A Method and System for Predicting Hydrogen-Doped Backfire Limit in Gas Stoves Based on Flame Image Features

By extracting flame image features and using deep learning models, a risk model for hydrogen-blended gas stoves with backfire was constructed. This solved the problems of high cost and low accuracy in the assessment of the adaptability of hydrogen-blended gas stoves, and enabled rapid and accurate safety assessment.

CN123119366APending Publication Date: 2026-11-17CHONGQING UNIV +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611637260.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-22
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies suffer from high testing costs, long testing cycles, and low accuracy when assessing the suitability of gas stoves for hydrogen blending, making them unsuitable for rapid screening and safety assessment of large-scale in-service gas stoves.

Method used

By collecting flame images and video data of gas stoves, geometric shape, color distribution and dynamic change features are extracted to construct a gas stove backfire risk model. Convolutional neural networks and visual Transformer networks are used for feature extraction and fusion to predict the hydrogen-doped backfire limit of gas stoves.

Benefits of technology

It significantly reduces testing costs, improves evaluation efficiency and accuracy, is suitable for large-scale rapid screening of in-service equipment, provides a safe hydrogen doping ratio, and avoids the risk of backfire.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123119366A_ABST
    Figure CN123119366A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for predicting the flashback limit of hydrogen-blended gas stoves based on flame image features. The method includes acquiring flame images / videos and operating parameters of different gas stoves under conditions of gradually increasing hydrogen blending ratio and adjusting flow rate and primary air, monitoring until flashback or criticality occurs; after preprocessing, extracting flame geometry, color distribution, and dynamic change features, and normalizing them to construct a feature vector; combining experimental and image criteria to calibrate stable, critically unstable, and flashback states; inputting the feature vector and incorporating the hydrogen blending ratio to construct a flashback risk model; after training, using the hydrogen blending ratio as a variable input to the model, calculating the flashback risk change curve, determining the critical point based on a threshold, and inverting the maximum safe hydrogen blending ratio. This invention eliminates the need for high-ratio hydrogen blending experiments, relying solely on natural gas or low-hydrogen blending conditions. By rapidly sensing combustion characteristics and intelligently analyzing them, it predicts the flashback limit of hydrogen-blended gas stoves, reducing testing costs, improving evaluation efficiency, and enhancing prediction accuracy and reliability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of gas application safety and artificial intelligence technology, specifically relating to a method and system for predicting the hydrogen-induced backfire limit of a gas stove based on flame image features. Background Technology

[0002] With the energy structure shifting towards a low-carbon model, hydrogen energy, as a clean and efficient secondary energy source, is gradually being applied in urban gas systems. Blending hydrogen into existing natural gas pipelines is considered an important and feasible transitional path. This method can achieve large-scale hydrogen transmission, distribution, and utilization without significant modifications to existing infrastructure, demonstrating good engineering economics and promising prospects for widespread adoption. However, as an emerging technology, the adaptability of hydrogen blending into natural gas in end-use gas equipment remains unresolved. The significant differences in the physicochemical properties of hydrogen and natural gas mean that unstable combustion can lead to burner backfire, gas leaks, and other hazards, requiring careful attention. Currently, the following methods are primarily used to assess the adaptability of hydrogen blending in gas stoves:

[0003] (1) Gradual hydrogen blending test method: By gradually increasing the hydrogen blending ratio, a combustion test is conducted on each gas stove until backfire occurs, thereby determining its maximum allowable hydrogen blending ratio. Although the results of this method are relatively reliable, it has problems such as high testing cost, long experimental cycle, and large gas source consumption, making it difficult to be applied to the rapid screening of a large number of in-service gas stoves.

[0004] (2) Theoretical calculation method based on combustion mechanism: a combustion model is established using parameters such as flame propagation speed and primary air coefficient to predict backfire conditions. However, this type of method usually requires precise boundary conditions and a complex physical modeling process, and is highly dependent on the structural parameters of the equipment, making it difficult to directly apply to in-service gas stoves with complex structures and variable states.

[0005] (3) Experience or standard judgment method: The applicability of gas stoves is judged based on existing standards or experience. However, this type of method is usually based on traditional natural gas conditions and lacks systematic verification for hydrogen blending conditions, resulting in limited prediction accuracy. Summary of the Invention

[0006] The first objective of this invention is to provide a method for predicting the hydrogen-doped backfire limit of a gas stove based on flame image features, which can significantly reduce testing costs, improve prediction accuracy and reliability, and increase evaluation efficiency.

[0007] The first objective of this invention is achieved through the following technical measures: a method for predicting the hydrogen-induced flashback limit of a gas stove based on flame image features, characterized by comprising the following steps:

[0008] S1. Select gas stove samples of different types and service years and conduct combustion tests. During the test, gradually increase the hydrogen blending ratio of natural gas, and at the same time adjust the natural gas flow and primary air supply status to make the gas stove samples run stably under different working conditions. Simultaneously record flame images and video data as well as corresponding operating parameters, and continuously monitor the changes in combustion status until backfire occurs or approaches the backfire critical state.

[0009] S2. Preprocess the acquired flame images and video data;

[0010] S3. Extract various features of the flame's geometric shape, color distribution, and dynamic changes from the preprocessed flame image and video data, and normalize these features to construct a unified feature vector.

[0011] S4. Based on experimental observation criteria and image feature criteria, the combustion state of flames in flame images and video data in the form of feature vectors is calibrated and divided into stable combustion state, critical unstable state and tempering state.

[0012] S5. Using the feature vector obtained in step S3 as input, and introducing the hydrogen doping ratio as an auxiliary input parameter, construct a gas stove backfire risk model.

[0013] S6. Train and evaluate the accuracy of the gas stove backfire risk model;

[0014] S7. Based on the output of the gas stove backfire risk model, the hydrogen-blended backfire limit of the gas stove is calculated by inversion: the hydrogen blending ratio is input as a variable into the gas stove backfire risk model, and the hydrogen blending ratio is gradually changed to calculate the corresponding backfire risk value. The function relationship curve of backfire risk with hydrogen blending ratio is obtained, and the critical point of gas stove transitioning from stable combustion to backfire is determined by setting a backfire judgment threshold. The hydrogen blending ratio corresponding to the critical point is the maximum safe hydrogen blending ratio of the gas stove, that is, the hydrogen-blended backfire limit of the gas stove is obtained.

[0015] This invention acquires flame images of gas stoves under natural gas or low hydrogen blending conditions, extracts flame morphological, color, and dynamic change characteristics, and quantitatively characterizes the flame state by combining backfire detection features and combustion mechanisms. It establishes a mapping relationship between flame image features and backfire risk, thereby achieving intelligent assessment of the combustion safety of hydrogen-blended gas stoves. This invention enables the prediction of hydrogen-blended gas stove backfire limits based solely on natural gas or low hydrogen blending conditions, without requiring high-proportion hydrogen blending experiments, through rapid sensing of combustion characteristics and intelligent analysis. It significantly reduces testing costs and improves assessment efficiency, prediction accuracy, and reliability. Applicable to in-service equipment, it has good engineering application value and promising prospects for widespread adoption.

[0016] In step S3 of the present invention,

[0017] In terms of geometric feature extraction, edge detection and contour extraction are performed on the flame region to obtain the flame outline. The flame height, width, and shape parameters are then calculated. Simultaneously, internal flame structural features are identified, including the presence and height variations of the inner flame, to reflect combustion intensity and flame propagation characteristics. In some implementations, the flame region can be segmented using binarization and morphological operations to improve the accuracy of contour extraction.

[0018] In terms of color feature extraction, the image is converted from the RGB color space to HSV or other color spaces suitable for flame recognition. The flame region is segmented by setting a threshold or using clustering methods. The ratio of blue flame regions to non-blue flame regions is calculated, and the color distribution features and brightness gradient changes of each channel are extracted to reflect the degree of combustion and flame temperature distribution characteristics. Preferably, the flame color distribution can be quantitatively described by statistically analyzing the color histogram or color moment features of the flame region pixels.

[0019] In terms of dynamic feature extraction, the changes in flame over time are analyzed based on continuous video frame data. Changes in the flame boundary are calculated using inter-frame differencing or optical flow methods to extract dynamic features such as flame oscillation frequency, amplitude, and boundary fluctuation amplitude, reflecting the instability of combustion. In some implementations, time series analysis methods can also be used to model the changing trends of flame features over time to further improve the sensitivity to the tempering development process.

[0020] In step S1 of this invention, the flame image is a frontal and side image capable of recording the flame's height, morphological structure, and inner and outer flame distribution characteristics. The flame video continuously acquires data during the combustion process to capture the dynamic changes in the flame. These dynamic changes include flame swaying, oscillation, and unstable development processes.

[0021] The flame images and videos described in this invention are collected during the stable combustion stage, the critical stage approaching tempering, and the instant tempering occurs, so that the constructed dataset can cover the entire tempering development process.

[0022] The image preprocessing described in this invention includes the following steps:

[0023] First, the original image is cropped (ROI extraction). The flame area is extracted by pre-setting the burner position or by using an image segmentation algorithm, and irrelevant information in the background is removed.

[0024] Secondly, background removal is performed. By constructing a background model or using color space segmentation methods, such as RGB or HSV threshold segmentation, non-flame areas and ambient light interference are eliminated to improve the accuracy of flame area recognition.

[0025] Furthermore, noise filtering is performed on the image, preferably using median filtering, Gaussian filtering, or bilateral filtering methods, to remove random noise generated during image acquisition while preserving flame edge information.

[0026] Then, the images are scaled to a fixed size to meet the input requirements of the deep learning model.

[0027] In addition, when processing video data, time-series processing is performed on consecutive frame images, including frame extraction and inter-frame difference calculation, in order to extract information on the dynamic changes of flames.

[0028] Multi-view acquisition was employed, using cameras positioned in front of and to the side of the gas stove to capture frontal and side images of the flame, thus comprehensively recording the flame's height, morphology, and the distribution characteristics of the inner and outer flames. High-frame-rate cameras were used to continuously capture video of the combustion process, recording the dynamic changes in the flame, including its swaying, oscillation, and unstable development.

[0029] The operating parameters described in this invention include hydrogen doping ratio, gas flow rate or nozzle velocity, primary air conditioning status or equivalence ratio, and information on the state and time of flashback.

[0030] The gas stove backfire risk model described in this invention adopts a multi-input fusion model structure, in which flame images and video data are used to extract features through convolutional neural networks or visual Transformer networks to obtain high-dimensional image feature vectors, while the operating parameters are encoded through a fully connected network. Subsequently, the two types of features are fused at the feature layer to form a unified feature representation for backfire risk prediction.

[0031] In step S6 of this invention, the collected and calibrated dataset is divided into a training set, a validation set, and a test set for training the gas stove backfire risk model, parameter tuning, and performance evaluation. During model training, an appropriate loss function is selected for optimization based on the output format, and the model parameters are iteratively updated using a gradient descent algorithm. During training, performance changes in the training and validation sets are monitored, and overfitting is suppressed by combining early-stop strategies or regularization methods to improve the model's generalization ability. After model training is completed, the accuracy of backfire risk prediction is evaluated using metrics such as mean squared error (MSE), mean absolute error (MAE), and prediction error range. The experimental observation criteria of this invention are as follows: observing the combustion process, when the flame enters the burner, the flame detaches from the nozzle and retracts downward, or abnormal combustion sounds and temperature rises occur inside the burner, it is determined to be a backfire state; when the flame does not show obvious backfire but shows obvious contraction, increased shaking, or unstable characteristics, it is determined to be a critical state approaching backfire; other normal and stable combustion states are calibrated as stable combustion.

[0032] Image feature criteria are used to quantitatively describe the tempering development process based on the feature vectors obtained in step S3: when the flame height decreases significantly, the inner flame structure weakens or disappears, the proportion of blue flame changes abruptly, and the fluctuation amplitude of the flame boundary increases significantly, the combustion state is determined to be unstable; when the flame outline shrinks towards the nozzle, the brightness of the flame root increases, and a clear trend of retreat occurs, these are used as image feature bases for the occurrence of tempering. In some implementations, the tempering development stage can be automatically identified by setting feature thresholds or constructing feature combination indicators.

[0033] The second objective of this invention is to provide a system that uses the above-described method for predicting the hydrogen-induced backfire limit of a gas stove based on flame image features.

[0034] The second objective of this invention is achieved through the following technical measures: a system using the above-mentioned method for predicting the hydrogen-doped flashback limit of a gas stove based on flame image features, characterized in that it comprises:

[0035] The image acquisition module is used to acquire flame images and video data;

[0036] The image preprocessing module is used to preprocess flame images and video data;

[0037] The feature extraction module is used to extract multidimensional features of flames from preprocessed flame images and video data;

[0038] The tempering risk prediction module is used to output the tempering risk prediction results;

[0039] The results output module is used to display the prediction results;

[0040] The image acquisition module, image preprocessing module, feature extraction module, tempering risk prediction module, and result output module are connected in sequence.

[0041] The system described in this invention is implemented through an embedded processing unit or an industrial computing platform and integrated with a camera device to form an integrated detection device. The prediction results can be visualized and output through a display terminal or a remote monitoring platform, including the flashback risk level, safe hydrogen doping range, and operating recommendations. When the flashback risk approaches the critical value, the system issues an early warning signal.

[0042] Compared with the prior art, the present invention has the following significant effects:

[0043] (1) In practical applications, when the flame image and corresponding operating parameters of the target gas stove under natural gas or low hydrogen blending conditions are input, the model can directly predict the trend of its backfire risk change, and quickly give the maximum safe hydrogen blending ratio through the above-mentioned inversion process, realizing safety assessment without the need for high-proportion hydrogen blending experiments. Through this hydrogen blending limit inversion method, the flame image features and backfire risk model are effectively transformed into safety boundary parameters with engineering significance, realizing the transformation from "state recognition" to "safety limit prediction", and providing a key technical means for the rapid assessment of the adaptability of gas stoves to hydrogen blending.

[0044] (2) This invention is applicable to large-scale rapid screening of in-service gas stoves. Multiple gas stoves can be continuously tested and evaluated through mobile testing equipment or fixed testing system without the need for step-by-step hydrogen doping experiments. It has the advantages of low cost, high efficiency, high accuracy and reliability of prediction, and strong scalability. It is applicable to the safety assessment of terminal equipment in urban gas systems and has good engineering application value and promotion prospects. Attached Figure Description

[0045] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0046] Figure 1 This is an overall flowchart of the method of the present invention. Detailed Implementation

[0047] The present invention will now be described in detail with reference to the embodiments and accompanying drawings to help those skilled in the art better understand the inventive concept of the present invention. However, the scope of protection of the claims of the present invention is not limited to the following embodiments. For those skilled in the art, all other embodiments obtained without creative effort without departing from the inventive concept of the present invention are within the scope of protection of the present invention.

[0048] like Figure 1 As shown, the present invention provides a method for predicting the hydrogen-doped flashback limit of a gas stove based on flame image features, comprising the following steps:

[0049] S1. Select gas stove samples of different types and service years and conduct combustion tests. During the test, gradually increase the hydrogen blending ratio of natural gas, and at the same time adjust the natural gas flow and primary air supply status to make the gas stove samples run stably under different working conditions. Simultaneously record flame images and video data as well as corresponding operating parameters, and continuously monitor the changes in combustion status until backfire occurs or approaches the backfire critical state.

[0050] Gas stove samples of different types and service lives were selected and combustion tests were conducted on a controlled experimental platform. The selected gas stoves covered different structural forms, rated heat loads, and degrees of aging to reflect typical differences in actual in-service gas stoves. During the experiment, the mixing ratio of natural gas and hydrogen was adjusted through a gas blending system. While gradually increasing the hydrogen blending ratio, the gas flow rate and primary air supply were adjusted to ensure stable operation of the gas stove under different operating conditions. Changes in its combustion state were continuously monitored until flashback occurred or approached the critical flashback state. Throughout the experiment, flame images and video data, along with corresponding operating parameters, were recorded simultaneously. These parameters included the hydrogen blending ratio, gas flow rate or nozzle velocity, primary air conditioning status or equivalence ratio, and the state and time of flashback occurrence. These operating parameters were used for experimental recording and result analysis and are not required as input parameters in actual model applications.

[0051] Flame images were acquired using a multi-view approach. Cameras were positioned in front of and to the side of the gas stove to capture frontal and side images of the flame, thus comprehensively recording its height, shape, structure, and the distribution characteristics of the inner and outer flames. High-frame-rate cameras were used to continuously capture video of the combustion process, recording the dynamic changes in the flame, including its swaying, oscillation, and unstable development.

[0052] During the combustion process of the gas stove, the focus was on collecting flame images and video data during the stable combustion stage, the critical stage approaching flashback, and the instant flashback occurred, ensuring that the constructed dataset could cover the entire flashback development process. Through the above experiments and image acquisition process, a dataset containing flame images and video data and corresponding state information under various gas stove types, operating conditions, and different hydrogen doping ratios was obtained, providing a reliable data foundation for subsequent flame feature extraction, flashback determination calibration, and prediction model construction.

[0053] S2. Preprocess the collected flame images and video data to improve the accuracy and stability of subsequent feature extraction and model training.

[0054] Flame image preprocessing includes the following steps:

[0055] First, the original image is cropped (ROI extraction). The flame area is extracted by pre-setting the burner position or by using an image segmentation algorithm, and irrelevant information in the background is removed.

[0056] Secondly, background removal is performed: by constructing a background model or using color space segmentation methods, such as RGB or HSV threshold segmentation, non-flame areas and ambient light interference are eliminated to improve the accuracy of flame area recognition.

[0057] Furthermore, noise filtering is performed on the image, preferably using median filtering, Gaussian filtering, or bilateral filtering methods, to remove random noise generated during image acquisition while preserving flame edge information.

[0058] Then, the images are scaled to a fixed size to meet the input requirements of the deep learning model.

[0059] In addition, when processing video data, time-series processing is performed on consecutive frame images, including frame extraction and inter-frame difference calculation, in order to extract information on the dynamic changes of flames.

[0060] In some implementations, data augmentation processing can also be performed on the image, including rotation, mirroring, brightness adjustment, and contrast adjustment, to expand the training dataset and improve the model's generalization ability.

[0061] Through the above preprocessing steps, standardized, noise-suppressed image data with prominent flame features are obtained, providing reliable input for subsequent flame feature extraction and flashback risk modeling.

[0062] S3. Extract various features of the flame's geometric shape, color distribution, and dynamic changes from the preprocessed flame images and video data, and normalize these features to construct a unified feature vector for use as input to the subsequent flashback risk prediction model.

[0063] After image preprocessing, feature extraction is performed on the flame images to obtain multidimensional flame features that characterize the combustion state and tempering development trend. The feature extraction process combines image processing methods with statistical analysis methods to quantitatively describe the flame from multiple dimensions, including geometric morphology, color distribution, and dynamic changes.

[0064] In terms of geometric feature extraction: edge detection and contour extraction are performed on the flame region to obtain the flame outline, and the flame height, flame width, and flame shape parameters are further calculated. Simultaneously, internal flame structural features are identified, including the presence of an inner flame and variations in its height, to reflect combustion intensity and flame propagation characteristics. In some implementations, the flame region can be segmented through binarization and morphological operations to improve the accuracy of contour extraction.

[0065] In terms of color feature extraction: the image is converted from the RGB color space to HSV or other color spaces suitable for flame recognition. The flame region is segmented by setting a threshold or using clustering methods. The ratio of blue flame regions to non-blue flame regions is calculated, and the color distribution features and brightness gradient changes of each channel are extracted to reflect the degree of combustion and flame temperature distribution characteristics. Preferably, the flame color distribution can be quantitatively described by statistically analyzing the color histogram or color moment features of the flame region pixels.

[0066] In terms of dynamic feature extraction, the changes in flame over time are analyzed based on continuous video frame data. Changes in the flame boundary are calculated using inter-frame differencing or optical flow methods to extract dynamic features such as flame oscillation frequency, amplitude, and boundary fluctuation amplitude, reflecting the instability of combustion. In some implementations, time series analysis methods can also be used to model the changing trends of flame features over time to further improve the sensitivity to the tempering development process.

[0067] In the aforementioned feature extraction process, various features can be normalized and a unified feature vector can be constructed for input into the subsequent backfire risk prediction model. Through the comprehensive extraction of multi-dimensional flame features, a complete characterization of the gas stove's combustion state can be achieved, providing crucial data support for backfire determination and hydrogen doping limit prediction.

[0068] S4. Using experimental observation criteria and image feature criteria, the combustion state of flames in flame images and video data in the form of feature vectors is calibrated, and the combustion state is divided into stable combustion state, critical unstable state and tempering state.

[0069] After extracting flame features, the combustion state of the gas stove is assessed and calibrated for flashback. Flashback assessment combines experimental observation criteria with image feature criteria to perform multi-dimensional identification and quantitative evaluation of the flame state, thereby improving the accuracy and reliability of the assessment.

[0070] Regarding experimental observation criteria: by directly observing the combustion process, when the flame enters the burner, the flame leaves the nozzle and retracts downward, or abnormal combustion sounds and temperature rises occur inside the burner, it is determined to be a flashback state; when the flame does not show obvious flashback, but shows obvious contraction, increased shaking, or unstable characteristics, it is determined to be a critical state close to flashback; other normal and stable combustion states are labeled as stable combustion.

[0071] Regarding image feature criteria: Based on the flame features extracted in step S3, the tempering development process is quantitatively described: when the flame height decreases significantly, the inner flame structure weakens or disappears, the proportion of blue flame changes abruptly, and the fluctuation amplitude of the flame boundary increases significantly, it can be determined that the combustion state tends to be unstable. Furthermore, when the flame outline shrinks towards the nozzle direction, the brightness at the flame root increases, and a clear tendency to recede appears, these can be used as image feature criteria for the occurrence of tempering. In some implementations, the tempering development stage can be automatically identified by setting feature thresholds or constructing feature combination indicators (judging and setting the parameter range based on the data of the processed image).

[0072] Based on the above criteria, the collected samples are calibrated, and the combustion state is divided into stable combustion state, critically unstable state, and tempering state, or further mapped to a continuous tempering risk index for subsequent model training. During the calibration process, time series information can be combined to annotate continuous images before and after tempering to ensure that the model can learn the characteristic evolution law in the tempering development process.

[0073] The tempering determination and calibration method, which integrates multiple criteria (including flame morphology, color, and other indicators), not only improves the accuracy of sample labeling but also enhances the reliability of model training, providing a crucial foundation for subsequent tempering risk prediction and hydrogen doping limit inversion. Furthermore, in some implementations, the tempering development process can be further analyzed by incorporating combustion mechanisms (e.g., which characteristic combinations are conditions for tempering to occur, and under which conditions tempering will not occur). The relationship between flame propagation characteristics and flow characteristics can be used to interpret and correct tempering trends, thereby improving the reliability of the determination results.

[0074] S5. Using the flame image features extracted in step S3 as input, and introducing the hydrogen doping ratio as an auxiliary input parameter, a gas stove backfire risk model is constructed.

[0075] After completing flame feature extraction and backfire determination calibration, a gas stove backfire risk model is constructed to predict the risk of backfire in gas stoves. The gas stove backfire risk model is based on deep learning methods and combines multi-source input information to achieve coupled modeling between flame image features and operating condition parameters, thereby improving the model's prediction accuracy and generalization ability.

[0076] During model construction, the flame image features extracted in step S3 are used as the main input, while the hydrogen doping ratio is introduced as an auxiliary input parameter to construct the model input data. Preferably, a multi-input fusion model structure is adopted, in which image data is used to extract features through a convolutional neural network or a visual Transformer network to obtain a high-dimensional image feature vector, while the running parameters are encoded through a fully connected network. Subsequently, the two types of features are fused at the feature layer to form a unified feature representation for backfire risk prediction.

[0077] In specific implementations, deep learning models, including Convolutional Neural Networks (CNNs), Residual Networks (ResNets), or Vision Transformers (ViTs), can be used to learn features from flame images. Spatial distribution features and local details of the flames can be extracted through multi-layer convolution or self-attention mechanisms. Simultaneously, techniques such as batch normalization, activation functions (e.g., ReLU), and Dropout can be introduced to improve the stability of model training and prevent overfitting.

[0078] Regarding the model output, it is set to regression output, which outputs a continuous flashback risk index to describe the degree to which the combustion state is close to flashback.

[0079] The flashback risk model constructed using the above method can realize the mapping relationship from flame images and operating parameters to flashback risk, providing a foundation for subsequent hydrogen doping limit prediction.

[0080] S6. Train and evaluate the accuracy of the gas stove backfire risk model;

[0081] First, the collected and labeled dataset is divided into training, validation, and test sets for model training, parameter tuning, and performance evaluation. During model training, an appropriate loss function is selected for optimization based on the output format, and the model parameters are iteratively updated using the gradient descent algorithm.

[0082] During training, performance changes on the training and validation sets are monitored, and overfitting is suppressed by combining early stopping strategies or regularization methods to improve the model's generalization ability. After model training is completed, the accuracy of tempering risk prediction is evaluated using metrics such as mean squared error (MSE), mean absolute error (MAE), and prediction error range.

[0083] Furthermore, by comparing the model's prediction results with the experimentally measured tempering conditions, the prediction error of the hydrogen doping limit was calculated to verify the effectiveness and reliability of the model under actual working conditions.

[0084] S7. Based on the output of the gas stove backfire risk model, perform an inverse calculation of the hydrogen-blended backfire limit of the gas stove: input the hydrogen blending ratio as a variable into the gas stove backfire risk model, and gradually change the hydrogen blending ratio to calculate the corresponding backfire risk value, obtain the functional relationship curve of backfire risk with hydrogen blending ratio, and determine the critical point of the gas stove transitioning from stable combustion to backfire by setting a backfire judgment threshold (based on the backfire judgment condition in step S4, this threshold is set manually). The hydrogen blending ratio corresponding to this critical point is the maximum safe hydrogen blending ratio of the gas stove, thus obtaining the hydrogen-blended backfire limit of the gas stove.

[0085] After obtaining the trained backfire risk prediction model, the hydrogen-blended backfire limit of the gas stove is inverted and calculated based on the model output. Specifically, by inputting the hydrogen blending ratio as a variable into the model, the hydrogen blending ratio is gradually changed, and the corresponding backfire risk value is calculated, thereby obtaining the functional relationship curve of backfire risk changing with the hydrogen blending ratio.

[0086] Based on this functional relationship, by setting a flashback detection threshold, the critical point at which the gas stove transitions from stable combustion to flashback is determined, and the corresponding hydrogen blending ratio is the maximum safe hydrogen blending ratio for that gas stove. In some implementations, interpolation methods or numerical solutions can be used to fit the flashback risk function, thereby improving the accuracy and stability of the critical hydrogen blending ratio calculation.

[0087] During the inversion process, the flashback determination results from step S4 can be used to correct the model output, making the hydrogen doping limit obtained from the inversion more consistent with actual combustion characteristics and combustion mechanisms. Furthermore, inversion analyses can be performed separately for different types of gas stoves and different operating conditions to obtain hydrogen doping limit ranges applicable to different application scenarios.

[0088] Furthermore, in practical applications, when the flame image and corresponding operating parameters of the target gas stove under natural gas or low hydrogen blending conditions are input, the model can directly predict the trend of its backfire risk and quickly give the maximum safe hydrogen blending ratio through the above inversion process, realizing safety assessment without the need for high hydrogen blending experiment conditions.

[0089] This hydrogen-doped limit inversion method effectively transforms flame image features and backfire risk models into engineering-significant safety boundary parameters, realizing the transformation from "state recognition" to "safety limit prediction" and providing a key technical means for the rapid assessment of the hydrogen-doped adaptability of gas stoves.

[0090] In practical engineering applications, the backfire risk prediction model trained as described above can be used to quickly assess the adaptability of gas stoves to hydrogen blending. Without requiring high-proportion hydrogen blending experiments, only flame images of the target gas stove under natural gas or low hydrogen blending conditions need to be acquired, and the combustion state can be analyzed and predicted using the model.

[0091] In practical applications, the real-time flame image of the gas stove during operation is first acquired through the image acquisition module. The image is then preprocessed and its features are extracted. The obtained features are input into the backfire risk prediction model to obtain the backfire risk value or risk level under the current operating conditions. Based on this, combined with the hydrogen doping limit inversion method in step S7, the backfire risk variation trend of the target gas stove under different hydrogen doping ratios is further calculated, thereby determining its maximum safe hydrogen doping ratio.

[0092] The above method can be integrated into an integrated gas stove hydrogen blending adaptability assessment system, which includes:

[0093] The image acquisition module is used to acquire flame images and video data;

[0094] The image preprocessing module is used to preprocess the completed image and video data of the flame;

[0095] The feature extraction module is used to extract multidimensional features of flames from preprocessed flame images and video data;

[0096] The tempering risk prediction module is used to output the tempering risk prediction results;

[0097] The results output module is used to display the prediction results;

[0098] The image acquisition and processing module, image preprocessing module, feature extraction module, tempering risk prediction module, and result output module are connected sequentially. The modules interact with each other via data interfaces.

[0099] In some implementations, the system can be implemented through an embedded processing unit or industrial computing platform, and integrated with camera equipment to form an integrated detection device, enabling real-time acquisition and online analysis of flame images. Prediction results can be visualized and output through a display terminal or remote monitoring platform, including information such as the flashback risk level, safe hydrogen doping range, and operational recommendations. Simultaneously, when the flashback risk approaches a critical value, the system can issue a warning signal, prompting the user to adjust the gas flow rate or air supply conditions to prevent flashback.

Claims

1. A method for predicting hydrogen blending tempering limit of gas stove based on flame image features, characterized in that Includes the following steps: S1. Select gas stove samples of different types and service years for combustion testing. During the test, gradually increase the hydrogen blending ratio of natural gas, while adjusting the natural gas flow rate and primary air supply status to ensure stable operation of the gas stove samples under different working conditions. Simultaneously record flame images and video data, as well as corresponding operating parameters, and continuously monitor changes in combustion status until flashback occurs or approaches the flashback critical state. Operating parameters include hydrogen blending ratio, gas flow rate or nozzle velocity, primary air conditioning status or equivalence ratio, and flashback occurrence status and time information. S2. Preprocess the acquired flame images and video data; S3. Extract various features of the flame's geometric shape, color distribution, and dynamic changes from the preprocessed flame image and video data, and normalize these features to construct a unified feature vector. S4. Using experimental observation criteria and image feature criteria, the combustion state of flames in flame images and video data in the form of feature vectors is calibrated and divided into stable combustion state, critical unstable state and tempering state. S5. Using the feature vector obtained in step S3 as input, and introducing the hydrogen doping ratio as an auxiliary input parameter, construct a gas stove backfire risk model. S6. Train and evaluate the accuracy of the gas stove backfire risk model; S7. Based on the output of the gas stove backfire risk model, the hydrogen-blended backfire limit of the gas stove is calculated by inversion: the hydrogen blending ratio is input as a variable into the gas stove backfire risk model, and the hydrogen blending ratio is gradually changed to calculate the corresponding backfire risk value. The function relationship curve of backfire risk with hydrogen blending ratio is obtained, and the critical point of gas stove transitioning from stable combustion to backfire is determined by setting a backfire judgment threshold. The hydrogen blending ratio corresponding to the critical point is the maximum safe hydrogen blending ratio of the gas stove, that is, the hydrogen-blended backfire limit of the gas stove is obtained.

2. The flame image feature-based hydrogen blending flare limit prediction method of claim 1, wherein: In step S3, Geometric feature extraction: By performing edge detection and contour extraction on the flame area, the flame outline is obtained, and the flame height, flame width and flame shape parameters are further calculated, while the internal structural features of the flame are identified. Extracting color distribution features: Convert the image from the RGB color space to a suitable color space, perform color segmentation on the flame area, calculate the ratio of blue flame area to non-blue flame area, and extract the color distribution features and brightness gradient changes of each channel. Extracting dynamic change features: Based on continuous video frame data, the changes of flame over time are analyzed, the changes of flame boundary are calculated, and dynamic features such as flame oscillation frequency, amplitude, and boundary fluctuation amplitude are extracted.

3. The method for predicting hydrogen blending tempering limit of gas stove based on flame image features according to claim 2, characterized in that: In step S1, the flame image is a frontal and side image that can record the height, shape and structure of the flame and the distribution characteristics of the inner and outer flames; the flame video continuously acquires the combustion process to capture the dynamic changes of the flame, including flame swaying, oscillation and unstable development process.

4. The flame image feature-based hydrogen blending flare limit prediction method according to claim 3, characterized in that: The operating parameters include hydrogen doping ratio, gas flow rate or nozzle velocity, primary air conditioning status or equivalence ratio, and information on the state and time of flashback.

5. The method of claim 4, wherein the method further comprises: determining the hydrogen blending flare limit based on the flame image features. The gas stove backfire risk model adopts a multi-input fusion model structure, in which flame images and video data are used to extract features through convolutional neural networks or visual Transformer networks to obtain high-dimensional image feature vectors, while operating parameters are encoded through fully connected networks. Then, the high-dimensional image feature vectors and the encoded operating parameters are fused at the feature layer to form a unified feature representation for backfire risk prediction.

6. The method for predicting hydrogen blending tempering limit of gas stove based on flame image features according to claim 5, characterized in that: In step S6, the collected and calibrated dataset is divided into a training set, a validation set, and a test set for training the gas stove backfire risk model, parameter tuning, and performance evaluation. During model training, an appropriate loss function is selected for optimization based on the output format, and the model parameters are iteratively updated using the gradient descent algorithm. During training, the performance changes of the training and validation sets are monitored, and overfitting is suppressed by combining an early stop strategy or regularization method. After the model training is completed, the accuracy of backfire risk prediction is evaluated using mean squared error, mean absolute error, and prediction error range indicators.

7. The method of claim 6, wherein the method further comprises: determining the hydrogen blending retrograde limit of the gas stove based on the flame image features. In step S2, firstly, the original image is cropped to extract the flame region, irrelevant information in the background is removed, then background removal is performed, and finally, noise filtering and scale normalization are performed on the image.

8. The method of claim 7, wherein the method further comprises: determining the hydrogen blending retrograde limit of the gas stove based on the flame image features. The experimental observation criteria are as follows: Observing the combustion process, when the flame enters the burner, detaches from the nozzle and retracts downwards, or abnormal combustion sounds and temperature rises occur inside the burner, it is determined to be a flashback state; when the flame does not show obvious flashback but shows obvious contraction, increased shaking, or unstable characteristics, it is determined to be a critical state approaching flashback; other normal and stable combustion states are labeled as stable combustion. The image feature criteria are based on the feature vector obtained in step S3 to quantitatively describe the flashback development process: when the flame height decreases significantly, the inner flame structure weakens or disappears, the proportion of blue flame changes abruptly, and the amplitude of flame boundary fluctuations increases significantly, the combustion state is determined to be unstable; when the flame outline contracts towards the nozzle, the brightness of the flame root increases, and there is an obvious retraction trend, flashback occurs.

9. A system for predicting the hydrogen blending backfire limit of a gas stove based on flame image features using the method of claim 1, characterized in that include: The image acquisition module is used to acquire flame images and video data; The image preprocessing module is used to preprocess flame images and video data; The feature extraction module is used to extract multidimensional features of flames from preprocessed flame images and video data; The tempering risk prediction module is used to output the tempering risk prediction results; The results output module is used to display the prediction results; The image acquisition module, image preprocessing module, feature extraction module, tempering risk prediction module, and result output module are connected sequentially.

10. The system of claim 9, wherein: The system is implemented through an embedded processing unit or industrial computing platform and integrated with camera equipment to form an integrated detection device. The prediction results are visualized and output through a display terminal or remote monitoring platform, including the flashback risk level, safe hydrogen doping range and operation recommendations. When the flashback risk approaches the critical value, the system issues an early warning signal.