CN113239991B - Flame image oxygen concentration prediction method based on regression generation countermeasure network - Google Patents
Flame image oxygen concentration prediction method based on regression generation countermeasure network Download PDFInfo
- Publication number
- CN113239991B CN113239991B CN202110467231.3A CN202110467231A CN113239991B CN 113239991 B CN113239991 B CN 113239991B CN 202110467231 A CN202110467231 A CN 202110467231A CN 113239991 B CN113239991 B CN 113239991B
- Authority
- CN
- China
- Prior art keywords
- model
- flame image
- regression
- oxygen concentration
- training
- 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.)
- Active
Links
- QVGXLLKOCUKJST-UHFFFAOYSA-N atomic oxygen Chemical compound [O] QVGXLLKOCUKJST-UHFFFAOYSA-N 0.000 title claims abstract description 52
- 239000001301 oxygen Substances 0.000 title claims abstract description 52
- 229910052760 oxygen Inorganic materials 0.000 title claims abstract description 52
- 238000000034 method Methods 0.000 title claims abstract description 38
- 238000012549 training Methods 0.000 claims abstract description 50
- 230000008569 process Effects 0.000 claims abstract description 10
- 238000011156 evaluation Methods 0.000 claims abstract description 8
- 238000007781 pre-processing Methods 0.000 claims abstract description 4
- 238000013527 convolutional neural network Methods 0.000 claims description 28
- 238000002474 experimental method Methods 0.000 claims description 12
- 238000011176 pooling Methods 0.000 claims description 12
- 230000006870 function Effects 0.000 claims description 9
- 238000005070 sampling Methods 0.000 claims description 9
- 238000002485 combustion reaction Methods 0.000 claims description 8
- 238000012545 processing Methods 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 6
- 239000007789 gas Substances 0.000 claims description 6
- CURLTUGMZLYLDI-UHFFFAOYSA-N Carbon dioxide Chemical compound O=C=O CURLTUGMZLYLDI-UHFFFAOYSA-N 0.000 claims description 4
- 230000008485 antagonism Effects 0.000 claims description 4
- 230000006872 improvement Effects 0.000 claims description 4
- 238000012360 testing method Methods 0.000 claims description 4
- UGFAIRIUMAVXCW-UHFFFAOYSA-N Carbon monoxide Chemical compound [O+]#[C-] UGFAIRIUMAVXCW-UHFFFAOYSA-N 0.000 claims description 3
- 238000013528 artificial neural network Methods 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 3
- 229910002091 carbon monoxide Inorganic materials 0.000 claims description 3
- 238000000638 solvent extraction Methods 0.000 claims description 3
- 229910002092 carbon dioxide Inorganic materials 0.000 claims description 2
- 239000001569 carbon dioxide Substances 0.000 claims description 2
- 238000010586 diagram Methods 0.000 description 5
- 210000002569 neuron Anatomy 0.000 description 4
- MWUXSHHQAYIFBG-UHFFFAOYSA-N Nitric oxide Chemical compound O=[N] MWUXSHHQAYIFBG-UHFFFAOYSA-N 0.000 description 3
- 238000013145 classification model Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000010606 normalization Methods 0.000 description 3
- 239000002912 waste gas Substances 0.000 description 3
- 239000000446 fuel Substances 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 238000003723 Smelting Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000033228 biological regulation Effects 0.000 description 1
- UBAZGMLMVVQSCD-UHFFFAOYSA-N carbon dioxide;molecular oxygen Chemical compound O=O.O=C=O UBAZGMLMVVQSCD-UHFFFAOYSA-N 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000001934 delay Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 239000000295 fuel oil Substances 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 239000003921 oil Substances 0.000 description 1
- 230000001105 regulatory effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Mathematical Optimization (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Pure & Applied Mathematics (AREA)
- Evolutionary Biology (AREA)
- Computational Mathematics (AREA)
- Mathematical Analysis (AREA)
- Computing Systems (AREA)
- Operations Research (AREA)
- Probability & Statistics with Applications (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
- Image Analysis (AREA)
Abstract
A flame image oxygen concentration prediction method based on regression generation countermeasure network belongs to the technical field of unbalanced learning of regression models. It comprises the following steps: 1. acquiring furnace flame image data; 2. preprocessing furnace flame image data and dividing a data set; 3. generating a minority interval sample in the furnace flame image dataset; 4. establishing and training a flame image oxygen concentration prediction model; 5. model performance evaluation. According to the invention, the flame image of a few intervals in the data set is expanded by using the regression generation countermeasure network based on gradient punishment, so that an unbalanced furnace flame image data set is balanced, the unbalanced learning problem existing in the training process of the flame image oxygen concentration prediction model is solved, the prediction accuracy of the regression model on the flame image oxygen concentration is improved, and the method has universality and universality.
Description
Technical Field
The invention belongs to the technical field of unbalanced learning of regression models, and particularly relates to a flame image oxygen concentration prediction method for generating an countermeasure network based on regression.
Background
In industrial combustion systems, combustion efficiency may be determined by measuring the heat conversion efficiency of fuel. In order to reduce the running cost and meet the requirements of environmental regulations, the combustion efficiency and the exhaust emission content should be controlled at proper levels. The oxygen and nitrogen oxide content in the exhaust gas is typically measured using a gas analyzer, but this approach can create delays and feedback controllers based on the oxygen content have a tendency to overcompensate. As an alternative, the current combustion state is reflected by sufficient information provided by the flame image measured on-line. Based on the deep learning technology, a convolutional neural network model capable of predicting the oxygen concentration in the furnace according to the flame image is established, and real-time monitoring and control are realized.
However, the acquired furnace flame image data is unbalanced, which makes training of Convolutional neural network (Convolutional NeuralNetwork, CNN) difficult. The unbalanced learning problem generally occurs in classification tasks, and when there is a large difference in the number of samples of each category in the data set, the category with small data amount is easily submerged in the category with large data amount, so that it is difficult for the classification model to accurately grasp the features to identify the different categories. Although flame image oxygen concentration prediction is a regression task, by sorting the acquired data, it was found that flame image data can be divided into 11 sections according to the oxygen concentration content, and the number of samples in each section is not uniform, which is also an unbalanced data set. When the number of samples in one interval is far smaller than that of other intervals, the prediction model is biased towards the interval with more samples, so that the prediction in the interval with less samples is inaccurate.
Aiming at the problem of training the model on the unbalanced data set, minority samples in the unbalanced data set can be amplified to enable the data set to regress and balance, such as random oversampling, synthesized minority oversampling technology, self-adaptive synthesized sampling and the like. The method is widely applied to solving the unbalanced learning problem of the classification model, but is not suitable for the regression model. Because the tag values of one class sample are the same in the dataset of the classification model, whereas in our flame image data the tag values of samples within one interval are different and consecutive. Only this continuously varying process is captured, so that the synthesized sample is more realistic.
Generating a countering network (Generative Adversarial Networks, GAN) is a powerful sample generation method. GAN learns the distribution of data from the training data by the countermeasure training of the discriminator and the generator, and can generate data similar to the training data. The GAN is utilized to generate a sample and balance data set of a few intervals, and the method is a feasible method for solving the unbalanced flame image oxygen concentration prediction model.
Disclosure of Invention
Aiming at the problems in the prior art, the invention aims to provide a flame image oxygen concentration prediction method based on regression generation countermeasure network, which has high prediction accuracy and universality.
The invention provides the following technical scheme: the method for predicting the oxygen concentration of the flame image based on the regression generation countermeasure network is characterized by comprising the following steps of: the method comprises the following steps:
(1) Acquiring furnace flame image data:
performing a combustion experiment in an experiment furnace, sampling the exhaust temperature, oxygen concentration, carbon dioxide concentration and carbon monoxide concentration of the filtered exhaust gas, capturing an image of flame in the experiment furnace, and transmitting the image to on-site processing equipment, wherein the oxygen concentration and the flame image in the experiment furnace are automatically recorded in the running processing process of the on-site processing equipment;
(2) Preprocessing of furnace flame image data and data set partitioning:
in order to accelerate the convergence rate of the model, reduce the training time of the model, normalize the data, and divide the data set into training set and test set;
(3) Generating a minority interval sample in the furnace flame image dataset:
establishing a regression generation countermeasure network RGAN-GP model based on gradient penalty, taking a few interval images in a furnace flame image data set as training samples, training the regression generation countermeasure network RGAN-GP model, and generating a few interval samples and balancing a flame image data set after the regression generation countermeasure network RGAN-GP model is trained;
(4) Establishing and training a flame image oxygen concentration prediction model:
establishing a flame image oxygen concentration prediction model, taking the balanced flame image data set as a training set, and training the flame image oxygen concentration prediction model;
(5) Model performance evaluation:
and (3) introducing an evaluation index Root Mean Square Error (RMSE) and a relative improvement Rate (RIMP) to evaluate the flame image oxygen concentration prediction model trained in the step (4).
The regression-based flame image oxygen concentration prediction method for generating the countermeasure network is characterized in that the process of the step (3) is as follows:
step 3.1: establishing a regression generation countermeasure network model based on gradient penalty:
the regression generation countermeasure network RGAN-GP model comprises a generator G, a discriminator D and a regression model R, wherein the generator G, the discriminator D and the regression model R are all neural networks, the generator G is used for capturing the distribution of real data and generating samples similar to the real data, the discriminator is used for judging whether the input of the generator G is the real data or the generated data, the regression model R is used for constraining the generation model G, the samples generated by the generation model correspond to the condition variables of the generator G, and the loss functions of the discriminator R, the generator G and the regression model R are as follows:
wherein: p (P) data (x) Representing a probability distribution of the real data x; p (P) z (z) represents a probability distribution of noise z; g (z|c) represents the generated data under the condition variable c;representing the sampling distribution +.>Epsilon represents an interpolation parameter; d () represents the output of the arbiter, +.>Representing the gradient 2 norms of the discriminator; r () tableShowing the output of the regression model; e represents the desire; λ represents a gradient penalty coefficient; alpha represents the constraint coefficient of the regression model to the generator;
step 3.2: training a regression based on gradient penalty using a few interval samples to generate an countermeasure network model and generate new samples, the specific process is as follows:
taking a few interval samples in an original training set as a training set, training regression to generate an antagonism network RGAN-GP model, and alternately training a generator G, a regression model R and a discriminator D, wherein the generator G can finally generate samples which enable the discriminator D to be incapable of discriminating true and false; after training the regression generation countermeasure network RGAN-GP model is completed, a generator G is used for generating a certain number of samples, the samples are mixed with the original data set, and the unbalance rate of the number of samples in each interval in the data set is 1:2, i.e., the interval sample size of the least samples and the interval sample size of the most samples is 1:2 or less;
the regression-based flame image oxygen concentration prediction method for generating the countermeasure network is characterized in that the process of the step (4) is as follows:
step 4.1: establishing a flame image oxygen concentration prediction model:
a convolutional neural network is used for predicting oxygen concentration, a convolutional neural network CNN regression model is established, the convolutional neural network CNN regression model consists of convolutional layers and fully-connected layers, a layer of maximum pooling layer is arranged behind each convolutional layer and is used for reducing calculation amount, the sizes of cores of all the convolutional layers and the pooling layer are 2 multiplied by 2, the step sizes of the convolutional layers are 1, the step sizes of the pooling layer are 2, a ReLU activation function is used for all the convolutional layers and the fully-connected layers except for the last fully-connected layer,
step 4.2: training a flame image oxygen concentration prediction model:
and (3) training a convolutional neural network CNN regression model by using a regression generated flame image data set after the balance of the antagonism network RGAN-GP, wherein the optimizer selects Momentum, and the learning rate is set to be 0.01.
By adopting the technology, compared with the prior art, the invention has the following beneficial effects:
according to the invention, the flame image of a few intervals in the data set is expanded by using the regression generation countermeasure network based on gradient punishment, so that an unbalanced furnace flame image data set is balanced, the unbalanced learning problem existing in the training process of the flame image oxygen concentration prediction model is solved, the prediction accuracy of the regression model on the flame image oxygen concentration is improved, and the method has universality and universality.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a schematic diagram of an experimental system for acquiring flame images according to the present invention;
FIG. 3 is a schematic diagram of flame image data obtained in accordance with the present invention;
FIG. 4 is a schematic diagram of an RGAN-GP model established in the present invention;
FIG. 5 is a schematic diagram of a few interval samples generated by the RGAN-GP model of the present invention;
FIG. 6 is a schematic diagram of a flame image oxygen concentration prediction model established by the invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be further described in detail below with reference to the accompanying drawings and examples of the present invention. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
On the contrary, the invention is intended to cover any alternatives, modifications, equivalents, and variations as may be included within the spirit and scope of the invention as defined by the appended claims. Further, in the following detailed description of the present invention, certain specific details are set forth in order to provide a better understanding of the present invention. The present invention will be fully understood by those skilled in the art without the details described herein.
Referring to fig. 1-6, a regression-based method for predicting oxygen concentration of a flame image for generating a countermeasure network includes the steps of:
(1) Acquiring furnace flame image data:
the furnace flame image experiment system is shown in fig. 2. An experiment furnace with a burner model number of 5514-6 is used, and industrial heavy oil is used as fuel for performing a combustion experiment. The oil regulating valve adopts an industrial servo motor to carry out positioning control. The combustion air comes from an air compressor that is directly driven using a variable frequency drive. And sending the waste gas sampled and filtered from the chimney to a waste gas analyzer, and measuring the exhaust temperature of the waste gas and the concentrations of oxygen, carbon dioxide and carbon monoxide. The sampling frequency of the gas analyzer is one sampling point per second. Through the inspection window of the experimental furnace, images of the flame inside the furnace were captured using a charge coupled device (Charge Coupled Device, CCD) camera, with the image sampling rate set to 5 seconds/frame. The output signal of the CCD camera is sent to the site computer through an IEEE-1394a interface. During operation, both the oxygen concentration in the furnace and the flame image are automatically recorded.
The experiment totally acquired 8512 flame images, each flame image having an oxygen concentration label measured by a gas analyzer. FIG. 3 shows an acquired image of flames and their oxygen concentration, with the flames becoming increasingly larger and brighter as the oxygen concentration decreases.
(2) Preprocessing of furnace flame image data and data set partitioning:
step 2.1: and (3) data normalization processing:
the number of pixels of the acquired digital image of the furnace flame is 658×492, and the image is reduced to a three-channel image of 100×100 in order to prevent the overflow of the memory and reduce the calculation amount.
In order to accelerate the convergence rate of the model and reduce the training time, the data is normalized, and the formula is as follows:
wherein x is normalized data; a is the collected original data; a, a min Is the minimum value in the original data; a, a max Is the maximum value in the original data.
Step 2.2: dividing the data set:
dividing the flame image dataset of the smelting furnace into a training set and a testing set according to the proportion of 3:1, and confirming a few intervals, namely that the number of samples in one interval in the training set is obviously smaller than that in other intervals.
(3) Generating a minority interval sample in the furnace flame image dataset:
step 3.1: establishing regression generation countermeasure network model based on gradient penalty
The regression generation countermeasure network RGAN-GP model consists of a generator G, a discriminator D and a regression model R, which are all neural networks. The generator G is used to capture a distribution of the real data, generating samples similar to the real data. The arbiter D is used to determine whether its input is real data or generated data. The regression model R is used to constrain the generative model G such that the samples generated by the generative model correspond to its conditional variables. The loss functions of the discriminator D, the generator G, the regression model R are as follows:
wherein: p (P) data (x) Representing a probability distribution of the real data x; p (P) z (z) represents a probability distribution of noise z; g (z|c) represents the generated data under the condition variable c;representing the sampling distribution +.>Epsilon represents an interpolation parameter; d () represents the output of the arbiter, +.>Representing the gradient 2 norms of the discriminator; r () represents the output of the regression model; e represents the desire; λ represents a gradient penalty coefficient; alpha represents the constraint coefficients of the regression model to the generator.
The generator structure of RGAN-GP is shown in FIG. 4 (a), and the structures of the discriminator D and the regression model R are shown in FIG. 4 (b). The generator G consists of 2 deconvolution layers, the inputs of which are a 100-dimensional noise vector and 1-dimensional condition variables. The input is projected into a small space and is deconvoluted twice to give an output of size 100 x 3. The generator G first deconvolution layer contains 64 convolution kernels and the second deconvolution layer contains 3 convolution kernels. The arbiter D consists of 3 convolutional layers, 1 fully connected layer, the input of which is a vector of size 100 x 3. The first convolution layer of the arbiter D contains 10 convolution kernels, the second convolution layer contains 64 convolution kernels, and the third convolution layer contains 128 convolution kernels. All hidden layers of the arbiter D are embedded in the condition variables except the first convolution layer of the arbiter D. The first convolution layer of the arbiter D has a convolution kernel size of 2 x 2 and a step size of 1. All deconvolution layers and convolution kernel sizes of the convolution layers except the first convolution layer of the discriminator D are 5 x 5, with a step size of 2. The activation function of the hidden layer of the generator G and the first convolution layer of the discriminator D is ReLU, and the activation function of the rest of the hidden layers of the discriminator D is Leaky-ReLU. Generator G uses batch normalization and arbiter D uses layer normalization. The regression model R consists of a 2-layer convolution, 2-layer full-join. Each convolution layer is followed by a layer of max-pooling to reduce the computational effort. The first convolution layer contains 10 convolution kernels and the second convolution layer contains 16 convolution kernels, the first fully-connected layer contains 100 neurons and the second fully-connected layer contains 3 neurons. The sizes of the cores of all the convolution layers and the pooling layers are 2 multiplied by 2, the step sizes of the convolution layers are 1, and the step sizes of the pooling layers are 2. All convolution layers and full connection layers, except the last full connection layer, use the ReLU activation function. The first convolution layer of the regression model R shares parameters with the first convolution layer of the arbiter D.
Step 3.2: training gradient penalty based regression using a few interval samples to generate an countermeasure network model and generate new samples:
and taking a few interval samples in the original training set as a training set to train the RGAN-GP model. The optimizer selects Adam, and the learning rate is set to 0.0002. By alternately training the generator G, the regression model R, and the discriminator D, the final generator G can generate a sample that makes the discriminator D unable to discriminate between true and false. The generator G has now learned the distribution of real samples, and can produce samples that are spurious.
After the RGAN-GP model training is completed, a generator G is used for generating a certain number of samples, and after the samples are mixed with the original data set, the number of samples in each interval in the data set is the same or similar. The resulting minority interval samples are shown in fig. 5.
(4) Establishing and training a flame image oxygen concentration prediction model:
step 4.1: establishing a flame image oxygen concentration prediction model:
the convolutional neural network was used to predict oxygen concentration and a CNN regression model was established as shown in fig. 6. The CNN regression model consists of 2 layers of convolutions, 2 layers of full-joins. Each convolution layer is followed by a layer of max-pooling to reduce the computational effort. The first convolution layer contains 10 convolution kernels and the second convolution layer contains 16 convolution kernels, the first fully-connected layer contains 100 neurons and the second fully-connected layer contains 3 neurons. The sizes of the cores of all the convolution layers and the pooling layers are 2 multiplied by 2, the step sizes of the convolution layers are 1, and the step sizes of the pooling layers are 2. All convolution layers and full connection layers, except the last full connection layer, use the ReLU activation function.
Step 4.2: training a flame image oxygen concentration prediction model:
and training a CNN regression model by using the flame image data set with balanced data volume of each interval, and selecting Momentum by an optimizer, wherein the learning rate is set to be 0.01.
(1) Model performance evaluation:
step 5.1: root mean square error RMSE evaluation
The root mean square error is defined as follows:
wherein: n represents the total amount of the test set samples; f (x) i ) Representing a predicted class of input samples xi; yi represents the input sample x i Is a true category of (c). The smaller the RMSE, the descriptionThe better the predictive performance of the regression model.
Step 5.2: evaluation of RIMP value of relative improvement Rate
The relative rate of increase can be expressed as:
wherein: RMSE base Representing the root mean square error before model lifting; RMSE imp Representing the root mean square error after model lifting. The higher the relative improvement rate, the more the performance of the model improves, and the more the superiority of the method can be explained.
The oxygen concentration range and the number of training samples in each section are shown in table 1, and it can be seen that section 11 is a small number of sections.
Table 1 oxygen concentration ranges and sample numbers in each section
The performance of the CNN regression model trained using the RGAN-GP balanced dataset, the CNN regression model trained on the original unbalanced dataset, and the CNN regression model trained by the random oversampling method were compared to obtain the results shown in table 2. Experimental results show that the RGAN-GP CNN regression model obtains the best performance, the RMSE is improved by 6.72% relative to the CNN regression model of unbalanced data, and the method is superior to a random oversampling method, and the effectiveness and the superiority of the method are proved.
TABLE 2 comparison of different CNN regression models
According to the method, the flame image of a few intervals in the data set is expanded by adopting regression generation countermeasure network based on gradient punishment, the data set is balanced, the prediction accuracy of the regression model on the oxygen concentration of the flame image is improved, and the method has universality.
The foregoing description of the preferred embodiments of the invention is not intended to be limiting, but rather is intended to cover all modifications, equivalents, and alternatives falling within the spirit and principles of the invention.
Claims (2)
1. The method for predicting the oxygen concentration of the flame image based on the regression generation countermeasure network is characterized by comprising the following steps of: the method comprises the following steps:
(1) Acquiring furnace flame image data:
performing a combustion experiment in an experiment furnace, sampling the exhaust temperature, oxygen concentration, carbon dioxide concentration and carbon monoxide concentration of the filtered exhaust gas, capturing an image of flame in the experiment furnace, and transmitting the image to on-site processing equipment, wherein the oxygen concentration and the flame image in the experiment furnace are automatically recorded in the running processing process of the on-site processing equipment;
(2) Preprocessing of furnace flame image data and data set partitioning:
in order to accelerate the convergence rate of the model, reduce the training time of the model, normalize the data, and divide the data set into training set and test set;
(3) Generating a minority interval sample in the furnace flame image dataset:
establishing a regression generation countermeasure network RGAN-GP model based on gradient penalty, taking a few interval images in a furnace flame image data set as training samples, training the regression generation countermeasure network RGAN-GP model, and generating a few interval samples and balancing a flame image data set after the regression generation countermeasure network RGAN-GP model is trained; the method comprises the following steps:
step 3.1: establishing a regression generation countermeasure network model based on gradient penalty:
the regression generation countermeasure network RGAN-GP model comprises a generator G, a discriminator D and a regression model R, wherein the generator G, the discriminator D and the regression model R are all neural networks, the generator G is used for capturing the distribution of real data and generating samples similar to the real data, the discriminator is used for judging whether the input of the generator G is the real data or the generated data, the regression model R is used for constraining the generation model G, the samples generated by the generation model correspond to the condition variables of the generator G, and the loss functions of the discriminator R, the generator G and the regression model R are as follows:
wherein: p (P) data (x) Representing a probability distribution of the real data x; p (P) z (z) represents a probability distribution of noise z; g (z|c) represents the generated data under the condition variable c;representing the sampling distribution +.>Epsilon represents an interpolation parameter; d () represents the output of the arbiter, +.>Representing the gradient 2 norms of the discriminator; r () represents the output of the regression model; e represents the desire; λ represents a gradient penalty coefficient; alpha represents the constraint coefficient of the regression model to the generator;
step 3.2: training a regression based on gradient penalty using a few interval samples to generate an countermeasure network model and generate new samples, the specific process is as follows:
taking a few interval samples in an original training set as a training set, training regression to generate an antagonism network RGAN-GP model, and alternately training a generator G, a regression model R and a discriminator D, wherein the generator G can finally generate samples which enable the discriminator D to be incapable of discriminating true and false; after training the regression generation countermeasure network RGAN-GP model is completed, a generator G is used for generating a certain number of samples, the samples are mixed with the original data set, and the unbalance rate of the number of samples in each interval in the data set is 1:2, i.e., the interval sample size of the least samples and the interval sample size of the most samples is 1:2 or less;
(4) Establishing and training a flame image oxygen concentration prediction model:
establishing a flame image oxygen concentration prediction model, taking the balanced flame image data set as a training set, and training the flame image oxygen concentration prediction model;
(5) Model performance evaluation:
and (3) introducing an evaluation index Root Mean Square Error (RMSE) and a relative improvement Rate (RIMP) to evaluate the flame image oxygen concentration prediction model trained in the step (4).
2. The regression-based flame image oxygen concentration prediction method of claim 1, wherein the process of step (4) is:
step 4.1: establishing a flame image oxygen concentration prediction model:
a convolutional neural network is used for predicting oxygen concentration, a convolutional neural network CNN regression model is established, the convolutional neural network CNN regression model consists of convolutional layers and fully-connected layers, a layer of maximum pooling layer is arranged behind each convolutional layer and is used for reducing calculation amount, the sizes of cores of all the convolutional layers and the pooling layer are 2 multiplied by 2, the step sizes of the convolutional layers are 1, the step sizes of the pooling layer are 2, a ReLU activation function is used for all the convolutional layers and the fully-connected layers except for the last fully-connected layer,
step 4.2: training a flame image oxygen concentration prediction model:
and (3) training a convolutional neural network CNN regression model by using a regression generated flame image data set after the balance of the antagonism network RGAN-GP, wherein the optimizer selects Momentum, and the learning rate is set to be 0.01.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110467231.3A CN113239991B (en) | 2021-04-28 | 2021-04-28 | Flame image oxygen concentration prediction method based on regression generation countermeasure network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110467231.3A CN113239991B (en) | 2021-04-28 | 2021-04-28 | Flame image oxygen concentration prediction method based on regression generation countermeasure network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113239991A CN113239991A (en) | 2021-08-10 |
| CN113239991B true CN113239991B (en) | 2024-02-27 |
Family
ID=77129779
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110467231.3A Active CN113239991B (en) | 2021-04-28 | 2021-04-28 | Flame image oxygen concentration prediction method based on regression generation countermeasure network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113239991B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113657240A (en) * | 2021-08-12 | 2021-11-16 | 浙江工业大学 | Unbalanced-like pearl classification method based on multi-view image amplification |
| CN113935460B (en) * | 2021-09-27 | 2023-08-11 | 苏州大学 | Intelligent diagnosis method for mechanical faults under unbalanced-like data set |
| CN114611597A (en) * | 2022-03-07 | 2022-06-10 | 山东大学 | Lithium ion battery health state estimation method and system based on RGAN |
| CN114429589B (en) * | 2022-04-07 | 2022-07-26 | 北京理工大学 | Hydrogen leakage concentration distribution prediction method and system |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103454390A (en) * | 2013-03-27 | 2013-12-18 | 深圳信息职业技术学院 | Method and device for measuring concentration of dissolved oxygen |
| CN112700402A (en) * | 2020-12-09 | 2021-04-23 | 浙江工业大学 | Boiler high-temperature oxygen concentration prediction method based on image and self-coding generation network |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10991087B2 (en) * | 2017-01-16 | 2021-04-27 | Praxair Technology, Inc. | Flame image analysis for furnace combustion control |
-
2021
- 2021-04-28 CN CN202110467231.3A patent/CN113239991B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103454390A (en) * | 2013-03-27 | 2013-12-18 | 深圳信息职业技术学院 | Method and device for measuring concentration of dissolved oxygen |
| CN112700402A (en) * | 2020-12-09 | 2021-04-23 | 浙江工业大学 | Boiler high-temperature oxygen concentration prediction method based on image and self-coding generation network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113239991A (en) | 2021-08-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113239991B (en) | Flame image oxygen concentration prediction method based on regression generation countermeasure network | |
| CN107577910B (en) | Vehicle exhaust concentration inversion method based on deep neural network | |
| CN111814956B (en) | Multi-task learning air quality prediction method based on multi-dimensional secondary feature extraction | |
| CN109815904B (en) | Fire identification method based on convolutional neural network | |
| CN116204794B (en) | Method and system for predicting dissolved gas in transformer oil by considering multidimensional data | |
| CN113405106B (en) | Artificial intelligence control method for garbage incineration process | |
| WO2022124276A1 (en) | Indoor air quality prediction method and indoor air quality detection system | |
| CN111862065B (en) | Power transmission line diagnosis method and system based on multitask deep convolutional neural network | |
| CN113011660A (en) | Air quality prediction method, system and storage medium | |
| CN117744704B (en) | Flue gas pollution source acquisition monitoring system, method and readable storage medium | |
| CN110059723B (en) | Robust smoke detection method based on integrated deep convolutional neural network | |
| CN116952654B (en) | Environment monitoring and early warning system for administrative supervision | |
| CN116977256A (en) | Training method, device, equipment and storage medium for defect detection model | |
| CN115097796B (en) | Quality control system and method for simulating big data and correcting AQL value | |
| CN116051496A (en) | Real-time sewer defect detection method | |
| CN114066819B (en) | Environmental corrosion severity identification method based on convolutional neural network deep learning | |
| CN115962428A (en) | Real-time online intelligent interpretability monitoring and tracing method for gas pipe network leakage | |
| CN113468823B (en) | Optical module damage detection method and system based on machine learning | |
| CN113703382B (en) | Workpiece identification system of foreroom pre-vacuumizing multipurpose atmosphere nitriding furnace | |
| TWI735385B (en) | Method and sintering factory for predicting sulfur oxide | |
| CN113008998B (en) | Concealed engineering internal defect judgment method based on PCNN | |
| CN114638809A (en) | Multi-scale micro-defect detection method based on PA-MLFPN workpiece surface | |
| CN112862180A (en) | Denitration system inlet NOx concentration prediction method | |
| CN113313204A (en) | Deep learning-based garbage incineration state identification method and incineration control method | |
| Yang et al. | Prediction of pollutant diffusion trend using CNN-GRU hybrid network in IoT air monitoring system |
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 |