CN112733876A

CN112733876A - Soft measurement method for nitrogen oxides in urban solid waste incineration process based on modular neural network

Info

Publication number: CN112733876A
Application number: CN202011168762.4A
Authority: CN
Inventors: 乔俊飞; 段滈杉; 蒙西; 汤健
Original assignee: Beijing University of Technology
Current assignee: Beijing University of Technology
Priority date: 2020-10-28
Filing date: 2020-10-28
Publication date: 2021-04-30

Abstract

A modular neural network-based soft measurement method for nitrogen oxides (NOx) in an urban solid waste incineration process belongs to the field of solid waste treatment, and can effectively control the emission of NOx by detecting the toxic gas, namely the nitrogen oxides (NOx) generated in MSWI in real time. In the industrial field, a high-precision instrument, namely a smoke emission continuous monitoring system is adopted to detect the concentration of NOx in smoke emission, the measurement result is greatly influenced by the environment, and the equipment maintenance cost is high. Firstly, decomposing a task by using a fuzzy c-mean algorithm, and decomposing the task into different subtasks; secondly, respectively designing soft measurement submodels by adopting a radial basis function neural network aiming at different subtasks, and establishing a nonlinear relation between characteristic variables and NOx; and finally, integrating the output of the sub-networks through a cascade neural network. The effectiveness of the proposed method is verified by adopting a benchmark experiment and actual data of a certain MSWI factory.

Description

Soft measurement method for nitrogen oxides in urban solid waste incineration process based on modular neural network

Technical Field

The invention belongs to the field of solid waste treatment.

Background

With the development of social economy and the continuous improvement of the living standard of peopleHigh, urban solid waste (MSW) is increasing at 8% of the global annual growth rate, causing a great impact on the ecological environment. MSW incineration (MSWI) is one of the currently used solid waste treatment methods, and not only can realize volume reduction of solid waste, but also can change waste into valuable. Nitrogen oxides (NOx, mainly including NO and NO)₂) Is one of typical emissions in the MSWI exhaust gas and is also a main component causing environmental pollution. The concentration of NOx in MSWI tail gas in China is higher than that in some countries of European Union, so that the reduction of the NOx emission in the MSWI tail gas is particularly important. At present, the main method for controlling NOx emission reduces the generation of NOx through a denitration device, and accurately detects the concentration of NOx in real time, thereby having important significance for controlling the emission of NOx in MSWI plants and improving the environmental protection requirement.

In actual operation, a Continuous flue gas emission monitoring system (CEMS) is usually used for detecting the concentration of NOx in exhaust emission, most enterprises adopt imported equipment and entrust professional companies to maintain, when the CEMS breaks down, professional maintenance personnel can arrive at the site at a high speed, the fault troubleshooting difficulty is high, the time is long, and the data missing time caused by the fault troubleshooting is long, so that a novel method for realizing online NOx measurement has important significance for guaranteeing economy, safety and college operation of MSWI plants. The soft measurement technology estimates the variable to be measured by establishing a mechanism model or a data driving model, is easy to realize, and has been successfully applied to a plurality of industrial processes. Currently, researchers have used soft-sensing techniques for NOx detection. The urban solid waste incineration process comprises complex physical and chemical reactions, NOx concentration is related to various process variables such as components of garbage entering the furnace, temperature and air quantity in the furnace, working conditions in the furnace in the incineration process are variable, and a single neural network is difficult to establish an accurate soft measurement model.

The modularized neural network simulates the 'modularized' structural feature and the 'divide-and-conquer' functional feature of a brain network, and consists of a plurality of sub-networks, each sub-network processes one sub-task in the global task, and the capability of processing complex tasks can be effectively improved.

In the solid waste incineration process, NOx in tail gas emission is mainlyThree important sources are that the rapid NOx is generated by the chemical reaction between organic and inorganic nitrogen-containing compounds contained in the solid waste and oxygen in the incineration process, and the thermal NOx is generated by N in primary air and secondary air₂And thirdly, the fuel type NOx is generated by oxidizing at high temperature, and is generated by cracking combustion-supporting fuel (gasoline and the like) at high temperature. In order to suppress the generation of pollutants, an SNCR (Selective non-catalytic reduction) system is practically used as a denitration device for converting NOx into N by spraying urea₂And the flue gas containing a small amount of NOx passes through the reactor, the mixed gas and the bag-type dust collector to realize the absorption and filtration of NOx and other pollutants. The concentration of NOx and other pollutant gases is detected by adopting a CEMS system in an MSWI factory, tail gas meeting the emission standard is discharged into the atmosphere through a chimney, if the emission exceeds the standard, the tail gas is punished according to the regulation, and at present, the concentration of NOx in the smoke discharged by a newly-built MSWI factory executes the content regulated in a table 1. Therefore, accurate detection of NOx is of great significance for controlling the generation of NOx and reducing industrial cost.

TABLE 1 NOx concentration Limit in MSWI plant

Disclosure of Invention

A modular neural network based MSWI process NOx emission concentration soft measurement method is presented herein. Firstly, decomposing a global task by adopting a Fuzzy C-means (FCM) algorithm, reducing task complexity, then establishing a corresponding sub-network by using a Radial Basis Function (RBF) neural network according to decomposed subtasks, then integrating the sub-model by a Fully Connected Cascaded (FCC) neural network, and finally verifying the effectiveness of the method by a benchmark experiment and real data of a certain domestic MSWI power plant.

The invention provides a soft measurement model of NOx emission concentration based on a modular neural network, which consists of four parts of data preprocessing, task decomposition, submodule construction and output integration. First, power is generated from MSWIThe factory collects the original process data, denoted by X; carrying out normalization, outlier elimination and feature selection to obtain a data set Q, wherein Q belongs to R^N×P(ii) a Then, the FCM algorithm is adopted to carry out task decomposition on the data set Q to obtain C groups of data clusters with different operation conditions, and the C groups of data clusters are used

Representing the data clusters under the 1 st, 2 nd and C-th operating conditions after clustering, wherein C represents the number of the operating conditions, namely the number of the data clusters; next, for each operating condition, a corresponding RBF sub-module is constructed, denoted RBF _1, RBF _2, …, RBF _ C, respectively, to "divide and conquer" establish a non-linear mapping between the input process variable and NOx at the different operating conditions. Finally, integrating the outputs of a plurality of simultaneously activated submodels by adopting an FCC neural network, and comprehensively evaluating the outputs of all the submodels to obtain the final output, wherein Y is_trainAnd Y_testRespectively represent

In the MSWI process, in order to eliminate dimension influence among process variables, accelerate the solution speed of the sub-network and improve the training precision, a Z-score standardization method is adopted in the MSWI process, and the formula (1) shows.

Wherein x is^nor_mData, x, representing the normalized mth process variable affecting NOx emission concentration^mRepresents the m-th process variable, μ, in the raw data set collected^mAnd σ^mRespectively, mean and standard deviation of the mth process variable.

Since CEMS is susceptible to the complex environment within an incinerator, the collected NOx concentration data often contains outliers which are either well above the limits given in Table 1 or well below normal values, which are problematic for subsequent data analysis, and therefore the Rajda criterion is used therein^[17]Namely, the abnormal value in the original data is removed by the method of three times of standard deviation higher than the NOx concentration data, as shown in the formula (2).

|y_o-μ^NOx|≥3σ^NOx (2)

Wherein, y_oRepresents the NOx concentration data, μ, corresponding to the o-th original data sample in the sample^NOxAnd σ^NOxNOx data y representing the mean and standard deviation of NOx concentration data, respectively, which will satisfy equation (2)_oIf the sample is regarded as an abnormal value, the o-th sample should be removed from the data set, the preprocessed data set is represented as S, and S is ═ S₁,s₂,...,s_N]^T，s₁,s₂,...,s_NRespectively representing the 1 st, 2 nd and Nth samples obtained after the abnormal value is removed, wherein N represents the size of the sample after the abnormal value is removed.

In the MSWI, the reaction process is complex, a plurality of process variables are selected according to prior knowledge and mechanism knowledge, the variables are correlated and coupled, and in order to improve the calculation efficiency and the accuracy of a soft measurement model, an mRMR algorithm based on Mutual Information (MI) is adopted in the MSWI^[18]The latent variables are selected, as the name implies, the algorithm aims to maximize the correlation between the process variables and the NOx concentration, minimizing the redundancy between the process variables, and the pre-processed data sets S are characterized herein by MI, which is defined as formula (3), to measure the relationship between the process variables in each data set S.

Wherein I represents the ith process variable sⁱAnd the jth process variable s^jDegree of correlation between, p(s)ⁱ) And p(s)^j) Are respectively sⁱAnd s^jP(s) as a function of the probability densityⁱ,s^j) For the ith process variable sⁱAnd the jth process variable s^jSo that maximizing the correlation between the process ith process variable and the NOx concentration can be expressed as

Where D represents the degree of correlation of each process variable in the data set S with the NOx concentration, SⁱAnd s^NOx is the ith process variable and the NOx concentration variable respectively, and S is a data set after process variable normalization and NOx abnormal value elimination. Similarly, minimizing redundancy between process variables can be expressed as

Wherein R represents the degree of correlation between the process variables, and the evaluation criterion phi of the mrMR algorithm obtained by combining the formula (4) and the formula (5) is

maxΦ(D,R),Φ＝D_-R (6)

The data set after feature selection is expressed as Q epsilon R^N×FN denotes the sample size and F denotes the number of process variables selected.

To prevent overfitting, the data was divided into three parts, with a division ratio of 2: 1: 1, each with Q₁、Q₂And Q₃It is shown that the three subdata sets are each of size N₁，N₂And N₃When Q is equal to Q₁∪Q₂∪Q₃，N＝N₁+N₂+N₃And N denotes the size of the data set before division. First of all with Q₁Performing task decomposition to obtain process variable subsets under different operation conditions; then training a corresponding RBF sub-network model for the subset under each working condition, and establishing a mapping relation between the process variable and NOx under each operating condition; then using Q₂Testing the submodels established under each operating condition, training an FCC neural network by utilizing the test output of the submodels to NOx, realizing output integration under different operating conditions, and comprehensively considering the prediction capability of the submodels under different operating conditions; finally using Q₃And testing the RBF submodel and the FCC neural network to finish the test of the whole NOx emission concentration soft measurement model.

Task decomposition is the premise of realizing modular design, and process data is classified into different subsets through task decomposition. In MSWI, the NOx concentration emission characteristic is greatly influenced by the components of the garbage entering the furnace, air volume and temperature, and related influence variables also have different distribution characteristics, so that the influence variables of the NOx concentration are clustered by adopting an FCM algorithm.

The set of samples to be clustered is Q₁By q₁,q₂,...,

1 st, 2 nd and Nth in the representation set₁Obtaining the membership degree of each sample point to all class centers by an FCM algorithm through optimizing an objective function, wherein the objective function of the FCM algorithm is defined as:

wherein U is a group consisting of U_cnThe formed membership matrix, V represents the clustering center of the data clusters under each operating condition, C represents the number of the data clusters, namely the number of the operating conditions, C represents any type of data in the C-th type of data clusters, namely any operating condition, and C is 3 and N is selected in the task decomposition process because the working conditions of the soft measurement of the NOx concentration can be divided into 3₁Represents the number of process variable samples, u_cnRepresents Q₁Membership degree and u of the nth sample to the c-type operation condition_cnE (0,1), r represents a weighted index, and most researchers take the value r as 2 to simplify the calculation, so the value of r is selected as 2, v is selected as_cRepresents the c-th cluster center point, q_nRepresenting a sample Q to be clustered₁The nth sample of (1). The sum of the membership degrees of each sample point under each operating condition is 1, so that the constraint condition satisfied by equation (7) is:

in order to solve the minimum value of the objective function, a Lagrangian multiplier lambda is introduced to construct a Lagrangian function:

formula (9) is respectively for u_cn、v_cAnd λ is derived and the equation is given as 0:

combining equations (10) (11) (12), the membership value and the cluster center for each operating condition are calculated as:

wherein u is_cnMembership value, v, indicating that the nth sample belongs to class c operating conditions_cAnd the process variables influencing the change of the concentration of NOx, such as primary air, secondary air, the temperature of a hearth, the amount of urea sprayed by an SNCR system and the like, present different distribution changes under different operating conditions, so that the FCM algorithm clusters the process variables under the same operating condition according to the similarity of the characteristic space distribution of the process variables, thereby obtaining the data clusters with different operating condition distributions in the MSWI process. From equations (13) and (14), the fuzzy slavery of the input process variable for each operation condition data cluster can be obtainedAttribute matrix U:

wherein C represents the number of operating conditions, N₁The size of a sample to be clustered is represented, the degree of similarity of each sample vector influencing the concentration of the NOx and the distribution of each operation condition data cluster is displayed by a membership degree matrix, and the higher the degree of similarity is, the higher the membership degree is, the degree of similarity u of the sample vector is_cnThe larger the process variable is, the higher the probability that the process variable belongs to the class c operation condition is, and the sum of the membership degrees of the same sample vector under each operation condition is 1, as shown in formula (8). In order to accurately define the working condition class to which the sample vector belongs, a membership threshold t is adopted to perform 'soft division' on data, as shown in a formula (16). When u is satisfied_cn>t, then the corresponding sample q_nThe sub-model belongs to the class c, if the value of t is overlarge, a small number of sample vectors form a class, the function significance of the sub-model established under each operation condition is enhanced, and the generalization of the sub-model to some critical points, namely the sample vectors on the edges of the operation conditions, is reduced; if the value of t is too small, the difference between the operation conditions is too small, the function significance of each sub-model is reduced, and the initial purpose of 'divide and conquer' is violated. In order to balance the two problems, the influence of different t values on the module performance is compared through a trial and error method, and finally the value of t is determined to be 0.36.

{cluster(q_n)＝c|u_cn>t} (16)

Wherein, cluster (·) is used for judging the sample q_nThe operating conditions of (a) belong to the category. After the task decomposition, the influencing variables of the NOx emission concentration can be clustered into C operating conditions.

Wherein, FCM ((-)) represents the adopted FCM algorithm, t represents a membership threshold, and Q₁Represents a set of data to be clustered,

representing the data clusters under the 1 st, 2 nd and C th operating conditions obtained after clustering

After data clusters under C working conditions are obtained through clustering, different subsets are processed in a manner of dividing and treating by adopting a modular neural network. Because the RBF network has a simple structure and can approximate any nonlinear function, the newrb function in matlab2014a is adopted to construct a nonlinear mapping relation between the process variables and NOx under different MSWI operating conditions, and the function is used as a training stopping mechanism according to preset precision and hidden node number.

The RBF network consists of three parts, namely an input layer, a hidden layer and an output layer. The input layer transmits process data affecting NOx into the network for data clusters under the c-th condition

By using

And

respectively representing the 1 st, 2 nd and P th process variables under the nth training sample, and taking the process variables as the input of the RBF network; the hidden layer passes through H basis functions

Completing the nonlinear transformation from the input space to the hidden space, realizing the high-dimensional mapping of MSWI process variables, and recording the number of nodes of the hidden layer as H; and carrying out linear weighted combination on the output layer in a new space to obtain a NOx test result. For the

Arbitrary sample of

The kernel function of the RBF neural network can be represented by the following formula

Wherein the content of the first and second substances,

represents the nth process variable, theta, under the c-th MWSI operating condition_hAnd σ_hRespectively the center and width of the h-th radial basis function,

represent to

This sample, the output of the h-th hidden layer node, is the actual output value of NOx emission concentration

Is shown as

Wherein, w₀Is a deviation, w_hAnd (H ═ 1., H) is a connection weight between the hidden layer and the output layer, a least square algorithm is adopted in a newrb function to optimize parameters, and Mean Square Error (MSE) is used as a model performance index, as shown in formula (20).

Wherein

True value, N, of NOx concentration for the nth training sample_cFor operation in the same category asNumber of samples of process variables under operating conditions. In order to ensure the calculation time and efficiency, the MSE is preset to be 0.001, and the maximum node number H of the hidden layer _max40, then the value of H may be based on the predetermined MSE and H_maxDetermining when the RBF network reaches MSE or H_maxAnd when the number of the nodes of the hidden layer is not increased.

When a process variable activates multiple submodels simultaneously, indicating that the process variable complies with the operating conditions of multiple MSWI, it is necessary to comprehensively evaluate the ability of each submodel to test NOx concentration. As each submodel has the characteristic of 'divide and conquer' and has obvious difference in respective functions, the non-linear mapping relation is formed between the NOx concentration test result and the true value, in order to better integrate the output of the submodel, the text is integrated by adopting the FCC neural network, and the training of the FCC neural network is carried out by the second part of data set Q₂Is done, assuming in data set Q₂The nun groups of process variables in the system activate a plurality of RBF sub-modules, and the number of the activated modules is recorded as mod (mod)<C)。

Wherein the content of the first and second substances,

the number of the sub-modules under different operation conditions activated by the process variable is equal to the number of input nodes of the FCC neural network, so that the number of the input nodes of the FCC neural network is mod, b represents the bias of the network, the input of the bias can be set to be +1, and rho represents the input weight, namely rho ═ rho₁₁,ρ₁₂,...,ρ_1F；...；ρ_mod,1,ρ_mod,2,...,ρ_mod,F；ρ_b,1,...,ρ_b,f,...,ρ_b,F]，ρ₁₁,ρ₁₂,...,ρ_1FRepresenting the weight between the 1 st input node and each network node; rho_mod,1,ρ_mod,2,...,ρ_mod,FRepresenting weights between the mod input node and the respective network nodes; rho_b,1,...,ρ_b,f,...,ρ_b,FRepresents the offsets of the 1 st, F-th and F-th nodes; w denotes the internal weight between the network nodes, i.e.W＝[w₁₂,w₁₃,...,w_1F；w₂₃,...,w_2F；...；w_F-1,F]When psi is ═ ρ, W]All weights are denoted by ψ.

Indicating the integrated output of the modules at multiple operating conditions for different NOx concentration prediction capabilities of the same process variable (nth process variable), y_nThe real value of the NOx corresponding to the sample is represented, F represents the number of network nodes, the larger the F value is, the more complex the network structure is in consideration of the depth of the FCC network, the certain similarity of the measurement results of the NOx under different operation conditions is in consideration, the number of the samples input into the FCC network is limited, and after multiple experiments, F is selected to be 4. Selecting tanh (-) as the activation function of the first F-1 neurons, tanh (-) being expressed as

Therein net_fThe input of the F-th node in the network is represented, the final neuron F carries out linear summation on all the inputs to obtain the final output, and the FCC network outputs the integrated output of the NOx predicted value

Is composed of

Where f (-) represents a non-linear mapping between the sub-module output and the FCC network versus NOx prediction.

The invention has the main effects that: (1) the provided modularized neural network captures the internal distribution characteristics of input variables through task decomposition, establishes NOx emission concentration measurement models under different operating conditions, and finally integrates output results of all modules through an FCC neural network, thereby realizing soft measurement of the concentration of NOx emitted by tail gas in the MSWI process; (2) different from the traditional weighted integration method, the FCC neural network is adopted to integrate the output of the submodule, the nonlinear relation between the NOx truth value and the submodule NO measured value is reflected by fewer neurons, and the precision of the soft measurement model is improved.

Drawings

FIG. 1 is a data segmentation chart of the present invention

FIG. 2 is a schematic diagram of the RBF neural network topology of the present invention

FIG. 3 is a diagram of the topology of the FCC neural network of the present invention

FIG. 4 is a graph of the sinE function fitting training and test results of the present invention

FIG. 5 is a sinE function fitting test error curve of the present invention

FIG. 6 shows the soft measurement of NOx concentration in MSWI of the present invention

FIG. 7 shows the soft measurement model test error of the present invention

Detailed Description

The simulation experiment is divided into two parts, firstly, the performance of the modular neural network model is verified by utilizing a reference experiment (sinE function fitting), then, the model is subjected to an industrial experiment by adopting actual data of an MSWI factory, and the effectiveness of the algorithm is reflected by comparing with the existing method.

Combining the selection result of the mRMR algorithm with the solid waste incineration mechanism, 20 characteristics were finally determined in the text, as shown in table 2.

TABLE 2 feature selection results

After the selection of the characteristics of the mRMR algorithm, the solid waste incineration mechanism is combined, the NOx emission is considered from the generation part and the elimination part, the analysis of the generation process is mainly related to the temperature and the air volume, the analysis of the elimination process is mainly related to the urea injection amount in the SNCR system and the adsorption amount of the activated carbon in the flue gas purification system, and the final 20 characteristics are obtained after the analysis.

The sinE function is expressed as

y＝0.8*sin(10x)*e^-0.2x x∈(0,1) (27)

The training data was 1500 sets and the test data was 500 sets. The fitting effect of the model on the training set and the test set using the modular neural network model presented herein and the prediction error of the model on the test set are shown in fig. 4 and 5, respectively.

The sinE function increases the oscillation attenuation along with the variable x, input variables are grouped into three types based on the FCM algorithm, three submodules are correspondingly constructed, and the membership threshold t is 0.36. In fig. 4, data activating only a single RBF subnetwork is indicated by a circle, while data activating two RBF subnetworks is indicated by an asterisk. The test error curve shown in fig. 5 indicates that the model can fit the reference function well.

The paper carries out industrial tests based on actual data of the MSWI factory in Beijing in 2019. And 4000 groups of samples are selected, wherein 2000 groups of data are used for training the RBF sub-network, 1000 groups of data are used for testing the RBF sub-network and training the FCC integrated network, and finally 1000 groups of data are used as a test set. Firstly, preprocessing input variables of a training set, dividing a data set into three types after task decomposition, correspondingly constructing three sub-modules, and finally integrating NOx output of the sub-modules based on an FCC network. The results obtained on the test set and training set are shown in fig. 6, and the test error is shown in fig. 7.

The mean square error of the training of the NOx soft measurement model established based on the modular neural network is 20.0096, and the mean square error of the testing is 18.0307. As can be seen from fig. 6 and 7, the soft measurement model output better fits the actual NOx concentration. In addition, the method is compared with other methods, namely a single RBF neural network and an MNN based on weighted integration, and the comparison result is shown in the table 3.

TABLE 3 comparison of NOx predictions in MSWI

Compared with a single RBF neural network, the modular neural network shows higher test precision and embodies the superiority of 'divide-and-conquer' of the modular neural network; meanwhile, the measurement accuracy of the soft measurement model based on FCC output integration is higher than that of the MNN based on weighting integration, because the FCC neural network can better capture the nonlinear mapping relation between data, the soft measurement capability of different MSWI operating conditions on the same sample is comprehensively analyzed, and the effectiveness of the FCC neural network is reflected.

Claims

1. The soft measurement method of nitrogen oxides in the urban solid waste incineration process based on the modular neural network is characterized by comprising the following steps of:

the method comprises four parts of data preprocessing, task decomposition, submodule construction and output integration;

first, raw process data is collected from the MSWI power plant, denoted by X; carrying out normalization, abnormal value elimination and feature selection to obtain a data set Q, wherein the size of a sample contained in the Q is N, and the number of process variables is P; then, the FCM algorithm is adopted to carry out task decomposition on the data set Q to obtain C groups of data clusters with different operation conditions, and the C groups of data clusters are used

Representing the data clusters under the 1 st, 2 nd and C-th operating conditions after clustering, wherein C represents the number of the operating conditions, namely the number of the data clusters; then constructing corresponding RBF sub-modules according to each operating condition, respectively expressing the RBF sub-modules by RBF _1, RBF _2, … and RBF _ C, and establishing a nonlinear mapping relation between input process variables and NOx under different operating conditions; finally, integrating the outputs of a plurality of simultaneously activated submodels by adopting an FCC neural network, and comprehensively evaluating the outputs of all the submodels to obtain the final output, wherein Y is_trainAnd Y_testRespectively, are shown.

2. The method of claim 1, wherein:

the data preprocessing comprises the following steps:

adopting a Z-score standardization method as shown in a formula (1);

wherein x is^nor_mData, x, representing the normalized mth process variable affecting NOx emission concentration^mRepresents the m-th process variable, μ, in the raw data set collected^mAnd σ^mRespectively representing the mean and standard deviation of the mth process variable;

removing abnormal values in the original data by adopting a Rajda criterion, namely a method of three times of standard deviation higher than NOx concentration data, as shown in a formula (2);

|y_o-μ^NOx|≥3σ^NOx (2)

wherein, y_oRepresents the NOx concentration data, μ, corresponding to the o-th original data sample in the sample^NOxAnd σ^NOxNOx data y representing the mean and standard deviation of NOx concentration data, respectively, which will satisfy equation (2)_oIf the sample is regarded as an abnormal value, the o-th sample should be removed from the data set, the preprocessed data set is represented as S, and S is ═ S₁,s₂,...,s_N]^T，s₁,s₂,...,s_NRespectively representing the 1 st, 2 nd and Nth samples obtained after the abnormal value is removed, wherein N represents the size of the sample after the abnormal value is removed;

carrying out feature selection on the preprocessed data set S, and measuring the relation among process variables in each data set S through MI, wherein the MI is defined as shown in a formula (3);

wherein I represents the ith process variable sⁱAnd the jth process variable s^jDegree of correlation between, p(s)ⁱ) And p(s)^j) Are respectively sⁱAnd s^jP(s) as a function of the probability densityⁱ,s^j) For the ith process variable sⁱAnd j (th)Process variable s^jSo that the correlation between the maximised process ith process variable and the NOx concentration is expressed as

Where D represents the degree of correlation of each process variable in the data set S with the NOx concentration, SⁱAnd s^NOxRespectively an ith process variable and an NOx concentration variable, and S is a data set subjected to process variable normalization and NOx abnormal value elimination; similarly, minimizing redundancy between process variables is expressed as

maxΦ(D,R),Φ＝D-R (6)

The data set after feature selection is represented as Q, and the sample size and the number of the selected process variables are represented by N and F respectively;

dividing the data Q into three parts, wherein the division ratio is 2: 1: 1, each with Q₁、Q₂And Q₃It is shown that the three subdata sets are each of size N₁，N₂And N₃When Q is equal to Q₁∪Q₂∪Q₃，N＝N₁+N₂+N₃N denotes the size of the dataset before partitioning, first using Q₁Performing task decomposition to obtain process variable subsets under different operation conditions; then training a corresponding RBF sub-network model for the subset under each working condition, and establishing a mapping relation between the process variable and NOx under each operating condition; then using Q₂Testing the submodels established under each operating condition, and training an FCC neural network by utilizing the test output of the submodels to NOx; finally using Q₃Testing RBF submodels andand the FCC neural network completes the test of the whole soft measurement model of the NOx emission concentration.

3. The method of claim 1, wherein:

classifying the process data into different subsets through task decomposition; in MSWI, the NOx concentration emission characteristic is greatly influenced by the components of the garbage entering the furnace, air volume and temperature, so that the influence variable of the NOx concentration is clustered by adopting an FCM algorithm;

the set of samples to be clustered is Q₁By using

wherein U is a group consisting of U_cnThe formed membership matrix, V represents the clustering center of the data clusters under each operating condition, C represents the number of the data clusters, namely the number of the operating conditions, C represents any type of data in the C-th type of data clusters, namely any operating condition, and C is selected to be 3 and N is selected in the decomposition process because the operating conditions of the soft measurement of the NOx concentration are 3₁Represents the number of process variable samples, u_cnRepresents Q₁Membership degree and u of the nth sample to the c-type operation condition_cnE (0,1), r represents a weighted index with a value of 2, v_cRepresents the c-th cluster center point, q_nRepresenting a sample Q to be clustered₁The nth sample of (a); the sum of the membership degrees of each sample point under each operating condition is 1, so that the constraint condition satisfied by equation (7) is:

wherein u is_cnMembership value, v, indicating that the nth sample belongs to class c operating conditions_cRepresenting the clustering center of the data cluster of the c-th operation condition, clustering the process variables under the same operation condition by an FCM algorithm according to the similarity of the characteristic spatial distribution of the process variables, thereby obtaining the data clusters with different operation condition distributions in the MSWI process; according to the formula (13) and the formula (14), obtaining a fuzzy membership matrix U of the input process variable relative to each operation condition data cluster:

wherein C represents the number of operating conditions, N₁The size of a sample to be clustered is represented, the degree of similarity of each sample vector influencing the concentration of the NOx and the distribution of each operation condition data cluster is displayed by a membership degree matrix, and the higher the degree of similarity is, the higher the membership degree is, the degree of similarity u of the sample vector is_cnThe larger the process variable is, the higher the probability that the process variable belongs to the class c operation condition is, and the sum of membership degrees of the same sample vector under each operation condition is 1, as shown in formula (8); in order to accurately define the working condition category to which the sample vector belongs, the value of a membership threshold t to the data is 0.36;

{cluster(q_n)＝c|u_cn>t} (16)

wherein, cluster (·) is used for judging the sample q_nThe class to which the operating condition of (2) belongs; after task decomposition, the influence variables of the NOx emission concentration can be clustered into C operating conditions;

After data clusters under C working conditions are obtained through clustering, processing different subsets in a manner of dividing and treating by adopting an RBF network;

the RBF network consists of three parts, namely an input layer, a hidden layer and an output layer; input layerTransmitting process data affecting NOx into the network as a first partial data set Q₁After task decomposition, the current input data set is obtained

When the temperature of the water is higher than the set temperature,

and

Completing the nonlinear transformation from the input space to the hidden space, realizing the high-dimensional mapping of MSWI process variables, and recording the number of nodes of the hidden layer as H; the output layer is subjected to linear weighted combination in a new space to obtain a NOx test result; for the

Arbitrary sample of

Wherein the content of the first and second substances,

represent to

Is shown as

Wherein, w₀Is a deviation, w_h(H1., H) is a connection weight between the hidden layer and the output layer, a least square algorithm is adopted in a newrb function to optimize parameters, and a Mean Square Error (MSE) is used as a model performance index, as shown in formula (20);

wherein

True value, N, of NOx concentration for the nth training sample_cThe number of samples of the process variable under the same type c operation condition; setting MSE to 0.001 and the maximum node number H of the hidden layer_maxIs 40, then the value of H is based on the preset MSE and H_maxDetermining when the RBF network reaches MSE or H_maxWhen the number of the nodes of the hidden layer is not increased;

training of FCC networks from a second partial data set Q₂Is done, assuming in data set Q₂The nun groups of process variables in the system activate a plurality of RBF sub-modules, and the number of the activated modules is recorded as mod (mod)<C)，

The number of input nodes of the FCC network is mod, b represents the bias of the network, and its input can be setLet's be +1, ρ represents the input weight, i.e., ρ ═ ρ₁₁,ρ₁₂,...,ρ_1F；...；ρ_mod,1,ρ_mod,2,...,ρ_mod,F；ρ_b,1,...,ρ_b,f,...,ρ_b,F]，ρ₁₁,ρ₁₂,...,ρ_1FRepresenting the weight between the 1 st input node and each network node; rho_mod,1,ρ_mod,2,...,ρ_mod,FRepresenting weights between the mod input node and the respective network nodes; rho_b,1,...,ρ_b,f,...,ρ_b,FRepresents the offsets of the 1 st, F-th and F-th nodes; w denotes the internal weight between network nodes, i.e. W ═ W₁₂,w₁₃,...,w_1F；w₂₃,...,w_2F；...；w_F-1,F]When psi is ═ ρ, W]All weights are denoted by ψ;

indicating the integrated output of the module's ability to predict different NOx concentrations for the same process variable, i.e., the nth process variable, at multiple operating conditions_nRepresenting the true value of NOx corresponding to the sample, wherein F represents the number of network nodes and is 4;

selecting tanh (-) as the activation function of the first F-1 neurons, tanh (-) being expressed as

Is composed of