Gas leakage quantitative calibration method and system based on multispectral feature fusion

Abstract

The application relates to the technical field of infrared thermal imaging, in particular to a gas leakage quantitative calibration method and system based on multispectral feature fusion, which comprises the following steps: configuring a multiband blackbody radiation source; building a simulated gas cloud environment and acquiring configuration parameters in real time, and classifying and collecting the configuration parameters to build a four-dimensional parameter space; acquiring infrared thermal imaging data corresponding to the simulated gas cloud environment, and associating the infrared thermal imaging data with the four-dimensional parameter space to form a structured sample set; sharing the structured sample set by using a double-channel 3D-CNN to extract radiation intensity features and cloud shape features, and splicing the radiation intensity features and the cloud shape features to obtain a fusion feature vector; performing nonlinear mapping on the fusion feature vector based on deep learning to output a concentration prediction value; and introducing an atmospheric transmission correction factor to correct the concentration prediction value to obtain an accurate concentration value. The application has the effect of improving the accuracy of gas leakage concentration calibration.

Description

Technical Field

[0001] This application relates to the technical field of infrared thermal imaging, and in particular to a method and system for quantitative calibration of gas leaks based on multispectral feature fusion. Background Technology

[0002] In industrial sectors such as petrochemicals, energy pipelines, and power equipment, gas leak monitoring is a crucial step in ensuring production safety, preventing explosions, and protecting the ecological environment. Among these, methane (… Sulfur hexafluoride () Due to the flammable, explosive, or strong greenhouse effect characteristics of gases such as flammable and explosive gases, the need for accurate quantitative monitoring of their leaks is particularly urgent.

[0003] Infrared thermal imaging technology has become one of the mainstream technologies for gas leak monitoring due to its advantages such as non-contact measurement, high real-time performance, and long-distance monitoring capabilities. Its core principle is to utilize the absorption characteristics of gas molecules to infrared radiation of specific wavelengths, and to capture the radiation difference between gas clouds and the background environment through an infrared thermal imager, thereby inverting the gas concentration.

[0004] However, existing infrared thermal imaging gas concentration calibration technology is limited in practical industrial applications due to the limitations of the calibration methods and interference from complex environments, making it difficult to achieve high-precision quantitative monitoring. The specific technical status is as follows:

[0005] The industry commonly uses the single-point blackbody calibration method, which only calibrates the temperature reading. However, the actual absorption process of infrared radiation by a gas follows the Lambert-Beer law, but its absorption effect changes with factors such as the length of the radiation transmission path and the motion state of gas molecules, forming a nonlinear mapping relationship between concentration and radiation. Single-point blackbody calibration can only correct the temperature measurement deviation of the equipment itself and cannot solve the nonlinear mapping process of concentration and radiation in the gas absorption path.

[0006] There are many interfering factors in the environment, such as water vapor / dust scattering, which can cause irregular drift in the intensity of the radiation signal received by the infrared thermal imager, thereby causing distortion of the grayscale value of the infrared image.

[0007] The shape of gas clouds is significantly affected by factors such as wind speed, wind direction, and ambient temperature, resulting in high deviations in the final concentration inversion. Summary of the Invention

[0008] To improve the accuracy of gas leak concentration calibration, this application provides a gas leak quantitative calibration method and system based on multispectral feature fusion.

[0009] Firstly, this application provides a quantitative calibration method for gas leakage based on multispectral feature fusion, employing the following technical solution:

[0010] A quantitative calibration method for gas leakage based on multispectral feature fusion includes the following steps:

[0011] Configure multi-band blackbody radiation sources;

[0012] A simulated gas cloud environment is built and configuration parameters are acquired in real time. The configuration parameters are then classified and collected to construct a four-dimensional parameter space.

[0013] Infrared thermal imaging data corresponding to the simulated gas cloud environment is acquired and correlated with the four-dimensional parameter space to form a structured sample set;

[0014] The structured sample set is shared using a dual-channel 3D-CNN to extract radiation intensity features and cloud morphology features, which are then concatenated to obtain a fused feature vector.

[0015] The fused feature vector is subjected to a deep learning-based nonlinear mapping to output a concentration prediction value;

[0016] An atmospheric transport correction factor is introduced to correct the concentration prediction value to obtain an accurate concentration value.

[0017] In some embodiments, a simulated gas cloud environment is constructed and configuration parameters are acquired in real time. These configuration parameters are then categorized and collected to construct a four-dimensional parameter space, including the following steps:

[0018] Configure the wind speed according to the adjustable turbulence generator, and configure the real-time temperature according to the temperature control air curtain;

[0019] Obtain the actual concentration value in the simulated gas cloud environment and the current position of the thermal infrared imaging device;

[0020] Each of the configuration parameters is collected and classified into a four-dimensional parameter space including gas concentration, observation distance, ambient temperature, and turbulence intensity. Based on the preset classification data volume, the corresponding data range is configured for each dimension parameter.

[0021] In some embodiments, acquiring infrared thermal imaging data corresponding to a simulated gas cloud environment and correlating it with the four-dimensional parameter space to form a structured sample set includes the following steps:

[0022] Acquire several frames of infrared thermal imaging images corresponding to the configuration parameters;

[0023] Calculate the grayscale matrix of each infrared thermal imaging image and construct a three-dimensional tensor with the corresponding frame time series;

[0024] The structured sample set is obtained by associating several of the three-dimensional tensors with the four-dimensional parameter space.

[0025] In some embodiments, a dual-channel 3D-CNN is used to share the structured sample set to extract radiance features, including the following steps:

[0026] The grayscale matrix is ​​corrected for distance attenuation based on the observation distance;

[0027] The grayscale matrix is ​​calibrated for radiometric grayscale based on the ambient temperature.

[0028] Feature extraction is performed on the corrected grayscale matrix based on 3D convolutional layers. Specifically...

[0029] The first layer extracts the local radiation intensity differences corresponding to the three-dimensional tensor, the second layer extracts the time variation trend of the radiation intensity, and the third layer extracts the spatial distribution of the radiation intensity.

[0030] In some embodiments, a dual-channel 3D-CNN is used to share the structured sample set to extract cloud morphology features, including the following steps:

[0031] The grayscale matrix is ​​binarized to obtain a cloud spatial mask, and the corresponding morphological parameter sequence is calculated based on the cloud spatial mask.

[0032] Feature extraction is performed on the morphological parameter sequence based on 3D convolutional layers. Specifically...

[0033] The first layer extracts local morphological changes, the second layer extracts the spatial evolution of morphology, and the third layer extracts the correlation between the morphological parameter sequence and the turbulence intensity.

[0034] In some embodiments, an atmospheric transport correction factor is introduced to correct the concentration prediction value to obtain an accurate concentration value, including the following steps:

[0035] Obtain the preset calibration transmittance;

[0036] Calculate the real-time, field-measured transmittance;

[0037] The atmospheric transport correction factor is calculated based on the calibrated transmittance and the field-measured transmittance, and the accurate concentration value is calculated using the following formula:

[0038] ,

[0039] in, This is represented as the concentration prediction value output by a nonlinear mapping processed by deep learning. The atmospheric transport correction factor, Characterized as at wavelength of The temperature is The calibrated transmittance at that time Characterized as at wavelength of The temperature is The actual transmittance measured on-site at that time.

[0040] In some embodiments, calculating real-time, field-measured transmittance includes the following steps:

[0041] Based on the actual water vapor absorption band acquired by the thermal infrared imaging device, the first gray value is calculated.

[0042] Obtain the second grayscale value of the preset multi-band blackbody radiation source at the same temperature and the same band;

[0043] The measured transmittance in the field is calculated based on the first gray value and the second gray value using the following formula.

[0044] ,

[0045] in, Characterized by the first grayscale value, Characterized as the second grayscale value, Characterized by the laboratory-calibrated water vapor transmittance calibration coefficient.

[0046] In some embodiments, the following steps are also included:

[0047] The wind speed is monitored in real time and it is determined whether it exceeds the wind speed threshold.

[0048] If it is not greater than, then load the standard model;

[0049] If it is greater than that, then the turbulence enhancement mode is activated;

[0050] Under the standard model, a standard weight set is applied where the radiation intensity weight is greater than or equal to the cloud morphology weight.

[0051] When the turbulence enhancement mode is activated, a high turbulence weight set is applied, where the radiation intensity is less than the cloud morphology weight.

[0052] In some embodiments, an atmospheric transport correction factor is introduced to correct the concentration prediction value to obtain an accurate concentration value, and the method further includes the following steps:

[0053] The atmospheric transport correction factor is dynamically adjusted to correct the concentration prediction value based on the volatility of the first gray value in the continuous frame time series.

[0054] If the volatility is greater than a preset value, the correction frequency is increased;

[0055] If the volatility is less than a preset value, the correction frequency is reduced.

[0056] Secondly, this application provides a gas leakage quantitative calibration system based on multispectral feature fusion, employing the following technical solution:

[0057] A gas leak quantitative calibration system based on multispectral feature fusion includes:

[0058] Configured with a multi-band blackbody radiation source;

[0059] Gas cloud simulation chamber, used to build a simulated gas cloud environment and acquire configuration parameters in real time;

[0060] An environmental sensor array is used to classify and collect the configuration parameters to construct a four-dimensional parameter space;

[0061] Infrared thermal imagers are used to acquire infrared thermal imaging data corresponding to simulated gas cloud environments.

[0062] The feature extraction engine is used to associate the infrared thermal imaging data with the four-dimensional parameter space to form a structured sample set, and to use a dual-channel 3D-CNN to share the structured sample set to extract radiation intensity features and cloud morphology features and stitch them together to obtain a fused feature vector.

[0063] A concentration inversion model is used to perform a deep learning-based nonlinear mapping on the fused feature vector to output a concentration prediction value, and an atmospheric transport correction factor is introduced to correct the concentration prediction value to obtain an accurate concentration value.

[0064] The technical solutions provided by the embodiments of this application have the following technical effects:

[0065] This invention achieves a calibration method that integrates physical experimental models and data-driven approaches to accurately model the nonlinear relationship between concentration and radiation. Through multi-channel feature extraction, radiation intensity features and cloud morphology features are integrated into the concentration inversion process to achieve adaptive calibration for different diffusion patterns. The invention also improves the environmental robustness of the calibration process by dynamically compensating for the influence of water vapor and dust on radiation transmission through atmospheric transport correction factors. Attached Figure Description

[0066] Figure 1 This is a schematic diagram illustrating the steps of a gas leakage quantitative calibration method based on multispectral feature fusion provided in this embodiment.

[0067] Figure 2 This is a schematic diagram of a gas leakage quantitative calibration system based on multispectral feature fusion provided in an embodiment of this application.

[0068] Figure 3 This is a logical schematic diagram of adjusting the turbulence mode in the gas leakage quantitative calibration method based on multispectral feature fusion provided in the embodiments of this application. Detailed Implementation

[0069] To better understand the purpose, technical solutions, and advantages of this application, it has been described and illustrated below with reference to the accompanying drawings and embodiments. However, those skilled in the art should understand that this application can be implemented without these details. In some cases, to avoid obscuring various aspects of this application due to unnecessary description, well-known methods, processes, systems, components, and / or circuits already described at a higher level will not be elaborated upon. It will be apparent to those skilled in the art that various modifications can be made to the embodiments disclosed in this application, and the general principles defined in this application can be applied to other embodiments and application scenarios without departing from the principles and scope of this application. Therefore, this application is not limited to the illustrated embodiments, but conforms to the broadest scope consistent with the scope of protection claimed in this application.

[0070] It should be noted that the descriptions of these embodiments are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0071] In the description of this application, "several" means one or more, "more than" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0072] In the description of this application, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples.

[0073] like Figure 1 As shown in the figure, this application discloses a method for quantitative calibration of gas leakage based on multispectral feature fusion, including the following steps:

[0074] The S100 is equipped with a multi-band blackbody radiation source.

[0075] In this embodiment, a dual-band independent temperature control design is adopted to address the issue that single-band calibration cannot cover the needs of multi-gas absorption, specifically, considering the waveband characteristics of different gas absorption.

[0076] It includes wavelength ranges of 3-5 μm and 8-12 μm, with the 3-5 μm range primarily used for monitoring methane (…). ), propane ( Hydrocarbon gases, such as sulfur hexafluoride (SF6), have a strong absorption peak around 3.3 μm; the 8-12 μm range is mainly used for monitoring SF6. ),carbon dioxide( (and other gases.)

[0077] Meanwhile, by employing a platinum resistance temperature sensor and a PID temperature control algorithm, the independent temperature control accuracy of the dual-band can reach ±0.1℃, and a stable blackbody radiation signal can be output, providing an accurate reference source for the subsequent calibration of the radiation transfer model.

[0078] The S200 constructs a simulated gas cloud environment and acquires configuration parameters in real time. It then classifies and collects these configuration parameters to build a four-dimensional parameter space.

[0079] A simulated gas cloud environment was built using a gas cloud simulation chamber that includes an adjustable turbulence generator and a temperature-controlled air curtain. This simulates the complex environment in an industrial setting and provides a realistic and controllable dataset for subsequent model training.

[0080] After configuring the simulation environment, the corresponding configuration parameters are obtained in real time and classified to construct a data space containing four-dimensional parameters. A radiation transmission matrix is ​​established through the four-dimensional parameter space to provide sample support for the subsequent configuration of training datasets in the full scenario.

[0081] S300 acquires infrared thermal imaging data corresponding to a simulated gas cloud environment and correlates it with a four-dimensional parameter space to form a structured sample set.

[0082] Infrared thermal imaging data corresponding to simulated gas clouds is acquired and associated with the corresponding four-dimensional parameter space configuration. Several frames of infrared thermal imaging images are acquired under each parameter combination, and a correlation is formed with the actual concentration value. Finally, a structured data sample set is constructed to extract different feature information based on several sample data.

[0083] S400 utilizes a dual-channel 3D-CNN to share a structured sample set to extract radiation intensity features and cloud morphology features, which are then concatenated to obtain a fused feature vector.

[0084] Based on the structured sample set, the radiation intensity features and cloud morphology features are extracted using two channels of 3D-CNN. The infrared thermal imaging signal during gas leakage contains two non-negligible dimensions: "radiation energy attenuation change" caused by distance and ambient temperature, and "spatial morphology distribution" affected by wind speed and gas characteristics. Current technology cannot fully characterize gas concentration information by extracting a single feature without considering interference factors such as wind speed, wind direction, and environment.

[0085] By combining the absorption capacity of gas for infrared radiation with the spatial distribution and diffusion state of gas through bilateral extraction and merging of vectors, the model's learning and reasoning capabilities become more comprehensive and accurate.

[0086] S500 performs a deep learning-based nonlinear mapping on the fused feature vectors to output concentration predictions.

[0087] The deep learning module learns the mapping relationship between the four-dimensional parameter space formed by the real configuration parameters of the laboratory and the feature vector corresponding to the actual measured gas concentration, and establishes a nonlinear model from "raw grayscale data" to "preliminary concentration estimate", and finally outputs the preliminary estimated concentration value "without atmospheric transmission correction".

[0088] S600 introduces an atmospheric transport correction factor to correct the concentration prediction value to obtain an accurate concentration value.

[0089] During the transmission of infrared radiation from gas clouds to thermal imaging equipment, it is absorbed or scattered by media such as water vapor, dust, and carbon dioxide in the atmosphere (i.e., atmospheric attenuation). The core logic of concentration inversion is to "infer the gas concentration from the amount of radiation intensity attenuation". Therefore, atmospheric attenuation will cause the radiation intensity of the thermal infrared imaging equipment to deviate from the actual absorption level of the gas. So, in order to make the final result more accurate, it is necessary to quantify and eliminate the radiation attenuation caused by non-target gases through atmospheric transmission correction.

[0090] The above methods enable the integration of physical experimental models and data-driven calibration methods, achieving accurate modeling of the nonlinear relationship between concentration and radiation. Through multi-channel feature extraction, radiation intensity features and cloud morphology features are integrated into the concentration inversion process, enabling adaptive calibration for different diffusion patterns. The atmospheric transport correction factor dynamically compensates for the influence of water vapor and dust on radiation transport, improving the environmental robustness of the calibration process.

[0091] In other embodiments, a simulated gas cloud environment is constructed and configuration parameters are acquired in real time. The configuration parameters are then categorized and collected to construct a four-dimensional parameter space, including the following steps:

[0092] S210, with wind speed configured according to the adjustable turbulence generator and real-time temperature configured according to the temperature control air curtain.

[0093] The adjustable turbulence generator employs a combination of a multi-fan array and an airflow guide plate, enabling continuous adjustment of wind speed and direction from 0 to 15 m / s, covering common industrial wind conditions from no wind (0 m / s) to strong winds (15 m / s). Simultaneously, by controlling the fluctuation range of fan speed, it achieves different levels of turbulence intensity, accurately simulating the diffusion morphology of gas clouds under different turbulent conditions.

[0094] The temperature-controlled air curtain employs semiconductor temperature control technology, with temperature-controlled air curtains installed at the inlet and outlet of the simulation chamber to achieve precise control of the internal ambient temperature between -30 and 60°C. This setup can simulate various scenarios, including daily environments (pipeline environments of -5 to 30°C), severe winter cold (northern pipeline environments of -30°C), and high summer temperatures (tank area exposure environments of 60°C), ensuring the environmental adaptability of the calibration data.

[0095] S220: Obtain the actual concentration value in the simulated gas cloud environment and the current position of the thermal infrared imaging device.

[0096] The true concentration value is obtained by precisely controlling the gas injection rate through a high-precision gas command controller and collecting the corresponding true concentration.

[0097] The cloud simulation chamber is equipped with an infrared thermal imager that can move along the direction of the gas outlet of the simulation chamber, and collect the current position in real time based on the position of the infrared thermal imaging device on the guide rail. At the same time, it is necessary to ensure that the observation angle is perpendicular to the gas diffusion direction.

[0098] S230 collects and classifies each configuration parameter into a four-dimensional parameter space including gas concentration, observation distance, ambient temperature, and turbulence intensity, and configures the corresponding data range for each dimension parameter based on the preset classification data volume.

[0099] The "Gas Concentration" dimension is configured based on actual concentration values, and is set to cover the entire range from trace leaks (0.1% LEL, below the safety threshold) to high-concentration leaks (100% LEL, approaching explosion risk). The gradient point interval increases with increasing concentration to ensure monitoring accuracy for low-concentration leaks. The data volume under this dimension is >20 gradient points, and the data range is 0.1%-100% LEL.

[0100] The "Observation Distance" dimension is configured based on the current location, and is set to cover common monitoring distances in industrial sites (such as 1m for close-range inspection of tank areas and 20m for long-range monitoring of pipelines) to avoid the problem of decreased accuracy at long distances caused by calibration at a single distance. The data volume under this dimension is 4 points, with data ranges of 1m, 5m, 10m, and 20m respectively.

[0101] The "Ambient Temperature" dimension is configured based on real-time temperature, and is set to cover extreme temperatures (extreme high and low temperatures) in industrial settings, while also covering the ambient temperature baseline and the freezing point critical temperature. This dimension contains four data points with ranges of -20℃, 0℃, 25℃, and 50℃.

[0102] The "turbulence intensity" dimension is configured based on wind speed, and is set to cover common and major wind conditions in industrial sites. Low turbulence corresponds to light wind environments, medium turbulence corresponds to medium wind environments, and high turbulence corresponds to strong wind environments. This dimension contains three data points, with data ranges of low (δ=0.1, corresponding to wind speeds less than 2 m / s), medium (δ=0.3, corresponding to wind speeds between 2 and 8 m / s), and high (δ=0.5, corresponding to wind speeds greater than 8 m / s).

[0103] The core function of the gas cloud simulation chamber is to reproduce a controllable industrial environment. Through turbulence generators, temperature-controlled air curtains, and guide rails, parameters such as observation distance, temperature, and wind speed are precisely controlled. Finally, precise control based on "real concentration" is a crucial step. Only when the four dimensions of concentration, distance, temperature, and turbulence are all controllable and true values ​​can the acquired infrared thermal imaging data accurately correspond to the influence of a single variable, providing a clean sample for the model to extract subsequent dual-channel features.

[0104] In other embodiments, acquiring infrared thermal imaging data corresponding to a simulated gas cloud environment and correlating it with a four-dimensional parameter space to form a structured sample set includes the following steps:

[0105] S310 acquires several frames of infrared thermal imaging images corresponding to the configured parameters.

[0106] S320 calculates the grayscale matrix of each infrared thermal imaging image and constructs a three-dimensional tensor with the corresponding frame time series.

[0107] S330 associates several three-dimensional tensors with the thought parameter space to obtain a structured sample set.

[0108] In the structured sample set, the three-dimensional tensor [H,W,T] (H is the image height, W is the image width, and T is the number of time series frames) corresponding to the infrared thermal imaging contains information on the spatial grayscale distribution and temporal variation of the gas cloud layer. The observation distance is used to correct for the attenuation effect of distance on radiation intensity (radiation intensity is inversely proportional to the square of the distance). Ambient temperature is used to correct for the interference of ambient background radiation on the radiation signal of the gas cloud (the higher the ambient temperature, the stronger the background radiation, and the lower the radiation contrast of the gas cloud). Turbulence intensity is used to guide the weighting of morphological features (the higher the turbulence intensity, the greater the weight of morphological features, to highlight the influence of morphological changes on concentration).

[0109] In other embodiments, a dual-channel 3D-CNN is used to share a structured sample set to extract radiation intensity features, including the following steps:

[0110] S410 performs distance attenuation correction on the grayscale matrix based on the observation distance.

[0111] S411 performs radiometric grayscale calibration on the grayscale matrix based on ambient temperature.

[0112] S412 extracts features from the corrected grayscale matrix based on 3D convolutional layers.

[0113] Specifically,

[0114] S413: The first layer extracts the local radiation intensity differences corresponding to the three-dimensional tensor; the second layer extracts the time variation trend of the radiation intensity; and the third layer extracts the spatial distribution of the radiation intensity.

[0115] First, the radiation intensity characteristics are based on the original grayscale matrix values ​​and their changes over time. Combined with parameters such as observation distance and ambient temperature, the absorption intensity of the gas to infrared radiation is extracted and analyzed (e.g., the greater the grayscale difference, the stronger the absorption and the higher the indirect reaction concentration).

[0116] Specifically, the grayscale value is first corrected for distance attenuation based on the observation distance (the formula is: corrected grayscale value = original grayscale value * ...). This compensates for the natural attenuation of infrared radiation with distance;

[0117] Then, based on the ambient temperature, subtract the gray value corresponding to the ambient background radiation at the same temperature (calibrated by the radiation gray value of a multispectral blackbody at the same ambient temperature).

[0118] Finally, after distance correction and temperature calibration, a radiance grayscale matrix contributed solely by the gas cloud is obtained.

[0119] Meanwhile, feature extraction is performed on the corrected radiation grayscale matrix through three 3D convolutional layers (3*3*3 kernel size, 1*1*1 stride). The first layer extracts local radiation intensity differences, such as the grayscale difference between the center and edge of the gas cloud; the second layer extracts the temporal trend of radiation intensity, such as the decay rate of grayscale values ​​during cloud diffusion, as well as the fluctuation amplitude over time; the third layer extracts the spatial distribution pattern of radiation intensity, such as the morphology of the high grayscale region at the center of the cloud, and finally outputs a 256-dimensional radiation intensity feature vector.

[0120] In other embodiments, a dual-channel 3D-CNN is used to share a structured sample set to extract cloud morphology features, including the following steps:

[0121] S420 performs binarization on the grayscale matrix to obtain the cloud spatial mask and calculates the corresponding morphological parameter sequence based on the cloud spatial mask.

[0122] S421 extracts features from morphological parameter sequences based on 3D convolutional layers.

[0123] Specifically,

[0124] S422: The first layer extracts local morphological changes, the second layer extracts the spatial evolution of morphology, and the third layer extracts the correlation between morphological parameter sequences and turbulence intensity.

[0125] The cloud morphology characteristics are calculated based on the spatial distribution of the original gray-scale matrix to determine the morphological parameters of the cloud and correct the influence of diffusion morphology on concentration inversion.

[0126] First, morphological preprocessing is performed, and the grayscale matrix is ​​binarized (based on adaptive threshold segmentation to distinguish gas clouds from the background) to obtain the spatial mask matrix of the cloud, where 1 represents gas and 0 represents the background. Furthermore, an opening operation (erosion followed by dilation) can be performed on the spatial mask matrix to remove isolated small pixels caused by dust and noise, while preserving the complete cloud outline.

[0127] The area (number of pixels, reflecting the scale of leakage), circularity, and eccentricity of a cloud are calculated based on a spatial mask matrix.

[0128] The gray-level gradient of the cloud edge is calculated based on the Sobel operator, and the mean (reflecting edge sharpness, with large gradients in coatings under high turbulence) and variance (reflecting the degree of edge irregularity) of the edge gradient are statistically analyzed.

[0129] Finally, parameters such as the area expansion rate and center offset of the cloud are calculated using the time-series mask to form the final morphological parameter sequence.

[0130] Simultaneously, using the same 3D convolutional structure as the radiation intensity feature channel, feature extraction is performed on the morphological parameter sequence. The first layer extracts local morphological changes, such as the degree of concavity and convexity of cloud edges; the second layer extracts the temporal evolution of morphology, such as the breaking speed of clouds under turbulent conditions; and the third layer extracts the correlation features between morphology and turbulence intensity, such as the irregularity of clouds under high turbulence. The final output is a 256-dimensional cloud morphology feature vector.

[0131] After feature extraction from both channels is complete, batch normalization is used to eliminate dimensional differences. The normalized radiation intensity and cloud morphology features are then concatenated using the Concatenate() function to obtain a full-dimensional feature vector that considers "radiation-morphology-environment". After concatenation, a fully connected layer and a ReLU activation function are added to perform a non-linear transformation on the fused features, ultimately outputting optimized fused features as input for subsequent deep learning estimation.

[0132] In other embodiments, an atmospheric transport correction factor is introduced to correct the concentration prediction value to obtain an accurate concentration value, including the following steps:

[0133] S610, obtain the preset calibrated transmittance.

[0134] S620 calculates real-time, on-site measured transmittance.

[0135] S630 calculates the atmospheric transport correction factor based on calibrated transmittance and field-measured transmittance, and then calculates the accurate concentration value using the following formula:

[0136] .

[0137] in, This is represented as the concentration prediction value output by a nonlinear mapping processed by deep learning. For atmospheric transport correction factor, Characterized as at wavelength of The temperature is The calibrated transmittance at that time Characterized as at wavelength of The temperature is The actual transmittance measured on-site at that time.

[0138] By introducing an atmospheric transport correction factor, the concentration estimation results from deep learning are combined with physical mechanisms to address the discrepancy between the field environment and the laboratory calibration environment.

[0139] Specifically, the concentration prediction value is first obtained by deep learning. This value is obtained by inputting a 512-dimensional fused feature vector extracted from multiple features into a fully connected layer and outputting a preliminary concentration estimate. This value is trained based on a laboratory-calibrated dataset and reflects the nonlinear mapping relationship between the fused features and the concentration, but does not take into account the transmission loss in the field environment.

[0140] Characterized by calibrated transmittance, which is determined in the laboratory calibration process under conditions free from additional environmental interference, through a multispectral blackbody at different wavelengths. Different temperatures The radiation signal was calculated using Lambert-Beer's law. The specific calculation process is as follows: ,in, For gas at wavelength ,temperature The absorption coefficient below, To calibrate the actual concentration inside the chamber, This represents the radiation transmission path length. The calibrated transmittance is a fixed value stored in the system database as a reference for transmittance.

[0141] The transmittance is represented by the actual measured transmittance, which reflects the degree of attenuation of infrared radiation by water vapor and dust at the site. It is obtained by collecting the water vapor absorption band at the site and inverting it through infrared thermal imaging equipment. This value is the object of real-time calculation, reflecting the influence of water vapor and dust at the site on radiation transmission.

[0142] The final concentration C is the result of converting the "laboratory calibration result" to the "actual concentration on site" by multiplying the estimated value obtained by deep learning with the atmospheric transport correction factor, thus correcting the calculation error caused by on-site environmental interference.

[0143] For example, consider the logic of correction using atmospheric transport correction factors:

[0144] If the on-site environment is highly disturbed (e.g., high humidity, high dust levels), then Too small (severe radiation attenuation). If the ratio is greater than 1, the measured value will be greater than the predicted value, which means that the preliminary estimate is underestimated due to radiation attenuation and needs to be amplified to the true concentration by a correction factor.

[0145] If the environment is clean (dry and windless), then The value is close to The value of ) Since the value is close to 1, it indicates that the preliminary estimate is close to the true concentration, and the correction range is small.

[0146] In other embodiments, calculating the real-time field-measured transmittance includes the following steps:

[0147] S621, based on the thermal infrared imaging equipment, collects the actual water vapor absorption band and calculates the first gray value.

[0148] S622, obtain the second grayscale value of the preset multi-band blackbody radiation source at the same temperature and the same band.

[0149] S623, the measured transmittance on site is calculated based on the first gray value and the second gray value combined with the following formula.

[0150] ,

[0151] in, Represented as the first grayscale value, Represented as the second grayscale value, Characterized by the laboratory-calibrated water vapor transmittance calibration coefficient.

[0152] When calculating the on-site measured transmittance, the signal was obtained by inverting the water vapor absorption band (6.5μm, which has strong water vapor absorption and does not overlap with the absorption bands of most monitored gases) acquired by infrared thermal imaging equipment. The first gray value represents the gray value of the on-site water vapor in the 6.5μm band, and the second gray value represents the gray value of the multispectral blackbody at the same temperature and in the same 6.5μm band.

[0153] In other embodiments, the following steps are also included:

[0154] The S700 monitors wind speed in real time and determines whether it exceeds the wind speed threshold.

[0155] If S710 is not greater than, then load the standard model.

[0156] If S720 is greater than this, the turbulence enhancement mode is activated.

[0157] S730, in the standard model, is a standard weight set with a radiation intensity weight greater than or equal to the cloud morphology weight.

[0158] S740, when activating the turbulence enhancement mode, loads a high turbulence weight set whose radiation intensity is less than the cloud morphology weight.

[0159] To achieve accurate adaptation for gas concentration inversion under different turbulence intensities and to solve the problems of morphological interference and concentration deviation caused by turbulence in existing technologies, this application further implements different processing modes under different turbulence states in its embodiments:

[0160] First, the core indicator of turbulence intensity is wind speed, so we first collect wind speed data of the calibration environment in real time as the basis for judging turbulence intensity.

[0161] Secondly, the collected wind speed is compared with the preset wind speed (2 m / s in this application). When the wind speed is less than the preset value, it corresponds to a low turbulence environment. At this time, the gas cloud diffusion state is stable, and the shape has little interference with the concentration inversion. In this case, the standard model is loaded and standardized processing is performed in subsequent feature extraction and concentration inversion. When the wind speed exceeds the preset value, it corresponds to a medium / high turbulence environment. At this time, the gas cloud is significantly disturbed by the airflow, and the shape is flat, broken or rapidly diffused. The shape has a large interference with the concentration inversion. It is necessary to start the turbulence enhancement mode. In this mode, customized corrections are required for subsequent feature extraction and concentration inversion.

[0162] Finally, the difference between the different modes lies in the different weight ratios among multiple features during subsequent feature extraction.

[0163] When learning the mapping relationship between "radiation intensity-concentration" under "low turbulence + stable morphology" in the standard model, the model parameters focus more on the absolute value and spatial distribution of gray values. When extracting features, the focus is on radiation intensity features while the weight of morphological features is relatively small, reducing unnecessary morphological corrections, or the two features are given the same weight.

[0164] When the turbulence enhancement mode is activated, the correlation between "morphological change rate and concentration" under "medium / high turbulence + dynamic morphology" is learned to specifically extract the dynamic fragmentation features of the cloud. During feature extraction, the weight of morphological features is increased to focus on capturing dynamic morphological information such as the degree of fragmentation at the cloud edge and the diffusion rate, while the time dimension features are enhanced, and the weight of radiation intensity features is appropriately reduced.

[0165] During the calibration process, the system continuously monitors wind speed changes (re-evaluating the threshold every 0.5 seconds). If the wind speed drops from >2m / s to ≤2m / s (e.g., when the gust ends), it automatically switches back from "turbulence enhancement mode" to "standard mode" and updates the model weights simultaneously. At the same time, it sends the "grayscale data - actual concentration" sample under the current turbulence state back to the database and fine-tunes the weight parameters of the two modes periodically (e.g., monthly) to improve the smoothness of mode switching and calibration accuracy.

[0166] By employing a simple and efficient logic of "wind speed monitoring → threshold judgment → mode switching," complex turbulent interference is transformed into a quantifiable and dynamically adaptable calibration strategy, addressing the pain point of existing technologies where "fixed models cannot cope with variable turbulence." Its practical effects are not only reflected in a significant improvement in concentration inversion accuracy, but also in enhanced engineering practicality in complex industrial scenarios (such as oil storage tank areas and open pipelines) through lightweight design and rapid response. This is a key technical aspect of this invention's breakthrough in overcoming the bottleneck of turbulence adaptability.

[0167] In other embodiments, an atmospheric transport correction factor is introduced to correct the concentration prediction value to obtain an accurate concentration value, and the method further includes the following steps:

[0168] S640 dynamically adjusts the correction frequency of the atmospheric transport correction factor for the concentration prediction value based on the volatility of the first gray value in a continuous frame time series.

[0169] S641, if the volatility is greater than the preset value, the correction frequency will be increased.

[0170] S642, if the volatility is less than the preset value, the correction frequency is reduced.

[0171] Since the measured transmittance changes dynamically with the environment, it is real-time. In order to ensure timely compensation for amplitude attenuation caused by rain, fog and humidity changes, this embodiment further adjusts the correction frequency dynamically based on the fluctuation of the transmittance. This avoids the problem of redundant calculations and reduced inversion speed when the environment is stable if a fixed frequency is used, and the problem of increased deviation due to untimely correction when encountering sudden environmental changes.

[0172] The relative standard deviation is calculated by calculating the mean gray value and standard deviation of each pixel in the first gray value of multiple frames in the sliding window. The fluctuation rate is judged by the difference in the relative standard deviation. The fluctuation rate can intuitively reflect the changes in environmental humidity (such as a sudden increase in temperature - increased water vapor absorption - decrease in gray value), and is not affected by the concentration of the target gas. It can purely characterize the environmental stability.

[0173] When volatility is less than the preset value, it indicates that the environment is stable, and the correction frequency can be appropriately reduced to reduce the amount of calculation, such as extending the correction frequency from 0.5s to 2s; when volatility is not less than the preset value, it indicates that the environment is abrupt, and the correction frequency needs to be appropriately increased to ensure timely correction, such as shortening the correction frequency from 1s to 0.5s.

[0174] This application also discloses a gas leak quantitative calibration system based on multispectral feature fusion, characterized in that it includes:

[0175] Configured with a multi-band blackbody radiation source;

[0176] Gas cloud simulation chamber, used to build a simulated gas cloud environment and acquire configuration parameters in real time;

[0177] An environmental sensor array is used to classify and collect configuration parameters to construct a four-dimensional parameter space;

[0178] Infrared thermal imagers are used to acquire infrared thermal imaging data corresponding to simulated gas cloud environments.

[0179] The feature extraction engine is used to associate infrared thermal imaging data with a four-dimensional parameter space to form a structured sample set, and to use a dual-channel 3D-CNN to share the structured sample set to extract radiation intensity features and cloud morphology features and stitch them together to obtain a fused feature vector.

[0180] The concentration inversion model is used to perform a deep learning-based nonlinear mapping on the fused feature vector to output a concentration prediction value, and an atmospheric transport correction factor is introduced to correct the concentration prediction value to obtain an accurate concentration value.

[0181] The implementation principle is as follows:

[0182] This invention achieves a calibration method that integrates physical experimental models and data-driven approaches to accurately model the nonlinear relationship between concentration and radiation. Through multi-channel feature extraction, radiation intensity features and cloud morphology features are integrated into the concentration inversion process to achieve adaptive calibration for different diffusion patterns. The invention also improves the environmental robustness of the calibration process by dynamically compensating for the influence of water vapor and dust on radiation transmission through atmospheric transport correction factors.

[0183] It should be understood that although the steps in the flowcharts in the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless otherwise expressly stated herein, there is no strict order in which these steps are performed, and they may be performed in other orders.

[0184] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for quantitative calibration of gas leakage based on multispectral feature fusion, characterized in that, Includes the following steps: Configure multi-band blackbody radiation sources; A simulated gas cloud environment is built and configuration parameters are acquired in real time. The configuration parameters are then classified and collected to construct a four-dimensional parameter space. Infrared thermal imaging data corresponding to the simulated gas cloud environment is acquired and correlated with the four-dimensional parameter space to form a structured sample set; The structured sample set is shared using a dual-channel 3D-CNN to extract radiation intensity features and cloud morphology features, which are then concatenated to obtain a fused feature vector. The fused feature vector is subjected to a deep learning-based nonlinear mapping to output a concentration prediction value; An atmospheric transport correction factor is introduced to correct the concentration prediction value to obtain an accurate concentration value.

2. The gas leakage quantitative calibration method based on multispectral feature fusion according to claim 1, characterized in that, The process of constructing a simulated gas cloud environment and acquiring configuration parameters in real time, and then classifying and collecting these parameters to build a four-dimensional parameter space, includes the following steps: Configure the wind speed according to the adjustable turbulence generator, and configure the real-time temperature according to the temperature control air curtain; Obtain the actual concentration value in the simulated gas cloud environment and the current position of the infrared thermal imager; Each of the configuration parameters is collected and classified into a four-dimensional parameter space including gas concentration, observation distance, ambient temperature, and turbulence intensity. Based on the preset classification data volume, the corresponding data range is configured for each dimension parameter.

3. The gas leakage quantitative calibration method based on multispectral feature fusion according to claim 2, characterized in that, Acquiring infrared thermal imaging data corresponding to a simulated gas cloud environment and correlating it with the four-dimensional parameter space to form a structured sample set includes the following steps: Acquire several frames of infrared thermal imaging images corresponding to the configuration parameters; Calculate the grayscale matrix of each infrared thermal imaging image and construct a three-dimensional tensor with the corresponding frame time series; The structured sample set is obtained by associating several of the three-dimensional tensors with the four-dimensional parameter space.

4. The gas leakage quantitative calibration method based on multispectral feature fusion according to claim 3, characterized in that, The method of extracting radiation intensity features by sharing the structured sample set using a dual-channel 3D-CNN includes the following steps: The grayscale matrix is ​​corrected for distance attenuation based on the observation distance; The grayscale matrix is ​​calibrated for radiometric grayscale based on the ambient temperature. Feature extraction is performed on the corrected grayscale matrix based on 3D convolutional layers. Specifically... The first layer extracts the local radiation intensity differences corresponding to the three-dimensional tensor, the second layer extracts the time variation trend of the radiation intensity, and the third layer extracts the spatial distribution of the radiation intensity.

5. The gas leakage quantitative calibration method based on multispectral feature fusion according to claim 3, characterized in that, The method of extracting cloud morphology features by sharing the structured sample set using a dual-channel 3D-CNN includes the following steps: The grayscale matrix is ​​binarized to obtain a cloud spatial mask, and the corresponding morphological parameter sequence is calculated based on the cloud spatial mask. Feature extraction is performed on the morphological parameter sequence based on 3D convolutional layers. Specifically... The first layer extracts local morphological changes, the second layer extracts the spatial evolution of morphology, and the third layer extracts the correlation between the morphological parameter sequence and the turbulence intensity.

6. The gas leakage quantitative calibration method based on multispectral feature fusion according to claim 3, characterized in that, To obtain an accurate concentration value, an atmospheric transport correction factor is introduced to correct the predicted concentration value, including the following steps: Obtain the preset calibration transmittance; Calculate the real-time, field-measured transmittance; The atmospheric transport correction factor is calculated based on the calibrated transmittance and the field-measured transmittance, and the accurate concentration value is calculated using the following formula: ， in, This is represented as the concentration prediction value output by a nonlinear mapping processed by deep learning. The atmospheric transport correction factor, Characterized as at wavelength of The temperature is The calibrated transmittance at that time Characterized as at wavelength of The temperature is The actual transmittance measured on-site at that time.

7. The gas leakage quantitative calibration method based on multispectral feature fusion according to claim 6, characterized in that, Calculating real-time, field-measured transmittance includes the following steps: Based on the actual water vapor absorption band acquired by the infrared thermal imager, the first gray value is calculated. Obtain the second grayscale value of the preset multi-band blackbody radiation source at the same temperature and the same band; The measured transmittance in the field is calculated based on the first gray value and the second gray value using the following formula. ， in, Characterized by the first grayscale value, Characterized as the second grayscale value, Characterized by the laboratory-calibrated water vapor transmittance calibration coefficient.

8. The gas leakage quantitative calibration method based on multispectral feature fusion according to claim 2, characterized in that, It also includes the following steps: The wind speed is monitored in real time and it is determined whether it exceeds the wind speed threshold. If it is not greater than, then load the standard model; If it is greater than that, then the turbulence enhancement mode is activated; Under the standard model, a standard weight set is applied where the radiation intensity weight is greater than or equal to the cloud morphology weight. When the turbulence enhancement mode is activated, a high turbulence weight set is applied, where the radiation intensity is less than the cloud morphology weight.

9. The gas leakage quantitative calibration method based on multispectral feature fusion according to claim 7, characterized in that, The method of correcting the predicted concentration value by introducing an atmospheric transport correction factor to obtain an accurate concentration value also includes the following steps: The atmospheric transport correction factor is dynamically adjusted to correct the concentration prediction value based on the volatility of the first gray value in the continuous frame time series. If the volatility is greater than a preset value, the correction frequency is increased; If the volatility is less than a preset value, the correction frequency is reduced.

10. A gas leak quantitative calibration system based on multispectral feature fusion, characterized in that, include: Configured with a multi-band blackbody radiation source; Gas cloud simulation chamber, used to build a simulated gas cloud environment and acquire configuration parameters in real time; An environmental sensor array is used to classify and collect the configuration parameters to construct a four-dimensional parameter space; Infrared thermal imagers are used to acquire infrared thermal imaging data corresponding to simulated gas cloud environments. The feature extraction engine is used to associate the infrared thermal imaging data with the four-dimensional parameter space to form a structured sample set, and to use a dual-channel 3D-CNN to share the structured sample set to extract radiation intensity features and cloud morphology features and stitch them together to obtain a fused feature vector. A concentration inversion model is used to perform a deep learning-based nonlinear mapping on the fused feature vector to output a concentration prediction value, and an atmospheric transport correction factor is introduced to correct the concentration prediction value to obtain an accurate concentration value.

Images