CO 2 Concentration Detection Model Training Method, Device and Thermal Power Carbon Emission Monitoring System

Abstract

The present invention discloses a method and device for training a CO2 concentration detection model and a thermal power carbon emission monitoring system. The method includes collecting CO2 with different concentrations and corresponding characteristic information as labeled sample data; training a generative adversarial network using the labeled sample data to obtain a trained generator and restorer; using the generator to generate a latent representation of pseudo data according to random noise, and then converting the latent representation of the pseudo data into pseudo data with the same dimension as the labeled sample data through the restorer; training a CO2 concentration detection model using the pseudo data and the labeled sample data. The CO2 concentration detection model training method based on the generative adversarial network and semi-supervised learning in the present invention can realize the learning and establishment of the detection model with as little data as possible, reduce the data acquisition cost, and the trained model is more accurate and can adapt to diverse detection conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial carbon emission monitoring, and particularly relates to a method and device for training a CO 2 concentration detection model and a thermal power carbon emission monitoring system. Background Technique

[0002] With the continuous acceleration of the industrialization process, the impact of industrial production on the environment has become increasingly prominent, posing a threat to the ecological environment and human health. The large-scale emission of anthropogenic greenhouse gases has become one of the important reasons for the increase in extreme climate events. Among numerous emission sources, carbon emissions from the power industry are a major emitter of greenhouse gases in China, and thermal power units, especially coal-fired power units, are the main emission sources of carbon dioxide (CO 2 ), and will surely become the key targets for emission reduction. The statistical calculation of carbon emissions is a crucial link among them. Therefore, in the field of industrial emission monitoring, the real-time online monitoring technology for CO 2 is gradually becoming a research hotspot.

[0003] Due to the advantages of high precision, high sensitivity, and non-contact nature, the optical gas detection method is widely used. Among them, the Tunable Diode Laser Absorption Spectroscopy (TDLAS) technology scans the absorption spectrum lines of the gas to be measured through the tunable characteristics of the laser, obtains the harmonic signal of the concentration information of the gas to be measured, and then realizes the detection of the gas concentration. Since the laser intensity will be significantly attenuated after passing through the gas absorption region, the detection of the attenuation amount can be used to detect the concentration of the gas to be measured.

[0004] To achieve the above gas concentration detection, the traditional method performs a series of calculations on the obtained harmonic signal to obtain the gas concentration. The calculation process is complex, and the calculation results are affected by many factors. The method based on machine learning provides another idea. By learning the relationship model between the optical intensity change signal obtained by scanning the absorption spectrum lines of the gas to be measured and the concentration of the gas to be measured, accurate and rapid concentration detection can be realized. For example, in existing research, in the patent application document with the publication number CN109724941A based on the TDLAS technology, 5 statistical characteristic parameters: mean value, standard deviation, sum of squared deviations, coefficient of variation, and maximum deviation are extracted from the difference signal between the absorbed laser and the original laser signal, and the CO 2 gas concentration is used as the output to train the RBF neural network to achieve CO 2Accurate measurement of gas concentration and error analysis; however, the manual feature selection method adopted in this solution may lead to information loss, resulting in a small applicable range of the measurement model. At the same time, the problem of insufficient labeled data caused by the small concentration values of the sample gas in actual model training is not considered. Although the patent application document with the publication number CN112504970A uses a generative adversarial network for dataset expansion, this solution also adopts the method of manual feature extraction, and the forward and backward correlation existing in the spectral signal is not considered in model construction, which may lead to the generated pseudo-data by the generative adversarial network not being real enough. Summary of the Invention

[0005] The technical problem to be solved by the present invention is how to realize the learning and establishment of a CO 2 concentration detection model with as little data as possible, while reducing the data acquisition cost and improving the generalization ability of the CO 2 concentration detection model and the accuracy of carbon concentration detection.

[0006] The present invention solves the above technical problems through the following technical means:

[0007] A training method for a CO 2 concentration detection model is proposed, including:

[0008] Collect different concentrations of CO 2 and the corresponding feature information as labeled sample data;

[0009] Use the labeled sample data to train the generative adversarial network to obtain a trained generator and restorer;

[0010] Use the generator to generate the latent representation of pseudo-data according to random noise, and then use the restorer to convert the latent representation of the pseudo-data into pseudo-data with the same dimension as the labeled sample data;

[0011] Use the pseudo-data and the labeled sample data to train the CO 2 concentration detection model.

[0012] Further, the collection of different concentrations of CO 2 and the corresponding feature information as labeled sample data includes:

[0013] Under different concentration standard gases, collect different concentrations of CO 2 and the corresponding feature information as labeled sample data;

[0014] Among them, the feature information includes laser signals, the temperature of the gas, and the flow rate of the gas.

[0015] Further, the generative adversarial network includes a generator, a discriminator, an embedder, and a restorer. The input of the generator is noise data, and its output is connected to the discriminator. The input of the embedder is labeled sample data, and its output is connected to the discriminator and the restorer respectively;

[0016] The generator is used to generate a latent representation of pseudo data based on the noise data;

[0017] The embedder is used to generate a latent representation of the labeled sample data based on the labeled sample data;

[0018] The discriminator is used to discriminate whether the output of the generator and the output of the embedder are labeled sample data or pseudo data;

[0019] The restorer is used to convert the latent representation of the labeled sample data or the latent representation of the pseudo data into data with the same dimension as the labeled sample data.

[0020] Further, training the generative adversarial network with the labeled sample data to obtain a trained generator and restorer includes:

[0021] Using the labeled sample data as the input of the embedder, training the embedder and the restorer, and the training objective is the first optimization objective:

[0022]

[0023] Where: and are the network parameters corresponding to the embedder and the restorer respectively; is a hyperparameter used to balance the reconstruction loss function and the supervision loss function ;

[0024] Using the noise data as the input of the generator, training the generator and the discriminator, and the training objective is the second optimization objective:

[0025]

[0026] Where: , are the network parameters corresponding to the generator and the discriminator respectively, is a hyperparameter used to balance the unsupervised loss function and the supervision loss function ;

[0027] Iteratively train until the preset first iteration number is reached to obtain a trained generator and restorer.

[0028] Further, the loss functions used in training the generative adversarial network include:

[0029]

[0030]

[0031]

[0032] wherein the gas temperature and flow rate are defined as static features, and represent the vector of static features and the vector of time features respectively; and represent the instantiated random vectors of static features and time features respectively, , ; and represent the specific values of static features and time features respectively; is the reconstructed static feature, is the time feature at the represents the tuple taking the expectation by sampling from the distribution ; represents the tuple taking the expectation by sampling from the synthetic data distribution ; represents the sequence length, represents the time instant, represents the random noise; represents the output of the generator for the time feature; is the true classification of the static feature, is the synthetic classification of the static feature; is the true classification of the time feature at the is the synthetic classification of the time feature at the is the true embedding of the static feature, is the true embedding of the time feature at the is the two - norm; , , represent the reconstruction loss function, the unsupervised loss function, and the supervised loss function respectively.

[0033] Furthermore, the generator, discriminator, embedder, and restorer all adopt a 3 - layer gated recurrent network.

[0034] Furthermore, the CO 2The concentration detection model includes a multi-scale enhanced feature extraction module, a prediction module, and an auto-encoder module. The input of the multi-scale enhanced feature extraction module is pseudo data and labeled sample data, and the output of the multi-scale enhanced feature extraction module is respectively connected to the prediction module and the auto-encoder module;

[0035] The multi-scale enhanced feature extraction module is used to obtain multi-scale enhanced features corresponding to the input data based on the input data;

[0036] The prediction module is used to predict the CO 2 concentration based on the multi-scale enhanced features;

[0037] The auto-encoder module is used to obtain reconstructed data with the same dimension as the labeled sample data based on the multi-scale enhanced features.

[0038] Further, the multi-scale enhanced feature extraction module includes a multi-scale hybrid enhancement sub-module and a feature extraction sub-module. The multi-scale hybrid enhancement sub-module includes an equal number of average pooling layers and multi-layer perceptrons. Each average pooling layer is connected in sequence, each multi-layer perceptron is connected in sequence, and the output of the th average pooling layer and the output of the th multi-layer perceptron are added and used as the input of the th multi-layer perceptron. The mixed data obtained by adding the time feature of the input data of the first average pooling layer and the output feature of the first multi-layer perceptron is used as the input of the feature extraction sub-module;

[0039] The feature extraction sub-module includes a Fourier transform layer, a first convolutional layer, a batch normalization layer BN, an activation function ReLU, an inverse Fourier transform layer, and a second convolutional layer connected in sequence;

[0040] The mixed data is used as the input of the Fourier transform layer. The output of the Fourier transform layer is sequentially connected to the first convolutional layer, the batch normalization layer BN, and the activation function ReLU and then connected to the inverse Fourier transform layer. The mixed data and the output feature of the inverse Fourier transform layer are added and used as the input of the second convolutional layer. The output of the second convolutional layer is the multi-scale enhanced feature.

[0041] Further, training the CO 2 concentration detection model using pseudo data and labeled sample data includes:

[0042] Training the multi-scale enhanced feature extraction module and the prediction module using the labeled sample data, and the training objective is the third optimization objective:

[0043]

[0044] Where: 、 They are the network parameters corresponding to the multi-scale enhanced feature extraction module and the prediction module respectively; denotes the true label of the -th labeled sample data, and 2 is the total amount of labeled sample data used for training the CO denotes the output result of the prediction module for the -th labeled sample data; is the two-norm;

[0045] The autoencoder module is trained using pseudo data, and the training objective is the fourth optimization objective:

[0046]

[0047] In the formula: are the network parameters corresponding to the autoencoder module; denotes the -th pseudo data; denotes the output result of the autoencoder module for the -th pseudo data, and is the total amount of pseudo data used for training the CO 2 concentration detection model;

[0048] When the iterative training reaches the preset second iteration number, a trained CO 2 concentration detection model is obtained.

[0049] In addition, the present invention also proposes a CO 2 concentration detection model training device, including:

[0050] A data collection unit for collecting different CO 2 concentration information and corresponding feature information as labeled sample data;

[0051] A first training unit for training a generative adversarial network using the labeled sample data to obtain a trained generator and restorer;

[0052] A pseudo data generation unit for using the generator to generate a latent representation of pseudo data based on random noise and then converting the latent representation of the pseudo data into pseudo data with the same dimension as the labeled sample data through the restorer;

[0053] A second training unit for training the CO 2 concentration detection model using the pseudo data and the labeled sample data.

[0054] In addition, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned CO 2 concentration detection model training method is implemented.

[0055] In addition, the present invention also provides a thermal power carbon emission monitoring system, including a laser, an absorption cell, a photodetector, a signal processor, an environmental detector, and a CO 2 concentration detector. The absorption cell is arranged on the output laser light path of the laser, and an air path is provided on the absorption cell for the gas to be measured to enter the absorption cell. The environmental detector is arranged at the air inlet of the air path. The output of the photodetector is connected to the CO 2 concentration detector through the signal processor, and the output of the environmental detector is connected to the CO 2 concentration detector;

[0056] CO 2 The concentration detector deploys a CO 2 concentration detection model trained by the above-mentioned CO 2 concentration detection model training method.

[0057] Further, the environmental detector is used to detect the temperature and flow rate of the gas to be measured.

[0058] Further, the input end of the laser is connected to a wavelength controller, and the wavelength controller is used to control the laser to output laser light with a set wavelength.

[0059] The advantages of the present invention are as follows:

[0060] (1) Since there are few effective labeled sample data in actual model training, for example, in the thermal power carbon emission scenario, the present invention trains the generative adversarial network based on a small amount of labeled sample data. Since the labeled sample data is sequential data and has temporal correlation before and after, the pseudo data generated by the trained generator and restorer conforms to its temporal correlation and is closer to the real data distribution; then, the pseudo data and the existing small amount of labeled sample data are used to form augmented data, and the augmented data is used for semi-supervised learning of the CO 2 concentration detection model; the CO 2 concentration detection model training method based on the generative adversarial network and semi-supervised learning can realize the learning and establishment of the detection model with as little data as possible, reduce the data acquisition cost, and at the same time improve the generalization ability of the CO 2 concentration detection model. Moreover, the detection result of the trained CO 2 concentration detection model is more accurate and can adapt to various detection conditions.

[0061] (2) The present invention utilizes prior information to design a corresponding generative adversarial network architecture and the loss function adopted in the training of the generative adversarial network, enabling the trained generator and restorer to generate pseudo-data with the same dimension and similar distribution as the real data, which is closer to the real data distribution, thereby improving the accuracy of the training of the CO 2 concentration detection model.

[0062] (3) The multi-scale enhanced feature extraction module is adopted in the CO 2 concentration detection model designed by the present invention, which plays a role in enhancing the features of the input. Since the sequence data contains coarse-grained patterns and fine-grained patterns, which complement each other and jointly reflect the complete information in the sequence, this multi-scale enhanced feature extraction module obtains different granularity patterns through average pooling, then mixes them, and obtains enhanced features through time-frequency feature fusion for more detailed analysis, which is helpful for the training of the CO 2 concentration detection model.

[0063] Additional aspects and advantages of the present invention will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present invention. Brief Description of the Drawings

[0064] Figure 1 is a schematic flowchart of a method for training a CO 2 concentration detection model proposed in an embodiment of the present invention;

[0065] Figure 2 is a schematic structural diagram of a generative adversarial network in an embodiment of the present invention;

[0066] Figure 3 is a schematic structural diagram of a CO 2 concentration detection model in an embodiment of the present invention;

[0067] Figure 4 is a schematic structural diagram of a multi-scale enhanced feature extraction module in an embodiment of the present invention;

[0068] Figure 5 is a schematic structural diagram of a device for training a CO 2 concentration detection model proposed in an embodiment of the present invention;

[0069] Figure 6 is a schematic structural diagram of a thermal power carbon emission monitoring system proposed in an embodiment of the present invention. Detailed Embodiments

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0071] As Figure 1 shown, a method for training a CO 2 concentration detection model is proposed in the first embodiment of the present invention. The method includes the following steps:

[0072] S10. Collect CO at different concentrations 2 and corresponding characteristic information as labeled sample data;

[0073] Specifically, in this embodiment, through standard gas detection, under standard gases of different concentrations, the characteristic information corresponding to CO at different concentrations 2 is collected, and the corresponding standard gas concentration is used as a label and the corresponding characteristic information to form a labeled sample, and a data set is composed of each labeled sample data, that is, real sample data. .

[0074] It should be noted that in order to perform carbon emission detection in thermal power enterprises, standard gases with known concentrations can be detected, for example, to obtain corresponding characteristic information, so as to obtain labeled sample data.

[0075] S20. Use the labeled sample data to train a generative adversarial network to obtain a trained generator and restorer;

[0076] S30. After the generator generates a latent representation of pseudo data according to random noise, the restorer converts the latent representation of the pseudo data into pseudo data with the same dimension as the labeled sample data;

[0077] It should be noted that considering that the number of effective labeled samples in the thermal power carbon emission scenario is small, resulting in poor generalization ability of the trained model, therefore, in this embodiment, a generative adversarial network is used to expand the number of sample data, realize the enhancement of labeled sample data, and during the training process of the generative adversarial network, the labeled sample data is used to train the generative adversarial network. Since the labeled sample data is sequence data and has temporal correlation before and after, it provides temporal information when generating pseudo data for noise data, so that the pseudo data generated by the trained generator and restorer has the same dimension and similar distribution as the real sample data, and is closer to the real data distribution, thereby improving the accuracy of training the CO 2 concentration detection model in the thermal power carbon emission scenario.

[0078] It should be understood that when those skilled in the art detect some harmful substances contained in other industrial waste gases, waste waters, and solid wastes, they can also refer to the data acquisition method provided in this method to achieve data augmentation for model training.

[0079] S40. Train the CO concentration detection model using the pseudo data and the labeled sample data. 2 Concentration detection model for training.

[0080] It should be noted that in this embodiment, the pseudo data is used as the unlabeled sample data to form an unlabeled data set. In the data set, Randomly select 80% of the data to form a labeled data set. The data set, The remaining 20% of the data in the data set forms a test set. The unlabeled data set; Together with the labeled data set, Carry out the training of the CO concentration detection model. After the training is completed, test and verify the CO concentration detection model on the test set. 2 Concentration detection model for training, and after the training is completed, test and verify the CO concentration detection model on the test set. On the CO concentration detection model, 2 Concentration detection model for testing and verification.

[0081] In this embodiment, by introducing generative adversarial learning and semi-supervised learning, the relationship between the CO gas concentration and the change in the laser signal caused by it is learned with as little real detection data as possible, and a CO concentration detection model is established. A method based on generative adversarial is used to synthesize pseudo data for the training of the concentration detection model, which increases the generalization ability of the model and improves the accuracy of the detection result in the case of few detection data due to few sample gas concentration values. 2 Gas concentration and the relationship between the resulting laser signal changes, establish a CO concentration detection model; 2 Adopt a generative adversarial-based method to synthesize pseudo data for the training of the concentration detection model, increase the generalization ability of the model, and improve the accuracy of the detection result in the case of few detection data due to few sample gas concentration values.

[0082] As a further preferred technical solution, the step S10: Collect different concentrations of CO and the corresponding characteristic information as labeled sample data, specifically including the following steps: 2 And the corresponding characteristic information as labeled sample data, specifically including the following steps:

[0083] Under different concentration standard gases, collect different concentrations of CO and the corresponding characteristic information as labeled sample data. 2 And the corresponding characteristic information as labeled sample data.

[0084] Among them, the characteristic information includes the laser signal, the temperature of the gas, and the flow rate of the gas.

[0085] It should be noted that in this embodiment, under different concentration standard gases, collect the laser signal, gas temperature, and gas flow rate data corresponding to different concentrations of CO as characteristics. 2 The corresponding laser signal, gas temperature, and gas flow rate data as characteristics, The corresponding standard gas concentration as a label, Form labeled sample data. , and consists of labeled sample data to form a data set .

[0086] Since in practical applications, the concentration of CO 2 is also related to environmental factors. In this embodiment, in addition to selecting laser signals, gas temperature and flow rate are also selected as characteristic information related to the concentration of CO 2 to construct labeled sample data, and the labeled sample data is used for data augmentation, thereby further ensuring that the synthesized pseudo-data is closer to the real data distribution.

[0087] As a further preferred technical solution, as Figure 2 shown, the generative adversarial network includes a generator, a discriminator, an embedder, and a restorer. The input of the generator is noise data, and the output is connected to the discriminator. The input of the embedder is labeled sample data, and the output is connected to the discriminator and the restorer respectively;

[0088] The generator is used to generate a latent representation of pseudo-data based on the noise data;

[0089] The embedder is used to generate a latent representation of the labeled sample data based on the labeled sample data;

[0090] The discriminator is used to discriminate whether the output of the generator and the output of the embedder are labeled sample data or pseudo-data;

[0091] The restorer is used to convert the latent representation of the labeled sample data or the latent representation of the pseudo-data into data with the same dimension as the labeled sample data.

[0092] Specifically, since the input labeled sample data is sequence data and there is temporal correlation before and after it, in order to generate pseudo-data that conforms to its temporal correlation, this embodiment uses prior information to design a corresponding generative adversarial network architecture and loss function. Among them, the generative adversarial network architecture includes a generator , discriminator , embedder and restorer , and the network parameters of the corresponding generator , discriminator , embedder and restorer are used one by one , , and are represented. The generator , discriminator , embedder and restorer The network of

[0093] The generator takes random noise as input and outputs a latent representation of fake data;

[0094] The discriminator is used to discriminate whether the outputs of the generator and the embedder are real data or fake data. The generator and the discriminator achieve an adversarial effect through alternating training, continuously improving their respective generation and discrimination abilities, and finally making the adversarial game between the generator and the discriminator reach the Nash equilibrium;

[0095] The embedder takes real labeled sample data as input and outputs a latent representation of the labeled sample data;

[0096] The restorer takes the latent representation of real labeled sample data as input and outputs reconstructed data with the same dimension as the labeled sample data.

[0097] After the training of the generative adversarial network is completed, the generator is used to generate a latent representation of fake data according to random noise, and then the restorer is used to convert the latent representation of fake data into fake data with the same dimension and similar distribution as the labeled sample data.

[0098] As a further preferred technical solution, in step S20: training the generative adversarial network with labeled sample data to obtain a trained generator and restorer specifically includes the following steps:

[0099] S21. Construct a training dataset using an unlabeled dataset composed of fake data and a labeled dataset ;

[0100] S22. Use the labeled sample data as the input of the embedder, and train the embedder and the restorer with the training objective being the first optimization objective:

[0101]

[0102] where: and are the network parameters corresponding to the embedder and the restorer respectively; is a hyperparameter used to balance the reconstruction loss function and the supervision loss function ;

[0103] S23. Use the noise data as the input of the generator, and train the generator and the discriminator with the training objective being the second optimization objective:

[0104]

[0105] Wherein: and are the network parameters corresponding to the generator and the discriminator respectively, is the hyperparameter used to balance the unsupervised loss function and the supervised loss function ;

[0106] S24. Repeat the above steps S22 to S23, and when the generative adversarial network is iteratively trained to reach a preset first number of iterations, a trained generator , discriminator , embedder and restorer are obtained.

[0107] It should be noted that after the training is completed, the trained generator and restorer are used to obtain synthetic data, that is, pseudo data, to complete the expansion of the dataset . The specific process of obtaining the synthetic data is as follows: Random noise is input into the trained generator to obtain its latent representation, and then it enters the trained restorer to reconstruct synthetic data with the same dimension and similar distribution as the real data according to the latent representation.

[0108] As a further preferred technical solution, the loss function used in training the generative adversarial network includes:

[0109]

[0110]

[0111]

[0112] Wherein: The gas temperature and flow rate are defined as static features, and represent the vector of static features and the vector of time features respectively; and represent the instantiated random vectors of static features and time features respectively, , ; and represent the specific values of static features and time features respectively; is the reconstructed static feature, is the reconstructed time feature at time represents the tuple Sampling the expectation from the distribution and denoting the tuple Sampling the expectation from the synthetic data distribution and denoting the sequence length denoting the time instant denoting the random noise; denoting the output of the generator for the time feature; being the true classification of the static feature being the synthetic classification of the static feature; being the true classification of the time feature at time instant being the synthetic classification of the time feature at time instant; being the true embedding of the static feature being the true embedding of the time feature at time instant; being the two - norm; and and respectively denote the reconstruction loss function, the unsupervised loss function, and the supervised loss function.

[0113] It should be noted that since the carbon concentration in the flue gas of thermal power carbon emissions is related to the flue gas flow rate and temperature, in this embodiment, the temperature and flow rate of the gas are selected as static features and jointly used for training the concentration detection model to improve the accuracy of the model training results.

[0114] In this embodiment, during the construction of the generative adversarial network, the static features, time features, and the time - series correlation in the data are fully considered. The loss calculations for the static features and time features are respectively performed in the loss function, making the synthetic data of the model more in line with the characteristics of the real data. The advantages of each loss function set in this embodiment are as follows: (1) Promote the generator to generate samples as close as possible to the real data distribution. (2) The discriminator improves its ability to distinguish real data and pseudo - data by optimizing the loss function. (3) Optimizing the loss function ensures a balanced state between the generator and the discriminator, avoiding unstable training caused by over - optimization of one of them.

[0115] As a further preferred technical solution, the CO 2 concentration detection model includes a multi - scale enhanced feature extraction module, a prediction module, and an auto - encoding module. The input of the multi - scale enhanced feature extraction module is pseudo - data and labeled sample data, and the output of the multi - scale enhanced feature extraction module is respectively connected to the prediction module and the auto - encoding module;

[0116] The multi-scale enhanced feature extraction module is used to obtain the multi-scale enhanced features corresponding to the input data based on the input data;

[0117] The prediction module is used to predict CO 2 concentration;

[0118] The auto-encoder module is used to obtain the reconstructed data with the same dimension as the labeled sample data based on the multi-scale enhanced features.

[0119] Specifically, as Figure 3 shown, the CO 2 concentration detection model includes a multi-scale enhanced feature extraction module , a prediction module and an auto-encoder module , and the network parameters corresponding to each module are represented by , and respectively; among them, 's structure includes a multi-scale hybrid enhancement sub-module and a feature extraction sub-module; and 's structure is a fully connected layer. It should be explained here that Figure 3 represents the process and structure during the training of the CO 2 concentration detection model. During training, the multi-scale enhanced feature extraction module is used in two iterative steps respectively, so it can be considered that the CO 2 concentration detection model contains two multi-scale enhanced feature extraction modules with the same structure, but the parameters of the two modules will be different.

[0120] Among them, the multi-scale hybrid enhancement sub-module is used to perform multi-scale hybrid enhancement on the input data to obtain the hybrid data with the same dimension as the input data; specifically, the multi-scale hybrid enhancement sub-module includes the same number of average pooling layers and multi-layer perceptrons. Each average pooling layer is connected in sequence, each multi-layer perceptron is connected in sequence, and the output of the th average pooling layer and the output of the th multi-layer perceptron are added and used as the input of the th multi-layer perceptron. The hybrid data obtained by adding the time feature of the input data of the first average pooling layer and the output feature of the first multi-layer perceptron is used as the input of the feature extraction sub-module.

[0121] Specifically, the multi-scale hybrid enhancement sub-module performs multi-scale hybrid enhancement on the input data to obtain the hybrid data with the same dimension as the input data. Assuming that the time feature of the input data of the multi-scale hybrid enhancement sub-module is , the specific calculation process is as follows:

[0122]

[0123]

[0124] Wherein: represents the average pooling operation, is the output of the th average pooling, represents a multi-layer perceptron, and are intermediate variables of the feature output,

[0125] and

[0126] is the mixed data obtained with the same dimension as the input data. Further, the feature extraction sub-module is used to extract features from the mixed data to obtain multi-scale enhanced features. Specifically, the feature extraction sub-module includes a Fourier transform layer PPT, a first convolutional layer Conv, a batch standard normalization layer BN, an activation function ReLU, an inverse Fourier transform layer Inv FFT, and a second convolutional layer Conv connected in sequence.

[0127] The mixed data is used as the input of the Fourier transform layer PPT. The output of the Fourier transform layer PPT is connected to the inverse Fourier transform layer Inv FFT through the first convolutional layer Conv, the batch standard normalization layer BN, and the activation function ReLU in sequence. The mixed data is added to the output features of the inverse Fourier transform layer Inv FFT and used as the input of the second convolutional layer Conv. The output of the second convolutional layer Conv is the multi-scale enhanced feature. Figure 4 It should be noted that the structure of the multi-scale enhanced feature extraction module is as

[0128] shown. Its output has the same dimension as the input, and it plays a role in enhancing the features of the input. Since the sequence contains coarse-grained patterns and fine-grained patterns, they complement each other and jointly reflect the complete information in the sequence. Therefore, this module obtains different granularity patterns through average pooling and then mixes them to obtain enhanced features for more detailed analysis.

[0129] In this embodiment, by designing a concentration detection model structure based on multi-scale enhanced feature extraction, different granularity features in the data can be captured, richer feature information can be obtained, the robustness of the model can be enhanced, it can adapt to diverse detection conditions, and the calculation amount is small, thereby realizing more efficient, accurate, and low-cost online real-time monitoring of carbon emissions. 2 As a further preferred technical solution, in step S40: training the CO

[0130] S41. Train the multi-scale enhanced feature extraction module and the prediction module using the labeled sample data, and the training objective is the third optimization objective:

[0131]

[0132] Where: and are the network parameters corresponding to the multi-scale enhanced feature extraction module and the prediction module respectively; represents the true label of the th labeled sample data, , is the total amount of labeled sample data used for training the CO 2 concentration detection model; represents the output result of the prediction module for the th labeled sample data; is the two-norm;

[0133] S42. Train the auto-encoder module using the pseudo data, and the training objective is the fourth optimization objective:

[0134]

[0135] Where: is the network parameter corresponding to the auto-encoder module; represents the th pseudo data; represents the output result of the auto-encoder module for the th pseudo data, , is the total amount of pseudo data used for training the CO 2 concentration detection model;

[0136] S43. Repeat the above steps S41 - S42 for iterative training until the preset second number of iterations is reached, and obtain the trained CO 2 concentration detection model.

[0137] It should be noted that after the iterative training is completed, use the test set to 2 perform a performance test on the CO 2 concentration detection model; if the performance meets the set requirements, the multi-scale enhanced feature extraction module and the prediction module in the CO 2 concentration detection model can be used for CO 2 concentration detection; if the performance does not meet the set requirements, adjust the network hyperparameters and re-train the CO

[0138] In addition, as Figure 5 shown, a training device for a CO 2 concentration detection model is proposed in the second embodiment of the present invention. The device includes:

[0139] A data collection unit 10, configured to collect different CO 2 concentration information and corresponding feature information as labeled sample data;

[0140] A first training unit 20, configured to train a generative adversarial network using the labeled sample data to obtain a trained generator and a restorer;

[0141] A pseudo-data generation unit 30, configured to generate a latent representation of pseudo-data by the generator according to random noise and then convert the latent representation of the pseudo-data into pseudo-data with the same dimension as the labeled sample data through the restorer;

[0142] A second training unit 40, configured to train a CO 2 concentration detection model using the pseudo-data and the labeled sample data.

[0143] As a further preferred technical solution, the data collection unit 10 is specifically configured to:

[0144] Collect different CO 2 concentrations and corresponding feature information as labeled sample data under different concentration standard gases;

[0145] Wherein, the feature information includes a laser signal, the temperature of the gas, and the flow rate of the gas.

[0146] As a further preferred technical solution, the first training unit 20 includes:

[0147] A first training subunit, configured to use the labeled sample data as the input of an embedder, and train the embedder and the restorer with a training target being a first optimization target:

[0148]

[0149] In the formula: and are respectively network parameters corresponding to the embedder and the restorer; is a hyperparameter for balancing the reconstruction loss function and the supervision loss function ;

[0150] A second training subunit, configured to use the noise data as the input of the generator, and train the generator and the discriminator with a training target being a second optimization target:

[0151]

[0152] In the formula: and are the network parameters corresponding to the generator and the discriminator respectively, is the hyperparameter for balancing the unsupervised loss function and the supervised loss function ;

[0153] The first iterative training subunit is used to perform iterative training until a preset first number of iterations is reached, and a trained generator and restorer are obtained.

[0154] As a further preferred technical solution, the second training unit 40 includes:

[0155] The third training subunit is used to train the multi-scale enhanced feature extraction module and the prediction module with labeled sample data, and the training objective is the third optimization objective:

[0156]

[0157] In the formula: and are the network parameters corresponding to the multi-scale enhanced feature extraction module and the prediction module respectively; represents the true label of the th labeled sample data, , is the total amount of labeled sample data for training the CO 2 concentration detection model; represents the output result of the prediction module for the th labeled sample data; is the two-norm;

[0158] The fourth training subunit is used to train the autoencoder module with pseudo data, and the training objective is the fourth optimization objective:

[0159]

[0160] In the formula: is the network parameter corresponding to the autoencoder module; represents the th pseudo data; represents the output result of the autoencoder module for the th pseudo data, , is the total amount of pseudo data for training the CO 2 concentration detection model;

[0161] The second iterative training subunit is used to perform iterative training until the preset second iteration number is reached, and then obtain the trained CO concentration detection model. 2 Concentration detection model.

[0162] It should be noted that for other embodiments or specific implementation methods of the CO concentration detection model training device of the present invention, reference can be made to the above method embodiments, and details are not described herein again. 2

[0163] In addition, as shown in the figure, the third embodiment of the present invention also proposes a thermal power carbon emission monitoring system, including a laser, an absorption cell, a photodetector, a signal processor, an environmental detector, and a CO concentration detector. The absorption cell is arranged on the output laser light path of the laser, and an air path is provided on the absorption cell for the gas to be measured to enter the absorption cell. The environmental detector is arranged at the air inlet of the air path. The output of the photodetector is connected to the CO concentration detector through the signal processor, and the output of the environmental detector is connected to the CO concentration detector. Figure 6

[0164] 2 2 2 2 2 2

[0165]

[0166]

[0167]

[0168] The CO concentration detector deploys a CO concentration detection model for thermal power carbon emission monitoring, which is trained by the CO concentration detection model training method described in the first embodiment above.

[0165] As a further preferred technical solution, the environmental detector is used to detect the temperature and flow rate of the gas to be measured.

[0166] It should be noted that since the changes in the flow rate and temperature of the flue gas for thermal power carbon emissions may cause changes in the CO2 concentration detection results, in this embodiment, temperature sensors and flow sensors are provided on the flue gas path to detect the temperature and flow rate of the flue gas.

[0167] As a further preferred technical solution, the input end of the laser is connected to a wavelength controller, and the wavelength controller is used to control the laser to output laser with a set wavelength.

[0168] Specifically, in this embodiment, the laser used is a tunable semiconductor laser, which can emit infrared light of a specific wavelength that can be absorbed by carbon dioxide molecules. The gas to be measured enters the absorption cell through the gas path. The laser emitted by the laser passes through the absorption cell, and the carbon dioxide molecules therein absorb the light of the specific wavelength, causing the laser intensity to attenuate. The photodetector is used to measure the change in light intensity after passing through the absorption cell. The carbon dioxide molecules will absorb the light energy of the specific wavelength, resulting in a decrease in the light intensity passing through the detector. The photodetector can convert these changing optical signals into electrical signals. The signal processor processes the signals output by the photodetector, such as filtering, amplifying, and converting to extract and analyze the information related to carbon dioxide absorption. The wavelength controller precisely controls the output wavelength of the laser by adjusting the current and temperature of the laser, locking it near the specific absorption line of carbon dioxide to improve the detection sensitivity. The environmental detector obtains the gas flow rate and temperature of the gas sample to be measured through detection; CO 2 A CO 2 concentration detection model is set in the concentration detector, which is used to perform CO 2 concentration detection according to the laser signal of the gas sample to be measured, the gas flow rate and temperature of the gas sample to be measured, through the CO 2 concentration detection model, to achieve online real-time monitoring of carbon emissions; the CO 2 concentration detection model is obtained by jointly training with real data and augmented data based on the generative adversarial network and semi-supervised learning proposed in the above first embodiment.

[0169] In actual use, for the gas with the concentration to be measured, it is introduced into the online real-time carbon emission monitoring system designed in this embodiment through the air inlet. Through the laser generated in the absorption system, its spectral signal enters the CO 2 concentration detection module after signal processing. The CO 2 concentration detection model in it will output its corresponding concentration. Specifically, in the CO 2 concentration detection model, the signal is input into the multi-scale enhanced feature extraction module to obtain a feature representation, and then input into the prediction module to obtain the corresponding concentration value according to its feature representation.

[0170] As a further preferred technical solution, this embodiment can also use online data to optimize the network parameters of the CO 2 concentration detection module and perform online update of the CO 2 concentration detection module.

[0171] In addition, the fourth embodiment of the present invention also proposes a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the CO 2 concentration detection model training method as described in the above first embodiment.

[0172] It should be noted that the logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch instructions from the instruction execution system, apparatus, or device and execute the instructions), or in combination with these instruction execution systems, apparatuses, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in combination with an instruction execution system, apparatus, or device. More specific examples (non-exhaustive list) of computer-readable media include the following: an electrical connection portion with one or more wirings (electronic device), a portable computer diskette case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or other suitable processing as necessary, and then stored in a computer memory.

[0173] It should be understood that various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), and the like.

[0174] In the description of this specification, the description referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0175] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0176] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A CO2 concentration detection model training method, characterized in that: include: Collect different concentrations of CO2 and corresponding feature information as labeled sample data; Use labeled sample data to train the generative adversarial network to obtain trained generators and restorers; The generator generates the latent representation of the pseudo data according to random noise, and then the restorer converts the latent representation of the pseudo data into pseudo data with the same dimension as the labeled sample data. The CO2 concentration detection model is trained using pseudo data and labeled sample data. The CO2 concentration detection model includes a multi-scale enhanced feature extraction module, a prediction module and an autoencoder module, wherein the multi-scale enhanced feature extraction module includes a multi-scale hybrid enhancement submodule and a feature extraction submodule connected in sequence; the multi-scale hybrid enhancement submodule includes an equal number of average pooling layers and multi-layer perceptrons, each average pooling layer is connected in sequence, each multi-layer perceptron is connected in sequence, and the first The output of the average pooling layer is The outputs of the multi-layer perceptrons are added together as the The mixed data obtained by adding the time features of the input data of the first average pooling layer and the output features of the first multi-layer perceptron is used as the input of the feature extraction submodule; The feature extraction submodule includes a Fourier transform layer, a first convolution layer, a batch normalization layer BN, an activation function ReLU, an inverse Fourier transform layer and a second convolution layer connected in sequence; the multi-scale mixed enhancement submodule is used to perform multi-scale mixed enhancement on the input data to obtain mixed data with the same dimension as the input data; the feature extraction submodule is used to extract features from the mixed data to obtain multi-scale enhanced features.

2. The CO2 concentration detection model training method according to claim 1, characterized in that: The collecting of different concentrations of CO2 and corresponding characteristic information as labeled sample data includes: Under different concentrations of standard gas, different concentrations of CO2 and corresponding characteristic information are collected as labeled sample data; The characteristic information includes laser signal, gas temperature and gas flow rate.

3. The CO2 concentration detection model training method according to claim 1, characterized in that: The generative adversarial network includes a generator, a discriminator, an embedder and a restorer. The input of the generator is noise data, and the output is connected to the discriminator. The input of the embedder is labeled sample data, and the output is connected to the discriminator and the restorer respectively. The generator is used to generate latent representations of pseudo data based on noisy data; The embedder is used to generate a latent representation of the labeled sample data based on the labeled sample data; The discriminator is used to determine whether the output of the generator and the output of the embedder are labeled sample data or pseudo data; The restorer is used to convert the latent representation of the labeled sample data or the latent representation of the pseudo data into data with the same dimension as the labeled sample data.

4. The CO2 concentration detection model training method according to claim 2, characterized in that: The method of using labeled sample data to train a generative adversarial network to obtain a trained generator and restorer includes: Using labeled sample data as the input of the embedder, the embedder and restorer are trained and the training objective is the first optimization objective: Where: and are the network parameters corresponding to the embedder and restorer respectively; To balance the reconstruction loss function And the supervised loss function Hyperparameters of Using the noise data as the input of the generator, the generator and discriminator are trained and the training objective is the second optimization objective: Where: , are the network parameters corresponding to the generator and the discriminator respectively, is used to balance the unsupervised loss function And the supervised loss function Hyperparameters of When the iterative training reaches the preset first number of iterations, the trained generator and restorer are obtained.

5. The CO2 concentration detection model training method according to any one of claims 1 to 4, characterized in that: The loss functions used when training the generative adversarial network include: Where: The gas temperature and flow rate are defined as static characteristics, and Represent the vector of static features and the vector of temporal features respectively; and Represent the instantiated random vectors of static features and temporal features respectively, , ;by and Represent specific values ​​of static and temporal features respectively; is the reconstructed static feature, For reconstruction Moment time characteristics Representing tuples From the distribution The expected sampling is Representing tuples From synthetic data distribution The expected sampling is represents the sequence length, Indicates the time, represents random noise; Represents the output of the generator for time features; is the true classification of static features, Classify and synthesize static features; for The true classification of moment-to-moment temporal features, for Classification and synthesis of moment-to-moment temporal features; is the true embedding of static features, for True embedding of moment-to-moment temporal features; is the two-norm; , , They represent the reconstruction loss function, unsupervised loss function and supervised loss function respectively.

6. The CO2 concentration detection model training method according to claim 3, characterized in that: The generator, discriminator, embedder and restorer all adopt a 3-layer gated recurrent network.

7. The CO2 concentration detection model training method according to claim 1, characterized in that: The input of the multi-scale enhanced feature extraction module is pseudo data and labeled sample data, and the output of the multi-scale enhanced feature extraction module is connected to the prediction module and the autoencoder module respectively; The multi-scale enhanced feature extraction module is used to obtain multi-scale enhanced features corresponding to the input data based on the input data; The prediction module is used to predict CO2 concentration based on multi-scale enhanced features; The self-encoding module is used to obtain reconstructed data with the same dimension as the labeled sample data based on the multi-scale enhanced features.

8. The CO2 concentration detection model training method according to claim 1, characterized in that: The multi-scale hybrid enhancement submodule is used to perform the following calculations on a certain input data: Where: represents the average pooling operation, For the The output of the sub-average pooling, represents a multi-layer perceptron, , Output intermediate variables for features, It is the mixed data output by the multi-scale mixed enhancement submodule.

9. The CO2 concentration detection model training method according to claim 7, characterized in that: The method of training the CO2 concentration detection model using pseudo data and labeled sample data includes: The multi-scale enhanced feature extraction module and the prediction module are trained using labeled sample data, and the training objective is the third optimization objective: Where: , are the network parameters corresponding to the multi-scale enhanced feature extraction module and the prediction module respectively; Indicates The true labels of labeled sample data, , is the total amount of labeled sample data used to train the CO2 concentration detection model; Indicates that the prediction module is Output results of labeled sample data; is the two-norm; The self-encoding module is trained using pseudo data and the training objective is the fourth optimization objective: Where: are the network parameters corresponding to the autoencoder module; Indicates Pseudo data; Represents the self-encoding module for The output result of the pseudo data is: , is the total amount of pseudo data used to train the CO2 concentration detection model; When the iterative training reaches the preset second iteration number, a trained CO2 concentration detection model is obtained.

10. A CO2 concentration detection model training device, characterized in that: include: A data collection unit, used to collect different CO2 concentration information and corresponding feature information as labeled sample data; A first training unit is used to train the generative adversarial network using labeled sample data to obtain a trained generator and restorer; A pseudo data generating unit, used for generating a potential representation of pseudo data according to random noise using a generator, and then converting the potential representation of the pseudo data into pseudo data with the same dimension as the labeled sample data through a restorer; The second training unit is used to train the CO2 concentration detection model using pseudo data and labeled sample data. The CO2 concentration detection model includes a multi-scale enhanced feature extraction module, a prediction module and an autoencoder module, wherein the multi-scale enhanced feature extraction module includes a multi-scale hybrid enhanced submodule and a feature extraction submodule connected in sequence; the multi-scale hybrid enhanced submodule includes an equal number of average pooling layers and multi-layer perceptrons, each average pooling layer is connected in sequence, each multi-layer perceptron is connected in sequence, and the first The output of the average pooling layer is The outputs of the multi-layer perceptrons are added together as the The mixed data obtained by adding the time features of the input data of the first average pooling layer and the output features of the first multi-layer perceptron is used as the input of the feature extraction submodule; The feature extraction submodule includes a Fourier transform layer, a first convolution layer, a batch normalization layer BN, an activation function ReLU, an inverse Fourier transform layer and a second convolution layer connected in sequence; the multi-scale mixed enhancement submodule is used to perform multi-scale mixed enhancement on the input data to obtain mixed data with the same dimension as the input data; the feature extraction submodule is used to extract features from the mixed data to obtain multi-scale enhanced features.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 9 is implemented.

12. A thermal power carbon emission monitoring system, characterized in that: It includes a laser, an absorption cell, a photoelectric detector, a signal processor, an environmental detector and a CO2 concentration detector. The absorption cell is arranged on the output laser light path of the laser and a gas path is provided on the absorption cell for the gas to be measured to enter the absorption cell. The environmental detector is arranged at the air inlet of the gas path. The output of the photoelectric detector is connected to the CO2 concentration detector via the signal processor, and the output of the environmental detector is connected to the CO2 concentration detector. The CO2 concentration detector deploys a CO2 concentration detection model trained by the CO2 concentration detection model training method described in any one of claims 1-9.

13. The thermal power carbon emission monitoring system according to claim 12, characterized in that: The environmental detector is used to detect the temperature and flow rate of the gas to be tested.

14. The thermal power carbon emission monitoring system according to claim 12, characterized in that: The input end of the laser is connected to a wavelength controller, and the wavelength controller is used to control the laser to output laser light of a set wavelength.

Images