CN114611398A - Brain-like modular neural network-based soft measurement method for nitrogen oxides in urban solid waste incineration process - Google Patents
Brain-like modular neural network-based soft measurement method for nitrogen oxides in urban solid waste incineration process Download PDFInfo
- Publication number
- CN114611398A CN114611398A CN202210266639.9A CN202210266639A CN114611398A CN 114611398 A CN114611398 A CN 114611398A CN 202210266639 A CN202210266639 A CN 202210266639A CN 114611398 A CN114611398 A CN 114611398A
- Authority
- CN
- China
- Prior art keywords
- network
- module
- sample
- output
- node
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- MWUXSHHQAYIFBG-UHFFFAOYSA-N nitrogen oxide Inorganic materials O=[N] MWUXSHHQAYIFBG-UHFFFAOYSA-N 0.000 title claims abstract description 28
- 238000000034 method Methods 0.000 title claims abstract description 26
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 20
- 230000008569 process Effects 0.000 title claims abstract description 17
- 239000010813 municipal solid waste Substances 0.000 title claims abstract description 13
- 238000004056 waste incineration Methods 0.000 title claims abstract description 8
- 238000000691 measurement method Methods 0.000 title claims abstract description 6
- 238000005259 measurement Methods 0.000 claims abstract description 23
- 238000012360 testing method Methods 0.000 claims abstract description 18
- 238000012549 training Methods 0.000 claims description 39
- 210000002569 neuron Anatomy 0.000 claims description 26
- 238000004422 calculation algorithm Methods 0.000 claims description 15
- 230000009471 action Effects 0.000 claims description 14
- 239000011159 matrix material Substances 0.000 claims description 10
- 238000004364 calculation method Methods 0.000 claims description 8
- 230000006870 function Effects 0.000 claims description 8
- 238000010276 construction Methods 0.000 claims description 5
- 238000013461 design Methods 0.000 claims description 5
- 238000011156 evaluation Methods 0.000 claims description 4
- 238000002474 experimental method Methods 0.000 claims description 4
- 238000005192 partition Methods 0.000 claims description 4
- 238000009966 trimming Methods 0.000 claims description 4
- 210000004556 brain Anatomy 0.000 claims description 3
- 230000003213 activating effect Effects 0.000 claims description 2
- 230000003044 adaptive effect Effects 0.000 claims description 2
- 238000000354 decomposition reaction Methods 0.000 claims description 2
- 238000009795 derivation Methods 0.000 claims description 2
- 230000008030 elimination Effects 0.000 claims description 2
- 238000003379 elimination reaction Methods 0.000 claims description 2
- 238000013138 pruning Methods 0.000 claims description 2
- 238000010187 selection method Methods 0.000 claims 1
- 238000001514 detection method Methods 0.000 abstract description 3
- 238000007781 pre-processing Methods 0.000 abstract 1
- 101001095088 Homo sapiens Melanoma antigen preferentially expressed in tumors Proteins 0.000 description 4
- 102100037020 Melanoma antigen preferentially expressed in tumors Human genes 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 239000002910 solid waste Substances 0.000 description 3
- RZCJYMOBWVJQGV-UHFFFAOYSA-N 2-naphthyloxyacetic acid Chemical compound C1=CC=CC2=CC(OCC(=O)O)=CC=C21 RZCJYMOBWVJQGV-UHFFFAOYSA-N 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000003062 neural network model Methods 0.000 description 2
- 231100001143 noxa Toxicity 0.000 description 2
- 238000009270 solid waste treatment Methods 0.000 description 2
- ODUCDPQEXGNKDN-UHFFFAOYSA-N Nitrogen oxide(NO) Natural products O=N ODUCDPQEXGNKDN-UHFFFAOYSA-N 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000001133 acceleration Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 239000003344 environmental pollutant Substances 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 231100000719 pollutant Toxicity 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/02—Reliability analysis or reliability optimisation; Failure analysis, e.g. worst case scenario performance, failure mode and effects analysis [FMEA]
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A50/00—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE in human health protection, e.g. against extreme weather
- Y02A50/20—Air quality improvement or preservation, e.g. vehicle emission control or emission reduction by using catalytic converters
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Algebra (AREA)
- Computational Mathematics (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Mathematical Optimization (AREA)
- Medical Informatics (AREA)
- Pure & Applied Mathematics (AREA)
- Databases & Information Systems (AREA)
- Feedback Control In General (AREA)
- Investigating Or Analyzing Materials By The Use Of Fluid Adsorption Or Reactions (AREA)
Abstract
The invention relates to a brain-like modular neural network-based soft measurement method for nitrogen oxides in urban solid waste incineration process, which realizes NO soft measurementXThe method for accurately acquiring the concentration in real time comprises the following steps: firstly, acquiring data; preprocessing the acquired data, and determining an input variable and an output variable of the model; then, establishing a soft measurement model by adopting a brain-like modular neural network; and finally, the test data is used as the input of the model, and the validity of the model is verified. The invention effectively realizes NOXThe real-time accurate detection of the concentration has important theoretical significance and application value.
Description
Technical Field
The invention relates to Nitrogen Oxide (NO) in the process of burning urban solid wastesX) A soft measurement method; establishes NO based on Brain-like Modular Neural Network (BIMNN)XSoft measurement model, realizes to NOXAnd (4) accurately acquiring the concentration in real time. Not only belongs to the field of urban solid waste treatment, but also belongs to the field of intelligent modeling.
Background
With the rapid development of Chinese economy and the continuous acceleration of urbanization process, the urban solid waste yield increases day by dayCities face the crisis of "solid waste enclosure". The incineration treatment mode of solid wastes increasingly becomes the main mode of urban solid waste treatment in China. And NOXIs one of main pollutants generated in the process of incinerating urban solid wastes, and seriously influences the health of people and the quality of ecological environment. With the increasing requirements of environmental protection and treatment in China, the control of NOx emission becomes one of the key problems to be solved urgently in municipal solid waste incineration plants, and the real-time accurate detection of NOx is one of the important prerequisites for improving the denitration efficiency of the municipal solid waste incineration plants. Thus, realize NOXThe real-time accurate detection has important theoretical significance and application value.
Disclosure of Invention
The invention aims to provide a brain-like modular neural network-based urban solid waste incineration process NOXSoft measurement method, adopting brain-like modular neural network to establish NOXSoft measurement model, implementation to NOXAnd (4) accurately acquiring the concentration in real time.
The invention adopts the following technical scheme and implementation steps:
1. collecting data;
2. determining the input and output variables of the model: determining the input variable of the model by adopting a maximum correlation minimum redundancy (mRMR) algorithm, wherein the output variable of the model is NO at the current momentXConcentration;
the method for determining the input variables of the model by adopting the mRMR algorithm is as follows:
given two random variables a and b, the mutual information between the two random variables is calculated as follows:
wherein, I is mutual information between random variables a and b, and p (a) and p (b) are edge probability distribution of the random variables a and b respectively; p (a, b) is the joint distribution of random variables a and b;
firstly, based on mutual information, finding a feature subset S having the maximum correlation with a variable c to be measured:
wherein ,miThe method comprises the following steps of taking characteristic variables in a characteristic subset S, | S | is the number of the characteristic variables in the characteristic subset S, D is the correlation between the selected characteristic variables and a variable c to be measured, and if D is larger, the correlation between the selected characteristic variables and the variable c to be measured is higher;
considering that certain similarity exists among the selected feature variables, and the elimination of the 'redundant' feature does not affect the model performance, therefore, the redundancy among the features is calculated, and the 'mutually exclusive' feature is found:
wherein ,mi,njThe characteristic variables in the characteristic subset S are taken as R, the redundancy among the variables in the characteristic subset S is taken as R, and the smaller the R is, the lower the redundancy is;
in the mRMR algorithm application process, the maximum correlation index D and the minimum redundancy index R are generally unified into an evaluation function Φ ═ D-R, and then an optimal feature subset S is determined by finding the maximum value of the evaluation function:
maxΦ(D,R),Φ=D-R (4)
3. designed for NOXA brain-like modular neural network model for soft measurement of concentration;
(1) task decomposition
In order to measure and evaluate the modularization degree of a network and simulate the modularization characteristic of a brain network, a modularization index (MQ) facing a modularization neural network is provided; the modularization index consists of the density in the module and the sparsity among the modules, wherein the density in the module is calculated as follows:
wherein ,JCIs the density in the moduleSet degree, P is the number of modules in the current network, NlNumber of samples assigned to the l-th module, xiTo input samples, hl and rlRespectively representing the position and the action range of the core node of the ith module;
the sparsity between the modules is calculated based on the Euclidean distance:
wherein ,JSD (h) is the degree of sparsity between modulesl,hs) Representing the distance between the core node of the l-th module and the core node of the s-th module, and comprehensively considering the density J in the moduleCAnd degree of sparsity J between modulesSThe modularization index measurement mode is provided as follows:
therefore, the larger the value of MQ is, the higher the modularization degree of the network is, so a brain-like modularization partition method is proposed, and the main idea is as follows: firstly, distributing training samples through core nodes to determine whether to distribute the training samples to a current existing module or a newly added module; then, a new core node of an existing module is determined by seeking the maximum "modularity" degree of the network, and therefore, the modular structure construction can be divided into two cases: adding new modules and updating existing modules;
adding new module
At the initial moment, the number of modules of the whole network is 0;
when the first data sample enters the network, setting the first data sample as a core node of a first submodule:
h1=x1 (8)
wherein ,h1 and r1Respectively representing the location and range of action, x, of the first module core node1As input vector for the first training sample, dmaxFor training sample xiAnd xjThe maximum distance therebetween;
at time t, when the tth training sample enters the network, assuming that k modules exist, finding the core node closest to the sample:
wherein ,xtIs the input vector of the t-th training sample, hsDenotes the location of the core node of the s-th module, kminRepresenting distance training samples xtThe nearest core node;
if the t-th training sample is not in kminIn the action range of the core node, a new module is needed to learn the current sample, and the parameters of the core node corresponding to the new module are set as follows:
hk+1=xt (12)
wherein ,hk+1,rk+1Position and range of action of core node corresponding to newly added module, xtIs the input vector for the t-th training sample,the farthest distance from other core nodes to the core node of the newly added module;
② optimizing existing modules
Otherwise, the sample is considered to be classifiedIs kminIn the module, in order to make the network have the optimal modularization degree, according to the formula (7), modularization index values MQ of the whole network under the condition that the current sample and the original core node are respectively used as the core nodes are respectively calculatedtAnd
if it isIf the network modularization degree of the current input sample as the core node is considered to be higher, the sample is used for replacing the existing core node to become a new core node, and the initial parameter setting is as follows:
wherein ,respectively as a new core node and an action range of the module; n is a radical ofkIs the number of samples allocated to the kth module;
if it isThe location of the current core nodeKeeping unchanged, only adjusting the action range of the nodeNamely:
after all training samples are compared, the samples are distributed to different sub-modules, a partition structure is formed, the modularization degree of the current network can be considered to be the maximum, and then a sub-network needs to be constructed according to the task of each sub-module;
(2) subnetwork structure design
wherein ,xi,yiRespectively the input variables and the output variables of the model,representing the domain, V being the input vector xiN is the number of model input variables and output variable data;
an adaptive task-oriented radial basis function neural network (ATO-RBF) is adopted to construct a sub-network corresponding to each module, and the design of the sub-network comprises three parts: network structure growth, network structure pruning and network parameter adjustment;
(ii) network architecture growth
The center, radius and connection weight to the output layer of the node of the first hidden layer of the s-th module are based on the sample with the largest absolute outputSetting:
wherein ,for the sample with the largest absolute output,respectively representing the input variable and the output variable corresponding to the sample with the largest absolute output, rsIs the range of action of the s-th module, NsIs the number of samples assigned to the s-th module;the center and the radius of a first hidden layer node of the s-th module and the connection weight to the output layer are respectively;
at time tsAt time, training error vector e (t)s) Obtained by the following formula:
wherein ,yfIs the desired output of the f-th sample,is the f-th sample at time tsIs calculated by the following equation:
wherein H is the number of neurons in the hidden layer,is the function of the jth hidden layer node of the s-th module,for the jth hidden layer node of the s-th module to the output layer connection weight,the center and radius of the jth hidden layer node of the s-th module;
Then, adding an RBF neuron pairLearning is carried out on each sample, and the initial parameters of the newly added neurons are as follows:
wherein Andrespectively representing the center of the new neuron of the s-th module and the connection weight to the output layer;is a firstThe input variable corresponding to each of the samples,are respectively the firstExpected output sum corresponding to each sample at tsA network output value of a time;
when the following relation is satisfied, the influence of the existing neurons on the newly added neurons is small:
from equations (27) and (28), the radius of the newly added neuron is set as:
when a neuron is newly added, network parameters are adjusted through a second-order learning algorithm; when reaching the preset maximum structure JmaxOr period of timeInspection training precision E0The network structure growth process ends; in the course of the experiment Jmax=10,E0The training accuracy of the network is measured using Root Mean Square Error (RMSE) at 0.0001, and is calculated as follows:
wherein ,yi and yiThe expected output and the network output of the ith sample are respectively;
② trimming of network structure
In order to avoid redundancy of the network structure, it is proposed to measure the contribution value of hidden layer neurons based on the index of the connection weight:
wherein SI (j) is the contribution value of the jth hidden layer node,is the weight from the jth hidden layer node of the s-th module to the output layer;
finding the hidden layer node with the smallest contribution value:
wherein ,the hidden layer node with the minimum contribution value in the s-th module, and J is the number of the hidden layer nodes in the network;
therefore, deleting the hidden node with the smallest contribution, adjusting parameters through a second-order learning algorithm, and comparing the root mean square error value (RMSE _1) after the node is deleted with the root mean square error value (RMSE _0) when the node is not deleted, wherein the calculation formula of RMSE is shown as a formula (30); if RMSE _1 is less than or equal to RMSE _0, the selected nodes can be pruned under the condition of not sacrificing the network learning capability, and then the process is repeated, otherwise, the selected nodes cannot be deleted; at this time, the network structure trimming process is finished, and the sub-network construction is completed; adjusting parameters by using a second-order learning algorithm every time when the neurons are deleted;
adjusting network parameters
The second order learning algorithm is as follows:
θL+1=θL-(QL+μLE)-1gL (33)
wherein θ refers to parameters to be adjusted, including center, radius and connection weight, Q is a Hessian-like matrix, μ is a learning coefficient (0.01 in the experiment), E is an identity matrix, g is a gradient vector, and L is an iteration step number (set to 50 in the experiment);
in order to reduce memory requirements and calculation time, the calculation of the Hessian-like matrix Q and the gradient vector g is converted into Hessian-like sub-matrix summation and gradient sub-vector summation:
qzis a Hessian-like sub-matrix, ηzThe gradient subvectors can be calculated by the following formula:
wherein ,ezIs the desired output y of the z-th samplezAnd network prediction outputDifference of jzFor the Jacobian vector, the calculation is as follows:
according to the chain derivation rule, each component in the jacobian vector in equation (39) is calculated as follows:
wherein The center and radius of the jth neuron of the s-th module and the connection weight value x to the output layerzIs the input vector of the z-th training sample;
4、NOXsoft measurement of concentration;
taking test sample data as the input of the brain-like modular neural network, wherein the output of the model is NOXSoft measurements of concentration; and at time T, after the Tth test sample enters the BIMNN, searching a core node closest to the sample, and activating a sub-module to which the core node belongs:
wherein ,xTAs input vector for the Tth test sample, hsIs the core node of the s-th module, lactIs a distance of the Tth test sample xTThe nearest core node belongs to the sub-network, and A is the number of sub-network modules;
therefore, the actual output of BIMNN is the lactOutput of the sub-network:
wherein ,for the actual output of the BIMNN at time T,is the firstactThe actual output of the sub-network;
and (3) quantitatively evaluating the test precision by adopting a Root Mean Square Error (RMSE), an average percent error (MAPE) and a Correlation Coefficient (CC), wherein the RMSE is calculated according to a formula (30), and the MAPE and CC are calculated according to the following formulas:
in the formula ,yiAndexpected output and net output, N, respectively, for the ith sampleTThe number of samples to be tested.
The invention has the following obvious advantages and beneficial effects:
1 the invention is based on the classGood nonlinear mapping capability and generalization capability of brain modular neural network, and establishes stable and effective NOXSoft concentration measurement model for NOXThe concentration is accurately obtained in real time, and NO is generated in the process of burning urban solid wastesXEmission control is of great significance.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a diagram of a brain-like modular neural network architecture;
FIG. 3 is a diagram of the structure of an ATO-RBF neural network;
FIG. 4 is a graph of the training results of subnetwork 1;
FIG. 5 is a graph of the results of the training of subnetwork 2;
FIG. 6 is a graph of the results of the training of subnetwork 3;
FIG. 7 is a BIMNN soft measurement model test output diagram;
FIG. 8 is a BIMNN soft measurement test error graph.
Detailed Description
The invention uses training data to establish for NOXA brain-like modular neural network model for soft measurement of concentration; verifying NO output by brain-like modular neural network soft measurement model by using test data setXAccuracy of real-time concentration.
As an example, the effectiveness of the method provided by the invention is verified by using actual data from a solid waste incineration plant in a certain city of Beijing. After removing obvious abnormal data, 1000 groups of 96-dimensional experimental data are obtained. Based on the obtained data sample, adopting an mRMR algorithm to perform feature selection, selection and NOXThe 20 variables with larger correlation are used as input variables of the soft measurement model, and are shown in table 1 in detail.
TABLE 1
For 1000 groups of data after dimension reduction, 750 groups of data are used for establishing a soft measurement model, and the other 250 groups of data are used for testing the performance of the model;
(1) based on 750 groups of training data, adopting a brain-like modular partitioning method, dividing 750 groups of training data into three subsets, wherein the sample number of each subset is 215, 273 and 262 respectively; correspondingly, the BIMNN consists of 3 sub-networks, and the sub-networks are established by the sub-network construction method in step 3; fig. 4, 5, and 6 are graphs of training results of respective sub-networks, where X-axis: number of training samples, in units of units per sample, Y-axis: NOXConcentration in mg/Nm3;
(2) NO by brain-like modular neural network based on 250 sets of test dataXConcentration soft measurement, the test results are shown in fig. 7, X-axis: number of samples tested, in units of units per sample, Y-axis: NOXConcentration in mg/Nm3(ii) a Test error as shown in fig. 8, X-axis: number of training samples, in units of units per sample, Y-axis: NOXError in concentration measurement in mg/Nm3;
(3) The measurement accuracy is quantitatively evaluated by using the root mean square error RMSE, the average percentage error MAPE and the correlation coefficient CC, and the calculation result is that RMSE is 6.5031, MAPE is 3.8514% and CC is 0.9777.
It will be apparent to those skilled in the art that various modifications and variations can be made in the present invention without departing from the spirit or scope of the invention. It should be noted that the above-mentioned embodiments illustrate rather than limit the invention, and that modifications and variations can be made without departing from the spirit and scope of the invention as defined by the appended claims.
Claims (1)
1. A brain-like modular neural network-based soft measurement method for nitrogen oxides in an urban solid waste incineration process is characterized by comprising the following steps:
step 1, data acquisition;
step 2, determining model input and output variables;
model input variables passing maximum phaseDetermining by a minimum redundancy mRMR algorithm, wherein the output variable of the model is NO at the current momentXConcentration;
step 3, designing a brain-like modular neural network, and establishing a soft measurement model;
step 4, using the test data as the input of the model, wherein the output of the model is NO at the current momentXA concentration measurement;
in step 2, the input variable selection method based on the mRMR algorithm is as follows:
given two random variables a and b, the mutual information between the two random variables is calculated as follows:
wherein, I is mutual information between random variables a and b, and p (a) and p (b) are edge probability distribution of the random variables a and b respectively; p (a, b) is the joint distribution of random variables a and b;
firstly, based on mutual information, finding a feature subset S having the maximum correlation with a variable c to be measured:
wherein ,miThe method comprises the following steps of taking characteristic variables in a characteristic subset S, | S | is the number of the characteristic variables in the characteristic subset S, D is the correlation between the selected characteristic variables and a variable c to be measured, and if D is larger, the correlation between the selected characteristic variables and the variable c to be measured is higher;
considering that certain similarity exists among the selected characteristic variables, and the elimination of the 'redundant' characteristic does not influence the performance of the model; therefore, the redundancy between features is calculated, and the "mutually exclusive" features are found:
wherein ,mi,njThe characteristic variables in the characteristic subset S are taken as R, the redundancy among the variables in the characteristic subset S is taken as R, and the smaller the R is, the lower the redundancy is;
in the mRMR algorithm application process, the maximum correlation index D and the minimum redundancy index R are generally unified into an evaluation function Φ — D-R, and then the optimal feature subset S is determined by finding the maximum value of the evaluation function:
maxΦ(D,R),Φ=D-R (4)
in step 3, the design method of the soft measurement model based on the brain-like modular neural network is as follows:
(1) task decomposition
In order to measure and evaluate the modularization degree of the network and simulate the modularization characteristic of the brain network, a modularization index MQ facing to a modularization neural network is provided, the modularization index MQ consists of the density degree in the modules and the sparsity degree among the modules, wherein the density degree in the modules is calculated as follows:
wherein ,JCFor the density within a module, P is the number of modules in the current network, NlNumber of samples assigned to the l-th module, xiTo input samples, hl and rlRespectively representing the position and the action range of the core node of the first module;
the sparsity between the modules is calculated based on the Euclidean distance:
wherein ,JSD (h) is the degree of sparsity between modulesl,hs) Representing the distance between the core node of the l-th module and the core node of the s-th module, and comprehensively considering the density J in the moduleCAnd degree of sparsity J between modulesSThe modularization index measurement mode is provided as follows:
therefore, the larger the value of MQ is, the higher the modularization degree of the network is, so a brain-like modularization partition method is proposed, and the main idea is as follows: firstly, distributing training samples through core nodes to determine whether to distribute the training samples to a current existing module or a newly added module; then, determining a new core node of the existing module by seeking the maximum modularization degree of the network; thus, modular structure construction can be divided into two cases: adding new modules and updating existing modules;
adding new module
At the initial moment, the number of modules of the whole network is 0;
when the first data sample enters the network, setting the first data sample as a core node of a first sub-module:
h1=x1 (8)
wherein ,h1 and r1Respectively representing the location and range of action, x, of the first module core node1As input vector for the first training sample, dmaxFor training sample xiAnd xjThe maximum distance therebetween;
at time t, when the tth training sample enters the network, assuming that k modules exist, finding the core node closest to the sample:
wherein ,xtIs the input vector of the t-th training sample, hsDenotes the location of the core node of the s-th module, kminRepresenting distance training sample xtThe nearest core node;
if the t-th training sample xtIs not at kminIn the action range of the core node, a new module is needed to learn the current sample, and the parameters of the core node corresponding to the new module are set as follows:
hk+1=xt (12)
wherein ,hk+1,rk+1Position and range of action of core node corresponding to newly added module, xtIs the input vector for the t-th training sample,the farthest distance from other core nodes to the core node of the newly added module;
② optimizing existing modules
Otherwise, the sample is considered to be classified as kminIn the module, in order to make the network have the optimal modularization degree, according to the formula (7), modularization index values MQ of the whole network under the condition that the current sample and the original core node are respectively used as the core nodes are respectively calculatedtAnd
if it isThen the network modularization degree of the current input sample selected as the core node is considered to be higher, and the sample is used for replacing the existing core nodeThe new core node, the initial parameters set as follows:
wherein ,respectively as a new core node and an action range of the module; n is a radical ofkIs the number of samples assigned to the kth module;
if it isThe location of the current core nodeKeeping unchanged, only adjusting the action range of the nodeNamely:
after all training samples are compared, the samples are distributed to different sub-modules, a partition structure is formed, the modularization degree of the current network is considered to be the maximum, and then a sub-network needs to be constructed according to the task of each sub-module;
(2) subnetwork structure design
wherein ,xi,yiRespectively the input variables and the output variables of the model,representing the domain, V being the input vector xiN is the number of model input variables and output variable data;
an adaptive task-oriented radial basis function neural network (ATO-RBF) is adopted to construct a sub-network corresponding to each module, and the design of the sub-network comprises three parts: network structure growth, network structure pruning and network parameter adjustment;
(ii) network architecture growth
The center, radius and connection weight to the output layer of the node of the first hidden layer of the s-th module are based on the sample with the largest absolute outputSetting:
wherein ,for the sample with the largest absolute output,respectively representing the input variable and the output variable corresponding to the sample with the largest absolute output, rsIs the range of action of the s-th module, NsIs the number of samples assigned to the s-th module;the center and the radius of a first hidden layer node of the s-th module and the connection weight to the output layer are respectively;
at time tsAt time, training error vector e (t)s) Obtained by the following formula:
wherein ,yfIs the desired output of the f-th sample,is the f-th sample at time tsIs calculated by the following equation:
wherein H is the number of neurons in the hidden layer,is the function of the jth hidden layer node of the s-th module,for the jth hidden layer node of the s-th module to the output layer connection weight,the center and radius of the jth hidden layer node of the s-th module;
Then, adding a new RBF neuron pairLearning is carried out on each sample, and the initial parameters of the newly added neurons are as follows:
wherein Andrespectively representing the center of the new neuron of the s-th module and the connection weight to the output layer;is a firstThe input variable corresponding to each of the samples,are respectively the firstExpected output sum corresponding to each sample at tsA network output value of a time;
when the following relation is satisfied, the influence of the existing neurons on the newly added neurons is small:
from equations (27) and (28), the radius of the newly added neuron is set as:
when a neuron is newly added, network parameters are adjusted through a second-order learning algorithm; when reaching the preset maximum structure JmaxOr desired training accuracy E0The network structure growth process ends; in the course of the experiment Jmax=10,E00.0001, adoptThe root mean square error RMSE is used to measure the training accuracy of the network and is calculated as follows:
wherein ,yi and yiThe expected output and the network output of the ith sample are respectively;
② trimming of network structure
In order to avoid redundancy of the network structure, it is proposed to measure the contribution value of hidden layer neurons based on the index of the connection weight:
wherein SI (j) is the contribution value of the jth hidden layer node,is the weight from the jth hidden layer node of the s-th module to the output layer;
finding the hidden layer node with the smallest contribution value:
wherein ,the hidden layer node with the minimum contribution value in the s-th module, and J is the number of the hidden layer nodes in the network;
therefore, the hidden node with the smallest contribution is deleted, parameters are adjusted through a second-order learning algorithm, the root mean square error value (RMSE _1) after the node is deleted is compared with the root mean square error value (RMSE _0) when the node is not deleted, and the calculating formula of RMSE is shown as a formula (30); if RMSE _1 is less than or equal to RMSE _0, the selected nodes can be pruned under the condition of not sacrificing the network learning capability, and then the process is repeated, otherwise, the selected nodes cannot be deleted; at this time, the network structure trimming process is finished, and the sub-network construction is completed; adjusting parameters by using a second-order learning algorithm every time when the neurons are deleted;
network parameter adjustment
The second order learning algorithm is as follows:
θL+1=θL-(QL+μLE)-1gL (33)
wherein theta refers to parameters needing to be adjusted and comprises a center, a radius and a connection weight, Q is a Hessian-like matrix, mu is a learning coefficient, and 0.01 is taken; e is an identity matrix, g is a gradient vector, and L is an iteration step number set to be 50;
in order to reduce memory requirements and calculation time, the calculation of the Hessian-like matrix Q and the gradient vector g is converted into Hessian-like sub-matrix summation and gradient sub-vector summation:
qzis a Hessian-like sub-matrix, ηzThe gradient subvectors can be calculated by the following formula:
wherein ,ezIs the desired output y of the z-th samplezAnd network prediction outputDifference of jzFor the Jacobian vector, the calculation is as follows:
according to the chain derivation rule, each component in the jacobian vector in equation (39) is calculated as follows:
wherein The center and radius of the jth neuron of the s-th module and the connection weight value x to the output layerzIs the input vector of the z-th training sample;
in step 4, the test data is used as the input of the model, and the output of the model is NO at the current momentXA concentration measurement;
and at time T, after the Tth test sample enters the BIMNN, searching a core node closest to the sample, and activating a sub-module to which the core node belongs:
wherein ,xTAs input vector for the Tth test sample, hsIs the core node of the s-th module, lactIs a distance of the Tth test sample xTThe nearest core node belongs to the sub-network, and A is the number of sub-network modules;
therefore, the actual output of BIMNN is the lactOutput of the sub-network:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210266639.9A CN114611398B (en) | 2022-03-17 | 2022-03-17 | Soft measurement method for nitrogen oxides in urban solid waste incineration process based on brain-like modularized neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210266639.9A CN114611398B (en) | 2022-03-17 | 2022-03-17 | Soft measurement method for nitrogen oxides in urban solid waste incineration process based on brain-like modularized neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114611398A true CN114611398A (en) | 2022-06-10 |
CN114611398B CN114611398B (en) | 2023-05-12 |
Family
ID=81865131
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210266639.9A Active CN114611398B (en) | 2022-03-17 | 2022-03-17 | Soft measurement method for nitrogen oxides in urban solid waste incineration process based on brain-like modularized neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114611398B (en) |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112733876A (en) * | 2020-10-28 | 2021-04-30 | 北京工业大学 | Soft measurement method for nitrogen oxides in urban solid waste incineration process based on modular neural network |
CN113077039A (en) * | 2021-03-22 | 2021-07-06 | 北京工业大学 | Task-driven RBF neural network-based water outlet total nitrogen TN soft measurement method |
CN113780639A (en) * | 2021-08-29 | 2021-12-10 | 北京工业大学 | Urban solid waste incineration nitrogen oxide NOx emission prediction method based on multitask learning framework |
-
2022
- 2022-03-17 CN CN202210266639.9A patent/CN114611398B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112733876A (en) * | 2020-10-28 | 2021-04-30 | 北京工业大学 | Soft measurement method for nitrogen oxides in urban solid waste incineration process based on modular neural network |
CN113077039A (en) * | 2021-03-22 | 2021-07-06 | 北京工业大学 | Task-driven RBF neural network-based water outlet total nitrogen TN soft measurement method |
CN113780639A (en) * | 2021-08-29 | 2021-12-10 | 北京工业大学 | Urban solid waste incineration nitrogen oxide NOx emission prediction method based on multitask learning framework |
Also Published As
Publication number | Publication date |
---|---|
CN114611398B (en) | 2023-05-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Genaro et al. | A neural network based model for urban noise prediction | |
CN106920007B (en) | PM based on second-order self-organizing fuzzy neural network2.5Intelligent prediction method | |
CN106529818B (en) | Water quality assessment Forecasting Methodology based on Fuzzy Wavelet Network | |
CN110309609B (en) | Building indoor air quality evaluation method based on rough set and WNN | |
CN106656357B (en) | Power frequency communication channel state evaluation system and method | |
CN114154401B (en) | Soil erosion modulus calculation method and system based on machine learning and observation data | |
CN108304674A (en) | A kind of railway prediction of soft roadbed settlement method based on BP neural network | |
Assas et al. | Use of the artificial neural network and meteorological data for predicting daily global solar radiation in Djelfa, Algeria | |
Kavyanifar et al. | Coastal solid waste prediction by applying machine learning approaches (Case study: Noor, Mazandaran Province, Iran) | |
CN111967308A (en) | Online road surface unevenness identification method and system | |
CN113657023A (en) | Near-surface ozone concentration inversion method based on combination of machine learning and deep learning | |
CN111914465B (en) | Clustering and particle swarm optimization-based method for calibrating hydrologic parameters of non-data region | |
CN114611398A (en) | Brain-like modular neural network-based soft measurement method for nitrogen oxides in urban solid waste incineration process | |
CN117194918A (en) | Air temperature prediction method and system based on self-attention echo state network | |
Wang | A neural network algorithm based assessment for marine ecological environment | |
Bakhary et al. | Univariate artificial neural network in forecasting demand of low cost house in Petaling Jaya | |
Fotouhi et al. | A sensitivity analysis on the air quality index based on the pollutants concentrations in Tehran | |
Furze et al. | Implementing stochastic distribution within the utopia plane of primary producers using a hybrid genetic algorithm | |
CN116796185A (en) | Urban solid waste incineration process NO based on attention modularized neural network x Prediction method | |
CN113722970B (en) | Photovoltaic power ultra-short-term online prediction method | |
Dong et al. | Effluent BOD soft measurement based on MNN | |
CN116882446A (en) | Intelligent odor monitoring and early warning method in household garbage collection process | |
Kai et al. | An adaptive hybrid wavelet neural network and its application | |
Gupta et al. | Predictive Analysis of Air Pollution Using Collaborative Filtering Prediction Algorithm | |
Chen et al. | An approach to design of maintenance float systems |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |