Combustible gas imaging method

Abstract

The invention is suitable for the technical field of petrochemical industry production safety monitoring, and provides a combustible gas imaging method, and the method achieves the precise recognition and feature analysis of gas plume through designing a preprocessing module, a post-processing module and a plume parameter quantification module. The preprocessing module adopts dynamic histogram stretching to enhance the contrast of an infrared image, and combines adaptive pseudo-color mapping and motion consistency filtering to suppress jitter noise; the post-processing module repairs the fracture plume and smooths the edge through morphological closed operation, and removes noise interference in combination with connected domain analysis and area filtering; the plume parameter quantization module estimates the diffusion area based on the mask pixel number and the spatial resolution, predicts the diffusion direction through adjacent frame light flow tracking, and calibrates gas concentration grading through gray variance. The method is suitable for the fields of industrial safety, environmental monitoring and the like, and the accuracy and robustness of gas leakage detection are remarkably improved.

Description

Technical Field

[0001] The invention belongs to the technical field of petrochemical production safety monitoring, and in particular relates to a combustible gas imaging method. Background Art

[0002] As the development of marine oil and gas resources continues, the damage caused by gas leaks to offshore platforms is gaining increasing attention. Offshore platforms are relatively confined, making escape difficult. Furthermore, they contain numerous electrical devices and a large amount of hydrocarbons. Once gas reaches a certain concentration, it is highly likely to explode, posing a threat to both human safety and the normal operation of the platform. This paper proposes a combustible gas leak area detection technology based on AI algorithms. This technology enables imaging monitoring of combustible gas leaks, enabling regional and non-point source detection, thereby improving the safety management of production facilities.

[0003] Currently available algorithms focus on three key areas: suppressing environmental interference, retrieving gas parameters, and locating gas leak sources. For environmental interference suppression, BP neural network algorithms, adaptive Kalman filtering algorithms, and support vector machines can be used. Adaptive filtering algorithms are widely used in meteorology, energy, radar, and other fields due to their flexibility and stability. For gas parameter retrieving, ANN neural network algorithms, least squares algorithms, and extreme learning machines can be used. Neural network algorithms, with their self-learning, self-adaptive, and associative storage characteristics, are widely used in prediction and estimation, pattern recognition, and decision support. For gas leak source locating, particle swarm optimization, ant colony optimization, and differential evolution algorithms can be used. Differential evolution algorithms are widely used in function optimization and combinatorial optimization due to their self-organization, parallelism, and nonlinearity.

[0004] However, the internal recognition algorithms of existing combustible gas infrared imaging equipment have problems such as slow response, low recognition, and the need for repeated training. It is necessary to provide a combustible gas recognition imaging method with good positioning effect, high accuracy, good anti-noise performance, and no need for on-site training. Summary of the Invention

[0005] The purpose of the embodiments of the present invention is to provide a combustible gas imaging method, aiming to solve the problems raised in the above background technology.

[0006] The embodiment of the present invention is implemented as follows: a combustible gas imaging method is used to realize gas plume identification based on the BiseNetv2 algorithm framework. The method realizes the target function by designing the following modules:

[0007] The preprocessing module is used to preprocess the input image or video frame. For infrared data, it enhances contrast through dynamic histogram stretching and adaptive pseudo-color mapping; and uses motion consistency filtering to suppress jitter noise between video frames.

[0008] The post-processing module uses morphological closing operations to link broken gas plumes, fill discontinuous areas of the gas plume, and smooth the edges of adjacent plumes. It then combines connected domain analysis with area filtering to remove noise and remove discrete target noise points whose areas are much smaller than the gas plume diffusion area, further increasing the accuracy of gas reservation identification.

[0009] The plume parameter quantification module estimates the gas emission and diffusion area based on the number of mask pixels × spatial resolution. For motion vectors, the gas diffusion direction is predicted using adjacent frame optical flow tracking technology. For gas concentration, the gas leakage intensity classification is calibrated using the grayscale variance of the mask area.

[0010] A further technical solution is that the method comprises the following specific steps:

[0011] Step 1: Test various gases under different leakage pressures, leakage volumes, ambient temperatures, and weather conditions. After fully recording and summarizing the test data, clean the data and then train and deeply learn it under the BiseNetv2 algorithm framework to form the required algorithm.

[0012] Step 2: Use the segmentation model to label all the data collected in step 1, pre-process the images or video frames in the input data, and then input them into the BiseNetv2 algorithm for dual-branch reasoning;

[0013] Step 3: The input data undergoes dual-branch inference in the BiseNetv2 algorithm. In the detail branch, plume edges are extracted from the input video frame or image. In the semantic branch, background noise and interference in the input video frame or image are denoised to reduce background interference and refine the identification of gas plume features.

[0014] Step 4: After branch inference using the BiseNetv2 algorithm, BGA feature fusion is performed. The gas plume edges processed in the detail branch and semantic branch are contrasted and fused with the background denoised images to enhance the difference between the low-concentration gas and the background.

[0015] Step 5: After BGA fusion, the BiseNetv2 algorithm performs mask operations, performs hole filling, motion consistency filtering, and morphological closing operations on the image or video frame to complete the gas inhomogeneity areas, flicker noise, and broken gas plumes in the image; at the same time, the gas plume parameters are analyzed, and the gas diffusion area, motion vector, concentration gradient, and center rate curve are calibrated to obtain the gas emission, diffusion direction, leakage intensity, and gas flow state.

[0016] Step 6: Based on steps 1 to 5, a trained algorithm model is obtained and configured with the hardware for field monitoring. The combustible gas plume is captured using a mid-infrared camera. The gas plume features are calibrated and pseudo-colored using the algorithm's pre-processing module, post-processing module, and plume parameter quantification module. The gas diffusion area is determined, and gas visualization is achieved through the host computer program to monitor gas leaks in real time.

[0017] The present invention provides a method for imaging combustible gases, which has the following beneficial effects:

[0018] (1) Long detection distance: the maximum detection distance can reach 300m;

[0019] (2) Large detection range: CH4 gas concentration detection range is 0.08% to 100%, C2H6 gas concentration detection range is 0.008% to 4%;

[0020] (3) High detection sensitivity: gas concentration detection accuracy is less than 1‰;

[0021] (4) Fast response time: Through the algorithm trained by machine learning, the time from detecting gas plume to outputting pseudo-color on imaging is less than 10s, and the detection image can be continuously output to achieve real-time monitoring.

[0022] In addition to the above effects, this method also has the characteristics of good positioning effect, high segmentation accuracy and good anti-noise performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A schematic structural diagram of a combustible gas imaging method provided by an embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram of the structure of the model after training;

[0025] Figure 3 Demonstrate visualization of gas plume diffusion;

[0026] Figure 4 A schematic diagram of a combustible gas imaging method and hardware configuration provided by an embodiment of the present invention and conducted in Jilin Oilfield for verification and testing;

[0027] Figure 5 A schematic diagram of a combustible gas imaging method and hardware configuration provided by an embodiment of the present invention and conducted in a test verification at the Xinao Gas Company in Leizhou Peninsula. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0029] The specific implementation of the present invention is described in detail below with reference to specific embodiments.

[0030] An embodiment of the present invention provides a combustible gas imaging method, which realizes gas plume recognition based on the BiseNetv2 algorithm framework. Figure 1 The figure shows the logical flow chart of gas plume identification. According to this figure, the method mainly realizes the target function by designing the following modules.

[0031] Preprocessing module: Preprocesses the input image or video frame. For infrared data, it enhances contrast through dynamic histogram stretching and adaptive pseudo-color mapping; and uses motion consistency filtering to suppress jitter noise between video frames.

[0032] Post-processing module: Use morphological closing operations to link broken gas plumes, fill discontinuous areas of the gas plume, and smooth the edges of adjacent plumes. Then, through a combination of connected domain analysis and area filtering, noise is removed. Discrete target noise points with an area much smaller than the gas plume diffusion area are removed, further improving the accuracy of gas reservation identification.

[0033] Plume parameter quantification module: For the gas plume diffusion area, the gas emission amount and diffusion area are estimated by the number of mask pixels × spatial resolution; for the motion vector, the gas diffusion direction is predicted by the adjacent frame optical flow tracking technology; for the gas concentration, the gas leakage intensity classification is calibrated by the grayscale variance of the mask area.

[0034] In an embodiment of the present invention, a gas infrared image is input, and the sub-pixel edges of the gas plume are captured via the detail branch of the BiseNetv2 algorithm framework. Through three layers of convolution without downsampling, faint traces of plume diffusion are retained. The semantic branch of the BiseNetv2 algorithm framework is used to suppress environmental interference and identify the plume body. Through two-way downsampling, feature diversity is enhanced. The global pattern of gas diffusion is captured by combining average pooling with residuals. The processing results of the detail branch and the semantic branch are combined via the BGA fusion layer. The multiplication operation strengthens the difference between the low-contrast plume and the background. Bidirectional correction solves the problem of blurring in semi-transparent areas. Thermodynamic feature compensation is input to eliminate the interference of refractive index changes caused by temperature and humidity on the plume morphology. In addition, morphological closing operations are used to connect plume areas that are broken by turbulence, and temporal consistency filtering is used to suppress plume flicker noise caused by wind disturbance.

[0035] As a preferred embodiment of the present invention, the method includes the following specific steps:

[0036] Step 1: Test various gases, such as compressed air, carbon dioxide, nitrogen, oxygen, hydrogen, methane, ethane, propane, ethylene, and acetylene, under varying leak pressures, leak volumes, ambient temperatures, and weather conditions. The test data is fully recorded and summarized, cleaned, and finally trained and deep-learned within the BiseNetv2 framework to develop the required algorithm.

[0037] Step 2: Use the segmentation model to label all the data collected in step 1, preprocess the input data (images or video frames), and then input it into the BiseNetv2 algorithm for dual-branch reasoning.

[0038] Step 3: The input data undergoes dual-branch inference in the BiseNetv2 algorithm. In the detail branch, plume edges are extracted from the input video frame or image. In the semantic branch, background noise and interference in the input video frame or image are denoised to reduce background interference and refine the recognition of gas plume features.

[0039] Step 4: After branch inference using the BiseNetv2 algorithm, BGA feature fusion is performed. The gas plume edges processed in the detail branch and semantic branch are contrasted and fused with the background denoised images to enhance the difference between the low-concentration gas and the background.

[0040] Step 5: After BGA fusion, the BiseNetv2 algorithm performs masking operations, including hole filling, motion consistency filtering, and morphological closing operations on the image or video frame. This completes the image's uneven gas areas, flicker noise, and broken gas plumes. Simultaneously, the gas plume parameters are analyzed, and the gas diffusion area, motion vector, concentration gradient, and center rate curve are calibrated to determine the gas emission volume, diffusion direction, leakage intensity, and gas flow state.

[0041] Step 6: Based on steps 1 to 5, obtain the trained algorithm model (such as Figure 2 As shown), from left to right, the input is an image or video with a gas plume. The semantic branch and detail branch of the core feature extraction module extract the gas plume features in the image or video. The branch reasoning results are BGA-fused through the aggregation module. The image contrast is enhanced through loss function correction and the gas plume is pseudo-colored and displayed on the visualization platform. After hardware configuration, field monitoring is carried out, and the combustible gas plume is captured by a mid-infrared camera. The gas plume features are calibrated and pseudo-colored through the algorithm's pre-processing module, post-processing module and plume parameter quantification module. The gas diffusion area is determined, and gas visualization is achieved through the host computer program to monitor gas leaks in real time. As shown Figure 3 Figure 2 shows a visualization of gas plume diffusion.

[0042] As a preferred embodiment of the present invention, in order to prove the beneficial effects of the present invention, the present invention was configured with hardware and then tested and verified by a third party, and the measured results were consistent with the above-mentioned effective effects. Tests were carried out in Jilin Oilfield and Xinao Gas Company in Leizhou Peninsula. The configured prototype was placed about 50m away from the gas discharge. To ensure the safety of the test, the gas discharge was manually controlled. The prototype was connected to the host computer. After the test started, the host computer displayed the spread of the pseudo-colored gas plume and the gas content. The test results are as follows: Figure 4 and Figure 5 As shown, after testing, the process from capturing the gas plume characteristics to the pseudo-color display is about 7 seconds, and the gas concentration detection time is about 9 seconds. In addition, the upper computer can continuously output the picture to observe the gas diffusion direction. The CH4 concentration measured in Jilin Oilfield is 84%, and the CH4 concentration measured in Xin Ao Gas Company is 96%.

[0043] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A combustible gas imaging method, characterized in that: Gas plume recognition is implemented based on the BiseNetv2 algorithm framework, and the target function is achieved by designing the following modules; The preprocessing module is used to preprocess the input image or video frame. For infrared data, it enhances contrast through dynamic histogram stretching and adaptive pseudo-color mapping; and uses motion consistency filtering to suppress jitter noise between video frames. The post-processing module uses morphological closing operations to link broken gas plumes, fill discontinuous areas within the plume, and smooth adjacent plume edges. It then combines connected domain analysis with area filtering to remove noise, eliminating discrete target noise points whose areas are much smaller than the plume diffusion area. The plume parameter quantification module estimates the gas emission and diffusion area based on the number of mask pixels × spatial resolution. For motion vectors, the gas diffusion direction is predicted using adjacent frame optical flow tracking technology. For gas concentration, the gas leakage intensity classification is calibrated using the grayscale variance of the mask area.

2. The combustible gas imaging method according to claim 1, characterized in that: The method comprises the following specific steps: Step 1: Data collection and building the required algorithm based on the BiseNetv2 algorithm framework; Step 2: Use the segmentation model to label all the data collected in step 1, pre-process the images or video frames in the input data, and then input them into the BiseNetv2 algorithm for dual-branch reasoning; Step 3: The input data undergoes dual-branch inference in the BiseNetv2 algorithm. In the detail branch, plume edges are extracted from the input video frame or image. In the semantic branch, background noise and interference in the input video frame or image are denoised. Step 4: After branch inference using the BiseNetv2 algorithm, BGA feature fusion is performed to contrast and fuse the gas plume edges processed in the detail and semantic branches with the background denoised images to enhance the difference between the low-concentration gas and the background. Step 5: After BGA fusion, the BiseNetv2 algorithm performs mask operations, performs hole filling, motion consistency filtering, and morphological closing operations on the image or video frame to complete the gas inhomogeneity areas, flicker noise, and broken gas plumes in the image. At the same time, the gas plume parameters are analyzed, and the gas diffusion area, motion vector, concentration gradient, and center rate curve are calibrated to obtain the gas emission, diffusion direction, leakage intensity, and gas flow state. Step 6: Based on steps 1 to 5, a trained algorithm model is obtained and configured with the hardware for field monitoring. The combustible gas plume is captured using a mid-infrared camera. The gas plume features are calibrated and pseudo-colored using the algorithm's pre-processing module, post-processing module, and plume parameter quantification module. The gas diffusion area is determined, and gas visualization is achieved through the host computer program to monitor gas leaks in real time.

3. The combustible gas imaging method according to claim 2, characterized in that: The step 1 includes the following specific steps: Various gases are tested under different leakage pressures, different leakage volumes, different ambient temperatures, and different weather conditions. The test data are fully recorded and summarized, cleaned, and then trained and deeply learned under the BiseNetv2 algorithm framework to form the required algorithm.

Images