Torch combustion monitoring and optimization control system based on multispectral infrared and visible light imaging

Abstract

The invention discloses a torch combustion monitoring and optimization control system based on multispectral infrared and visible light imaging, and the system comprises a multispectral imaging module which is used for synchronously collecting an infrared image and a visible light image of a torch flame, and carrying out the spatial registration, so as to obtain the multispectral radiation data of the flame; the physical model calculation module is used for calculating a first combustion state parameter according to the multispectral radiation data based on a Planck radiation formula and a grey body radiation model; the artificial intelligence calculation module is used for calculating a second combustion state parameter according to the multispectral radiation data by adopting a pre-trained multi-scale visual Transform network; the data fusion and correction module is used for fusing and correcting the first combustion state parameter and the second combustion state parameter by adopting a Kalman filtering algorithm so as to output a final combustion state parameter; and the closed-loop control module is used for dynamically adjusting the operation parameters of the torch according to the final combustion state parameters so as to optimize the combustion efficiency.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of industrial combustion monitoring and control, and particularly relates to a flare combustion monitoring and optimization control system based on multispectral infrared and visible light imaging. BACKGROUND

[0002] Flares are important equipment in the chemical, oil, and natural gas industries for processing process waste gas and emergency release gas. Its core function is to convert harmful gases into relatively harmless carbon dioxide and water vapor through combustion, thereby reducing harm to the environment and human health. However, the combustion efficiency of flares directly affects the environmental impact of exhaust gases. Low combustion efficiency will result in the emission of a large amount of unburned hydrocarbons, carbon monoxide, particulate matter, and other pollutants, seriously polluting the atmospheric environment. In addition, fluctuations in combustion efficiency will also affect the operational stability of flare equipment and even cause safety accidents.

[0003] Flare systems are widely used in the chemical industry. Due to the particularity of flare design and operation, the determination of its destruction removal efficiency (DRE) is extremely challenging. There has been long-standing controversy about the total amount of flare emission pollutants, but there is still no accurate answer.

[0004] In 2010, the Texas Commission on Environmental Quality (TCEQ) commissioned the University of Texas at Austin (UT) to conduct a comprehensive study on flare combustion efficiency (CE) and destruction removal efficiency (DRE). The study used two sets of supplementary remote sensing monitoring systems: Telops' infrared hyperspectral imager (Hyper-Cam) and IMACC's passive / active Fourier transform infrared spectrometer (PFTIR / AFTIR). The study showed that the detection results of the IMACC system differed by about 2%-2.5% from the combustion efficiency CE value of the traditional "sampling measurement method", with data availability of 99%-100%; while the average difference of the Telops system was 19.9%, with data availability of only 39%. At the same time, both systems cannot provide real-time monitoring data, and the purchase cost is high (especially for Telops equipment). Therefore, there is currently a lack of a solution that can monitor flare efficiency in real time and provide feedback to operators to adjust operating conditions.

[0005] Currently, the measurement of flare combustion efficiency mainly relies on sampling analysis method, infrared spectral imaging technology, Fourier transform infrared spectroscopy technology, video imaging spectral radiometer, and thermocouple automatic disassembly system. These methods can achieve the measurement of combustion efficiency to some extent, but generally have problems such as insufficient measurement accuracy, poor real-time performance, high cost, limited applicability, and lack of closed-loop control capability.

[0006] For example, although sampling analysis is accurate, it has disadvantages such as response lag, complex operation, and inability to provide real-time feedback. Although infrared spectral imaging technology has advantages such as non-contact and remote measurement, it has problems such as expensive equipment, low data availability, and large measurement error. Although Fourier transform infrared spectral technology has high measurement accuracy, it also has problems such as expensive equipment, inability to provide real-time feedback, and complex installation. Although the video imaging spectral radiometer has a certain real-time performance, its measurement accuracy is still greatly affected by the combustion state. Although the thermocouple automatic disassembly system solves the problem of traditional thermocouple maintenance, this method is a contact measurement, and has disadvantages such as short equipment life, high maintenance cost, and limited measurement points.

[0007] Therefore, there is an urgent need for a torch combustion monitoring and optimization control system that can achieve high-precision, real-time measurement and closed-loop control capability to solve the problems in the prior art. SUMMARY

[0008] One of the purposes of the present application is to provide a torch combustion monitoring and optimization control system based on multispectral infrared and visible light imaging, which can achieve high-precision, real-time measurement of key parameters such as torch combustion temperature, combustion efficiency, and gas flow.

[0009] The purpose of the present application is to provide a torch combustion monitoring and optimization control system based on multispectral infrared and visible light imaging, which can dynamically adjust the torch operating parameters based on the measurement results, thereby optimizing the combustion efficiency, reducing harmful gas emissions, and improving environmental friendliness and industrial safety.

[0010] The purpose of the present application is to provide a torch combustion monitoring and optimization control system based on multispectral infrared and visible light imaging, which can achieve high-precision, real-time monitoring and closed-loop optimization control of the torch combustion state by constructing a physical model and artificial intelligence model dual-link collaborative mechanism, combining multispectral image fusion and Kalman filtering algorithm.

[0011] The purpose of the present application is to provide a torch combustion monitoring and optimization control system based on multispectral infrared and visible light imaging, wherein the physical model calculation module calculates the flame temperature and combustion efficiency based on the Planck radiation formula and gray body radiation model, combined with multispectral image data. For the measurement results of multiple spectral channels, the weighted Gaussian formula is used for correction to improve the calculation accuracy. The combustion efficiency calculation unit calculates the combustion efficiency based on the flame temperature and gas composition.

[0012] The purpose of the present application is to provide a torch combustion monitoring and optimization control system based on multispectral infrared and visible light imaging, wherein the data fusion and correction module uses Kalman filtering algorithm to fuse and correct the output of the physical model and artificial intelligence model, reduces the single measurement error, and improves the overall measurement accuracy. The error analysis and correction unit analyzes and corrects the measurement error to improve the measurement accuracy.

[0013] The purpose of the present application is to provide a torch combustion monitoring and optimization control system based on multispectral infrared and visible light imaging, which uses multispectral image fusion technology to obtain multispectral image information of the flame through the registration of the uncooled infrared camera and the megapixel zoom visible light lens, can maintain high image quality and measurement accuracy under different combustion states, and uses weighted Gaussian formula to correct the measurement results of multiple spectral channels based on Planck radiation formula and gray body radiation model, to improve the calculation accuracy.

[0014] The purpose of the present application is to provide a torch combustion monitoring and optimization control system based on multispectral infrared and visible light imaging, which proposes a multiscale visual Transformer network based on Transformer architecture for feature extraction and combustion parameter prediction of multispectral images, correction of multispectral system measurement error, and improvement of measurement accuracy and generalization ability.

[0015] The present application provides a torch combustion monitoring and optimization control system based on multispectral infrared and visible light imaging, characterized in that it comprises: A multispectral imaging module is used to synchronously acquire infrared images and visible light images of the torch flame, and perform spatial registration to obtain multispectral radiation data of the flame. A physical model calculation module is used to calculate first combustion state parameters based on Planck radiation formula and gray body radiation model according to the multispectral radiation data. An artificial intelligence calculation module uses a pre-trained multiscale visual Transformer network to calculate second combustion state parameters according to the multispectral radiation data. A data fusion and correction module is used to fuse and correct the first combustion state parameters and the second combustion state parameters using Kalman filtering algorithm to output final combustion state parameters. A closed-loop control module is used to dynamically adjust the operating parameters of the torch according to the final combustion state parameters to optimize the combustion efficiency.

[0016] The present application provides a torch combustion monitoring and optimization control system based on multispectral infrared and visible light imaging, which comprises a multispectral imaging module, a physical model calculation module, an artificial intelligence calculation module, a data fusion and correction module, a closed-loop control module, and a data storage and visualization module.

[0017] In one possible implementation, the multispectral imaging module includes an uncooled infrared camera, a visible light camera, and a multispectral registration unit. The uncooled infrared camera, operating in wavelengths covering 3-5 microns and 8-12 microns, is used to capture infrared radiation images of flames. The visible light camera is used to capture visible light images of the flames. The multispectral registration unit is used to spatially register the infrared radiation images with the visible light images to generate an aligned flame position array.

[0018] In one possible embodiment, the flare combustion monitoring and optimization control system includes a temperature calculation module, which integrates a multi-wavelength feature fusion model based on the Transformer architecture. The multi-wavelength feature fusion model includes an encoding layer, a multi-head attention mechanism layer, and a decoding layer. The encoding layer encodes the radiation intensity data of the red, green, and blue light bands and the infrared band into feature vectors. The multi-head attention mechanism layer realizes cross-spectral feature association through self-attention weight distribution. The decoding layer outputs self-corrected flame temperature distribution data. The Transformer model of the multi-wavelength feature fusion model adopts an end-to-end training method. The input end receives the original spectral data and the output end directly generates the corrected temperature distribution matrix.

[0019] In one possible embodiment, the physical model calculation module includes a weighted Gaussian correction unit, which establishes a weight matrix based on the response sensitivity coefficient of each spectral channel, and uses the Levenberg-Marquardt algorithm to perform nonlinear optimization on the first combustion state parameter based on multi-band calculation, so that the system measurement error is controlled within ±2%.

[0020] On the other hand, a method for optimizing and controlling flare combustion is provided. The method comprises the following steps: Step 1: Obtain flame radiation data through a multispectral imaging system; Step 2: Calculate the initial temperature distribution based on Planck's formula and gray body model; Step 3: Use weighted Gaussian algorithm and Levenberg-Marquardt algorithm to perform multispectral data fusion correction; Step 4: Calculate combustion efficiency using the Transformer model; Step 5: Input the corrected temperature data into the combustion optimization control module; Step 6: Generate control instructions based on the physical lookup table and flow meter data; Step seven, adjust the combustion parameters through the actuator and return to the monitoring step.

[0021] In a third aspect, the present application provides an electronic device comprising a processor and a memory, wherein the memory stores a computer program, and the processor implements the flare combustion optimization control method described in the second aspect by executing the computer program. The electronic device can be integrated into a local control cabinet of a flare device or deployed in a server cluster of a remote monitoring center.

[0022] Compared with the prior art, the present invention has at least one or more of the following advantages and positive effects: 1. High-precision measurement: By integrating infrared and visible light image fusion technology with physical models and artificial intelligence algorithms, the measurement accuracy of flame temperature and combustion efficiency is significantly improved; 2. Real-time and dynamic response: The system has real-time image acquisition and processing capabilities, can quickly respond to changes in combustion status, and achieve dynamic closed-loop control; 3. Closed-loop optimization control: Based on combustion efficiency feedback, dynamically adjust the flare operating parameters to improve combustion efficiency and reduce pollutant emissions; 4. High system robustness: The dual-link model fusion mechanism and Kalman filter algorithm are used to effectively improve the stability and reliability of the system under complex working conditions; 5. Environmental friendliness and energy saving: Through combustion optimization control, unburned hydrocarbons and harmful gas emissions are reduced, and energy utilization efficiency is improved; 6. Remote monitoring and data management: Support remote data access and visual display to facilitate industrial site management and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present application, a brief introduction to the drawings of the embodiments will be given below. Obviously, the drawings described below only relate to some embodiments of the present application, and are not intended to limit the present application.

[0024] Figure 1 This is a schematic diagram of the infrastructure of a flare combustion monitoring and optimization control system using multi-spectral infrared and visible light imaging according to some embodiments of the present application.

[0025] Figure 2 This is a schematic diagram of a dual-link architecture system in which a physical model and an artificial intelligence model work in parallel according to some embodiments of the present application.

[0026] Figure 3A It is a schematic diagram of an optimized system integration solution of a flare combustion monitoring and optimization control system according to some embodiments of the present application.

[0027] Figure 3B It is a schematic diagram of the operation mode of the improved PID controller according to some embodiments of the present application.

[0028] Figure 3C is a system integration and dual-flow fusion perception sketch according to some embodiments of the present application.

[0029] Figure 4 is a workflow sketch of a flare combustion monitoring and optimization control system according to some embodiments of the present application.

[0030] Figure 5 is a flare combustion monitoring and optimization control method sketch based on multispectral infrared and visible light imaging according to some embodiments of the present application.

[0031] In the figure: 10, multispectral imaging module; 101, uncooled infrared camera; 102, visible light camera; 103, multispectral registration unit; 104, dynamic spectral selection module; 111, visual embedding module; 112, infrared embedding module; 113, multi-head attention mechanism; 114, layer normalization module; 115, feedforward network module; 116, random dropout module; 117, fusion output module; 118, loss function module; 140, image acquisition and transmission module; 1041, tunable filter array; 1042, principal component analysis unit; 20, physical model calculation module; 201, Planck formula calculation unit; 202, weighted Gaussian correction unit; 203, combustion efficiency calculation unit; 30, closed-loop control module; 301, physical lookup table; 302, interpolation calculation unit; 303, difference calculation and basic control unit; 304, actuator interface; 305, mass flow meter; 306, combustion aid control valve; 307, improved PID controller; 307a, feedforward compensation unit; 307b, dynamic weight adjustment unit; 307c, multivariate coupling control interface; 307d, emission constraint optimization unit; 307e, hybrid control strategy generation unit; 310, multi-scale visual Transformer network; 3041, flow control interface; 40, auxiliary function module; 401, environmental compensation unit; 402, data storage unit; 403, environmental temperature and humidity sensor; 404, atmospheric transmittance calculation unit; 50, artificial intelligence calculation module; 60, data fusion and correction module; 601, physical link calibration subunit; 602, artificial intelligence link subunit; 603, adaptive parameter adjustment subunit; 604, dual-link output fusion subunit; 605, online calibration trigger subunit; 80, data storage and visualization module. DETAILED DESCRIPTION

[0032] In order to make the objectives, technical solutions and advantages of the present application clearer, the following will be combined with the drawings for a clear and complete description of the technical solutions in the embodiments of the present application. It should be appreciated that the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based upon the embodiments described in the present application, any other embodiments obtained by a person of ordinary skill in the art without creative effort should belong to the scope of the protection of the present application.

[0033] Reference to "an embodiment" in this application means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily mutually exclusive or alternative embodiments. It is expressly understood that the described embodiments of the application are merely examples and that one of ordinary skill in the art would be able to make changes and modifications thereto without the exercising of inventive faculty.

[0034] As described above, it should be emphasized that when the term "comprises / comprising" is used in this specification, it is specifically intended that inclusion of a feature, integer, step or component is not exclusive or in addition to one or more other features, integers, steps, components or groups of features, integers, steps, components. As used in this application, the singular forms "a," "an," and "the" include plural references unless the context clearly dictates otherwise.

[0035] Unless otherwise defined, all technical and scientific terms used in the present application have the same meanings as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of the present application is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. The use of the terms "including," "comprising," "having," "containing," "involving," "characterized by," "comprises," "comprising," "has," "including," "involving," "including" and other similar forms in the disclosure are intended to be open-ended and otherwise mean and convey that certain embodiments include, while other embodiments do not necessarily include, one or more of the features, elements, steps, components or groups thereof described or illustrated. The terms "first," "second," and the like, as used in the description and the appended claims, are used for distinguishing between similar elements and not necessarily for describing a specific sequential or chronological order. It is to be understood that the terms so used are interchangeable under appropriate circumstances such that the embodiments of the application described herein are, for example, capable of use in either order. Furthermore, the term "another" and "or" as used in the description and the appended claims are used expansively and open-endedly, such that "another" represents "any one or more other than some one or more of the prior elements" and "or" means "and / or" unless the context clearly indicates otherwise. The term "plurality" as used in the description and the appended claims means two or more unless the context clearly indicates otherwise.

[0036] Conceptual Definitions For the convenience of understanding, some concepts related to the embodiments of the present application are first described below.

[0037] Multi-spectral infrared and visible imaging system: refers to the core device that cooperates with the non-cooled infrared camera and the visible light zoom camera to obtain flame radiation data and calculate the temperature distribution. The system realizes the spatial alignment of the infrared and visible light images through the multi-spectral registration technology, and forms a flame position array.

[0038] Weighted Gaussian correction algorithm: an algorithm for fusing the measurement results of each waveband by introducing a weight matrix in view of the spectral response difference problem existing in multi-spectral vector measurement. Specifically, by establishing the response sensitivity coefficient matrix of each spectral channel, the Levenberg-Marquardt algorithm is used to nonlinearly optimize the multi-spectral temperature calculation results, so that the system measurement error reaches the same level as the VISR device, and the optimization is preferably within ±2%.

[0039] Physical lookup table: includes a database storing the mapping relationship of fuel type, combustion condition, exhaust component and optimal combustion parameter. When it is detected that the current working condition parameters match the known parameter combination in the physical lookup table, the corresponding optimal control parameters are automatically called to realize closed-loop feedback control.

[0040] In order to more clearly understand the present application, the system architecture and working process of the embodiments of the present application are exemplarily described below in conjunction with the drawings.

[0041] Embodiment one: torch combustion monitoring and optimization control system As shown in Figure 1 , the present embodiment provides a torch combustion monitoring and optimization control system based on multi-spectral infrared and visible light imaging, which realizes real-time monitoring and basic control of the torch combustion state. The torch combustion monitoring and optimization control system comprises a multi-spectral imaging module 10, a physical model calculation module 20, a closed-loop control module 30 and an auxiliary function module 40.

[0042] The multi-spectral imaging module 10 comprises a non-cooled infrared camera 101 and a visible light camera 102, wherein the focal length range of the visible light camera 102 is preferably 25-100mm, and the working waveband is 400-700nm. In the flame monitoring, the visible light radiation information in the range of 0.4-0.7μm is mainly obtained, which is used to collect the visible light image of the flame. In some optional embodiments, the working waveband of the non-cooled infrared camera 101 covers 3-5 microns and 8-12 microns, which is used to collect the infrared radiation image of the flame. The multi-spectral imaging module 10 further comprises a multi-spectral registration unit 103, which is used to generate a flame position array in combination with the image data collected by the non-cooled infrared camera 101 and the visible light camera 102.

[0043] In the embodiment, the non-cooled infrared camera 101 synchronously collects flame radiation data with the visible light camera 102, and the spatial alignment is performed through the multispectral registration unit 103 to generate a flame position array. The above scheme establishes a flame radiation characteristic database through multispectral position registration, and provides accurate multispectral radiation data for subsequent temperature calculation and combustion efficiency analysis. The physical model calculation module 20 includes a Planck formula calculation unit 201, a weighted Gaussian correction unit 202, and a combustion efficiency calculation unit 203, wherein the combustion efficiency calculation unit 203 is preferably a Transformer model. The physical model calculation module 20 receives the radiation intensity data of the red, green and blue light wave bands and the infrared wave band based on the Planck formula and the gray body radiation model, and obtains a flame temperature distribution matrix through multi-wavelength combination calculation.

[0044] Reference Figure 3A As shown in the partial optional embodiment, the Planck formula calculation unit 201 is based on the multispectral image data collected by the multispectral imaging module 10, the flame temperature calculation model is based on the Planck radiation formula and the gray body radiation model, the flame temperature is calculated based on the relationship between the single-wave radiation energy and the emissivity, wavelength, temperature and other parameters according to the Planck radiation law and the gray body radiation model, and the formula is as follows:

[0045] Among them, is the single-wave radiation energy, is the emissivity, is the wavelength, is the temperature, is the Planck constant, is the speed of light, is the Boltzmann constant.

[0046] For high-temperature flames, the Wien approximation formula can be used for simplified calculation:

[0047] Among them, the flame temperature calculation model is based on the Planck radiation formula, and for high-temperature flames, the Wien approximation formula can be used for simplified calculation. For the gray body radiation source such as flame, the actual radiation intensity is equal to the product of the spectral emissivity and the black body radiation brightness.

[0048] Further, in view of the spectral response difference problem existing in multispectral vector measurement, the weighted Gaussian correction unit 202 establishes a weight matrix according to the response sensitivity coefficients of each spectral channel:

[0049] The weight matrix includes infrared long-wave band weight, infrared middle-wave band weight, visible red light segment weight and visible green light segment weight, and the typical values are 0.4, 0.3, 0.2 and 0.1 respectively.

[0050] wherein The SNR of each waveband is dynamically adjusted (e.g., the infrared waveband weight The visible light waveband weight The multispectral temperature calculation result is non-linearly optimized by Levenberg-Marquardt algorithm, so that the system measurement error reaches the same level as the VISR device, such as preferably within ±2%.

[0051] In an optional embodiment, for monitoring the dynamic combustion process of the torch to be monitored, the weighted Gaussian correction unit 202 uses a Kalman filter to estimate the state of the multi-time sequence measurement data, so as to realize real-time updating of the flame temperature field of the torch.

[0052] The combustion efficiency calculation unit 203 calculates the combustion efficiency based on the flame temperature distribution and the combustion theory. The combustion efficiency calculation unit 203 obtains the ratio of the difference between the current average temperature and the ambient temperature to the difference between the theoretical combustion temperature and the ambient temperature, and the calculation formula is:

[0053] wherein, represents the combustion efficiency, T avg is the average temperature of the flame obtained by multispectral measurement, T env is the ambient temperature, T theoretical is the theoretical combustion temperature determined based on the fuel type and the combustion condition. When the combustion efficiency CE≥95%, it is determined as a high-efficiency combustion state, and when CE<85%, it is determined as a low-efficiency combustion state. Further, the combustion efficiency calculation unit 203 is integrated with a multispectral feature fusion model based on the Transformer architecture. The multispectral feature fusion model includes an encoding layer, a multi-head attention mechanism layer, and a decoding layer, wherein the encoding layer encodes the radiation intensity data of the red, green, and blue three-color light wavebands and the infrared waveband into a feature vector, the multi-head attention mechanism layer realizes cross-spectral feature association through self-attention weight distribution, and the decoding layer outputs the flame temperature distribution data after self-correction. The Transformer model of the multispectral feature fusion model receives the original spectral data at the input end and directly generates the corrected temperature distribution matrix at the output end.

[0054] Further, the closed-loop control module 30 includes a physical lookup table 301, an interpolation calculation unit 302, a difference calculation and basic control unit 303, and an actuator interface 304, and the closed-loop control module 30 is further connected with a mass flow meter 305 and a combustion aid control valve 306. The preferred measurement accuracy of the mass flow meter 305 reaches ±1% FS, and the output signal thereof adjusts the combustion aid flow through a PID (proportional-integral-derivative) controller.

[0055] The physical lookup table 301 of the closed-loop control module 30 stores the mapping relationship between fuel type, combustion condition, exhaust component and optimal combustion parameters, and the table structure includes fields such as fuel type, combustion efficiency range, optimal air-fuel ratio, fuel flow coefficient and combustion air flow, covering the optimal control parameters of different fuel types such as natural gas, liquefied gas and refinery gas under different combustion efficiency ranges. When it is detected that the current working condition parameters match the known parameter combination in the physical lookup table 301, the closed-loop control module 30 automatically calls the corresponding optimal control parameters to realize closed-loop feedback control. When it is detected that the current working condition parameters match the known parameter combination in the physical lookup table 301, the closed-loop control module 30 automatically calls the corresponding optimal control parameters to realize closed-loop feedback control. For example, in terms of dynamic combustion control, the closed-loop control module 30 calls the physical lookup table 301, matches the current fuel type (such as natural gas) and combustion condition (such as lean oxygen / rich oxygen), and generates optimal control parameters such as combustion air flow .

[0056] In an optional embodiment, the closed-loop control module 30 is also configured with an interpolation calculation unit 302. When the current working condition parameters do not match the completely corresponding parameter combination in the physical lookup table 301, the interpolation calculation unit 302 uses a bilinear interpolation algorithm to perform linear interpolation calculation based on the nearest two known parameter combinations to generate control parameters adapted to the current working condition. For example, if the current working condition does not match the parameters in the physical lookup table 301, the interpolation calculation unit 302 performs bilinear interpolation based on the nearest two known parameters, such as 、 , to generate adaptive parameters such as For extreme working conditions beyond the known parameter range, the control strategy of the nearest neighbor parameter combination is used for conservative control.

[0057] The difference calculation and basic control unit 303 is used to receive the target control parameters generated by the interpolation calculation unit 302, calculate the deviation from the current actual parameters, and generate control output using a proportional regulation algorithm with a proportional gain = 1.2. When the deviation exceeds ±10%, the unit limits the control output change rate to not more than 5% / s, ensuring system stability.

[0058] The actuator interface 304 is the output interface of the closed-loop control module 30, connected to the mass flow meter 305 and the combustion aid control valve 306, which converts the control command into standard signal output to realize accurate regulation of fuel flow and combustion air flow.

[0059] The auxiliary function module 40 further includes an environmental compensation unit 401 and a data storage unit 402. The environmental compensation unit 401 is connected to an environmental temperature and humidity sensor 403 and an atmospheric transmittance calculation unit 404. The atmospheric transmittance calculation unit 404 calculates the atmospheric attenuation coefficient in real time based on the MODTRAN radiation transfer model and feeds back to the physical model calculation module 20, realizing dynamic compensation of environmental factors on the measurement results. The data storage unit 402 stores historical data in a database, including time stamp, temperature distribution, combustion efficiency, control parameter and other information, providing data support for system operation analysis and fault diagnosis, etc.

[0060] Embodiment Two: Dual-link Enhanced Detection System Based on the infrastructure of Embodiment One, this embodiment further expands the system function. As shown in Figure 2 The embodiment further introduces an artificial intelligence calculation link to form a dual-link architecture in which the physical model and the artificial intelligence model work in parallel, significantly improving the measurement accuracy and system robustness. The enhanced monitoring system further includes an artificial intelligence calculation module 50, a data fusion and correction module 60, and a function expansion of the multi-spectral imaging module 10 based on Embodiment One.

[0061] The multi-spectral imaging module 10 further includes a dynamic spectrum selection module 104, which includes a tunable filter array 1041 and a principal component analysis unit 1042. The tunable filter array 1041 is preferably an acousto-optic tunable filter AOTF, with an adaptive wavelength range of 3-8 μm and a spectral resolution of less than or equal to 10 nm. The principal component analysis unit 1042 optimizes the spectral band combination in real time, dynamically selects the optimal spectral band combination according to the current fuel type (such as natural gas / liquefied gas / heavy hydrocarbon) and the combustion state (such as complete / incomplete), and in some optional embodiments, such as 4.3 μm for monitoring CO2 peak, 3.4 μm for monitoring CH group absorption band, and 1.6 μm for monitoring H2O absorption band.

[0062] Specifically, the principal component analysis unit 1042 calculates the contribution of each principal component through the principal component analysis algorithm, and the first and second principal components are linear combinations of each wavelength. According to different combustion states, the optimal waveband is selected, 4.3 μm and 1.6 μm waveband for complete combustion, and 3.4 μm and 2.0 μm waveband for incomplete combustion. Compared with the fixed waveband design, the dynamic spectrum selection module 104 optimizes the waveband configuration in real time through online spectrum analysis, reduces the interference of soot and improves the signal-to-noise ratio.

[0063] The artificial intelligence calculation module 50 adopts a multi-scale visual Transformer network for comprehensive processing of multi-spectral radiation data and joint estimation of combustion efficiency and gas concentration to obtain the second combustion state parameter. The multi-scale visual Transformer network includes a multi-scale feature extraction module, a pooling attention mechanism, a local and global attention mechanism, and a feedforward neural network.

[0064] The pooling attention mechanism reduces the resolution by pooling the query tensor, the key tensor, and the value tensor, and reduces the computational complexity by performing a pooling operation on the query tensor, the key tensor, and the value tensor. Specifically, the resolution is reduced by pooling the query tensor Q, and the computational complexity is reduced by pooling the key-value tensor (K, V), and the calculation formula is: (5) Where Q, K, and V represent the query tensor, the key tensor, and the value tensor, respectively, and d is the feature dimension. The pooling attention mechanism performs a pooling operation on Q, K, and V to significantly reduce the computational complexity while maintaining the feature expression capability.

[0065] The local and global attention mechanism includes local attention that performs self-attention calculation within an n x n window, global attention that extracts global features through a sparse uniform grid, and axial attention that handles long-range dependencies between channels, where n is a natural number greater than or equal to 4.

[0066] Further, the artificial intelligence calculation module 50 integrates a dynamic axial attention module that achieves a dynamic balance between spatial resolution and computational efficiency through learnable grid division parameters. Specifically, it introduces a grid division parameter a based on window attention. When a = 0.5, the feature map is divided into a 2 x 2 grid and the attention weight is calculated independently.

[0067] The data fusion and correction module 60 receives the output data of the physical model calculation module 20 and the artificial intelligence calculation module 50, respectively, and uses an adaptive algorithm to calibrate the two calculation links in real time. The data fusion and correction module 60 includes a physical link calibration subunit 601, an artificial intelligence link subunit 602, an adaptive parameter adjustment subunit 603, a dual-link output fusion subunit 604, and an online calibration triggering subunit 605.

[0068] The physical link calibration subunit 601 uses an improved Wien radiation formula, which adds an emissivity temperature dependence correction term and an environmental radiation compensation term based on the standard Wien approximation formula, and the calculation formula is: (6) Where for the corrected spectral radiance, for the temperature-dependent emissivity correction function, for the ambient radiation compensation term, h is the Planck constant, c is the speed of light, λ is the wavelength, k is the Boltzmann constant, and T is the absolute temperature. Further, the physical link calibration subunit 601 calculates the systematic deviation of the physical model based on the first-layer physical model output in combination with historical deviation data through a moving average algorithm, and realizes the calibration coefficient K phy is dynamically updated, and the calculation formula is

[0069] , (7) where K phy(t) is the calibration coefficient at the current time t, K phy(t-1) is the calibration coefficient at the previous time, is the learning rate, which is 0.01-0.05, T ref is the reference temperature value, T phy is the temperature value output by the physical model, (T ref -T phy ) is the temperature deviation, and the first-layer physical model output is calibrated in real time based on the above formula. Further, the physical link calibration subunit 601 uses an improved Kalman filter to fuse the calibrated physical model output with other sensor data, and the weight coefficient is dynamically adjusted according to the combustion stability. In some optional embodiments, the weight scheme adopts the following: when the combustion efficiency CE≥95%, the system gives priority to the artificial intelligence model output; when CE<85%, the system gives priority to the physical model output, to ensure measurement reliability.

[0070] The artificial intelligence link subunit 602 uses an xViT++ spatiotemporal attention network mechanism to monitor the stability of the artificial intelligence model output. For example, when the standard deviation of 10 consecutive sampling points exceeds 20K, the calibration program is started, and the bias parameters of the last layer of the network are fine-tuned through a backpropagation algorithm. The artificial intelligence link subunit 602 integrates a dynamic axial attention module, which realizes dynamic balance between spatial resolution and computational efficiency through learnable grid division parameters. Specifically, it includes introducing a grid division parameter α based on window attention. In some optional embodiments, when α=0.5, the feature map is divided into ≥2×2 grids and the attention weight is calculated independently.

[0071] The adaptive parameter adjustment subunit 603 as the core parameter management unit of the data fusion and correction module 60, the input end is connected feedback output end of physical link calibration subunit 601 and artificial intelligence link subunit 602 respectively, receive the combustion condition monitoring data from multispectral imaging module 10. The adaptive parameter adjustment subunit 603 according to the combustion efficiency variation amplitude, flame temperature fluctuation rate and double link calibration precision difference, dynamic adjustment key parameter of each calibration subunit. According to the change of combustion condition dynamic adjustment calibration parameter, when the combustion efficiency changes more than 5%, reevaluate the applicability of calibration coefficient, if necessary, trigger parameter reset.

[0072] In double link parallel processing, the double link output fusion subunit 604 adopts weighted average algorithm to fuse the output after calibration of the physical link calibration subunit 601 and the artificial intelligence link subunit 602, and the weight coefficient is dynamically allocated according to the historical accuracy of each link. According to the combustion efficiency CE, the weight is dynamically adjusted: when CE≥95%, w=0.3, focusing on artificial intelligence model; When CE<85%, w=0.7, focusing on physical model.

[0073] In some optional embodiments, the online calibration trigger subunit 605 is configured with a laser-induced fluorescence (LIF) calibration module, and the trigger conditions include: such as temperature fluctuation exceeds threshold (such as more than ±50K / s), or smoke concentration exceeds the preset threshold (such as 500mg / m³). When the temperature fluctuation exceeds the threshold, the laser-induced fluorescence (LIF) calibration module is automatically started for instantaneous spectral calibration.

[0074] Further explanation, the double link output fusion subunit 604 of the data fusion and correction module 60 is used to fuse and correct the first combustion state parameter output by the physical model calculation module 20 and the second combustion state parameter output by the artificial intelligence calculation module 50 by using Kalman filtering algorithm, to output the final combustion state parameter. The data fusion and correction module 60 realizes double model data fusion through Kalman filtering unit of double link output fusion subunit 604, specifically including: Establishing state space model, taking the first combustion state parameter output by the physical model calculation module 20 as the system state prediction value , taking the second combustion state parameter output by the artificial intelligence calculation module 50 as the observation value z k ; According to the combustion efficiency CE, the observation noise covariance matrix R is dynamically adjusted: when CE≥95%, reduce the value of R to make the Kalman gain K k Bias towards artificial intelligence model observation, when CE<85%, increase the value of R to make the Kalman gain K k Bias towards physical model prediction; The Kalman filter fusion unit adopts a Kalman filter algorithm to fuse and correct the outputs of the physical model calculation module 20 and the artificial intelligence calculation module 50, thereby reducing single measurement error and improving overall measurement accuracy.

[0075] The Kalman filter algorithm is used to fuse and update the measurement results of the multispectral system at different time points, The Kalman gain is calculated according to the observation matrix H, the observation noise covariance matrix R, and the current state covariance matrix P k|k-1 , and the optimal Kalman gain is calculated as follows: (8), The state update equation is as follows:

[0076] wherein, is the current state estimate, is the current observation value, is the Kalman gain, is the observation matrix, is the observation noise covariance matrix, is the state estimate covariance matrix.

[0077] The Kalman filter algorithm realizes the fusion and updating of the measurement results of the multispectral imaging module 10 at different time points, reduces single measurement error, and improves overall measurement accuracy. This method can control single measurement error within 1%, and overall error below 1%, so as to achieve a level close to and beyond that of the VISR device.

[0078] When the output deviation of the physical model and the artificial intelligence model exceeds the preset threshold, the data fusion and correction module 60 triggers the online calibration triggering subunit 605 to start the laser-induced fluorescence calibration program.

[0079] The artificial intelligence calculation module 50 and the physical model calculation module 20 constitute a double-link verification mechanism. The artificial intelligence calculation module 50 adopts an artificial intelligence multi-physical factor link based on the Transformer architecture and has high accuracy. The physical model calculation module 20 adopts a physical derivation link based on a radiation transmission model and has strong generalization. Through cross verification of the two links, the system can automatically select the output result that best meets the industrial demand. When the combustion efficiency is greater than or equal to 95%, the system gives priority to the artificial intelligence model output; when the combustion efficiency is less than 85%, the system gives priority to the physical model output, ensuring measurement reliability.

[0080] By the above technical solution, the embodiment compared to the first embodiment achieves significant performance improvement. The temperature measurement accuracy, combustion efficiency measurement accuracy and system response time can be improved. The environmental adaptability is significantly improved, and stable performance can be maintained in complex environments such as rain and fog weather, strong background light interference, etc.

[0081] Embodiment three: complete system integration and dual-flow fusion perception On the basis of the first embodiment and the second embodiment, the embodiment further provides an optimized system integration scheme. As shown in Figure 3A The multi-spectral infrared and visible light imaging torch combustion monitoring and optimization control system further includes a multi-spectral imaging module 10, a physical model calculation module 20, an artificial intelligence calculation module 50, a data fusion and correction module 60, a closed-loop control module 30, and a data storage and visualization module 80.

[0082] The multi-spectral imaging module 10 includes a non-cooled infrared camera 101, a visible light camera 102, a multi-spectral registration unit 103, and an image acquisition and transmission module 140. The working wavelength of the non-cooled infrared camera 101 is preferably 3-5 microns and 8-12 microns, which is used to acquire infrared radiation images of the flame. The visible light camera 102 uses a zoom lens with more than one million pixels, and the working wavelength is visible light, which is used to acquire visible light images of the flame. The lens focal length range is preferably 25-100mm, and the working wavelength is 400-700nm. In flame monitoring, the visible light radiation information in the range of 0.4-0.7μm is mainly acquired to acquire visible light images of the flame. The multi-spectral registration unit 103 is used to spatially register the infrared image and the visible light image to ensure consistent image coordinates. The image acquisition and transmission module 140 is used for image data acquisition, compression and transmission, and supports remote monitoring.

[0083] Further, the physical model calculation module 20 and the artificial intelligence calculation module 50 are connected in parallel to the output end of the multi-spectral imaging module 10, forming a double-link calculation architecture; the physical model calculation module 20 calculates the first combustion state parameter based on the Planck radiation formula and the gray body radiation model, the artificial intelligence calculation module 50 outputs the second combustion state parameter through a multi-scale visual Transformer network 310, and the temperature distribution matrix of the physical model calculation module 20 is used as the supervised learning input feature of the artificial intelligence calculation module 50. The input end of the data fusion and correction module 60 is connected to the output end of the physical model calculation module 20 and the artificial intelligence calculation module 50.

[0084] Further, referring to Figure 3B The closed-loop control module 30 further includes an improved PID controller 307, and the control target function is:

[0085] wherein is the temperature standard deviation, is the average temperature.

[0086] The control objective function further comprises an emission concentration penalty term:

[0087] wherein γ is a penalty coefficient, is the current nitrogen oxide concentration, is the emission limit value.

[0088] The input of the improved PID controller 307 is connected to the output of the data fusion and correction module 60, and the output of the improved PID controller 307 is connected to the mass flow meter 305 and the combustion auxiliary control valve 306 through the flow control interface 3041. The improved PID controller 307 further comprises a feedforward compensation unit 307a, a dynamic weight adjustment unit 307b, a multivariate coupling control interface 307c, an emission constraint optimization unit 307d, and a hybrid control strategy generation unit 307e, which realize information interaction through an internal data bus.

[0089] The input of the feedforward compensation unit 307a is connected to the predicted output of the physical lookup table 301, and a pre-adjustment control amount is generated based on the matching results of fuel type and combustion conditions: the higher the fuel heat value, the larger the pre-adjustment amount; the higher the environmental humidity (such as rainy days), the pre-adjustment amount is correspondingly reduced to compensate for the influence of humidity on combustion efficiency.

[0090] The input of the dynamic weight adjustment unit 307b is connected to the output of the data fusion and correction module 60 to extract the combustion efficiency CE parameter, which is used to dynamically adjust the proportional gain coefficient according to the fluctuation amplitude of the combustion efficiency. The dynamic weight adjustment unit 307b uses the combustion efficiency feedback value CE as a weight adjustment factor: when the combustion efficiency exceeds 95%, the proportional gain coefficient Kp is reduced to 80% of the baseline value to avoid excessive adjustment; when the flame temperature fluctuates violently (such as changing more than 150K per second), the integral time T_i is shortened to 50% of the baseline value to improve the response speed.

[0091] The multivariate coupling control interface 307c, whose output is respectively connected to the mass flow meter 305 and the combustion auxiliary control valve 306, uses a decoupling matrix D to synchronously adjust the fuel flow and the combustion air flow.

[0092] The emission constraint optimization unit 307d is used to trigger a step response of the combustion auxiliary control valve 306 when the corrected emission concentration exceeds the standard value, so that the excess air coefficient increases, and it executes the following control logic.

[0093] The emission concentration is derived indirectly by the multispectral imaging module 10 and the physical model calculation module 20, or the laser-induced fluorescence (LIF) calibration module of the online calibration trigger subunit 605 is used to calibrate the multispectral imaging module 10 instantaneously, so as to obtain the original emission concentration, and the corrected concentration is calculated by combining the environmental humidity correction factor and the atmospheric attenuation coefficient; When the corrected emission concentration exceeds the standard value, the step response of the combustion aid control valve 306 is triggered, so that the excess air coefficient is increased to the interval of 1.15-1.25; The stability index of the flame temperature distribution and the emission concentration penalty term are combined to generate a control target function:

[0094] Wherein, J represents the value of the control target function, and represents the overall performance index of the system; α represents a stability weight coefficient, which is used to adjust the importance of the temperature stability in the target function, and a typical value is 0.3-0.7; the stability index is based on the standard deviation of the flame temperature And the average temperature is calculated, which is used to normalize the temperature fluctuation degree, The higher the value is, the greater the combustion fluctuation is; β represents an emission concentration weight coefficient, which is used to adjust the importance of the emission control in the target function, and a typical value is 0.2-0.5, The square sum of the concentration ratios of various pollutants is used to comprehensively evaluate the emission compliance; γ represents an emission penalty coefficient, which is used to dynamically adjust the penalty strength when the emission exceeds the standard, and a typical value is 1.0-3.0; the emission penalty term is dynamically adjusted according to the difference between the current nitrogen oxide concentration And the emission limit value When the difference approaches or exceeds , the penalty coefficient γ is segmented to enhance the control strength; the optimal control instruction under the constraint condition is solved by using an optimization algorithm, so as to ensure that the fuel flow rate change rate does not exceed a safety threshold.

[0095] The mixed control strategy generation unit 307e sets a control mode switching threshold: when the predicted combustion efficiency output by the improved PID controller 307 is lower than 85% or the control instruction mutation rate exceeds the threshold, such as 10% / s, the control mode is automatically switched to the conservative control mode. At this time, the fuel flow is limited to 90%-95% of the matching parameter of the physical lookup table 301, the combustion air flow is dynamically adjusted according to the safe air-fuel ratio, and when the emission concentration exceeds the standard value, the fuel flow is further limited to 85%-90% of the matching value.

[0096] Through the above technical scheme, the improved PID controller 307 realizes the collaborative optimization of combustion stability and emission constraint, and can simultaneously ensure the operation safety in extreme working conditions through the conservative control mode.

[0097] Reference Figure 3A The data storage and visualization module 80 is connected to the output ends of the multispectral imaging module 10, the physical model calculation module 20, the artificial intelligence calculation module 50, and the data fusion and correction module 60, respectively, and its data storage unit stores historical data and a combustion efficiency trend chart, and the visualization and alarm unit generates a flame temperature field thermal map and triggers a combustion abnormality alarm signal.

[0098] Further disclosed is a closed-loop optimization process, the closed-loop control module 30 and the environmental compensation unit 401 establish a bidirectional data channel, dynamically correct the control parameters according to the environmental temperature and humidity and the atmospheric attenuation coefficient output by the atmospheric transmittance calculation unit 404, and online calibration is connected to the weighted Gaussian correction unit 202 of the physical model calculation module 20 and the model training and optimization unit of the artificial intelligence calculation module 50; Based on the above method, the multispectral infrared and visible light imaging torch combustion monitoring and optimization control system forms a closed-loop optimization process, that is, the registered image data collected by the multispectral imaging module 10 is processed in parallel by the physical model calculation module 20 and the artificial intelligence calculation module 50, and the final combustion parameters output by the data fusion and correction module 60 drive the closed-loop control module 30 to adjust the combustion parameters, and the adjusted environmental parameters are fed back to the multispectral imaging module 10 to form a closed-loop iteration, and at the same time, when the temperature fluctuation rate exceeds ±50K / s or the smoke concentration exceeds 500mg / m³, etc. Threshold value triggers the laser-induced fluorescence calibration program.

[0099] To solve the problem that the precision of the prior art is limited when processing single modal data, and it is difficult to maintain robustness under complex environmental lighting or shielding conditions, the multispectral imaging module 10 further includes a dual-flow fusion perception system based on Transformer, and its network topology is as shown in Figure 3C The system includes the following functional modules: Among them, the visual embedding module 111 (Visual Embedding, VE) is used to convert the input visible light image into a first feature vector sequence; the infrared embedding module 112 (Infrared Embedding, IE) is used to convert the input infrared image into a second feature vector sequence. By mapping the visible light image and infrared image captured by the visible light camera 102 and the uncooled infrared camera 101 to a "common semantic space" of the same dimension and the same sequence length, the modal differences are eliminated, the respective unique information is retained, and an aligned token sequence is provided for subsequent multi-head attention fusion. The multi-head attention mechanism 113 (Multi-head Attention, MHA) acts on the first feature vector sequence and the second feature vector sequence respectively to capture cross-pixel / cross-channel dependencies; the layer normalization module 114 (Layer Normalization, LN) is set before and after each MHA and each feedforward network to stabilize training; the feed-forward network module 115 (Feed-Forward Network, FFN) is used to perform nonlinear transformation on the LN output; the random dropout module 116 (Dropout, DO) is set after the multi-head attention mechanism 113 (MHA) and the feedforward network module 115 (FFN) to prevent overfitting; the fusion output module 117 (Output) is used to generate the target task result based on the fused feature vector sequence; the loss function module 118 (Loss Function, LF) is used to calculate the error between the predicted result and the true value during the training phase through gradient backpropagation and update the network parameters in a backpropagation manner.

[0100] The weighted Gaussian correction unit 202 performs weighted Gaussian correction on the measurement results of multiple spectral channels to improve the calculation accuracy. This unit adopts the same weight matrix calculation method as the weighted Gaussian correction unit 202 in the first embodiment, as shown in formula (3).

[0101] Specifically, in this embodiment, the artificial intelligence computing module 50 employs a multi-scale visual Transformer network 310 for feature extraction and combustion efficiency calculation. This network, based on the MultiScale-ViTs architecture, achieves hierarchical representation of multi-scale features by gradually expanding the channel width D and reducing the sequence length L. This design overcomes the accuracy loss caused by simplifying assumptions in traditional radiation multispectral physics formulas (such as assuming a constant emissivity and ignoring ambient radiation interference).

[0102] The multi-scale visual Transformer network 310 includes the following architecture variants: Group42: The input is divided into two parts, including an RGB wavelength channel and an infrared channel; Group43: input includes two RGB wavelength channels and one infrared channel; Group42s: contains one RGB wavelength channel, one infrared channel, and randomly samples one RGB channel through dropout; Group43s: contains two RGB wavelength channels, one infrared channel, and randomly samples one RGB channel through dropout; Group44: contains all RGB and infrared wavelength channels. The preferred embodiment of the present application adopts the Group43s architecture, which has good computational efficiency while maintaining high accuracy.

[0103] The MultiScale-ViTs architecture in the present embodiment adopts a parallel / serial hybrid connection method, which specifically includes: three different types of pooling attention organized through parallel and serial jump connections; three serial FFN modules to realize step-by-step transformation of features; and mixed use of Group31, Group32, and Group33 modules.

[0104] In practical applications, the network realizes high-precision temperature measurement through the following steps: using MultiScale-ViTs to calibrate the measurement data at a specific time; obtaining intermediate results through optimal fitting; updating the state in combination with Kalman filtering; and outputting the final Transformer link prediction value.

[0105] In the present embodiment, the artificial intelligence computing module 50 and the physical model computing module 20 constitute a double-link verification mechanism. The artificial intelligence computing module 50 adopts an artificial intelligence multi-physical factor link based on the above-mentioned Transformer architecture, which has high accuracy; the physical model computing module 20 adopts a physically derived link based on a radiation transmission model, which has strong generalization. Through cross-validation of the two links, the multi-spectral infrared and visible light imaging torch combustion monitoring and optimization control system can automatically select the output result that best meets the industrial requirements. When ≥95%, the system gives priority to the artificial intelligence model output; when CE<85%, the system gives priority to the physical model output, ensuring measurement reliability.

[0106] The above technical solution realizes a breakthrough in temperature measurement accuracy and control performance under the premise of ensuring industrial reliability through the construction of a "physically constrained AI + AI enhanced physics" two-way optimization closed loop, providing a reliable temperature field input for the subsequent closed-loop control module 30.

[0107] System workflow As Figure 4 shown, the workflow of the present system is further disclosed, including: At least one multi-spectral imaging module 10 is provided to collect visible light and infrared band information, The physical model calculation module 20 process includes: Step S11: The Planck formula calculation unit 201 receives the visible light and infrared band radiation intensity data collected by the multi-spectral imaging module 10, and calculates the initial temperature distribution based on formulas (1) and (2); Step S12: The multi-band results are fused based on the weighted Gaussian correction unit 202, and the weight matrix W is dynamically adjusted according to the signal-to-noise ratio of each band; Step S13: The combustion efficiency calculation unit 203 based on the Transformer architecture extracts cross-spectral features through a self-attention mechanism, outputs a corrected temperature distribution matrix, and outputs a first combustion state parameter; In some optional embodiments, the error of the corrected temperature distribution matrix in this step S3 is ≤±50K.

[0108] The artificial intelligence calculation module 50 process is executed in parallel, including: Step S21: The multi-scale visual Transformer network receives the same multi-spectral radiation data, and performs initial feature extraction through a multi-scale feature extraction module; Step S22: The pooling attention mechanism realizes resolution reduction through pooling query tensor Q, key tensor K, and value tensor V, and calculates according to the formula Attention(Q,K,V) = softmax(QK^T / √d)V; Step S23: The local and global attention mechanism performs self-attention calculation within an n×n window, and simultaneously extracts global features through a sparse uniform grid, and outputs a second combustion state parameter; The data fusion and correction module 60 process includes: Step S31: A state space model is established, and the first combustion state parameter output by the physical model calculation module 20 is taken as the system state prediction value The second combustion state parameter output by the artificial intelligence calculation module 50 is taken as the observation value z k ; Step S32: The observation noise covariance matrix R is dynamically adjusted according to the combustion efficiency CE, and when CE≥95%, the R value is reduced to make the Kalman gain K k biased towards the artificial intelligence model observation, and when CE<85%, the R value is increased to make the Kalman gain K k biased towards the physical model prediction; Step S33: The final combustion state parameter after fusion is output through the state update equation ; When the output deviation between the physical model and the artificial intelligence model exceeds the preset threshold, an online calibration process is performed: Step S41: When the temperature fluctuation exceeds the threshold, such as ±50K / s, or the soot concentration exceeds the threshold, such as 500mg / m³, the online calibration trigger subunit 605 automatically starts the laser-induced fluorescence calibration module; Step S42: The laser-induced fluorescence calibration module performs instantaneous spectral calibration on the multi-spectral imaging module 10, and updates the response sensitivity coefficients of each spectral channel; Step S43: Based on the calibration result, the weight matrix of the weighted Gaussian correction unit 202 is updated to ensure that the measurement accuracy is maintained within ±2%; Subsequently, the working process of the closed-loop control module 30 is entered, including: Step S51: Based on the physical lookup table 301, the optimal control parameters are generated by matching the current fuel type and combustion conditions; Step S52: The interpolation calculation unit 302 performs bilinear interpolation based on the nearest neighbor parameters to generate adaptive parameters; Step S53: The improved PID controller 307 performs multivariate coupled control according to the final combustion state parameters, and the feedforward compensation unit 307a generates a pre-adjustment control amount according to the matching results of the fuel type and the combustion conditions; The dynamic adjustment process of the improved PID controller 307 includes: Step S54: The dynamic weight adjustment unit 307b uses the combustion efficiency feedback value CE as a weight adjustment factor, when CE≥95%, the proportional gain coefficient Kp is reduced to 80% of the baseline value, and when CE<85%, the integral time Ti is shortened to 50% of the baseline value; Step S55: The emission constraint optimization unit 307d monitors the corrected emission concentration, and when it exceeds the standard value, triggers the step response of the combustion aid control valve 306, so that the excess air coefficient increases to the interval of 1.15-1.25; Step S56: The multivariate coupled control interface 307c uses the decoupling matrix D to synchronously adjust the fuel flow and the combustion air flow, and under normal working conditions, the coefficients are d11=1.0, d12=-0.15×theoretical air-fuel ratio, d21=0.08×theoretical air-fuel ratio, and d22=1.0; Finally, the execution and feedback process includes: Step S61: The actuator interface 304 sends control instructions to the mass flow meter 305 and the combustion aid control valve 306 to adjust the combustion aid flow, realizing closed-loop feedback control; Step S62: The adjusted combustion parameters are fed back to the multi-spectral imaging module 10 through the environmental compensation unit 401, forming a complete closed-loop iteration; Step S63: The data storage and visualization module 80 records system operation data in real time, generates a thermal map of the flame temperature field, and triggers an alarm signal when an abnormal combustion is detected. Step S64: Trigger the laser-induced fluorescence calibration procedure when the temperature fluctuation rate exceeds the threshold.

[0109] Through the above steps, high-precision, real-time monitoring and closed-loop optimization control of the combustion state of the detected object by the torch are achieved. Through the above complete workflow, the system forms a "collection-computation-fusion-control-feedback" closed-loop optimization mechanism, ensuring the continuous optimization of the combustion efficiency of the combustible material in the torch.

[0110] Further, with reference to Figure 5 In some optional embodiments of the present application, a torch combustion monitoring and optimization control method based on multispectral infrared and visible light imaging is provided, which includes the following steps: Step one, acquire flame radiation data through a multispectral imaging system; Step two, calculate the initial temperature distribution based on the Planck formula and the gray body model; Step three, perform multispectral data fusion correction using the weighted Gaussian algorithm and the Levenberg-Marquardt algorithm; Step four, calculate the combustion efficiency through a Transformer model; Step five, input the corrected temperature data into the combustion optimization control module; Step six, generate control instructions based on the physical lookup table and the flowmeter data; Step seven, adjust the combustion parameters through the actuator and return to the monitoring step.

[0111] Further, in step four "calculate the combustion efficiency through a Transformer model", the present embodiment adopts a multiscale visual Transformer network 310 as the specific implementation manner, which receives the fusion correction results of the weighted Gaussian algorithm and the Levenberg-Marquardt algorithm in step three as input.

[0112] The multiscale visual Transformer network 310 can adopt one of the following architecture variants: Group42 architecture: configure one RGB wavelength channel and one infrared channel input branch, suitable for resource-constrained scenarios; Group43 architecture: configure two RGB wavelength channels and one infrared channel input branch, suitable for scenarios requiring higher accuracy; Group42s architecture: increase the dropout random sampling mechanism based on Group42 to improve the robustness of the model; Group43s architecture: increase dropout random sampling mechanism on the basis of Group43, maintain high precision while having good computational efficiency; Group44 architecture: configure all RGB and infrared wavelength channels, suitable for scenarios that are not sensitive to computing resource requirements. Any of the above architecture variants are fully matched in data structure with the multispectral data fusion result of step three.

[0113] The combustion efficiency calculation process includes the following key calculation modules: Pooling attention mechanism, which reduces the computational complexity while maintaining feature expression ability; and Local and global attention mechanism, further including: local window attention, for capturing spatial local features within an n×n window, n being a natural number ≥4; global grid attention, for extracting long-range dependencies; and axial attention, for processing inter-channel correlations.

[0114] The output of the reconstruction result adopts a double-link check mechanism. Compare the reconstructed temperature field with the initial temperature distribution calculated based on the Planck formula in step two. When the combustion efficiency CE≥95%, the system gives priority to the artificial intelligence model output; when CE<85%, the system gives priority to the physical model output. Ensure that the output result meets the input requirements of the combustion optimization control module in step five.

[0115] The reconstructed temperature field data is transmitted to step five through a standardized interface to provide accurate temperature field input for subsequent combustion optimization control.

[0116] In summary, the present scheme uses multispectral image fusion technology to obtain multispectral image information of the flame through the registration of the uncooled infrared camera and the visible light camera. This technology can maintain high image quality and measurement accuracy under different combustion states. Based on the Planck radiation formula and the gray body radiation model, combined with multispectral image data, the flame temperature is calculated. For the measurement results of multiple spectral channels, a weighted Gaussian formula is used for correction to improve the calculation accuracy.

[0117] Through the above technical solutions, the present invention achieves the following technical effects: 1. Cost reduction: replace traditional high-cost equipment with a multispectral imaging system, reducing the hardware cost of the flare monitoring system by more than 40%; 2. Precision improvement: use the algorithm architecture combining the Transformer model and Kalman filtering to control the flame temperature measurement error within ±50K, achieving the same precision level as the VISR device; 3. Comprehensive coverage: through the collaborative work of the physical lookup table and the interpolation algorithm, it can cover 100% of the normal combustion conditions and 98% of the abnormal combustion conditions control requirements; 4. Wide applicability: The system supports real-time monitoring and optimal control of flare devices for various fuel types (natural gas, refinery gas, syngas, etc.) and different combustion conditions (lean / oxy-combustion), meeting the increasingly stringent environmental regulations.

[0118] The above describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection claimed by the present application is defined by the appended claims and their equivalents.

Claims

1. A torch combustion monitoring and optimization control system based on multi-spectral infrared and visible light imaging, characterized in that: include: Multispectral imaging module, used to synchronously collect infrared and visible light images of the torch flame and perform spatial registration to obtain multispectral radiation data of the flame; a physical model calculation module, configured to calculate a first combustion state parameter according to the multispectral radiation data based on the Planck radiation formula and the gray body radiation model; an artificial intelligence computing module, which uses a pre-trained multi-scale visual Transformer network to calculate a second combustion state parameter based on the multispectral radiation data; a data fusion and correction module, configured to fuse and correct the first combustion state parameter and the second combustion state parameter using a Kalman filter algorithm to output a final combustion state parameter; and The closed-loop control module is used to dynamically adjust the operating parameters of the flare according to the final combustion state parameters to optimize the combustion efficiency.

2. The flare combustion monitoring and optimization control system according to claim 1, characterized in that: The multispectral imaging module includes: An uncooled infrared camera, whose operating bands cover 3-5 microns and 8-12 microns, is used to collect infrared radiation images of the flame; a visible light camera for collecting visible light images of the flame; and A multispectral registration unit is used to spatially register the infrared radiation image with the visible light image to generate an aligned flame position array.

3. The flare combustion monitoring and optimization control system according to claim 2, characterized in that: The multispectral imaging module also includes a dynamic spectrum selection module, which includes a tunable filter array and a principal component analysis unit. The principal component analysis unit dynamically selects the optimal spectral band combination based on the current fuel type and combustion state to guide the tunable filter array to perform imaging.

4. The flare combustion monitoring and optimization control system according to claim 3, characterized in that: The system also includes an environmental compensation module, which is connected to an environmental temperature and humidity sensor and calculates the atmospheric attenuation coefficient in real time based on the MODTRAN radiation transfer model to dynamically compensate for the data collected by the multispectral imaging module.

5. The flare combustion monitoring and optimization control system according to claim 1, characterized in that: The physical model calculation module includes a weighted Gaussian correction unit, which establishes a weight matrix according to the response sensitivity coefficient of each spectral channel and uses the Levenberg-Marquardt algorithm to perform nonlinear optimization on the first combustion state parameter based on multi-band calculation.

6. The flare combustion monitoring and optimization control system according to claim 1, characterized in that: The multi-scale visual Transformer network includes a multi-scale feature extraction module, a pooling attention mechanism, and a local and global attention mechanism, where: The pooled attention mechanism achieves resolution reduction by pooling query tensors, key tensors, and value tensors, and reduces computational complexity by performing pooling operations on query, key, and value tensors. The local and global attention mechanism includes local attention that performs self-attention calculation within an n×n window and global attention that realizes global feature extraction through a sparse uniform grid, where n is a natural number ≥4.

7. The flare combustion monitoring and optimization control system according to claim 1, characterized in that: The data fusion and correction module realizes dual-model data fusion through the Kalman filter unit, specifically including: Establishing a state space model, using the first combustion state parameter output by the physical model calculation module as a system state prediction value, and using the second combustion state parameter output by the artificial intelligence calculation module as an observation value; The observation noise covariance matrix R is dynamically adjusted according to the combustion efficiency CE: when CE ≥ 95%, the R value is reduced to make the Kalman gain biased towards the artificial intelligence model observation; when CE < 85%, the R value is increased to make the Kalman gain biased towards the physical model prediction.

8. The flare combustion monitoring and optimization control system according to claim 1, characterized in that: The closed-loop control module includes: A physical lookup table storing a mapping relationship between fuel type, combustion conditions, and optimal combustion parameters; an interpolation calculation unit, which, when detecting that the current operating condition parameters are not matched in the physical lookup table, uses a bilinear interpolation algorithm to perform linear interpolation calculation based on the nearest neighbor known parameter combination in the physical lookup table to generate control parameters adapted to the current operating condition; and An improved PID controller is used to generate control instructions according to the final combustion state parameters.

9. The flare combustion monitoring and optimization control system according to claim 8, characterized in that: The improved PID controller comprises: A feedforward compensation unit, used to generate a pre-adjustment control amount based on the matching result between the fuel type and the combustion conditions; A dynamic weight adjustment unit, configured to dynamically adjust a proportional gain coefficient according to a fluctuation range of combustion efficiency; and The multivariable coupling control interface has its output ends connected to the mass flow meter and the combustion aid control valve respectively, which are used to synchronously adjust the fuel flow and the combustion air flow.

10. The flare combustion monitoring and optimization control system according to claim 1, characterized in that: The system forms a closed-loop optimization process: The registration image data collected by the multispectral imaging module is processed in parallel by the physical model calculation module and the artificial intelligence calculation module; The final combustion parameters output by the data fusion and correction module drive the closed-loop control module to adjust the combustion parameters; and the adjusted environmental parameters are fed back to the multispectral imaging module to form a closed-loop iteration.

Images