Flame stability discrimination method based on variational auto-encoder

Abstract

A flame stability discrimination method based on a variational auto-encoder comprises the following steps: acquiring a real-time flame image, and preprocessing the flame image; based on the preprocessed flame image, geometric parameters and physical parameters of the flame are extracted; performing data fusion processing on the flame parameters, and calculating frequency parameters of flame flicker; preliminarily sorting flame stable and unstable sample databases; importing a flame stable and unstable sample database into a variational automatic encoder, realizing data dimension reduction and classification effects, and outputting potential probability distribution; and establishing a category center distance index, namely a flame stability index, by using a Gaussian hybrid clustering model based on the potential probability distribution so as to monitor a stable state and an unstable state of the flame, and carrying out real-time adjustment on combustion in the unstable state. The flame stability is accurately represented according to the flame image, the flame image area can be tracked and displayed when combustion is unstable, and the real-time performance, accuracy and reliability of flame monitoring are effectively improved.

Description

Technical Field

[0001] The invention relates to industrial boiler combustion monitoring technology, and in particular to a flame stability discrimination method based on a variational autoencoder. Background Art

[0002] Affected by my country's energy distribution structure, coal-fired power generation still plays an important role in my country's power industry. Therefore, the safe and economical operation of thermal power units plays an important role in ensuring the stability of my country's power grid. Coal-fired boilers occupy an important position in power production and industrial heating. The flame stability during their combustion process is directly related to the safety of equipment operation and combustion efficiency. Traditional combustion monitoring methods mainly rely on the monitoring of single parameters such as temperature sensors and pressure sensors, but these methods cannot fully reflect the combustion state, especially under high load fluctuation conditions, the monitoring effect is not ideal. With the development of computer vision technology, combustion monitoring methods based on flame image analysis have gradually been applied to flame monitoring of industrial boilers. However, most of the existing flame monitoring methods can only detect simple features of the flame, such as area, brightness, etc., and cannot accurately and real-time evaluate the stability of the flame. In addition, the limitations of traditional image processing methods make it difficult for detection accuracy and response speed to meet actual needs. Therefore, how to combine multidimensional feature analysis to improve the accuracy and real-time performance of flame stability monitoring has become a hot topic of research.

[0003] The variational autoencoder (VAE) in deep learning is an unsupervised learning model that is good at processing complex and high-dimensional data. It can extract the multidimensional features of the flame by learning the latent variable space of the flame image, and then model and predict the dynamic changes of the flame. Through VAE, the complex features of the flame image data can be refined into latent variables with high recognition, so as to achieve more accurate monitoring of flame stability. Compared with traditional image processing technology, VAE can handle more complex combustion conditions and show better adaptability under high-frequency fluctuations and abnormal conditions. Combined with multidimensional feature analysis methods, not only can VAE be used based on conventional features such as brightness, shape, and color of flame images, but also deeper features such as the spatiotemporal dynamic pattern of flame changes and the volatility of combustion intensity can be extracted through VAE. This method can better capture the details of the changes in flame stability and evaluate the potential risks in the combustion process in real time. Through the combination of deep learning and computer vision technology, the combustion monitoring system will be able to better cope with complex and changeable conditions and ensure that coal-fired boilers can maintain stable and efficient operation under high load fluctuations. Summary of the invention

[0004] In order to solve the above problems, the present invention proposes a flame stability discrimination method based on variational autoencoder, and the specific technical solution is as follows: A flame stability discrimination method based on a variational autoencoder comprises the following steps: Step 1: Collect real-time flame images under different loads and working conditions, and pre-process the collected flame images to generate multi-dimensional image data of the flame; Step 2: Extracting geometric and physical parameters of the flame based on the preprocessed flame image; Step 3: Perform data fusion processing on the geometric parameters and physical parameters of the flame, and calculate the frequency parameters of the flame flickering; Step 4: Use chi-square distribution and different confidence intervals set by square prediction error statistics to achieve preliminary sorting of flame stable sample database and flame unstable sample database; Step 5: By importing the flame stability sample database and the flame instability sample database into the variational autoencoder, data dimension reduction and classification are achieved, and the potential probability distribution is output; Step 6: Based on the potential probability distribution, the Gaussian mixture clustering model is implemented to establish a category center distance index, namely the flame stability index, which is used to accurately characterize the stability of the flame; Step 7: According to the flame stability index, the stability of the flame is evaluated to monitor the stable and unstable states of the flame, and the unstable combustion is adjusted in real time.

[0005] Furthermore, the flame image acquisition frame rate is not less than 50 frames per second, and the image resolution is not less than 1920×1080 pixels.

[0006] Furthermore, image preprocessing includes adaptive filtering, image edge enhancement and region segmentation to accurately divide the main combustion area, unburned area and burnt-out area of ​​the flame.

[0007] Furthermore, the geometric parameters include flame areas, flame contours and flame front positions in different regions, and the physical parameters include flame brightness distribution, combustion frequency and temperature field distribution.

[0008] Furthermore, in step 3, the geometric parameters and physical parameters are reduced in dimension by principal component analysis, the brightness change of the flame is converted into the combustion frequency by Fourier transform, and then the flickering frequency parameter of the flame is calculated.

[0009] Further, in step 6, the flame stability index is used to accurately distinguish flame stability, wherein an RSI value between [0.6, 1] indicates that the flame is stable, an RSI value between (0.3, 0.6] indicates that the flame is unstable, and an RSI value between [0, 0.3] indicates that the flame is about to go out or has gone out.

[0010] Furthermore, in step 7, when the unstable state of the flame is detected, the flame area is marked and displayed based on the YOLO algorithm, the stability state of the flame is fed back in real time, and the combustion parameters are automatically adjusted to restore the flame stability.

[0011] Furthermore, the flame image is collected in real time by a high-speed camera, the collected flame image is pre-processed by an image processing module, and a combustion control module adjusts the combustion in real time according to the flame stability evaluation.

[0012] Furthermore, the image processing module has the functions of flame image preprocessing, parameter extraction, data fusion and flame stability index calculation.

[0013] Furthermore, the combustion control module is an integrated control system responsible for receiving the output data of the image processing module and adjusting the combustion equipment in real time according to the flame stability index.

[0014] Beneficial effects of the present invention: (1) Compared with the calculation speed of flame stability index of shallow features, the present invention can identify flame stability of different loads in real time without the need for label images.

[0015] (2) The present invention accurately characterizes the stability of the flame based on the flame image and can track and display the flame image area when the combustion is unstable. Compared with manually identifying the flame stability based only on the high-speed flame image, the flame stability identification accuracy is significantly improved.

[0016] (3) The present invention proposes a comprehensive flame stability index RSI to achieve quantitative identification of the flame stability degree and provide guidance for advanced combustion adjustments. Compared with flame stability indicators that rely on shallow features, the identification accuracy is significantly improved.

[0017] (4) The present invention adopts advanced image analysis and data fusion technology to effectively improve the real-time, accuracy and reliability of flame monitoring, and can realize intelligent monitoring and optimization in a variety of industrial combustion scenarios, providing strong support for safe, environmentally friendly and efficient combustion processes.

[0018] (5) The intelligent model constructed by the present invention not only has high prediction accuracy, but also has the robustness to overcome random interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flowchart of the flame stability determination method of the present invention. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances.

[0021] like Figure 1 As shown, a flame stability discrimination method based on variational autoencoder includes the following steps: Step 1: Collect real-time flame images under different loads and working conditions, and preprocess the collected flame images to generate multi-dimensional image data of the flame.

[0022] Step 2: Based on the preprocessed flame image, extract the geometric parameters and physical parameters of the flame.

[0023] Step 3: Perform multi-dimensional data fusion processing on the geometric parameters and physical parameters of the flame, and calculate the frequency parameters of the flame flickering; these parameters are important indicators of flame stability and can reflect the energy changes and instability during the combustion process.

[0024] Step 4: Use the chi-square distribution and different confidence intervals set by the squared prediction error (SPE) statistic to achieve preliminary sorting of the flame stable sample database and the flame unstable sample database.

[0025] Step 5: By importing the flame stability sample database and the flame instability sample database into the variational autoencoder, data dimension reduction and classification are achieved, and the potential probability distribution is output to simplify the data structure. Dimension reduction processing not only improves the computational efficiency, but also highlights the key features that affect flame stability, which is convenient for subsequent analysis.

[0026] Step 6: Based on the potential probability distribution, the Gaussian mixture clustering model is implemented to establish a category center distance index, namely the flame stability index (RSI), which is used to accurately characterize the stability of the flame. RSI can quantify the stability of the flame, allowing operators to intuitively understand the operating status of the flame and make corresponding adjustments.

[0027] Step 7: According to the flame stability index, the stability of the flame is evaluated to monitor the stable and unstable states of the flame, and the unstable combustion is adjusted in real time.

[0028] In one embodiment, the flame image acquisition frame rate is not less than 50 frames per second, and the image resolution is not less than 1920×1080 pixels. The flame image is acquired by a high-speed camera, which needs to be installed in a suitable position of the combustion equipment to ensure that it can fully capture the flame image. Under different loads and working conditions, the high-speed camera acquires the flame image in real time, and transmits the image data to the image processing module for subsequent processing. The acquired flame image is pre-processed by the image processing module, and the combustion is adjusted in real time by the combustion control module according to the flame stability evaluation.

[0029] The image processing module includes functions of image preprocessing, feature extraction, data fusion and RSI calculation, which are implemented using computer vision software and related algorithm libraries (such as OpenCV).

[0030] The combustion control module is an integrated control system, which is responsible for receiving the output data of the image processing module and adjusting the combustion equipment in real time according to the RSI value.

[0031] The combustion control module has data transmission function, real-time feedback function, adaptive function and comprehensive judgment function; specifically: Data transmission function: The high-speed camera transmits real-time image data to the image processing module through the Ethernet network interface, and the data processed by the image processing module is transmitted to the combustion control module through the interface.

[0032] Real-time feedback function: Establish a real-time monitoring mechanism to ensure that the combustion control module can receive the RSI value and alarm information from the image processing module in a timely manner.

[0033] Adaptive function: The acquisition frame rate of the high-speed camera can be automatically adjusted according to the working conditions to adapt to the flame state under different loads.

[0034] Comprehensive identification function: The image processing module combines adaptive filtering, edge enhancement and image segmentation technology to accurately divide the main combustion area, unburned area and burnt area of ​​the flame, thereby improving the accuracy of stability identification.

[0035] The detailed steps for flame stability determination and adjustment are as follows: In step 1, the flame image preprocessing process is: Flame image preprocessing includes adaptive filtering, edge enhancement and region segmentation, thereby generating multi-dimensional image data of the flame to accurately divide the main combustion area, unburned area and burnt area of ​​the flame, ensuring that the shape and changes of the flame can be clearly captured under complex working conditions, providing a basis for subsequent feature extraction. Among them: Adaptive filtering: Use the adaptive filtering algorithm to eliminate noise from the collected image; this step analyzes the local statistical characteristics of the image and dynamically adjusts the filtering parameters to ensure the clarity of the flame image.

[0036] Edge Enhancement: Edge enhancement technology is used to increase the contrast of the flame edge to ensure that the flame outline is clearly visible in the image.

[0037] Region segmentation: The flame is separated from the background through the region segmentation algorithm to generate multi-dimensional image data of the flame for subsequent feature extraction.

[0038] In step 2, the process of flame parameter extraction is: Geometric parameter extraction: Geometric parameters are extracted from the preprocessed flame image, including the flame area in different regions, flame contour and flame front position; these features are automatically calculated by image analysis software.

[0039] Physical parameter extraction: measure the brightness distribution, combustion frequency and temperature field distribution of the flame; the brightness distribution is obtained through grayscale histogram analysis, and the flickering frequency is obtained through time series analysis.

[0040] In step 3, the process of parameter fusion and calculation is: Through principal component analysis, geometric parameters and physical parameters are linearly combined to achieve multi-dimensional data fusion processing of these parameters; the brightness change of the flame is converted into combustion frequency through Fourier transform, and the frequency parameter of the flame flicker is calculated to be used as a key parameter for evaluating flame stability. Among them, the frequency parameter can often reveal the pulsation characteristics of the flame and help identify potential combustion instability.

[0041] In step 4, the process of sorting the flame database is: Chi-square distribution and SPE statistics are used to set different confidence intervals, and flame samples are preliminarily sorted to form a flame stable sample database and a flame unstable sample database.

[0042] Among them, the expression of squared prediction error (SPE) is:

[0043]

[0044] represents the control limit of SPE. When the flame combustion is unstable, the sample vector x consists of the normal part superimposed by the fault part, making SPE greater than .

[0045] The sample vector x is the flame's geometric parameters, physical parameters, frequency parameters and other parameters; the imported sample vector x data includes a flame stability sample database (normal part) and a flame instability sample database (fault part).

[0046] Hotelling statistic The expression is:

[0047] When normal data follows a multivariate normal distribution, The statistic is related to the F distribution, namely:

[0048] If the number of data points N of flame geometric parameters, physical parameters and frequency parameters is large, then The statistic can be approximated by a chi-square distribution with 1 degree of freedom, namely:

[0049] Set the above and The control limits (usually 95% confidence intervals are set to obtain the corresponding boundary values) can be used to distinguish between stable flame and unstable flame sample data.

[0050] In step 5, the dimension reduction and classification process of the data is: The variational autoencoder is mainly trained by taking flame images (flame stability samples) under stable combustion conditions as training samples. Due to the dimensionality reduction characteristics of the variational autoencoder itself, the flame image can be finally mapped to a multidimensional normal distribution. By adjusting the learning, the obtained multidimensional normal distribution fits the probability distribution under stable combustion conditions. By setting conditions (inputting labeled data, where the data is the flame image and the label is the stable and unstable state of the flame), the above variational autoencoder can input data and labels to obtain probability distributions of different types of conditions, that is, the variational autoencoder outputs a potential probability distribution with a stable flame probability distribution and an unstable flame probability distribution, forming the mean and variance of the normal distribution.

[0051] In step 6, the flame stability index (RSI) is calculated as: The Gaussian mixture model is mainly drawn based on the mean and variance of the normal distribution of the variational autoencoder output conditions. Since the probability distributions of flame images under stable and unstable combustion conditions are different, the distinction between stable flame images and unstable flame images can be achieved by the flame images falling into the above different probability distributions.

[0052] In this way, the Gaussian mixture model can obtain the flame stability probability distribution F1 and the flame instability probability distribution F2, as well as the probability distribution F3 of the flame image to be tested. By calculating the KL1 divergence of F1 and F3 and the KL2 divergence of F2 and F3, the distance between the probability distribution F3 of the image to be tested and the two probability distributions F1 and F2 is obtained.

[0053]

[0054] The flame stability index (RSI) ranges from [0, 1]. The RSI value is used to accurately distinguish flame stability, specifically: The RSI value is between [0.6, 1], indicating a stable flame; RSI values ​​between (0.3, 0.6] indicate unstable flame; An RSI value between [0, 0.3] indicates that the flame is about to go out or has gone out.

[0055] In step 7, the flame abnormality state monitoring and adjustment process is: Abnormal flame status monitoring: monitor the RSI value through the set threshold. When the RSI value is lower than the set value, the flame status is judged to be abnormal.

[0056] When the flame state is assessed to be abnormal, the real-time flame tracking algorithm is used to trigger an alarm, and the feedback flame information is transmitted to the combustion control module in real time to optimize the combustion parameters. Through real-time monitoring and feedback mechanism, the system can respond quickly to changes in flame state to ensure safety and combustion efficiency.

[0057] Abnormal flame state adjustment process: When the unstable state of the flame is detected, the flame area is marked and displayed based on the YOLO algorithm, the stability state of the flame is fed back in real time, and the combustion parameters are automatically adjusted to restore the flame stability.

[0058] The YOLO model detects and labels the flame area. The process includes data preparation, model training, flame detection and result display. In order to realize flame area detection, a large number of image data sets containing flame areas and annotating their locations are required. The steps of data preparation are as follows: (1) Collect image samples containing flames. The samples should contain various flame types, backgrounds, brightness, and different flame shapes to ensure the generalization ability of the model.

[0059] (2) Use labeling tools (such as LabelImg) to label the flame area. Usually, the flame is labeled as the "fire" category and a labeling file (such as a text file in YOLO format) is generated, which contains the target category and bounding box location. In order to enhance the robustness of the model, the data can be enhanced by image rotation, scaling, flipping, etc.

[0060] (3) Select a suitable YOLO series model, set the number of categories, and import the above dataset for training. The detection results will display the bounding box and confidence score of the flame area.

[0061] (4) The detected flame area information is marked on the image to facilitate real-time flame monitoring applications. Libraries such as OpenCV can be used to draw a bounding box on the image and display the flame area.

[0062] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A flame stability discrimination method based on variational autoencoder, characterized in that: The following steps are involved: Step 1: Collect real-time flame images under different loads and working conditions, and pre-process the collected flame images to generate multi-dimensional image data of the flame; Step 2: Extracting geometric and physical parameters of the flame based on the preprocessed flame image; Step 3: Perform data fusion processing on the geometric parameters and physical parameters of the flame, and calculate the frequency parameters of the flame flickering; Step 4: Use chi-square distribution and different confidence intervals set by square prediction error statistics to achieve preliminary sorting of flame stable sample database and flame unstable sample database; Step 5: By importing the flame stability sample database and the flame instability sample database into the variational autoencoder, data dimension reduction and classification are achieved, and the potential probability distribution is output; Step 6: Based on the potential probability distribution, the Gaussian mixture clustering model is implemented to establish a category center distance index, namely the flame stability index, which is used to accurately characterize the stability of the flame; Step 7: According to the flame stability index, the stability of the flame is evaluated to monitor the stable and unstable states of the flame, and the unstable combustion is adjusted in real time.

2. The flame stability determination method according to claim 1, characterized in that: The flame image acquisition frame rate is not less than 50 frames per second, and the image resolution is not less than 1920×1080 pixels.

3. The flame stability determination method according to claim 2, characterized in that: Image preprocessing includes adaptive filtering, image edge enhancement and region segmentation to accurately divide the main combustion area, unburned area and burnt area of ​​the flame.

4. The flame stability determination method according to claim 1, characterized in that: The geometric parameters include flame areas, flame contours and flame front positions in different regions, and the physical parameters include flame brightness distribution, combustion frequency and temperature field distribution.

5. The flame stability determination method according to claim 4, characterized in that: In step 3, the geometric parameters and physical parameters are reduced in dimension by principal component analysis, the brightness change of the flame is converted into the combustion frequency by Fourier transform, and then the flickering frequency parameters of the flame are calculated.

6. The flame stability determination method according to claim 1, characterized in that: In step 6, the flame stability index is used to accurately distinguish flame stability, wherein an RSI value between [0.6, 1] indicates a stable flame, an RSI value between (0.3, 0.6] indicates an unstable flame, and an RSI value between [0, 0.3] indicates that the flame is about to go out or has gone out.

7. The flame stability determination method according to claim 1, characterized in that: In step 7, when the unstable state of the flame is detected, the flame area is marked and displayed based on the YOLO algorithm, the stability state of the flame is fed back in real time, and the combustion parameters are automatically adjusted to restore the flame stability.

8. The flame stability determination method according to any one of claims 1 to 7, characterized in that: The flame image is collected in real time by a high-speed camera, and the collected flame image is pre-processed by an image processing module. A combustion control module adjusts the combustion in real time according to the flame stability assessment.

9. The flame stability determination method according to claim 8, characterized in that: The image processing module has the functions of flame image preprocessing, parameter extraction, data fusion and flame stability index calculation.

10. The flame stability determination method according to claim 9, characterized in that: The combustion control module is an integrated control system, which is responsible for receiving the output data of the image processing module and adjusting the combustion equipment in real time according to the flame stability index.

Images