Combustion stability quantitative characterization method and system based on combustion significant feature learning

Abstract

The invention relates to the technical field of thermal power generating unit combustor flame image processing, in particular to a combustion stability quantitative characterization method and system based on combustion significant feature learning, and the method comprises the steps: based on the combustion significant feature real-time probability distribution and the combustion significant feature average probability distribution, through the coincidence interval fitting between mathematical models, calculating the combustion stability of the combustion significant feature; the method solves the problem that in the prior art, due to interference of background image features, combustion saliency features are not obvious, the encoder extracts the combustion saliency features, interference of the background image features is weakened, and the combustion stability is improved. The accuracy of performing a downstream clustering classification task on the features extracted by a classical unsupervised algorithm is improved, and the combustion state recognition effect based on the flame image features is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of flame image processing of a thermal power unit burner, and specifically relates to a method and system for quantitatively characterizing combustion stability based on learning of combustion significant features. Background Art

[0002] Due to the rapid growth of electricity demand in China, the design of traditional coal-fired units is mainly for stable combustion at high loads, usually operating in the rated load range of 70%-100%. The peak shaving capacity is insufficient. To adapt to the changes in future power grid demand, the flexibility of traditional units is transformed to improve the low-load stable operation ability of thermal power units. And the low-load stable combustion technology is the core technology of the flexibility transformation of thermal power units, which directly determines the ability of the unit to participate in deep peak shaving. Usually, the minimum power of a thermal power unit is determined by the steam turbine and the boiler. The minimum load allowed by the boiler technology depends on the stability of boiler combustion. For the boiler to achieve stable combustion under low-load conditions, accurate combustion state monitoring is indispensable. Especially during low-load operation, a timely, accurate and reliable combustion state monitoring system is essential. By real-time data flame state feedback to guide combustion optimization, improve combustion stability, and reduce pollutant emissions. To sum up, the existing technology has the interference of background image features, resulting in the problem that the combustion significant features are not obvious. Summary of the Invention

[0003] The present invention aims at the problems existing in the prior art, and provides a method and system for quantitatively characterizing combustion stability based on learning of combustion significant features.

[0004] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0005] In the first aspect, an embodiment of the present invention provides a method for quantitatively characterizing combustion stability based on learning of combustion significant features, including:

[0006] Based on a set of combustion flame images collected by flame acquisition, a significant feature encoder model is generated through training by comparing a variational autoencoder model;

[0007] Based on a set of real-time stable combustion images, through the input of the significant feature encoder model, the real-time probability distribution of combustion significant features is output;

[0008] Based on a set of historical stable combustion images, through the input of the significant feature encoder model, the average probability distribution of combustion significant features is output;

[0009] Based on the real-time probability distribution of combustion significant features and the average probability distribution of combustion significant features, through the fitting of the overlapping interval between mathematical models, the deviation degree S of the feature distribution from the stable combustion distribution is obtained to determine the difference or similarity between the state data of the real-time probability distribution of combustion significant features and the average probability distribution of combustion significant features, and is used to evaluate the furnace combustion stability.

[0010] Preferably, the training of the contrastive variational autoencoder model includes:

[0011] Based on the combustion flame image set, by extracting the stable combustion image set, an objective training set is obtained;

[0012] Based on the combustion flame image set, by extracting the fire extinguishing image set, a background training set is obtained;

[0013] According to the model training requirements, determine the structure and various hyperparameters of the contrastive variational autoencoder model;

[0014] Based on the objective training set and the background training set, perform model training and synchronously calculate the loss function of the model;

[0015] Based on the calculation result of the loss function, through the judgment of loss convergence, end the model training.

[0016] Preferably, the model training includes:

[0017] Take the stable combustion flame image of the objective training set as the input of the specific feature encoder branch and the background feature encoder branch of the contrastive variational autoencoder model. Through the first loss function, obtain the combustion significant feature and the input background feature, splice the combustion input interval, where the combustion input interval is (input background feature - combustion significant feature), and reconstruct the image through the first shared decoder;

[0018] Take the background training set as the input of the background feature encoder branch of the contrastive variational autoencoder model. Through the second loss function, obtain the reconstructed background feature, where the reconstructed image has the same dimension size as the input image, splice the reconstructed output interval, where the reconstructed output interval is (0 - reconstructed background feature), and reconstruct the image through the shared decoder;

[0019] Based on the total loss function of the sum of the first loss function and the second loss function, train the contrastive variational autoencoder model until the loss function value drops to the expected range or the number of loops reaches the expected value.

[0020] Preferably, the first loss function:

[0021]

[0022] Among them, the is the reconstruction loss; the D KL (q φz (z|x)||p(z)) is the KL divergence loss of the first shared decoder branch; the (q φs(s|x)||p(s)) is the KL divergence loss of the second shared decoder branch, and γ·TC is the total correlation term loss for increasing the independence of significant features and background features; Loss1 is the first loss value.

[0023] Preferably, the second loss function:

[0024]

[0025] Loss2 is the second loss value.

[0026] Preferably, the coincidence interval fitting between the mathematical models includes an interval fitting function;

[0027] The interval fitting function:

[0028]

[0029] Among them, is the average probability distribution of the combustion significant features, and q φs (s|x i ) is the real-time probability distribution of the combustion significant features, and S is the coincidence degree of the feature distribution deviating from the stable combustion distribution, where S is a dimensionless value between the interval [0,1].

[0030] Preferably, after obtaining the coincidence degree S of the feature distribution deviating from the stable combustion distribution, stability processing is further included;

[0031] The stability processing includes:

[0032] Based on the coincidence degree S of the feature distribution deviating from the stable combustion distribution, through the moving average processing of the S value mapping model, the combustion stability index is obtained.

[0033] Preferably, the moving average processing of the S value mapping model includes a moving average function;

[0034] The moving average function:

[0035]

[0036] Among them, Stabiltiy is the combustion stability index; a is the upper limit of the S value setting in the fire extinguishing condition, and b is an adjustable parameter, where by adjusting the size of b, flexible adjustment can be made according to specific requirements and scenarios.

[0037] Preferably, the flame acquisition includes:

[0038] Based on the collected combustion flame video, the combustion flame video is converted into a combustion flame image set;

[0039] Based on the combustion flame image set, resize each image to a square flame image;

[0040] Based on the square flame images, obtain the average image or median image of all frame images within a predetermined time to form a predetermined combustion flame image set;

[0041] Based on the predetermined combustion flame image set, generate a combustion flame image set through an image enhancement algorithm to improve the image contrast.

[0042] Meanwhile, the present invention also provides a combustion stability quantitative characterization system based on combustion significant feature learning, including:

[0043] Based on the above stability quantitative characterization method, the stability quantitative characterization system includes:

[0044] Model generation module: used to generate a significant feature encoder model by training a contrastive variational autoencoder model based on the combustion flame image set collected by the flame acquisition;

[0045] Real-time probability acquisition module: used to output the real-time probability distribution of combustion significant features by inputting the real-time stable combustion image set through the significant feature encoder model;

[0046] Average probability acquisition module: used to output the average probability distribution of combustion significant features by inputting the historical stable combustion image set through the significant feature encoder model;

[0047] Distribution coincidence degree acquisition module: used to obtain the feature distribution deviation from the stable combustion distribution coincidence degree S by fitting the coincidence interval between the real-time probability distribution of combustion significant features and the average probability distribution of combustion significant features through a mathematical model, so as to determine the difference or similarity between the state data of the real-time probability distribution of combustion significant features and the average probability distribution of combustion significant features, and to evaluate the furnace combustion stability.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] This combustion stability quantitative characterization method includes four main steps to determine the difference or similarity between the state data of the real-time probability distribution of combustion significant features and the average probability distribution of combustion significant features, and to evaluate the furnace combustion stability; based on the above four steps, the present invention provides a feature learning method based on a contrastive variational autoencoder, enabling the encoder to extract combustion significant features, weakening the interference of background image features, improving the accuracy of the features extracted by classical unsupervised algorithms for downstream clustering and classification tasks, and improving the combustion state recognition effect based on flame image features. Description of the Drawings

[0050] Figure 1 is the system flow chart of the embodiment of the present invention;

[0051] Figure 2 is the model training flow chart of the embodiment of the present invention;

[0052] Figure 3 is the specific flow chart of the embodiment of the present invention;

[0053] Figure 4 is the schematic diagram of data distribution comparison of the embodiment of the present invention;

[0054] Figure 5 is the structural diagram of the contrast variational autoencoder network of the embodiment of the present invention;

[0055] Figure 6 is the diagram of the change in the coal feeding rate of the burner in the first embodiment of the present invention;

[0056] Figure 7a is the combustion stability evaluation curve 1 in the first embodiment of the present invention;

[0057] Figure 7b is the combustion stability evaluation curve 2 in the first embodiment of the present invention;

[0058] Figure 8 is the shared feature encoder architecture and parameter table of the contrast variational autoencoder model in the first embodiment of the present invention;

[0059] Figure 9 is the specific feature encoder architecture and parameter table of the contrast variational autoencoder model in the first embodiment of the present invention;

[0060] Figure 10 is the decoder architecture and parameter table of the contrast variational autoencoder model in the first embodiment of the present invention;

[0061] Figure 11 is the discriminator architecture and parameter table of the contrast variational autoencoder model in the first embodiment of the present invention. Detailed implementation manners

[0062] It should be noted that the methods used in the present invention are all conventional methods unless otherwise specified; the raw materials and devices used are all conventional commercially available products unless otherwise specified, and their sources are not specifically limited.

[0063] The present invention will be further described below in conjunction with specific embodiments, but the protection scope of the present invention is not limited thereto.

[0064] Figure 1 is the system flow chart according to the embodiment of the present invention, Figure 2 is the model training flow chart according to the embodiment of the present invention, Figure 3 is the specific flow chart according to the embodiment of the present invention. As Figure 1 、2 As shown in FIGS. 3, the claim of the embodiment of the present invention includes: a quantitative characterization method for combustion stability based on learning of significant combustion characteristics, comprising:

[0065] Based on a set of combustion flame images collected by flame acquisition, through comparative variational autoencoder model training, a significant feature encoder model S101 is generated;

[0066] Based on a set of real-time stable combustion images, through input of the significant feature encoder model, a real-time probability distribution of combustion significant features S102 is output;

[0067] Based on a set of historical stable combustion images, through input of the significant feature encoder model, an average probability distribution of combustion significant features S103 is output;

[0068] Based on the real-time probability distribution of combustion significant features and the average probability distribution of combustion significant features, through fitting of the overlapping interval between mathematical models, a coincidence degree S of the feature distribution deviating from the stable combustion distribution is obtained to determine the difference or similarity between the state data of the real-time probability distribution of combustion significant features and the average probability distribution of combustion significant features, and is used to evaluate the furnace combustion stability S104.

[0069] As Figure 2 shown, the comparative variational autoencoder model training includes:

[0070] Based on a set of combustion flame images, by extracting a set of stable combustion images, a target training set S201 is obtained;

[0071] Based on a set of combustion flame images, by extracting a set of extinguished fire images, a background training set S202 is obtained;

[0072] According to the model training requirements, the structure and various hyperparameters of the comparative variational autoencoder model are determined S203;

[0073] Based on the target training set and the background training set, model training is carried out, and the loss function of the model is calculated synchronously S204;

[0074] Based on the calculation result of the loss function, through judgment of loss convergence, the model training is ended S205.

[0075] As Figure 1 、 3 shown, the model training includes:

[0076] Taking the stable combustion flame images of the target training set as the input of the specific feature encoder branch and the background feature encoder branch of the comparative variational autoencoder model, through the first loss function, combustion significant features and input background features are obtained, and the combustion input interval is spliced, where the combustion input interval is (input background feature - combustion significant feature), and the image is reconstructed through the first shared decoder;

[0077] Use the background training set as the input to the background feature encoder branch of the contrastive variational autoencoder model. Through the second loss function, reconstruct the background features. Among them, the reconstructed image has the same dimension size as the input image. Concatenate the reconstructed output interval, where the reconstructed output interval is (0 - reconstructed background features), and reconstruct the image through the shared decoder;

[0078] Train the contrastive variational autoencoder model based on the total loss function, which is the sum of the first loss function and the second loss function, until the value of the loss function drops to the expected range or the number of loops reaches the expected value.

[0079] The first loss function:

[0080]

[0081] Among them, the is the reconstruction loss; the D KL (q φz (z|x)||p(z)) is the KL divergence loss of the first shared decoder branch; the (q φs (s|x)||p(s)) is the KL divergence loss of the second shared decoder branch. The γ·TC is the total correlation term loss that increases the independence of the significant features and the background features; the Loss1 is the first loss value.

[0082] The second loss function:

[0083]

[0084] The Loss2 is the second loss value.

[0085] The coincidence interval fitting between the mathematical models includes an interval fitting function;

[0086] The interval fitting function:

[0087]

[0088] Among them, the is the average probability distribution of the burning significant features, the q φs (s|x i ) is the real-time probability distribution of the burning significant features. The S is the coincidence degree of the feature distribution deviating from the stable combustion distribution, where S is a dimensionless numerical value in the interval [0, 1].

[0089] After obtaining the coincidence degree S of the feature distribution deviating from the stable combustion distribution, stability processing is also included;

[0090] The stability processing includes:

[0091] Based on the coincidence degree S of the feature distribution deviating from the stable combustion distribution, the combustion stability index is obtained through the moving average processing of the S-value mapping model.

[0092] The moving average processing of the S-value mapping model includes a moving average function;

[0093] The moving average function:

[0094]

[0095] Among them, the Stability is the combustion stability index; the a is the upper limit of the S-value setting for the extinguishing condition, and the b is an adjustable parameter. Among them, by adjusting the size of b, flexible adjustment can be made according to specific requirements and scenarios.

[0096] Such as Figure 1 、 3 As shown in 4, the flame acquisition includes:

[0097] Based on the collected combustion flame video, the combustion flame video is converted into a combustion flame image set; based on the combustion flame image set, the size of each picture is adjusted to a square flame image;

[0098] Based on the square flame image, the average image or median image of all frame images within a predetermined time is obtained to form a predetermined combustion flame image set;

[0099] Based on the predetermined combustion flame image set, a combustion flame image set is generated through an image enhancement algorithm to improve the image contrast.

[0100] Such as Figure 1 、 2 As shown in 3, at the same time, the present invention also provides a combustion stability quantitative characterization system based on the learning of combustion significant features, including:

[0101] Based on the above stability quantitative characterization method, the stability quantitative characterization system includes:

[0102] Model generation module: used to generate a significant feature encoder model S101 through the training of a contrastive variational autoencoder model based on the combustion flame image set collected by flame acquisition;

[0103] Real-time probability acquisition module: used to output the real-time probability distribution S102 of combustion significant features through the input of the significant feature encoder model based on the real-time stable combustion image set;

[0104] Average probability acquisition module: used to output the average probability distribution S103 of combustion significant features through the input of the significant feature encoder model based on the historical stable combustion image set;

[0105] Distribution overlap degree acquisition module: It is used to obtain the overlap degree S of the characteristic distribution deviating from the stable combustion distribution based on the real-time probability distribution of combustion significant features and the average probability distribution of combustion significant features through the fitting of the overlap interval between mathematical models, so as to determine the difference or similarity between the state data of the real-time probability distribution of combustion significant features and the average probability distribution of combustion significant features, and is used to evaluate the furnace combustion stability S104.

[0106] Embodiment 1:

[0107] The following combines Figure 1 、 2 The working principle of the combustion stability quantitative characterization method shown in the embodiment is described.

[0108] As Figure 1 shown, this combustion stability quantitative characterization method includes four main steps. Among them, the first step: Based on the combustion flame image set collected by the flame, a significant feature encoder model is generated through the training of the contrastive variational autoencoder model; the second step: Based on the real-time stable combustion image set, through the input of the significant feature encoder model, the real-time probability distribution of combustion significant features is output; the third step: Based on the historical stable combustion image set, through the input of the significant feature encoder model, the average probability distribution of combustion significant features is output; the fourth step: Based on the real-time probability distribution of combustion significant features and the average probability distribution of combustion significant features, through the fitting of the overlap interval between mathematical models; among them, through the fourth step, the entire combustion stability quantitative characterization method is completed, and the overlap degree S of the characteristic distribution deviating from the stable combustion distribution is obtained to determine the difference or similarity between the state data of the real-time probability distribution of combustion significant features and the average probability distribution of combustion significant features, and is used to evaluate the furnace combustion stability; based on the above four steps, first, collect the video of the combustion flame of the burner, convert the video into an image set, and adjust the picture size to make it a square flame image with equal length and width, and take the average image or median image of all frame images in a short time, and use the image enhancement algorithm to improve the image contrast; secondly, use the stable combustion image set as the target training set and the fire extinguishing image set as the background training set, and send them into the contrastive variational autoencoder model for training; thirdly, after the model training is completed, the combustion significant related Gaussian distribution q φs (s|x) can be obtained through the specific feature encoder; finally, according to the above steps, the historical stable combustion image set is input into the model to obtain the average probability distribution of the stable combustion latent space The probability distribution q φs (s|x i ) of the real-time state data is obtained through the above steps, and the difference or similarity between the state data of is used to evaluate the furnace combustion stability. The comparison of the difference or similarity can generally be completed by calculating the overlap interval between two mathematical models in the Hilbert space; among them, Denote the normal distribution as q φs (s|x i ) denotes the real-time observation distribution. The calculation result S represents the coincidence degree of the two distributions, which is a dimensionless value within the interval [0, 1]. The S-value mapping model calculated by the above steps is optimized by the following formula to map the dimensionless value measuring combustion stability to the range of [0, 1]. In order to prevent the influence of flame flicker on the results, a moving average process is finally performed on the combustion stability index. Among them, S is the preliminary stability degree of the probability distribution deviating from the stable operating condition calculated in the above steps, which is a dimensionless value within the interval [0, 1]. The present invention provides a feature learning method based on a contrastive variational autoencoder, enabling the encoder to extract combustion saliency features, weakening the interference of background image features, improving the accuracy of the features extracted by classical unsupervised algorithms for downstream clustering and classification tasks, and enhancing the combustion state recognition effect based on flame image features.

[0109] The working principle of the combustion stability quantitative characterization method shown in the following combined with Figure 1 、 2 embodiments will be described.

[0110] As Figure 2 shown, the model training flowchart of this combustion stability quantitative characterization method includes five main steps. The first step: Based on the combustion flame image set, obtain the target training set by extracting the stable combustion image set. The second step: Based on the combustion flame image set, obtain the background training set by extracting the extinguished fire image set. The third step: Determine the structure and various hyperparameters of the contrastive variational autoencoder model according to the model training requirements. The fourth step: Based on the target training set and the background training set, perform model training and synchronously calculate the loss function of the model. The fifth step: Based on the calculation result of the loss function, end the model training through loss convergence judgment. Based on the above steps, first, preprocess the flame images, determine the model structure and various hyperparameters of the contrastive variational autoencoder model based on the target training set and the background training set. Then, use the training set pictures to train the model and calculate the loss function, and judge whether the loss converges. If it converges, the training can be ended and the model can be saved.

[0111] The working principle of the combustion stability quantitative characterization method shown in the following combined with Figure 2 、 3 、4 embodiments will be described.

[0112] The model training includes: using the stable combustion flame images of the target training set as the inputs of the specific feature encoder branch and the background feature encoder branch of the contrastive variational autoencoder model, obtaining the combustion significant features and the input background features through the first loss function, splicing the combustion input interval, where the combustion input interval is (input background feature - combustion significant feature), and reconstructing the image through the first shared decoder; using the background training set as the input of the background feature encoder branch of the contrastive variational autoencoder model, obtaining the reconstructed background features through the second loss function, where the reconstructed image has the same dimension as the input image, splicing the reconstructed output interval, where the reconstructed output interval is (0 - reconstructed background feature), and reconstructing the image through the shared decoder; training the contrastive variational autoencoder model based on the total loss function which is the sum of the first loss function and the second loss function until the loss function value drops to the expected range or the number of loops reaches the expected value; where, the first loss function:

[0113]

[0114] where, the is the reconstruction loss; the D KL (q φz (z|x)||p(z)) is the KL divergence loss of the first shared decoder branch; the (q φs (s|x)||p(s)) is the KL divergence loss of the second shared decoder branch, and the γ·TC is the total correlation term loss for increasing the independence of the significant features and the background features; the Loss1 is the first loss value; where, the second loss function:

[0115]

[0116] The Loss2 is the second loss value;

[0117] In summary, the training steps include: First, using the stable combustion flame images of the target data set as the inputs of the specific feature encoder branch and the background feature encoder branch of the contrastive variational autoencoder model, obtaining the combustion significant feature s and the background feature z, splicing [z, s] and reconstructing the image through the shared decoder, and the reconstructed image has the same dimension as the input image; using the background data set as the input of the background feature encoder branch of the contrastive variational autoencoder model, obtaining the background feature z, splicing [z, 0] and reconstructing the image through the shared decoder; Second, setting the input of the contrastive autoencoder model as the target data set graph x tg , and the output image as x' tg , and the model consists of the shared decoding process p θ (x|[z, s]), the significant feature encoder branch and the background encoder branch q φz (z|x), and the mathematical description is: p θ (x|[z,s]):[z,s]→x' tg ; The loss function is as follows:

[0118]

[0119] In TC, the first term is the reconstruction loss, the second and third terms are the KL divergence losses of the two encoder branches respectively, and TC is the total correlation term loss for increasing the independence of significant features and background features. The specific mathematical description is as follows:

[0120]

[0121] The above formula uses the density ratio trick to approximate the TC loss, and v (i) is the concatenation of [s (i) ,z (i) . D ψ (v (i) ) is the discriminator output probability. The discriminator loss is as follows:

[0122]

[0123] The above formula randomly exchanges the column vectors of s and z within a batch to obtain s’ and z’, and v' perm is [z’,s’].

[0124] Let the input of the contrastive autoencoder model be the background dataset graph x bg , and the output image be x' bg . The model consists of a shared decoding process p θ (x|[z,0]) and a background encoder branch q φz (z|x). The mathematical description is: The above loss function is as follows:

[0125]

[0126] Use the sum of Loss1 and Loss2 as the total loss function to train the contrastive variational autoencoder model until the loss function value drops to the expected range or the number of loops reaches the expected value.

[0127] The following combines Figure 1 , 2 to illustrate the working principle of the combustion stability quantitative characterization method shown in the embodiments.

[0128] The overlapping interval fitting between the mathematical models includes an interval fitting function;

[0129] The interval fitting function:

[0130]

[0131] Among them, the is the average probability distribution of significant combustion characteristics, and the q φs (s|x i ) is the real-time probability distribution of significant combustion characteristics. The S is the coincidence degree of the feature distribution deviating from the stable combustion distribution, where S is a dimensionless value in the interval [0, 1];

[0132] Based on the interval fitting function, the above-mentioned historical stable combustion image set is input into the model to obtain the average probability distribution of the stable combustion latent space The real-time state data is used to obtain the probability distribution q φs (s|x i ), and the difference or similarity between the state data of is used to evaluate the furnace combustion stability. The comparison of the difference or similarity can generally be completed in the Hilbert space by calculating the coincidence interval between two mathematical models. The calculation formula is as follows:

[0133]

[0134] The working principle of the quantitative characterization method of combustion stability shown in the following Figure 1 and 2 embodiments will be described.

[0135] After obtaining the coincidence degree S of the feature distribution deviating from the stable combustion distribution, stability processing is also included;

[0136] The stability processing includes:

[0137] Based on the coincidence degree S of the feature distribution deviating from the stable combustion distribution, through the moving average processing of the S-value mapping model, a combustion stability index is obtained.

[0138] The moving average processing of the S-value mapping model includes a moving average function;

[0139] The moving average function:

[0140]

[0141] Among them, the Stability is the combustion stability index; the a is the upper limit of the S value for the extinguishing condition, and the b is an adjustable parameter. By adjusting the size of b, it can be flexibly adjusted according to specific requirements and scenarios.

[0142] Optimize the S - value mapping model calculated in the above steps through the following formula, map the dimensionless value measuring combustion stability to the range of [0, 1]. To prevent the influence of flame flickering on the results, perform a moving average process on the combustion stability index finally.

[0143]

[0144] Among them, S is the preliminary stability degree deviating from the probability distribution of the stable operating condition calculated in the above steps, which is a dimensionless value between the interval [0, 1]; a is the upper limit of the S - value setting for the extinguishing condition, and b is an adjustable parameter. By adjusting the size of b, it can be flexibly adjusted according to specific requirements and scenarios.

[0145] The following combines Figure 1 、 2 to illustrate the working principle of the combustion stability quantitative characterization method shown in the embodiments.

[0146] The pre - processing of the flame image includes: The first step: Based on the collected combustion flame video, convert the combustion flame video into a set of combustion flame images; The second step: Based on the set of combustion flame images, adjust the size of each picture to a square flame image; The third step: Based on the square flame image, obtain the average image or median image of all frame images within a predetermined time to form a set of predetermined combustion flame images; Collect the video of the combustion flame of the burner, convert the video into an image set, adjust the picture size to make it a square flame image with equal length and width, take the average image or median image of all frame images within a short time, and use an image enhancement algorithm to improve the image contrast.

[0147] The following combines Figure 1 、 2 、3 to illustrate the working principle of the combustion stability quantitative characterization system shown in the embodiments.

[0148] Such as Figure 1As shown in the figure, this quantitative characterization method for combustion stability includes four main modules, and each main module corresponds to each of the above steps. Among them, the first module corresponds to the first step, that is: based on the combustion flame image set collected by the flame, through the training of the contrast variational autoencoder model, a significant feature encoder model is generated; the second module corresponds to the second step, that is: based on the real-time stable combustion image set, through the input of the significant feature encoder model, the real-time probability distribution of combustion significant features is output; the third module corresponds to the third step, that is: based on the historical stable combustion image set, through the input of the significant feature encoder model, the average probability distribution of combustion significant features is output; the fourth module corresponds to the fourth step, that is: based on the real-time probability distribution of combustion significant features and the average probability distribution of combustion significant features, through the fitting of the overlapping interval between mathematical models; among them, through the fourth step, the entire quantitative characterization method for combustion stability is completed, and the deviation degree S of the feature distribution from the stable combustion distribution overlap is obtained to determine the difference or similarity between the state data of the real-time probability distribution of combustion significant features and the average probability distribution of combustion significant features, and is used to evaluate the furnace combustion stability; based on the above four modules, first, collect the video of the combustion flame of the burner, convert the video into an image set, and adjust the size of the picture to make it a square flame image with equal length and width. Take the average image or median image of all frame images in a short time, and use the image enhancement algorithm to improve the image contrast; secondly, use the stable combustion image set as the target training set and the fire extinguishing image set as the background training set, and send them into the contrast variational autoencoder model for training; thirdly, after the model training is completed, the significant relevant Gaussian distribution q φs (s|x) of combustion can be obtained through the specific feature encoder; finally, according to the above steps, the historical stable combustion image set is input into the model to obtain the average probability distribution of the stable combustion latent space The probability distribution q φs (s|x i ) of the real-time state data is obtained through the above steps, and the difference or similarity between the state data of is used to evaluate the furnace combustion stability. The comparison of the difference or similarity can generally be completed by calculating the overlapping interval between two mathematical models in the Hilbert space; among them, represents the normal distribution, and q φs (s|x i) It represents the real-time observation distribution, and the calculation result S represents the coincidence degree of the two distributions. It is a dimensionless value within the interval [0, 1]; the S value mapping model calculated by the above steps is optimized by the following formula to map the dimensionless value measuring the combustion stability to the range of [0, 1]. In order to prevent the influence of flame flicker on the result, a moving average process is finally performed on the combustion stability index; where S is the preliminary stability degree of the probability distribution deviating from the stable operating condition calculated by the above steps, which is a dimensionless value within the interval [0, 1]. The present invention provides a feature learning method based on a contrastive variational autoencoder, enabling the encoder to extract combustion salient features, weakening the interference of background image features, improving the accuracy of the features extracted by classical unsupervised algorithms for downstream clustering and classification tasks, and enhancing the combustion state recognition effect based on flame image features.

[0149] Example 1:

[0150] As Figure 6 , 7, and 8 show, in this example, 10 consecutive hours of combustion flame images are selected for analysis, where Figure 6 the vertical axis in the figure shows the change in the coal feeding amount during this period. It can be seen from the figure that during the period from 6:00 to 11:27, the coal feeding amount is in the high-level interval of [45 t / h, 50 t / h]. In this experiment, the coal feeding amount of the burner is gradually reduced at 11:27, and the coal feeding rate is reduced to 35 t / h at 11:36. After maintaining the coal feeding amount unchanged for a period of time, the coal feeding amount is further reduced. At 12:00, the coal feeding amount changes to 0 t / h and the burner is closed for a period of time; the burner is restarted at 19:18 and the coal feeding amount is gradually increased. At 19:26, the coal feeding amount returns to the high-level interval of 45 t / h. During the subsequent data collection period, the coal feeding amount is stabilized within the high-level interval of [45 t / h, 52 t / h], and the combustion state in this interval also remains stable. According to the combustion principle, the pulverized coal concentration is the most important influencing factor for the ignition and combustion characteristics of the pulverized coal air flow, directly affecting the stability of boiler combustion. The coal feeding amount is selected as the relevant variable for the change of the boiler combustion state, and the coal feeding amount is used as an index to evaluate the combustion state and compared with the quantitative stability result Stability of the contrastive variational autoencoder model and the combustion stability calculation model later to verify the effectiveness of the feature extraction method.

[0151] As Figure 3 shown, the implementation steps of the combustion salient feature extraction based on the contrastive variational autoencoder and the combustion stability calculation method of the Stability calculation model:

[0152] (1) Extract the flame videos of the burner at 25 frames per second frame by frame to obtain an RGB image set with a resolution of 960×576×3, and compress the images to a size of 128×128×3 through the bilinear method; to eliminate the influence of flicker on camera imaging and obtain high-quality representative images, smooth the images, take the median of the channel values at the corresponding pixel positions of the 25 frames of pictures obtained per second, and obtain a representative image representing the flame combustion state within 1 second. In this case, after preprocessing the video images in the time period from 10:00 to 12:30, 9000 flame images of 128×128×3 are obtained in 2.5 hours as training data, and 9000 preprocessed images during the fire extinguishing stage with a coal feeding rate of 0 t / h are selected as the background data set.

[0153] (2) Train a contrastive variational autoencoder model. The network structure is as Figure 5 shown. The shared feature encoder, specific feature encoder, and shared decoder architectures and parameters in this case are as Figure 8 , 9 , 10, 11 shown. In the table, k is the size of the convolutional kernel, s is the stride, d is the number of convolutional kernels, scale is the negative slope of the LeakyRelu activation function of the discriminator network. The output layers of the shared decoder and discriminator require values in the range [0,1], so the sigmoid activation function is used in the last layer. The selected optimizer is Adam (Adaptive Moment Estimation), the batch size is 128, the learning rate is 0.001, and the loss function uses the common cross-entropy loss function. The number of training epochs Epoch is set to 100 rounds. After 100 rounds of training, the loss function converges to the expected value, and the model training is completed at this time.

[0154] (3) Input the stable combustion image set of the burner operating at high load into the significant feature encoder branch to learn the average probability of all stable combustion latent space combustion significant features, a 32-dimensional normal distribution Input the test image set into the significant feature encoder branch to learn the probability of the latent space combustion significant features, a 32-dimensional normal distribution q φs (s|x i ).

[0155] (4) Calculate the coincidence degree S between the test set distribution learned through the above steps and the stable combustion average probability distribution through the following formula, that is, the degree to which the test set feature distribution deviates from the stable combustion distribution.

[0156]

[0157] (5) Optimize the S - value mapping model calculated in the above steps through the following formula. In this case, a = 0.925 and b = 9.2419. Map the dimensionless value measuring combustion stability to the range of [0, 1]. To prevent the influence of flame flickering on the results, perform a moving average process on the combustion stability index finally. The combustion stability evaluation results are as Figure 7a 、 7b 。

[0158]

[0159] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, rather than limiting the protection scope of the present invention. Simple modifications or equivalent replacements made by those of ordinary skill in the art to the technical solution of the present invention do not depart from the essence and scope of the technical solution of the present invention.

Claims

1. A quantitative characterization method for combustion stability based on learning of combustion salient features, characterized in that: include: Based on the combustion flame image set collected by flame, a significant feature encoder model is generated by contrasting the variational autoencoder model training; Based on a real-time stable combustion image set, the significant feature encoder model is input to output the real-time probability distribution of combustion significant features; Based on the historical stable combustion image set, the significant feature encoder model is input and the average probability distribution of combustion significant features is output; Based on the real-time probability distribution of combustion significant characteristics and the average probability distribution of combustion significant characteristics, the overlap degree S of the characteristic distribution deviation from the stable combustion distribution is obtained by fitting the overlap interval between mathematical models, so as to determine the difference or similarity between the state data of the real-time probability distribution of combustion significant characteristics and the average probability distribution of combustion significant characteristics, which is used to evaluate the furnace combustion stability.

2. The method for quantitative characterization of combustion stability according to claim 1, characterized in that: The contrastive variational autoencoder model training includes: Based on the combustion flame image set, a target training set is obtained by extracting a stable combustion image set; Based on the burning flame image set, a background training set is obtained by extracting the fire extinguishing image set; Determine the structure and hyperparameters of the contrast variational autoencoder model according to the model training requirements; Based on the target training set and the background training set, the model is trained and the loss function of the model is calculated simultaneously; Based on the calculation results of the loss function, the model training is terminated by judging the loss convergence.

3. The method for quantitative characterization of combustion stability according to claim 2, characterized in that: The model training includes: The stable flame image of the target training set is used as the input of the specific feature encoder branch and the background feature encoder branch of the contrast variational autoencoder model, and the combustion significant features and the input background features are obtained through the first loss function, and the combustion input interval is spliced, where the combustion input interval is (input background features-combustion significant features), and the image is reconstructed through the first shared decoder; The background training set is used as the input of the background feature encoder branch of the contrastive variational autoencoder model, and the reconstructed background features are obtained through the second loss function, wherein the reconstructed image has the same dimension as the input image, and the reconstructed output interval is spliced, wherein the reconstructed output interval is (0-reconstructed background features), and the image is reconstructed through the shared decoder; Based on a total loss function which is the sum of the first loss function and the second loss function, the contrastive variational autoencoder model is trained until the loss function value drops to an expected range or the number of cycles reaches an expected value.

4. The method for quantitatively characterizing combustion stability according to claim 3, characterized in that: include: The first loss function: Among them, the is the reconstruction loss; KL (q φz (z|x)∥p(z)) is the KL divergence loss of the first shared decoder branch; the (q φs (s|x)||p(s)) is the KL divergence loss of the second shared decoder branch, the γ·TC is the total correlation loss for increasing the independence of the salient features and the background features; and the Loss1 is the first loss value.

5. The method for quantitatively characterizing combustion stability according to claim 3, characterized in that: include: The second loss function: The Loss2 is a second loss value.

6. The method for quantitative characterization of combustion stability according to claim 1, characterized in that: The coincidence interval fitting between the mathematical models includes an interval fitting function; The interval fitting function: Among them, the is the average probability distribution of significant combustion characteristics, and q φs (s|x i ) is the real-time probability distribution of significant combustion characteristics, and S is the overlap degree of characteristic distribution deviating from stable combustion distribution, where S is a dimensionless value in the interval [0,1].

7. The method for quantitative characterization of combustion stability according to claim 1, characterized in that: The obtaining of the characteristic distribution deviation from the stable combustion distribution overlap S also includes stability processing; The stability treatment includes: Based on the overlap degree S of the characteristic distribution deviating from the stable combustion distribution, the combustion stability index is obtained through the sliding average processing of the S value mapping model.

8. The method for quantitatively characterizing combustion stability according to claim 7, characterized in that: The S-value mapping model sliding average processing includes a sliding average function; The sliding average function: Among them, Stability is a combustion stability index; a is an upper limit for setting the S value of the fire extinguishing condition; b is an adjustable parameter, wherein by adjusting the size of b, it can be flexibly adjusted according to specific needs and scenarios.

9. The method for quantitative characterization of combustion stability according to claim 1, characterized in that: The flame collection comprises: Based on the collected burning flame video, the burning flame video is converted into a burning flame image set; Based on the burning flame image set, each image is resized into a square flame image; Based on the square flame image, an average image or a median image of all frame images within a predetermined time is obtained to form a predetermined burning flame image set; Based on the predetermined combustion flame image set, a combustion flame image set is generated through an image enhancement algorithm to improve the image contrast.

10. A combustion stability quantitative characterization system based on combustion salient feature learning, characterized in that: include: Based on the stability quantitative characterization method according to claim 1 to claim 9, the stability quantitative characterization system comprises: Model generation module: used to generate a significant feature encoder model by training a comparative variational autoencoder model based on a set of combustion flame images collected by flames; Real-time probability acquisition module: used to output the real-time probability distribution of combustion salient features based on the real-time stable combustion image set through the salient feature encoder model input; Average probability acquisition module: used for stable combustion image sets based on history, input through the salient feature encoder model, and output the average probability distribution of combustion salient features; Distribution overlap acquisition module: It is used to obtain the overlap S of the characteristic distribution deviation from the stable combustion distribution based on the real-time probability distribution of the combustion significant characteristics and the average probability distribution of the combustion significant characteristics through the overlap interval fitting between the mathematical models, so as to determine the difference or similarity between the state data of the real-time probability distribution of the combustion significant characteristics and the average probability distribution of the combustion significant characteristics, so as to evaluate the furnace combustion stability.

Images