Denoising diffusion model-based smelting furnace oxygen concentration prediction method

Abstract

The invention discloses a smelting furnace oxygen concentration prediction method based on a denoising diffusion model, and belongs to the technical field of energy process soft measurement modeling. The method comprises the following steps: (1) acquiring flame image data of a smelting furnace; (2) smelting furnace flame image data preprocessing and data set division; (3) establishing a prediction model based on U-Net; (4) constructing a joint prediction model integrated with DDPM, and training the model by using a flame image data set; according to the smelting furnace oxygen concentration prediction method based on the de-noising diffusion model, the U-Net feature map output process and the DDPM reverse de-noising process are combined, the feature extraction capacity of the model is improved, and therefore the prediction accuracy of the model using flame image data is improved.

Description

Technical Field

[0001] The invention relates to the field of soft-sensing modeling of energy processes, and in particular to a method for predicting oxygen concentration in a furnace based on a denoising diffusion model. Background Art

[0002] In industrial combustion systems, combustion efficiency can be determined by measuring the efficiency of fuel conversion to heat. In order to reduce operating costs and meet environmental regulations, combustion efficiency and exhaust gas emission content should be controlled at appropriate levels. Gas analyzers are usually used to measure the oxygen and nitrogen oxide content in exhaust gas, but this method will produce delays, and the feedback controller based on oxygen content has a tendency to overcompensate. The flame image measured online provides sufficient information to reflect the current combustion state. Therefore, a soft sensing model based on deep learning technology that can predict the oxygen concentration in the furnace based on the flame image has become a possible alternative to expensive hardware sensors.

[0003] Flames can be divided into the core, inner flame and outer flame. Different areas will show different properties with different oxygen concentrations. Therefore, semantic segmentation of flame images helps improve the model's ability to obtain information from different areas in flame images. U-Net (U shape Network) is a convolutional neural network for image segmentation. It uses a U-shaped network structure and a jump connection mechanism to effectively capture feature information at different levels, thereby improving the accuracy of image segmentation and detail retention. The powerful performance of U-Net makes it widely used in many fields. Among them, the Denoising Diffusion Probabilistic Model (DDPM) also uses a variant model of U-Net to achieve high-quality data generation. The DDPM model is a generative model that samples from standard Gaussian noise and gradually removes noise through a reverse process to generate new data. The model can be divided into two processes: the forward process and the reverse process. In the forward process, Gaussian noise is gradually added to the original data until the data is completely converted into standard Gaussian noise, while the reverse process gradually removes noise until the data is restored. On this basis, a neural network is trained to predict the noise that needs to be removed in each link of the reverse process. Considering that the process of U-Net model outputting feature maps after semantic segmentation based on the original image can be regarded as the process of extracting effective features after removing useless information from the original image, which is similar to the DDPM inverse denoising process to a certain extent, the present invention proposes a soft measurement modeling method for furnace oxygen concentration prediction based on a denoising diffusion model. Firstly, a soft measurement model is designed based on U-Net for oxygen concentration prediction. At the same time, the process of U-Net outputting feature maps is combined with the DDPM inverse denoising process to improve the feature extraction ability of the model, thereby improving the prediction accuracy of the model using flame image data. Summary of the invention

[0004] In order to solve the problem of difficulty in predicting key quality variables using image data in industrial processes, the present invention proposes a furnace oxygen concentration prediction method based on a denoising diffusion model, which combines the process of U-Net output feature maps with the DDPM inverse denoising process to improve the feature extraction ability of the model, thereby improving the prediction accuracy of the model using flame image data.

[0005] The technical solution adopted by the present invention to solve its technical problem is:

[0006] A method for predicting furnace oxygen concentration based on a denoising diffusion model comprises the following steps:

[0007] (1) Obtain furnace flame image data as original sample data.

[0008] (2) Preprocessing of original sample data and data set division: The original sample data obtained is divided into a training set and a test set. In order to speed up the model convergence and reduce the model training time, the training set data and the test set data are normalized.

[0009] (3) Construct a flame prediction model based on U-Net.

[0010] (4) Construct a joint prediction model incorporating DDPM: Construct a joint loss function based on the denoising diffusion model and train the model using the flame image dataset.

[0011] (5) Model performance evaluation: In order to evaluate the method of the present invention more objectively, the evaluation indicators root mean square error (RMSE) and maximum absolute error (MAE) are introduced.

[0012] Further, the process of (3) is as follows:

[0013] Step 3.1: Build a prediction model based on U-Net

[0014] U-Net is mainly composed of an encoder and a decoder with skip connections. Among them, the main function of the encoder is to map the image data into a latent vector, the main function of the decoder is to reconstruct the image data from the latent vector, and the function of the skip connection is to improve the fineness of the image data restored by the decoder. The present invention proposes a prediction model based on U-Net, which first uses U-Net to perform semantic segmentation on the flame image. The goal of this link is to remove noise and background while extracting the features of the flame image. Afterwards, the model will use the feature map output by U-Net to predict the oxygen concentration label. Here, if X represents the flame image data we collected for soft measurement modeling, is the corresponding oxygen concentration label, where n represents the number of data, and the oxygen concentration predicted by the model is Specific model loss function for:

[0015]

[0016] Further, the process of (4) is as follows:

[0017] Step 4.1: Build a denoising diffusion model

[0018] DDPM is a latent variable model consisting of two Markov chains: a forward denoising process and a backward denoising process. Formally, noise is added to the original data in the forward denoising process, while in the backward process, new data is generated from the noise through a series of denoising processes. Here, q(·) represents the forward process and p(·) represents the backward process. For the original data x0~q(x0), in a diffusion forward process containing T steps, the data x in the tth step is t is the data x at step (t-1) t-1 Add Gaussian noise to the basis of represents Gaussian distribution, it can be expressed as:

[0019]

[0020] in, is the variance sequence used in each step, its value is between 0 and 1, and Ι represents the unit matrix. If the diffusion step number T is long enough, the final x T It completely becomes random noise. For the entire diffusion process, it can be expressed in the form of a Markov chain as follows:

[0021]

[0022] where q(x 1:T |x0) represents the posterior probability of the forward diffusion process given the initial x0. Furthermore, another α t =1-β t , The following are some of the techniques for re-parameterization:

[0023]

[0024] The inverse process can be expressed as the joint probability distribution p θ (x 0:T ), the distribution q(xt|xt-1) is estimated through a neural network, and the inverse process can be expressed in the form of a Markov chain as:

[0025]

[0026] in p θ (x t-1 |x t ) is a parameterized Gaussian distribution with mean μ θ (x t ,t) and variance ∑ θ (x t ,t) is obtained by neural network training, and θ represents the trainable parameters in the neural network. The training goal of the neural network is to make the predicted noise consistent with the real noise added in the forward process. The target loss function is for:

[0027]

[0028] Where t∈[1,T], ε is Gaussian noise, is the trained neural network, represents the expectation. The final reverse process generates data as follows:

[0029]

[0030] Where z is the added random noise, and when t>1, When t=1, z=0.

[0031] σ t is the variance coefficient, which has

[0032] Step 4.2: Construct denoising loss function

[0033] Through step 4.2, the DDPM inverse process has been obtained, that is, the noise of the noisy data is gradually removed until the data is restored, as shown in formula (8). By shifting the terms in formula (8), we can get:

[0034]

[0035] Assuming that the feature map output by the U-Net module is DDPM, the inverse process performs an additional denoising on the original data. Considering that when t = 1, z = 0, then it is constructed according to formula (9):

[0036]

[0037] Among them, x0 represents the known original flame image data, and α0, σ0 and t0 are unknown parameters, and u0 is the U-Net output feature map. Since the U-Net model is a neural network model that needs to be trained, in order to simplify the calculation, the above unknown parameters are regarded as the trainable parameters of the U-Net model. φ (·) represents the U-Net model, ψ θ (·) represents the noise prediction neural network of DDPM, φ and θ are the trainable parameters in the U-Net model and the noise prediction neural network of DDPM, respectively, which can be constructed as follows:

[0038] F φ (x0)-η·ψ θ (x0)=x0 (11)

[0039] Where η represents the hyperparameter coefficient used to balance the two models, and the denoising loss function is obtained in formula (11):

[0040]

[0041] Step 4.3: Joint training

[0042] In order to improve the feature extraction capability of the model, the present invention proposes a furnace oxygen concentration prediction method based on a denoising diffusion model. The model includes a U-Net prediction model and a DDPM model. The U-Net prediction model first uses the flame image to output a feature map, and then predicts the oxygen concentration based on the feature map. In step 4.2, it is assumed that the feature map output by the U-Net model is the DDPM reverse process. An additional denoising is performed on the original data, and a denoising loss function is constructed. Finally, the DDPM model and the U-Net model are trained simultaneously, and the noise prediction neural network in the DDPM is used to improve the feature extraction capability of the U-Net model. The joint loss function is composed of the DDPM model loss function, the U-Net model loss function and the denoising loss function. At the same time, in order to balance these loss functions, a hyperparameter γ is introduced to balance the loss function:

[0043]

[0044] Wherein γ>0, the model proposed in the present invention is then trained on the training set and optimized using the joint loss function. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of the method of the present invention;

[0046] Figure 2 It is a comparison chart of the prediction results of the method of the present invention and other methods;

[0047] Figure 3 It is a comparison chart of absolute errors between the method of the present invention and other methods.

[0048] By adopting the above technology, compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] The present invention proposes a furnace oxygen concentration prediction method based on a denoising diffusion model, which combines the process of U-Net outputting feature maps with the DDPM inverse denoising process to improve the feature extraction capability of the model, thereby improving the prediction accuracy of the model using flame image data. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments of the specification. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0051] On the contrary, the present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention as defined by the claims. Further, in order to make the public have a better understanding of the present invention, some specific details are described in detail in the detailed description of the present invention below. Those skilled in the art can fully understand the present invention without the description of these details.

[0052] Reference Figure 1 to Figure 3 , a soft-sensor modeling method for predicting furnace oxygen concentration based on a denoising diffusion model, the method comprising the following steps:

[0053] (1) Obtaining furnace flame image data: Use an experimental furnace with a burner model of 5514-6 and industrial heavy oil as fuel to conduct combustion experiments. The filtered exhaust gas sampled from the chimney is sent to the exhaust gas analyzer to measure the exhaust temperature, oxygen, carbon dioxide and carbon monoxide concentrations of the exhaust gas. Through the observation window of the experimental furnace, a charge coupled device (CCD) camera is used to capture the image of the flame in the furnace. The output signal of the CCD camera is sent to the on-site computer through the IEEE-1394a interface. During the operation, the oxygen concentration and flame image in the furnace are automatically recorded;

[0054] (2) Preprocessing of original sample data and data set division: The original sample data obtained is divided into a training set and a test set. In order to speed up the model convergence and reduce the model training time, the training set data and the test set data are normalized;

[0055] (3) Establish and train a flame soft measurement model based on U-Net;

[0056] (4) Construct a joint prediction model integrating DDPM: construct a joint loss function based on the denoising diffusion model and train the model using the flame image dataset;

[0057] (5) Model performance evaluation: In order to evaluate the method of this embodiment more objectively, the evaluation indicators root mean square error (RMSE) and maximum absolute error (MAE) are introduced.

[0058] Specifically, the process of step (2) is:

[0059] Step 2.1: In order to speed up the training of the model, the image data is normalized to the maximum and minimum values, and the normalized data set is divided into a training set and a test set. The training set consists of 110 sets of data, the validation set consists of 44 sets of data, and the test set consists of 44 sets.

[0060] Specifically, the process of step (3) is:

[0061] Step 3.1: Build a prediction model based on U-Net

[0062] U-Net is mainly composed of an encoder and a decoder with jump connections. The main function of the encoder is to map image data into latent vectors, the main function of the decoder is to reconstruct image data from latent vectors, and the function of the jump connection is to improve the fineness of the image data restored by the decoder. This embodiment proposes a prediction model based on U-Net. The model first uses U-Net to perform semantic segmentation on the flame image. The goal of this link is to remove noise and background while extracting the features of the flame image. Afterwards, the model will use the feature map output by U-Net to predict the oxygen concentration label. Here, if X represents the collected flame image data for soft measurement modeling, is the corresponding oxygen concentration label, where n represents the number of data. The oxygen concentration predicted by the model based on the flame image is At this time, the loss function of the model is:

[0063]

[0064] Specifically, the process of step (4) is:

[0065] Step 4.1: Build a denoising diffusion model

[0066] DDPM is a latent variable model consisting of two Markov chains: a forward denoising process and a backward denoising process. Formally, noise is added to the original data in the forward denoising process, while in the backward process, new data is generated from the noise through a series of denoising processes. Here, q(·) represents the forward process and p(·) represents the backward process. For the original data x0~q(x0), in a diffusion forward process containing T steps, the data x in the tth step is t is the data x at step (t-1) t-1 Add Gaussian noise to the basis of represents Gaussian distribution, it can be expressed as:

[0067]

[0068] in, is the variance sequence used in each step, its value is between 0 and 1, and Ι represents the unit matrix. If the diffusion step number T is long enough, the final x T It completely becomes random noise. For the entire diffusion process, it can be expressed in the form of a Markov chain as follows:

[0069]

[0070] where q(x1: T |x0) represents the posterior probability of the forward diffusion process given the initial x0. Furthermore, another t =1-β t , The following are some of the techniques for re-parameterization:

[0071]

[0072] The inverse process can be expressed as the joint probability distribution p θ (x 0:T ), the distribution q(xt|xt-1) is estimated through a neural network, and the inverse process can be expressed in the form of a Markov chain as:

[0073]

[0074] in p θ (x t-1 |x t ) is a parameterized Gaussian distribution with mean μ θ (x t ,t) and variance ∑ θ (x t ,t) is obtained by neural network training, and θ represents the trainable parameters in the neural network. The training goal of the neural network is to make the predicted noise consistent with the real noise added in the forward process. The target loss function is for:

[0075]

[0076] Where t∈[1,T], ε is Gaussian noise, is the trained neural network, represents the expectation. The final reverse process generates data as follows:

[0077]

[0078] Where z is the added random noise, and when t>1, When t=1, z=0.

[0079] σ t is the variance coefficient, which has

[0080] Step 4.2: Construct denoising loss function

[0081] Through step 4.2, the DDPM inverse process has been obtained, that is, the noise of the noisy data is gradually removed until the data is restored, as shown in formula (21). By shifting the terms in formula (21), we can get:

[0082]

[0083] Assuming that the feature map output by the U-Net module is DDPM, the inverse process performs an additional denoising on the original data, then it is constructed according to formula (22):

[0084]

[0085] Among them, x0 represents the known original flame image data, and α0, σ0 and t0 are unknown parameters, and u0 is the U-Net output feature map. Since the U-Net model is a neural network model that needs to be trained, in order to simplify the calculation, the above unknown parameters are regarded as the trainable parameters of the U-Net model. φ (·) represents the U-Net model, ψ θ (·) represents the noise prediction neural network of DDPM, φ and θ are the trainable parameters in the U-Net model and the noise prediction neural network of DDPM, respectively, which can be constructed as follows:

[0086] F φ (x0)-η·ψ θ (x0) = x0 (24)

[0087] Where η represents the hyperparameter coefficient used to balance the two models. Based on formula (11), the denoising loss function is obtained:

[0088]

[0089] Step 4.3: Joint training

[0090] In order to improve the feature extraction capability of the model, this embodiment proposes a furnace oxygen concentration prediction method based on a denoising diffusion model. The model includes a U-Net prediction model and a DDPM model. The U-Net prediction model first uses the flame image to output a feature map, and then predicts the oxygen concentration based on the feature map. In step 4.2, it is assumed that the feature map output by the U-Net module is the DDPM reverse process. An additional denoising is performed on the original data, and a denoising loss function is constructed. Finally, the DDPM model and the U-Net model are trained simultaneously, and the noise prediction neural network in the DDPM is used to improve the feature extraction capability of the U-Net model. The joint loss function is composed of the DDPM model loss function, the U-Net model loss function and the denoising loss function. At the same time, in order to balance these loss functions, a hyperparameter γ is introduced to balance the loss function:

[0091]

[0092] Wherein γ>0, the model proposed in the present invention is then trained on the training set and optimized using the joint loss function.

[0093] Then, the model proposed in this embodiment is trained on the training set, hyperparameters are selected on the validation set, and finally the model effect is tested on the test set. Finally, the hyperparameters γ and η are selected as 0.006 and -1500 respectively.

[0094] Specifically, the process of step (5) is:

[0095] Step 5.1: Compare and evaluate the model results

[0096] To compare the model effects, support vector regression (SVR) and residual network (Res-Net) were used to compare with the method proposed in this embodiment. At the same time, in order to verify whether the joint loss function proposed in this embodiment works, the flame oxygen concentration prediction model based on U-Net (U-Net Soft Sensor, USS) proposed in this embodiment was used for a separate experiment. The final comparison results are shown in Table 1:

[0097] Table 1 Model performance comparison

[0098]

[0099] The results show that all indicators of the method proposed in the present invention are smaller than those of the SVR and Res-Net methods, indicating that its ability to predict oxygen concentration using flame images is better. At the same time, the results show that the use of the joint loss function proposed in this embodiment can further improve the model performance. From the results, the RMSE value decreased by 21.83% and the MAE value decreased by 17.94%, which proves the effectiveness and superiority of the method proposed in this embodiment.

[0100] This embodiment method proposes a furnace oxygen concentration prediction method based on a denoising diffusion model, which combines the process of U-Net outputting feature maps with the DDPM inverse denoising process to improve the feature extraction capability of the model, thereby improving the prediction accuracy of the model using flame image data.

[0101] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for predicting furnace oxygen concentration based on a denoising diffusion model, the method The following steps are involved: (1) Obtaining furnace flame image data as original sample data; (2) Preprocessing of original sample data and data set division: The original sample data obtained is divided into a training set and a test set; in order to speed up the model convergence and reduce the model training time, the training set data and the test set data are normalized; (3) Establish a prediction model based on U-Net; (4) Construct a joint prediction model integrating DDPM: construct a joint loss function based on the denoising diffusion model and train the model using the flame image dataset; (5) Model performance evaluation.

2. A method for predicting furnace oxygen concentration based on a denoised diffusion model as claimed in claim 1, It is characterized in that A soft sensing model based on U-Net is constructed for oxygen concentration prediction. The process of (3) is as follows: Step 3.1: Build a prediction model based on U-Net U-Net includes an encoder and a decoder with skip connections. The model first uses U-Net to perform semantic segmentation on the flame image to remove noise and background while extracting the features of the flame image. Then, the feature map output by U-Net is used to predict the oxygen concentration label. Let X represent the flame image data used for soft sensor modeling, is the corresponding oxygen concentration label, where n represents the number of data, and the oxygen concentration predicted based on the model is: Specific model loss function l soft for:

3. A method for predicting furnace oxygen concentration based on a denoised diffusion model as claimed in claim 1 or claim 2, It is characterized in that In (4), the process of establishing the denoising diffusion model is as follows: 1) Let q(·) represent the forward process and p(·) represent the reverse process. For the original data x 0 ~q(x 0 ), in a diffusion forward process consisting of T steps, the data x in step t t is the data x at step (t-1) t-1 Add Gaussian noise to the basis of represents Gaussian distribution, it can be expressed as: in, is the variance sequence used in each step, its value is between 0 and 1, Ι represents the unit matrix; for the entire diffusion process, it is expressed in the form of a Markov chain as: where q(x 1 : T |x 0 ) means that given the initial x 0 The posterior probability of the forward diffusion process when ; 2) Let α t =1-β t , The following are some of the techniques for re-parameterization: Using a neural network to estimate the distribution q(x t |x t-1 ), the reverse process is expressed in the form of a Markov chain as: in p θ (x t-1 |x t ) is a parameterized Gaussian distribution with mean μ θ (x t ,t) and variance ∑ θ (x t ,t) is obtained by neural network training, θ represents the trainable parameters in the neural network; the training goal of the neural network is to make the predicted noise consistent with the real noise added in the forward process, and the target loss function l diffusion for: Where t∈[1,T], ε is Gaussian noise, is the trained neural network, represents expectations; The final reverse process generates data as follows: Where z is the added random noise, and when t>1, When t=1, z=0; σ t is the variance coefficient, which has 4. A method for predicting furnace oxygen concentration based on a denoised diffusion model as claimed in claim 3, It is characterized in that In (4), the process of constructing the denoising loss function is as follows: 1) By shifting the terms in formula (8), we can obtain: Assume that the feature map output by the U-Net module is DDPM. The inverse process performs an additional denoising on the original data, and considers that when t = 1, z = 0, it is constructed according to formula (9): where x 0 Represents the known original flame image data, α 0 , σ 0 ,t 0 is an unknown parameter, u 0 Output feature map for U-Net; 2) Treat the unknown parameters as trainable parameters of the U-Net model and use F φ (·) represents the U-Net model, ψ θ (·) represents the noise prediction neural network of DDPM, φ and θ are the trainable parameters in the U-Net model and the noise prediction neural network of DDPM, respectively, which can be constructed as follows: F φ (x 0 )-η·ψ θ (x 0 )=x 0 (11) Where η represents the hyperparameter coefficient used to balance the two models; the denoising loss function l is obtained based on formula (11): re : l re =||F φ (x)-η·ψ θ (x)|| 2 。 (12) 5. A method for predicting furnace oxygen concentration based on a denoised diffusion model as claimed in claim 4, It is characterized in that In (4), the model training adopts joint training, and the specific process is as follows: The DDPM model and the U-Net model are trained simultaneously, and the noise prediction neural network in the DDPM is used to improve the feature extraction ability of the U-Net model. The joint loss function is composed of the DDPM model loss function, the U-Net model loss function and the denoising loss function. At the same time, the hyperparameter γ is introduced to balance the loss function: l=l diffusion +l soft +γl re (13) Where γ>0, then the model proposed in the present invention is trained on the training set and optimized using the joint loss function.

Images