Hydrogen combustion tempering monitoring system and method based on multi-modal data fusion

Abstract

The invention discloses a hydrogen combustion tempering monitoring system and method based on multi-modal data fusion, and relates to the technical field of combustion monitoring and diagnosis. The data acquisition module is used for acquiring physical parameters and image data in the hydrogen combustion process; the data preprocessing module preprocesses the physical parameters and the image data; the feature extraction module extracts features from the preprocessed physical parameters and image data; the multi-modal feature fusion module splices the extracted features to form a joint feature vector and performs dimension reduction; the tempering state judgment module is used for constructing a tempering state judgment model, outputting a tempering state occurrence probability value and judging a combustion state according to a set safety threshold value; and the sound-light alarm module selects to give an alarm based on the result. Multi-modal features are extracted based on physical parameters and image data and fused to form a joint feature vector, a tempering state discrimination model is constructed according to a random forest algorithm, a sound-light alarm module can be triggered in real time in combination with a set safety threshold, and real-time and accurate discrimination of a tempering phenomenon is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of combustion monitoring and diagnosis, and particularly to a hydrogen combustion flashback monitoring system and method based on multi-modal data fusion. Background Art

[0002] With the increasing concern about global climate change, low-carbon development has become an important and urgent requirement in the energy field. Hydrogen fuel is considered to be one of the cleanest and most prominent carbon-neutral fuels in gas turbine operation due to its characteristics such as pollution-free, high efficiency, and recyclability. Therefore, low-carbon fuels such as hydrogen-containing, hydrogen-rich, or pure hydrogen have currently received extensive attention.

[0003] Due to the active chemical properties and fast diffusion rate of hydrogen, the hydrogen combustion technology faces huge challenges, and one of the most prominent problems is flashback. The flashback phenomenon refers to the movement of the flame from the reaction area of the combustion chamber upstream to the gas premixing area, which is the result of the coupling of factors such as combustion speed, local flow field conditions, and combustion instability. Simply put, flashback occurs when the flame speed is higher than the gas flow speed. However, in the premixing section of the burner, the flame propagation speed is slow, and flashback can still occur through the boundary layer back into the combustion chamber. Flashback not only affects the combustion efficiency but may also cause serious safety problems. Therefore, monitoring the flashback phenomenon is of great significance for improving combustion safety and optimizing the combustion process.

[0004] Traditional flashback monitoring methods generally use temperature sensors or pressure sensors alone, which have certain limitations. They can only measure the local temperature or pressure at their installation positions and cannot reflect the overall state of the combustion chamber. If flashback occurs in the sensor blind area, it will lead to monitoring failure. In addition, the measurement results of the sensors are easily affected by environmental factors, which may cause measurement errors and reduce the monitoring accuracy.

[0005] High-frequency planar laser-induced fluorescence technology (PLIF) is a cutting-edge optical measurement technology in the field of combustion diagnosis, with the advantages of high spatio-temporal resolution and high sensitivity. At the initial stage of the flashback phenomenon, the flame characteristics will change, accompanied by changes in the flame structure, flame position, and dynamic characteristics, which provides a new opportunity for flashback monitoring.

[0006] In summary, by monitoring the OH free radicals in the hydrogen combustion products through the PLIF technique, the flame characteristics can be directly captured, providing more comprehensive combustion state information. The physical parameters (such as temperature and pressure) and OH-PLIF images respectively reflect different aspects of the combustion process. The temperature and pressure data can reflect the macroscopic state of combustion, and the image data can reflect the microscopic structure and local characteristics of combustion. Therefore, considering the combination of traditional monitoring means and high-frequency optical diagnostic techniques helps to provide high-frequency and multi-dimensional information, which is of great significance for achieving accurate monitoring of combustion chamber flashback. Summary of the Invention

[0007] To solve the deficiencies in the background technology, the present invention provides a hydrogen combustion flashback monitoring system and method based on multi-modal data fusion. It extracts multi-modal features from physical parameters and image data and fuses them to form a joint feature vector. A flashback state discrimination model is constructed according to the random forest algorithm, and the acoustic and optical alarm module can be triggered in real time in combination with a set safety threshold to achieve real-time and accurate discrimination of the flashback phenomenon.

[0008] To achieve the above object, the present invention adopts the following technical solutions:

[0009] A hydrogen combustion flashback monitoring system based on multi-modal data fusion includes a data acquisition module, a data preprocessing module, a feature extraction module, a multi-modal feature fusion module, a flashback state discrimination module, and an acoustic and optical alarm module connected in sequence, where:

[0010] The data acquisition module is composed of a sensor system and an OH-PLIF system, which respectively obtain the physical parameters and image data of the hydrogen combustion process. The physical parameters include temperature and pressure data, and the image data is the OH-PLIF image of the OH free radical distribution;

[0011] The data preprocessing module preprocesses the obtained physical parameters and image data respectively. For the physical parameters, discrete wavelet transform is used to remove noise and perform normalization processing. For the image data, median filtering is used for noise reduction, and then the background region and flame region of the image target position are segmented by the maximum inter-class variance method;

[0012] The feature extraction module extracts features from the preprocessed physical parameters and image data respectively. For the physical parameters, the mean, variance, slope, and peak value of the temperature and pressure data are calculated as the extracted features. For the image data, the shape feature, color feature, and texture feature of the flame are obtained according to the flame region as the extracted features;

[0013] The multi-modal feature fusion module consists of two parts: feature splicing and feature dimensionality reduction. In the feature splicing part, the extracted features of physical parameters and image data are spliced to form a joint feature vector. In the feature dimensionality reduction part, the principal component analysis method is used to reduce the dimensionality of the joint feature vector;

[0014] The tempering state discrimination module uses the random forest algorithm to construct a tempering state discrimination model. Based on the dimensionality-reduced joint feature vector, the data of normal combustion and tempering states are labeled and used as a data set for input. The model is trained and tested, and the probability value of the occurrence of the tempering state is output. The real-time collected joint feature vector is input into the tempering state discrimination model, and it is determined whether it belongs to the normal combustion or tempering state according to the set safety threshold;

[0015] The acoustic-optic alarm module is based on the output result of the tempering state discrimination module. If it is determined to be in the tempering state, an alarm is issued.

[0016] Further, the data acquisition frequencies of the sensor system and the OH-PLIF system are synchronized to ensure the time consistency of the physical parameters and image data.

[0017] Further, a set of temperature sensors and pressure sensors are installed at the inner wall and the outlet of the fuel nozzle of the sensor system to monitor and collect the temperature and pressure data during the hydrogen combustion process in real time.

[0018] Further, the OH-PLIF system includes a laser, an optical system, a high-speed camera, and a timing control system. The timing controller ensures the synchronization of the laser and the high-speed camera through a digital signal generator, so that the laser beam emitted by the laser is vertically incident on the combustion chamber through the optical system, and the OH-PLIF image of the OH radical distribution during the hydrogen combustion process is captured in real time.

[0019] A hydrogen combustion tempering monitoring method based on multi-modal data fusion includes the following steps:

[0020] Step 1: Data acquisition

[0021] When igniting, the sensor system and the OH-PLIF system are started simultaneously to collect physical parameters and image data;

[0022] Step 2: Data preprocessing

[0023] For the preprocessing of physical parameters, first, the discrete wavelet transform is selected to decompose the data into approximation coefficients and detail coefficients. The detail coefficients of each layer are threshold processed to remove the noise components. The inverse wavelet transform is performed using the processed detail coefficients and approximation coefficients to reconstruct the denoised temperature and pressure data. Then, the denoised temperature and pressure data are normalized respectively;

[0024] For the preprocessing of image data, first, the median filtering method is used for noise reduction. A fixed window of size 3×3 is selected to traverse each pixel of the image. The pixel values within the window are sorted, and the median value is taken to replace the current pixel value. This operation is repeated until all pixels are processed. Then, the Otsu method is used for image segmentation on the denoised image. The number of pixels n at each gray level of the image is counted. i , and the probability at each gray level is calculated. In the formula, N is the total number of pixels at each gray level. Let the gray level k be the threshold to divide the image into the background region and the flame region. They are respectively calculated according to and to calculate the mean probability of the background region and the flame region. In the formula, L is the range of gray levels of the image. Then, they are respectively calculated according to and to calculate the mean of the background region and the flame region. Then, according to σ 2 (k) = ω0(k)·ω1(k)·(μ0(k) - μ1(k)) 2 calculate the inter-class variance. Traverse all possible thresholds k to find the threshold k that maximizes the inter-class variance. * , and use the threshold k * to binarize the image. The pixels with gray levels less than or equal to the threshold k * are set as the background region, and the pixels with gray levels greater than the threshold k * are set as the flame region;

[0025] Step 3: Feature extraction

[0026] For the feature extraction of physical parameters, calculate the mean, variance, slope, and peak value of the temperature and pressure data respectively as the extracted features;

[0027] For the feature extraction of image data, according to the segmented flame region, extract the shape features, color features, and texture features of the flame during the hydrogen combustion process as the extracted features, where:

[0028] The shape features include the flame area, flame perimeter, and flame circularity. The formula for calculating the flame area is In the formula, I(i, j) represents the value of any pixel (i, j) inside the flame region, row 1 and column 1 respectively represent the number of rows and columns inside the flame region. The formula for calculating the flame perimeter is In the formula, B(i, j) represents the value of any pixel (i, j) on the boundary of the flame region, row 2 and column 2 respectively represent the number of rows and columns on the boundary of the flame region. The formula for calculating the flame circularity is The value range is [0, 1];

[0029] The described color features include the mean gray value and the gray variance. The calculation formula for the mean gray value is The calculation formula for the gray variance is In the formula, I(i) represents the gray value of the i-th pixel, and K represents the total number of pixels in the flame area of the image;

[0030] The described texture features are extracted through a gray-level co-occurrence matrix to statistically describe the gray value distribution of pixel pairs in the image. The directions of 0°, 45°, 90°, and 135° pixel pairs are selected, the statistical distance is 1 pixel, the flame area is traversed, the gray value distribution of pixel pairs that meet the direction and distance conditions is statistically calculated, a gray-level co-occurrence matrix is constructed and the matrix is normalized so that each element represents a probability. Based on the normalized gray-level co-occurrence matrix calculate the texture features, where P(i,j) represents the number of occurrences of pixel pairs with gray values of i and j;

[0031] Step Four: Multimodal Feature Fusion

[0032] For the feature splicing part, represent the feature vector of the physical parameters as F temp-pressure =[f1,f2,...,f M , represent the feature vector of the image data as F OH-PLIF =[g1,g2,...,g K , then the joint feature vector is F joint =[f1,f2,...,f M ,g1,g2,...,g K . Standardize each feature in the joint feature vector so that the mean is 0 and the variance is 1. Then, calculate the covariance matrix and perform eigenvalue decomposition to obtain eigenvalues and eigenvectors. Sort the eigenvalues from largest to smallest, and the corresponding eigenvectors are the principal components. Select the principal components with a cumulative contribution rate of 95% according to the eigenvalue magnitudes, and project the standardized feature matrix onto the selected principal components for dimensionality reduction;

[0033] Step Five: Tempering State Discrimination and Acousto-optical Alarm

[0034] The tempering state discrimination module uses the random forest algorithm to construct a tempering state discrimination model. Based on the result of multi-modal feature fusion, the combustion state of each sample is labeled, where 1 represents the tempering state and 0 represents normal combustion. The data set is divided into a training set and a test set. The parameters of the random forest are set, and the training set is used to train the tempering state discrimination model, and the test set is used to evaluate the tempering state discrimination model. The trained tempering state discrimination model is loaded into the system, and the jointly acquired feature vector is input in real time, and the probability value of the occurrence of the tempering state is output. According to the set safety threshold, it is determined whether it is normal combustion or tempering state. If the output probability value is greater than the safety threshold, it is determined as the tempering state, and at the same time, the sound and light alarm module gives an alarm, otherwise it is determined as normal combustion.

[0035] Further, calculating the texture feature based on the normalized gray-level co-occurrence matrix in step three includes:

[0036] Contrast, the calculation formula is

[0037] Energy, the calculation formula is

[0038] Homogeneity, the calculation formula is

[0039] Correlation, the calculation formula is In the formula, μ i and μ j respectively represent the average gray values of pixel pairs i and j, and σ i and σ j respectively represent the standard deviations of the gray values of pixel pairs i and j.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention uses a sensor system to collect the temperature and pressure in the combustion chamber as physical parameters, and at the same time captures the OH radical distribution image during the hydrogen combustion process through the OH-PLIF system as image data. Multiple features of the two physical parameters and image data are extracted respectively, and the multi-modal features are fused to form a jointly acquired feature vector. The tempering state discrimination model constructed according to the random forest algorithm combines the set safety threshold to trigger the sound and light alarm module in real time, realizing real-time and accurate discrimination of the tempering phenomenon, improving the accuracy of hydrogen combustion tempering monitoring, and providing a reference for adjusting combustion parameters. Description of the Drawings

[0041] Figure 1 is a schematic diagram of the hydrogen combustion tempering monitoring system of the present invention;

[0042] Figure 2 is a schematic diagram of the data acquisition module in the present invention. Detailed Embodiments

[0043] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0044] A hydrogen combustion flashback monitoring system based on multi-modal data fusion, which combines Figure 1 As shown in the figure, it includes a data acquisition module, a data preprocessing module, a feature extraction module, a multi-modal feature fusion module, a flashback state discrimination module, and an acoustic-optic alarm module that are connected in sequence.

[0045] The data acquisition module is composed of a sensor system and an OH-PLIF system. As shown in Figure 2 the figure, physical parameters and image data of the hydrogen combustion process are obtained by the two respectively. The sensor system installs a set of temperature sensors and pressure sensors at the inner wall and the outlet of the fuel nozzle to monitor and collect the temperature and pressure data during the hydrogen combustion process in real time as physical parameters; the OH-PLIF system includes a laser, an optical system, a high-speed camera, and a timing control system. The timing controller ensures the synchronization of the laser and the high-speed camera through a digital signal generator, so that the laser beam emitted by the laser is vertically incident on the combustion chamber through the optical system, and the OH-PLIF image of the OH radical distribution during the hydrogen combustion process is captured in real time as image data. It should be noted that the data acquisition frequencies of the sensor system and the OH-PLIF system should also be synchronized to ensure the time consistency of the physical parameters and the image data.

[0046] The data preprocessing module sets different preprocessing methods for the acquired physical parameters and image data, and stores the preprocessed physical parameters and image data. For the physical parameters, discrete wavelet transform is used to remove noise and perform normalization processing, and the temperature and pressure data after removing noise are normalized to the range of [0,1] to eliminate the dimension gap; for the image data, median filtering is used for noise reduction and combined with the maximum inter-class variance method, and the image data is segmented according to the distribution characteristics of the OH radicals to segment the background area and the flame area of the image target position.

[0047] The feature extraction module sets different extraction methods for the preprocessed physical parameters and image data. For the physical parameters, the mean, variance, slope, and peak value of the temperature and pressure data are calculated respectively as extraction features; for the image data, the shape feature, color feature, and texture feature of the flame during the hydrogen combustion process are obtained according to the segmented flame area as extraction features.

[0048] The multi-modal feature fusion module consists of two parts: feature splicing and feature dimensionality reduction, aiming to obtain richer multi-modal features and provide input data for the hydrogen combustion flashback state discrimination module. The feature splicing part splices the extracted features of physical parameters and image data to form a joint feature vector; the feature dimensionality reduction part uses the principal component analysis method to reduce the dimensionality of the joint feature vector to reduce redundant information and improve the computational efficiency of the model.

[0049] The flashback state discrimination module constructs a flashback state discrimination model using the random forest algorithm. Based on the dimensionality-reduced joint feature vector, the data of normal combustion and flashback states are labeled and used as a data set for input. The model is trained and tested, and the probability value of the occurrence of the flashback state is output. The real-time collected joint feature vector is input into the flashback state discrimination model, and it is determined whether it belongs to normal combustion or flashback state according to the set safety threshold.

[0050] The acoustic-optic alarm module generates a high-level signal based on the output result of the flashback state discrimination module. If it is determined to be in the flashback state, the signal is transmitted to the acoustic-optic alarm through a wired method. After detecting the high-level signal, the acoustic-optic alarm gives an alarm.

[0051] It should be noted that in the system of the present invention, the temperature and pressure data and OH-PLIF images collected by the data acquisition module are all transmitted to the computer through transmission lines. A common cable is used to connect the flashback state discrimination module and the acoustic-optic alarm module, and the data transmission between the remaining modules is completed within the computer.

[0052] As Figures 1 to 2 shown, a hydrogen combustion flashback monitoring method based on multi-modal data fusion includes the following steps:

[0053] Step 1: Data acquisition

[0054] A set of temperature sensors and pressure sensors are installed at the inner wall of the fuel nozzle and the outlet of the fuel nozzle respectively. The optical system is adjusted so that the laser beam emitted by the laser is vertically incident into the combustion chamber. When igniting, the sensor system and the high-speed camera are started simultaneously to collect the physical parameters and image data during the hydrogen combustion process, and they are stored in the computer.

[0055] Step 2: Data preprocessing

[0056] For the preprocessing of physical parameters, for the temperature and pressure data, wavelet transform is first used to remove noise. The discrete wavelet transform is selected to decompose the data into approximation coefficients and detail coefficients. The detail coefficients of each layer are processed by thresholding to remove the noise components. The processed detail coefficients and approximation coefficients are used for inverse wavelet transform to reconstruct the denoised temperature and pressure data. Then, the denoised temperature and pressure data are respectively processed according to Normalize, where X is the original temperature or pressure data, and X max is the maximum value in the temperature or pressure data, and X min is the minimum value in the temperature or pressure data, and X norm is the normalized temperature or pressure data;

[0057] For the preprocessing of image data, first, use the median filtering method for noise reduction. Select a fixed window of size 3×3 to traverse each pixel of the image, sort the pixel values within the window, take the median to replace the current pixel value, and repeat the operation until all pixels are processed. Then, use the Otsu method for image segmentation on the denoised image. Count the number of pixels n i at each gray level of the image, and calculate the probability of each gray level. Where N is the total number of pixels at each gray level. Let the gray level k be the threshold to divide the image into the background region and the foreground region (flame region), and calculate the probability means of the background region and the foreground region respectively according to and . Where L is the range of gray levels of the image. Then calculate the means of the background region and the foreground region respectively according to and . Then calculate the between-class variance according to σ 2 (k) = ω0(k)·ω1(k)·(μ0(k) - μ1(k)) 2 . Traverse all possible thresholds k to find the threshold k that maximizes the between-class variance * , and use the threshold k * to binarize the image. Pixels with gray levels less than or equal to the threshold k * are set as the background region, and pixels with gray levels greater than the threshold k * are set as the foreground region.

[0058] Step 3: Feature extraction

[0059] For the feature extraction of physical parameters, calculate the mean, variance, slope, and peak value of the temperature and pressure data respectively as the extracted features. Among them:

[0060] The formula for the mean is reflecting the central tendency of the data. The formula for the variance is reflecting the volatility of the data. The formula for the slope is reflecting the rising or falling speed of the data. The formula for the peak value is P = max(x i ), reflecting the extreme situation of the data. Where x i represents the i-th data point, and M represents the total number of data points;

[0061] For the feature extraction of image data, the shape features, color features, and texture features of the flame during the hydrogen combustion process are extracted based on the segmented flame region as the extraction features, where:

[0062] The shape features mainly include the flame area, flame perimeter, and flame circularity. The formula for calculating the flame area is which refers to the total number of pixels in the flame region of the image. In the formula, I(i,j) represents the value of any pixel (i,j) inside the flame region, row 1 and column 1 represent the number of rows and columns inside the flame region respectively. The formula for calculating the flame perimeter is which refers to the boundary length of the flame region in the image. In the formula, B(i,j) represents the value of any pixel (i,j) on the boundary of the flame region, row 2 and column 2 represent the number of rows and columns of the flame region boundary respectively. The formula for calculating the flame circularity is which refers to an index indicating the degree of closeness of the flame region in the image to a circle, with a value range of [0,1]. The closer the value is to 1, the closer it is to a circle;

[0063] The color features mainly include the gray mean and gray variance, and statistics will be extracted from the color information of the image respectively. The formula for calculating the gray mean is which refers to the average value of the color intensities of all pixels in the image. The formula for calculating the gray variance is which refers to the variance of the color intensities of all pixels in the flame region and reflects the degree of color fluctuation. In the formula, I(i) represents the gray value of the i-th pixel, and K represents the total number of pixels in the flame region of the image;

[0064] The texture features are mainly described by extracting the gray value distribution of pixel pairs in the statistical image through the gray-level co-occurrence matrix. Select the directions of 0°, 45°, 90°, and 135° pixel pairs, with a statistical distance of 1 pixel. Traverse the flame region, count the gray value distribution of pixel pairs that meet the direction and distance conditions, construct the gray-level co-occurrence matrix, and through normalize the matrix so that each element represents a probability. In the formula, P(i,j) represents the number of occurrences of pixel pairs with gray values of i and j. Calculate the texture features based on the normalized gray-level co-occurrence matrix, including:

[0065] Contrast, the formula for which is reflecting the degree of local change in the image;

[0066] Energy, the formula for which is reflecting the uniformity of the image;

[0067] Homogeneity, the formula for which is reflecting the local smoothness of the image;

[0068] Correlation, calculated by the formula reflects the linear correlation of the image. In the formula, μ i and μ j respectively represent the average gray values of pixel pairs i and j, and σ i and σ j respectively represent the standard deviations of the gray values of pixel pairs i and j;

[0069] All the extracted features of the above physical parameters and image data are stored as feature vectors for subsequent analysis.

[0070] Step Four: Multimodal Feature Fusion

[0071] For the feature splicing part, it means connecting the feature vectors from different data sources in a certain order to form a joint feature vector of a higher dimension. Denote the feature vector of the physical parameters as F temp-pressure = [f1, f2,..., f M , and the feature vector of the image data as F OH-PLIF = [g1, g2,..., g K , then the joint feature vector is F joint = [f1, f2,..., f M , g1, g2,..., g K . Based on the feature splicing result, according to standardize each feature in the joint feature vector so that the mean is 0 and the variance is 1. Then, calculate the covariance matrix through , where Z represents the standardized feature matrix. Perform eigenvalue decomposition on the covariance matrix Σ, Zυ i = λ i υ i , to obtain the eigenvalues λ i and the eigenvectors υ i . Sort the eigenvalues λ i from largest to smallest, and the corresponding eigenvectors υ i are the principal components. Select the principal components with a cumulative contribution rate of 95% according to the eigenvalue magnitudes, and project the standardized feature matrix Z onto the selected principal components for dimensionality reduction.

[0072] Step Five: Tempering State Discrimination and Acousto - optic Alarm

[0073] The tempering state discrimination module mainly uses the random forest algorithm to construct a tempering state discrimination model. Based on the results of the above-mentioned multi-modal feature fusion, each sample is labeled with a combustion state, where 1 represents the tempering state and 0 represents normal combustion. The data set is divided into a training set and a test set. Set the parameters of the random forest, including the number of trees, the maximum depth, etc. Use the training set to train the tempering state discrimination model. The tempering state discrimination model will construct multiple decision trees and train them with randomly selected features and samples. After the tempering state discrimination model is constructed, use the test set to evaluate the model performance. To perform real-time discrimination of the tempering state, load the trained tempering state discrimination model into the system, input the jointly collected feature vectors in real time into the tempering state discrimination model, and the tempering state discrimination model will output the probability value of the occurrence of the tempering state. Determine whether it belongs to normal combustion or the tempering state according to the set safety threshold. If the output probability value is greater than the safety threshold, it is determined to be the tempering state; otherwise, it is determined to be normal combustion. If it is determined to be the tempering state, the controller of the sound and light alarm module outputs a high-level signal to trigger the sound and light alarm, and the sound and light alarm emits a high-frequency beeping sound and warning lights.

[0074] The present invention innovatively combines physical parameters with image data, overcomes the limitations of a single data source, significantly improves the accuracy and robustness of tempering state discrimination, is applicable to real-time monitoring and early warning in complex combustion environments, and has important engineering application value.

[0075] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, in any aspect, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, it is intended to embrace all changes within the meaning and scope of the equivalent conditions of the claims in the present invention. Any reference signs in the claims should not be regarded as limiting the claims involved.

[0076] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A hydrogen combustion flashback monitoring system based on multi-modal data fusion, characterized in that: It includes a data acquisition module, a data preprocessing module, a feature extraction module, a multi-modal feature fusion module, a backfire state discrimination module, and an acoustic-optic alarm module connected in sequence, where: The data acquisition module consists of a sensor system and an OH-PLIF system, which respectively obtain the physical parameters and image data of the hydrogen combustion process. The physical parameters include temperature and pressure data, and the image data is the OH-PLIF image of the OH radical distribution; The data preprocessing module preprocesses the obtained physical parameters and image data respectively. For the physical parameters, discrete wavelet transform is used to remove noise and perform normalization processing. For the image data, median filtering is used for noise reduction, and then the background region and the flame region of the image target position are segmented by the maximum inter-class variance method; The feature extraction module extracts features from the preprocessed physical parameters and image data respectively. For the physical parameters, the mean, variance, slope, and peak value of the temperature and pressure data are calculated as the extracted features. For the image data, the shape feature, color feature, and texture feature of the flame are obtained based on the flame region as the extracted features; The multi-modal feature fusion module consists of two parts: feature splicing and feature dimensionality reduction. The feature splicing part splices the extracted features of the physical parameters and image data to form a joint feature vector. The feature dimensionality reduction part uses the principal component analysis method to reduce the dimensionality of the joint feature vector; The backfire state discrimination module uses the random forest algorithm to construct a backfire state discrimination model. Based on the reduced joint feature vector, the data of normal combustion and backfire state are labeled and used as a data set for input. The model is trained and tested, and the probability value of the occurrence of the backfire state is output. The real-time collected joint feature vector is input into the backfire state discrimination model, and it is determined whether it belongs to normal combustion or backfire state according to the set safety threshold; The acoustic-optic alarm module alarms based on the output result of the backfire state discrimination module. If it is determined to be in the backfire state, an alarm is issued.

2. The hydrogen combustion flashback monitoring system based on multi-modal data fusion according to claim 1, wherein: The data acquisition frequencies of the sensor system and the OH-PLIF system are synchronized to ensure the time consistency of the physical parameters and image data.

3. The hydrogen combustion flashback monitoring system based on multi-modal data fusion according to claim 2, characterized in that: A set of temperature sensors and pressure sensors are installed at the inner wall and the outlet of the fuel nozzle of the sensor system to monitor and collect the temperature and pressure data during the hydrogen combustion process in real time.

4. A hydrogen combustion flashback monitoring system based on multi-modal data fusion according to claim 2, characterized in that: The OH-PLIF system includes a laser, an optical system, a high-speed camera, and a timing control system. The timing controller ensures the synchronization of the laser and the high-speed camera through a digital signal generator, so that the laser beam emitted by the laser is vertically incident on the combustion chamber through the optical system, and the OH-PLIF image of the OH radical distribution during the hydrogen combustion process is captured in real time.

5. A hydrogen combustion flashback monitoring method based on multi-modal data fusion, characterized in that: According to the hydrogen combustion backfire monitoring system described in claim 1, its monitoring method includes the following steps: Step 1: Data acquisition When igniting, the sensor system and the OH-PLIF system are started simultaneously to collect physical parameters and image data; Step 2: Data preprocessing For the preprocessing of physical parameters, first select the discrete wavelet transform to decompose the data into approximation coefficients and detail coefficients, perform threshold processing on the detail coefficients of each layer to remove the noise components, use the processed detail coefficients and approximation coefficients for inverse wavelet transform to reconstruct the denoised temperature and pressure data, and then normalize the denoised temperature and pressure data respectively; For the preprocessing of image data, first, the median filtering method is used for noise reduction. A fixed window of size 3×3 is selected to traverse each pixel of the image. The pixel values within the window are sorted, and the median is taken to replace the current pixel value. The operation is repeated until all pixels are processed. Then, the maximum inter-class variance method is used for image segmentation on the denoised image. The number of pixels n at each gray level of the image is counted. i , and the probability at each gray level is calculated. In the formula, N is the total number of pixels at each gray level. Let the gray level k be the threshold to divide the image into the background region and the flame region. Respectively, according to and calculate the mean probability of the background region and the flame region. In the formula, L is the range of gray levels of the image. Then, respectively, according to and calculate the mean of the background region and the flame region. Then, according to σ 2 (k) = ω0(k)·ω1(k)·(μ0(k) - μ1(k)) 2 calculate the inter-class variance. Traverse all possible thresholds k to find the threshold k that maximizes the inter-class variance. * , and use the threshold k * to binarize the image. Pixels with a gray level less than or equal to the threshold k * are set as the background region, and pixels with a gray level greater than the threshold k * are set as the flame region; Step 3: Feature extraction For the feature extraction of physical parameters, calculate the mean, variance, slope and peak value of the temperature and pressure data respectively as the extracted features; For the feature extraction of image data, extract the shape features, color features and texture features of the flame during the hydrogen combustion process according to the segmented flame region as the extracted features, where: The shape features include the flame area, flame perimeter, and flame circularity. The calculation formula for the flame area is where I(i,j) represents the value of any pixel (i,j) inside the flame region, row 1 and column 1 respectively represent the number of rows and columns inside the flame region. The calculation formula for the flame perimeter is where B(i,j) represents the value of any pixel (i,j) on the boundary of the flame region, row 2 and column 2 respectively represent the number of rows and columns on the boundary of the flame region. The calculation formula for the flame circularity is The value range is [0,1]; The color features include the mean gray value and the gray variance, and the calculation formula for the mean gray value is The calculation formula for the gray variance is where I(i) represents the gray value of the i-th pixel, and K represents the total number of pixels in the flame region of the image; The texture features are used to describe the flame texture by extracting the gray value distribution of pixel pairs in the statistical image through the gray-level co-occurrence matrix. The directions of 0°, 45°, 90°, and 135° pixel pairs are selected, the statistical distance is 1 pixel, the flame area is traversed, the gray value distribution of pixel pairs that meet the direction and distance conditions is statistically calculated, the gray-level co-occurrence matrix is constructed and the matrix is normalized so that each element represents a probability. Based on the normalized gray-level co-occurrence matrix calculate the texture features, where P(i,j) represents the number of occurrences of pixel pairs with gray values of i and j; Step 4: Multimodal feature fusion For the feature splicing part, represent the feature vector of the physical parameters as F temp-pressure = [f1, f2,..., f M , represent the feature vector of the image data as F OH-PLIF = [g1, g2,..., g K , then the combined feature vector is F joint = [f1, f2,..., f M , g1, g2,..., g K . Standardize each feature in the combined feature vector so that the mean is 0 and the variance is 1. Then, calculate the covariance matrix and perform eigenvalue decomposition to obtain eigenvalues and eigenvectors. Sort the eigenvalues from largest to smallest, and the corresponding eigenvectors are the principal components. Select the principal components with a cumulative contribution rate of 95% according to the eigenvalue magnitudes, and project the standardized feature matrix onto the selected principal components for dimensionality reduction; Step 5: Flashback state discrimination and sound and light alarm The flashback state discrimination module uses the random forest algorithm to construct a flashback state discrimination model, annotates the combustion state of each sample based on the result of multimodal feature fusion, 1 represents the flashback state, 0 represents normal combustion, divides the data set into a training set and a test set, sets the parameters of the random forest, uses the training set to train the flashback state discrimination model, uses the test set to evaluate the flashback state discrimination model, loads the trained flashback state discrimination model into the system, inputs the jointly collected feature vector in real time, outputs the probability value of the occurrence of the flashback state, and determines whether it belongs to normal combustion or flashback state according to the set safety threshold. If the output probability value is greater than the safety threshold, it is determined as the flashback state, and at the same time the sound and light alarm module gives an alarm, otherwise it is determined as normal combustion.

6. The hydrogen combustion flashback monitoring method based on multimodal data fusion according to claim 5, characterized in that: The calculation of the texture feature based on the normalized gray-level co-occurrence matrix in Step 3 includes: Contrast ratio, calculated by the formula Energy, the calculation formula is Homogeneity, the calculation formula is Correlation, the calculation formula is In the formula, μ i and μ j respectively represent the average gray values of pixel pairs i and j, and σ i and σ j respectively represent the standard deviations of the gray values of pixel pairs i and j.

Images