression model of the logistic regression, as follows:

in the formula is the weight, x is the input sample data, and y is the probability that the sample is the positive class of the classifier;

The loss function is used to measure the difference between the output of the model and the real output. In logistic regression, the value of the loss function is equal to the total probability that the sample is a certain category. The formula is as follows:

In the formula, is the weight, N is the number of samples,/> is the probability that the sample is a positive class, y _n is the sample category label, 0 or 1.

According to the idea of maximum likelihood estimation, it is necessary to find the optimal ω to achieve the maximum value of the loss function. At this time, the stochastic gradient descent method is used to randomly generate an initial value of ω, and then continuously iterate through the following formula to obtain the optimal ω :

In the formula, for/> Initial value./> for/> new value;

will ask for The value is substituted into the mathematical model of logistic regression to calculate the category probability score of each sample, and the category with the highest probability score is used as the final category of the sample; the onevsall idea is also used to extend the logistic regression to achieve multi-classification, assuming that the data has N categories , use logistic regression to build an independent binary classifier for each category in N categories. For classifier i, the sample with label==i is set as the positive class, the remaining samples are set as the negative class, and so on. Input the sample data to be predicted, get the probability p that all classifiers judge it to be the corresponding positive class, and take the sample type corresponding to the largest probability in p as the final prediction type.

The sample categories are voted on respectively based on the classification model established by support vector machine, random forest and logistic regression, and the category with the majority of votes (≥2) is used as the final category of the sample.

Step S104: Use the samples in the sub-region and on the classification boundary as modeling samples, screen out characteristic wavelengths with high sensitivity and correlation, and establish a regression sub-model.

In the embodiment of the present invention, since the probability of classifier error is greater at the regional boundary, each sub-model includes samples distributed on the boundary to avoid larger prediction errors caused by classification errors. The stable variable replacement method (SVP) is used ) selects the characteristic wavelength, establishes the optimal variable subset, and uses the least squares vector machine to establish a sub-regression model. The SVP is based on the evolutionary principles of intraspecific competition and survival of the fittest, taking into account the stability, replacement degree and relationship between variables. Statistics related to model performance were used to evaluate the variables, and the subset of variables with the smallest RMSE mean and relatively low standard deviation value was regarded as the optimal variable; for each sub-region, SVP selected the unique variables for nitrite and nitrate respectively. Subset. Models built with specialized subsets of variables can be adapted to the properties of the target ions, resulting in better performance. And use the LSSVMlabv1_8_R2009b_R2011a toolbox in MATLAB to build a least squares support vector machine model, use the RBF kernel function, and also use grid search to find the best regularization parameters and kernel parameters to obtain the sub-regression model of each sub-region.

In the embodiment of the present invention, the stable variable substitution method (SVP) is used to establish models for the nitrate and nitrite components in each sub-region to select the optimal characteristic wavelength subset; first, Monte Carlo sampling is used to obtain the sample space and variables In the sub-data set of the space, the stability of each variable is calculated and sorted in the sub-data set of the sample space, and the variables with high stability are regarded as elite variables, and the rest are normal variables. The calculation formula of stability S _j is:

In the formula, b _ij is the regression coefficient of the j-th variable of the i-th sample, is the average regression coefficient of the j-th variable, and M is the total number of samples.

Then perform variable substitution analysis in the sub-data set of the variable space, calculate the degree of substitution of each variable and sort the variables with high degree of substitution as important variables; the calculation formula of the degree of substitution PD _j is:

PD _j =PCE _j -SCE _j (9)

In the formula, PCE _j is the mean root mean square error of the model built with multiple wavelength subsets that do not contain j variables, and SCE _j is the mean square error of the model built with the remaining multiple wavelength subsets that contain j variables. Root error mean.

Merge the elite and important variables into a new subset of variables and repeat the above process. N iterations obtain N variable subsets, and finally use cross-validation to select the variable subset with the smallest mean root mean square error and relatively low standard deviation as the optimal subset.

Use the LSSVMlabv1_8_R2009b_R2011a toolbox to train 4 least squares support vector machine (LSSVM) regression submodels. LSSVM is an SVM whose loss function is a quadratic loss function. Its objective function is as follows:

In the formula, x _i is the sample vector, y _i is the sample classification label, ω is a vector whose dimension is equal to the characteristic dimension of the sample, b is a real number, n is the total number of samples, C is the penalty factor, and ξ _i represents relaxation. variable, is a nonlinear mapping function that maps sample space to high-dimensional feature space.

Use the RBF kernel function as follows:

At this time, the final model structure of LSSVM is:

The model parameters [α ₁ α ₂ … α _n ] in the formula can be obtained by solving the LSSVM objective function using the Lagrangian method.

Where α = [α ₁ , α ₂ , ..., α _n ] is the Lagrange multiplier.

In the LSSVMlabv1_8_R2009b_R2011a toolbox, the LSSVM model can be established by initializing the model parameters using the tunelssvm function, and can output the optimal penalty factor C and kernel parameter g found using grid search, where the initial values of C and g are set to 100 and 0.01, the tunelssvm function is as follows: model=tunelssvm(model_ori, optfun, costfun, costfun_args), set its input parameters to costfun='crossvalidatessvm'; costfun_args={10,'mse'}; optfun='gridsearch'; model_ori=initlssvm (trnX, trnY, 'function estimation', c, g, 'RBF_kernel'), then use the trainlssvm function to establish a regression model, output the model structure, and use it as an important input of the simlssvm function to output predictions for unknown samples ValueY.

Step S105, obtain the spectral data of the sample to be tested, determine the category of the sample to be tested according to the relationship model, and use the regression sub-model corresponding to the category of the sample to be tested to perform analysis and prediction, and obtain the medium nitrate and nitrite of the sample to be tested. concentration.

As shown in Figure 3, the embodiment of the present invention provides an algorithm framework diagram for the content analysis of the sample to be tested, obtains the spectral data of the sample to be tested, that is, the spectral data, and then uses the support vector machine (SVM) classification model, random forest classification ( RF) model and logistic regression model are used for classification (LR) to obtain three categories i, j, k, and the majority category among the three categories i, j, k is selected as the category of the sample to be tested. The present invention uses three classifiers A joint classifier is established to vote on the sample categories so that the predicted categories match the real categories. The categories of voting inconsistencies among the three classification models are shown in Table 2. It can be seen that due to the existence of the voting mechanism in the present invention, the correct classification is finally obtained. result.

Table 2 Determination of categories with inconsistent votes from three base classifiers

In the embodiment of the present invention, after obtaining category l, the stable variable replacement method is used to select a subset of variables corresponding to the category area, and the least squares support vector is used to establish a regression model, and finally the predicted values of nitrate and nitrite concentrations are obtained .

In the embodiment of the present invention, leave-one-out cross-validation is used as the evaluation strategy, using four classic parameters: average relative error (ARE), maximum relative error (MRE), root mean square error of prediction (RMSEP) and coefficient of determination (R ² ) To evaluate the performance of the established model, all procedures in this example are completed in MATLAB.

As shown in Table 3, the mixed machine learning model of the present invention is compared with the prediction and analysis results of the mixed solution concentration using a single machine learning model. The single machine learning model first uses SVP to select the characteristic wavelength, and then uses LSSVM to build the model.

Table 3. Detection results using different algorithms

As can be seen from Table 3, using the prediction method of the hybrid machine learning model of the present invention, the results show that the average relative error of nitrate dropped from 6.25% to 1.64%, the maximum relative error dropped from 39.96% to 5.01%, and the average relative error of nitrite It dropped from 12.37% to 4.58%, and the maximum relative error dropped from 79.81% to 9.23%. As shown in Figures 4 and 5, respectively, the effects of single model and hybrid model in predicting nitrate and nitrite concentrations provided by the embodiments of the present invention are compared. Although the average relative error predicted by a single model is small (<10%) when the concentration of the analyte is relatively high, its prediction error increases greatly when the analyte concentration is lower than 0.4 mg NL ^-1 ; while the mixed model of the present invention For the prediction method of the machine learning model, no matter how the analyte concentration changes within the modeling area, the average relative error of the hybrid modeling is always controlled below 5%, and the performance is more stable.

Embodiments of the present invention provide a hybrid machine learning model that combines classification and regression algorithms at the same time. This model can solve the problem of uneven accuracy in predicting nitrate and nitrite by a single model. In addition, the classification system was optimized by building a joint classifier using support vector machine, random forest, and logistic regression. Experimental results show that compared with other direct spectroscopy methods using a single model, this method significantly reduces the maximum relative error in predicting nitrate and nitrite concentrations and improves the prediction accuracy for low-concentration components. It should be understood that the method of the present invention is not only applicable to the mixed solution of nitrate and nitrite at a certain concentration ratio prepared in this embodiment, but can also be applied to any concentration range with nitrate and nitrite as the main components. of any water sample.

The above-mentioned embodiments only express several implementation modes of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the patent scope of the present invention. It should be noted that, for those of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the scope of protection of the patent of the present invention should be determined by the appended claims.

Other embodiments of the disclosure will be readily apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the disclosure that follow the general principles of the disclosure and include common knowledge or customary technical means in the technical field that are not disclosed in the disclosure. . It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the disclosure being indicated by the following claims.

It should be understood that although the steps in the flowcharts of various embodiments of the present invention are shown in sequence as indicated by arrows, these steps are not necessarily executed in the order indicated by arrows. Unless explicitly stated in this article, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least some of the steps in each embodiment may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but may be executed at different times. The order of execution is not necessarily sequential, but may be performed in turn or alternately with other steps or sub-steps of other steps or at least part of the stages.

Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing relevant hardware through computer programs. The programs can be stored in a non-volatile computer-readable storage medium. , when the program is executed, it may include the processes of the above-mentioned method embodiments. Any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and/or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Synchlink DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

The technical features of the above-described embodiments can be combined in any way. To simplify the description, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, All should be considered to be within the scope of this manual.

Claims (8)

1. The method for simultaneously detecting the nitrate and nitrite contents in water based on the hybrid machine learning model is characterized by comprising the following steps of:

s1, preparing a series of mixed solution samples of nitrate and nitrite with different nitrogen contents, and measuring spectral data of the samples;

s2, forming a two-dimensional plane by using the nitrogen content of the nitrate and the nitrite in the sample, acquiring the optimal critical concentration, dividing the two-dimensional plane into four sub-areas, and obtaining four types of samples by taking the sample in each sub-area as one type of sample;

s3, establishing a relation model of the nitrogen content of the nitrate and nitrite corresponding to each type in the four types of samples and the corresponding spectrum data;

s4, taking samples in the subareas and on the classification boundaries as modeling samples, screening characteristic wavelengths with high sensitivity and correlation, and establishing a regression sub-model; and

in the step S4, a stable variable displacement method is adopted to select characteristic wavelengths, and the establishment of the optimal variable subset is specifically as follows:

obtaining a sub-data set of a sample space and a variable space by Monte Carlo sampling, and counting in the sub-data set of the sample spaceCalculating the stability of each variable to obtain an elite variable with high stability and stability S _j The calculation formula is as follows:in b _ij Regression coefficient of jth variable for ith sample,/>The average value of regression coefficients of the j-th variable is represented by M, and the total number of samples is represented by M;

performing variable displacement analysis in the sub-data set of the variable space, calculating the degree of displacement, and obtaining important variables with high degree of displacement, wherein the degree of displacement PD _j The calculation formula is as follows: PD (potential difference) device _j ＝PCE _j -SCE _j In PCE _j Root mean square error mean value (SCE) of model built for each of multiple wavelength subsets without j-variable _j Root mean square error mean values of the models respectively established for the plurality of remaining wavelength subsets containing j variables;

merging the elite variables and the important variables, and obtaining an optimal variable subset by using a cross verification method;

the regression sub-model is:

wherein,x _i is a sample vector, σ is the bandwidth of the gaussian kernel, i.e., the kernel parameter, [ bα ] ₁ α ₂ … α _n ]The method is constant and can be obtained by solving a least square support vector machine objective function by a Lagrangian method;

s5, acquiring spectrum data of the sample to be detected, determining the type of the sample to be detected according to the relation model, and adopting a regression sub-model corresponding to the type of the sample to be detected to conduct analysis and prediction to obtain the concentration of nitrate and nitrite in the sample to be detected.

2. The method for simultaneously detecting nitrate and nitrite contents in water based on a hybrid machine learning model according to claim 1, wherein the step S3 is specifically:

and training the nitrogen content of the nitrate and the nitrite corresponding to each type of sample in the four types of samples and the corresponding spectrum data to obtain a support vector machine classification model, a random forest classification model and a logistic regression model.

3. The method for simultaneously detecting nitrate and nitrite contents in water based on a hybrid machine learning model according to claim 2, wherein the obtaining a support vector machine classification model specifically comprises:

the objective function of the support vector classification model is:

s.t.y _i (ω ^T x _i +b)≥1-ξ _i ，ξ _i ≥0,i＝1,2,...,l

the x is _i Is a sample vector, y _i Is a sample classification label, ω is a vector whose dimension is equal to the characteristic dimension of the sample, b is a real number, n is the total number of samples, C is a penalty factor, ζ _i Represents a relaxation variable;

and selecting a Gaussian kernel function as a kernel function of a support vector machine, wherein the function expression is as follows:

in which x is _i ,x _j The eigenvector representing the sample in low dimensional space, σ, is the bandwidth of the gaussian kernel, i.e., the kernel parameter.

4. The method for simultaneously detecting nitrate and nitrite contents in water based on a hybrid machine learning model according to claim 2, wherein the obtaining a random forest classification model specifically comprises:

sampling the sample to obtain a self-service sample, namely constructing a CART tree, extracting a plurality of features from each node of the CART tree, and calculating the base index of each feature to obtain classification features with classification capability; the method for calculating the matrix index D of the sample is as followsThe C is _k Number of K-th category;

and classifying according to the classification characteristics to obtain a tree structure with completely split nodes.

5. The method for simultaneously detecting nitrate and nitrite contents in water based on a hybrid machine learning model according to claim 2, wherein the obtaining a logistic regression model specifically comprises:

the logistic regression model is:

in the middle ofAs the weight, x is input sample data, y is the probability that the sample is the positive class of the classifier;

the loss function of the model is:

in the method, in the process of the invention,is weight, N is number of samples, +.>For the probability that the sample is of positive class, y _n Is a sample class label, 0 or 1.

6. The method for simultaneously detecting nitrate and nitrite contents in water based on the hybrid machine learning model according to claim 2, wherein the determining the category of the sample to be detected according to the relation model in the step S5 is specifically:

and classifying by adopting a support vector machine classification model, a random forest classification model and a logistic regression model to obtain three categories, and selecting the category which is the majority of the three categories as the category of the sample to be detected.

7. The method for simultaneous detection of nitrate and nitrite content in water based on a hybrid machine learning model according to claim 1, wherein the conditions for determining the spectral data in steps S1 and S5 are:

the spectrum scanning range is 190-400nm, and the spectrum scanning interval is 1nm.

8. The method for simultaneous detection of nitrate and nitrite content in water based on a hybrid machine learning model as claimed in claim 1, wherein the optimal critical concentration in step S2 is 0.4mg nl ^-1 。