Combustible gas cloud imaging method based on refrigeration type infrared camera off-site sample learning

Abstract

The invention relates to a gas leakage detection technology based on deep learning, in particular to a combustible gas cloud imaging method based on refrigeration type infrared camera off-site sample learning. The combustible gas cloud imaging method comprises the following steps of S1, making a data set; s2, training a network model; s3, feature extraction and classification; s4, optimizing a loss function and adjusting the model; s5, outputting detection and classification results; the method is mainly used for accurately identifying gas leakage areas and classifying the gas leakage areas, and is widely applied to the fields of industrial safety, environmental monitoring and clean energy, so that the accuracy and robustness of gas leakage detection are improved.

Description

Technical Field

[0001] The present invention relates to a gas leakage detection technology based on deep learning, and in particular to combining plume model features with a YOLOv8 network for accurately identifying and classifying gas leakage areas. Background Art

[0002] Gas leak detection has always been an important task in the field of industrial and environmental monitoring, especially in the natural gas, petroleum, chemical and other industries. Traditional detection methods rely on sensor arrays, infrared imaging or chemical sensors, but these methods have certain limitations in detection accuracy, response speed and adaptability to complex environmental changes.

[0003] In recent years, detection methods based on computer vision and deep learning have gradually gained attention, especially the YOLO series of models, which are widely used in various visual tasks due to their efficient target detection capabilities. However, the traditional YOLO model still has a high false detection and missed detection rate when dealing with small objects or complex backgrounds (such as clouds, smoke, etc.).

[0004] To this end, researchers have tried to combine physical models (such as gas plume models) to further improve detection accuracy, especially when dealing with the diffusion pattern of gas leaks. Summary of the invention

[0005] The gas leak detection and classification network proposed in this paper based on YOLOv8 and plume model feature fusion aims to improve the accuracy and robustness of gas leak detection, especially in complex environments. To this end, we proposed a multi-task joint optimization scheme that combines the constraints of target detection, image segmentation, classification tasks and plume model features to improve the detection effect of deep learning networks by fusing physical model features.

[0006] To achieve the above-mentioned purpose, the following is a detailed implementation scheme of the present invention:

[0007] 1. A combustible gas cloud imaging method based on off-site sample learning of a refrigerated infrared camera, characterized in that it comprises the following steps:

[0008] S1 makes a dataset:

[0009] The collected data is labeled, and then image normalization and data enhancement are performed.

[0010] S2 network model training:

[0011] The dataset prepared in step 1 is input into the improved YOLOv8 network for training, and an additional classification branch is added to YOLOv8 to distinguish gas leakage from background. At the same time, a feature map of gas leakage diffusion is generated through the plume model simulation module and fused with the detection results of YOLOv8.

[0012] S3 feature extraction and classification:

[0013] The gas injection and diffusion process is modeled through the plume model to generate a gas diffusion feature map. The output feature map of the plume model is spliced ​​with the gas region segmentation mask image generated by YOLOv8, and the classification branch is further processed to output a binary classification probability value to determine whether the image area is a gas leak.

[0014] S4 loss function optimization and model adjustment

[0015] Multiple loss terms are added to the loss function to design the loss function and improve the gas leak detection accuracy and robustness of the network.

[0016] S5 detection and classification result output:

[0017] The trained and optimized YOLOv8 model can output the detection box and classification results of the gas leakage area.

[0018] Preferably, in step S2, the improved YOLOv8 network includes:

[0019] This scheme adds a classification branch to YOLOv8. It classifies the candidate boxes and segmentation masks generated by the target detection and image segmentation branches in combination with the physical features extracted by the plume model to distinguish gas leakage from the background. At the same time, the target boxes and segmentation masks generated by the detection branch of YOLOv8 are fused with the gas diffusion feature map generated by the plume model to provide richer input information, thereby improving the accuracy of gas leakage judgment. In addition, the YOLOv8 backbone network adopts an improved convolutional neural network architecture to enhance feature extraction capabilities and ensure the robustness of the network at multiple scales.

[0020] Preferably, in step S2, the gas jet diffusion characteristics and plume model include:

[0021] Assume that the gas concentration distribution C(x,y) can be represented by the following physical model:

[0022]

[0023] in:

[0024] C(x,y,t) is the concentration of the gas at the position (x,y) at time t;

[0025] Q is the gas leakage rate;

[0026] σ x and σ y are the standard deviations of gas diffusion along the x and y directions, respectively, reflecting the directionality and range of gas diffusion;

[0027] x and y represent spatial coordinates.

[0028] Assume that a detection box is output through the YOLOv8 network, and a region R = {(x i ,y i )}, we can determine whether the object is a gas by calculating the symmetry of the gas concentration distribution in the area. Here, (x i ,y i ) are the pixel coordinates in the image.

[0029] Use the gradient operator to calculate the concentration gradient in the target area:

[0030]

[0031] Gas leakage usually manifests itself as diffusion in one direction being more obvious than in another direction, resulting in a concentration distribution diagram showing certain directional differences. Directional characteristics of the target area concentration diagram:

[0032]

[0033] Where D asymmetry is a value that indicates a difference in symmetry or directionality.

[0034] Preferably, in step S3, the sub-step of plume model feature extraction includes:

[0035] The gas concentration distribution maps generated by the plume model are used as auxiliary feature inputs of the network. These feature maps contain the directionality and spatial distribution of gas diffusion. Subsequently, these feature maps are processed using a convolutional neural network (CNN) to extract feature vectors containing gas diffusion patterns and concatenate them with the visual features extracted by the network.

[0036] Preferably, in step S4, the sub-steps of loss function design include:

[0037] The above symmetry evaluation and concentration gradient calculation are introduced into the loss function, and a jet diffusion consistency loss is designed.

[0038]

[0039] in:

[0040] D asymmetry It is a value that evaluates the asymmetry of the target area and reflects the injection characteristics of the gas;

[0041] It is the gradient of the gas concentration map, reflecting the directionality of gas diffusion;

[0042] α and β are weight coefficients used to balance the contribution of symmetry and gradient terms;

[0043] τ is a threshold value used to determine whether the gas has significant jet diffusion characteristics. If it exceeds this value, it is considered to be gas.

[0044] To ensure the consistency between network features and plume model features, a plume model feature consistency loss is designed.

[0045]

[0046] The jet diffusion consistency loss L jet and plume model characteristics consistency L align Introduced into the total loss function, the loss function becomes:

[0047] L total =λ det L det +λ seg L seg +λ cls L cls +λ align L align +λ jet L jet

[0048] Among them, L det , L seg , L cls They are target detection loss, segmentation loss, classification loss and,λ det ,λ seg ,λ cls ,λ align ,λ jet is the corresponding weight coefficient.

[0049] The beneficial effects of the present invention are:

[0050] Through the above scheme, the network can not only make full use of the efficient detection capability of YOLOv8, but also combine the physical information provided by the plume model to propose an efficient gas leak detection method. The network can not only detect the location and shape of gas leaks, but also improve the classification accuracy of gas leaks by introducing the features of the plume model, especially in distinguishing between gas and easily confused background objects such as clouds, further improving the accuracy and robustness of gas leak detection, especially in complex scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a specific implementation flow chart of the present invention. DETAILED DESCRIPTION

[0052] For ease of understanding, here we combine Figure 1 , the specific working mode of the present invention is further described as follows:

[0053] The specific implementation process of the present invention is as follows:

[0054] 1. A combustible gas cloud imaging method based on off-site sample learning of a refrigerated infrared camera, characterized in that it comprises the following steps:

[0055] S1 makes a dataset:

[0056] Firstly, a cooled medium-wave infrared camera is used to collect 320×256 grayscale image data, which includes the gas leakage area and background, and annotates them (including the location, shape and category label of the leakage area). Then, the image is preprocessed, including normalization and data enhancement through rotation, scaling, translation, etc., to improve the diversity of training data.

[0057] S2 network model training:

[0058] In the present invention, the YOLOv8 network is an important part of the gas leakage detection method.

[0059] The improvement of the YOLOv8 network is that an additional classification branch is added to the YOLOv8 network, which is responsible for judging whether the area is a gas leak based on the characteristics of the gas leak area, and is trained through multi-layer convolution and fully connected networks, and the loss function is modified to include target detection loss, segmentation loss, classification loss, and plume model feature consistency loss. In addition, the YOLOv8 backbone network adopts an improved convolutional neural network architecture to enhance feature extraction capabilities and ensure the robustness of the network at multiple scales.

[0060] The target box and segmentation mask generated by the detection branch of YOLOv8 are fused with the gas diffusion feature map generated by the plume model to provide richer input information, thereby improving the accuracy of gas leakage judgment.

[0061] Input the dataset prepared in step 1 into the YOLOv8 network for training. The input image is a 320×256 grayscale image.

[0062] Among them, the gas jet diffusion characteristics and plume model include:

[0063] Assume that the gas concentration distribution C(x,y) can be represented by the following physical model:

[0064]

[0065] in:

[0066] C(x, y, t) is the concentration of the gas at the position (x, y) at time t;

[0067] Q is the gas leakage rate;

[0068] σ x and σ y are the standard deviations of gas diffusion along the x and y directions, respectively, reflecting the directionality and range of gas diffusion;

[0069] x and y represent spatial coordinates.

[0070] Assume that a detection box is output through the YOLOv8 network, and a region R = {(x i ,y i )}, we can determine whether the object is a gas by calculating the symmetry of the gas concentration distribution in the area. Here, (x i ,y i ) are the pixel coordinates in the image.

[0071] Calculate the concentration gradient in the area. The gas injection characteristics will show that the concentration is higher in the center area and gradually decreases outwards, and there is a certain directionality. Use the gradient operator to calculate the concentration gradient of the target area:

[0072]

[0073] Then the uniformity of the gradient direction is calculated, and the gas jet flow will have obvious characteristics of expansion on one side and contraction on the other side.

[0074] The asymmetry of the concentration distribution in the target area is calculated to further determine whether it meets the characteristics of gas jet.

[0075] Gas leakage usually manifests itself as diffusion in one direction being more obvious than in another direction, resulting in a concentration distribution diagram showing certain directional differences. Directional characteristics of the target area concentration diagram:

[0076]

[0077] Where D asymmetry is a value that indicates a difference in symmetry or directionality. If this value is large, it means that the target object has significant jet diffusion characteristics. From the above formula, it can be seen that gas leakage will form an asymmetric diffusion pattern, that is, the gas concentration gradually becomes thinner at a distance from the leakage source, resulting in a contraction of the concentration on one side and an expansion of the concentration on the other side.

[0078] S3 feature extraction and classification:

[0079] In the gas leak detection task, the plume model is a physics-based mathematical model used to describe the diffusion process after a gas leak. The model generates a gas concentration distribution map by considering factors such as the intensity of the gas leak source, wind speed, air pressure, and temperature, revealing the diffusion law of the gas in space and time. This concentration distribution map provides key information for determining the location and nature of the gas leak.

[0080] The gas concentration distribution maps generated by the plume model are used as auxiliary input features of the network. These feature maps show the directionality and spatial distribution of gas diffusion, which helps to distinguish gas leaks from other background interference (such as clouds, smoke, etc.) in visual information. The process of generating feature maps relies on sensor data (such as wind speed, humidity, temperature, etc.) and environmental parameters to provide accurate gas diffusion information. The output feature map of the plume model is spliced ​​with the gas region segmentation mask image generated by YOLOv8 to provide additional physical information support for the network.

[0081] In the network, the spliced ​​feature map is processed by a convolutional neural network (CNN) to extract a feature vector containing the gas diffusion pattern. This feature vector is then used as the input of the classification branch, which uses a fully connected network to perform calculations and output a binary classification result to determine whether the image area is a gas leak. By splicing these physical model features with the visual features extracted by the network itself, the accuracy of the classification decision is enhanced.

[0082] S4 loss function optimization and model adjustment

[0083] The above symmetry evaluation and concentration gradient calculation are introduced into the loss function to design a jet diffusion consistency loss. The goal of this loss is to minimize the diffusion asymmetry of the target area to ensure that the network can accurately distinguish between gas and other similar objects.

[0084]

[0085] in:

[0086] D asymmetry It is a value that evaluates the asymmetry of the target area and reflects the injection characteristics of the gas;

[0087] It is the gradient of the gas concentration map, reflecting the directionality of gas diffusion;

[0088] α and β are weight coefficients used to balance the contribution of symmetry and gradient terms;

[0089] τ is a threshold value used to determine whether the gas has significant jet diffusion characteristics. If it exceeds this value, it is considered to be gas.

[0090] In order to ensure the consistency between the network features and the plume model features, a plume model feature consistency loss is designed. This loss is calculated by calculating the feature map F generated by YOLOv8 YOLO The characteristic map F generated by the plume model plume The pixel-by-pixel difference between them is used to optimize the network output so that the gas diffusion characteristics learned by the network are consistent with the physical model:

[0091]

[0092] The jet diffusion consistency loss L jet and plume model characteristics consistency L align Introduced into the total loss function, the loss function becomes:

[0093] L total =λ det L det +λ seg L seg +λ cls L cls +λ align L align +λ jet L jet

[0094] Among them, L det , L seg , L cls They are target detection loss, segmentation loss, classification loss and,λ det ,λ seg ,λ cls ,λ align ,λ jet is the corresponding weight coefficient.

[0095] S5 detection and classification result output:

[0096] The trained and optimized YOLOv8 model can output the detection box of the gas leakage area, which includes the location, size and category confidence (gas or non-gas) of the gas leakage. The network determines whether the area is a gas leakage through the classification branch. If the classification result is "gas", the coordinates and category information of the gas leakage area are output; if it is "non-gas", the area is ignored.

[0097] The final output results include two aspects: one is the gas leak detection box, which identifies the specific location and size of the leak; the other is the classification result, which determines whether the area is a gas leak through the probability value of the binary classification, providing a basis for further emergency response and decision-making.

[0098] Of course, for those skilled in the art, the present invention is not limited to the details of the above exemplary embodiments, but also includes the same or similar methods that can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive, and the scope of the present invention is limited by the attached claims rather than the above description.

[0099] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0100] The technical parts not described in detail in the present invention are all well-known technologies.

Claims

1. A combustible gas cloud imaging method based on off-site sample learning of a refrigerated infrared camera, characterized in that: The following steps are involved: S1 makes a dataset: The collected data is labeled, and then image normalization and data enhancement are performed. S2 network model training: The dataset prepared in step 1 is input into the improved YOLOv8 network for training, and an additional classification branch is added to YOLOv8 to distinguish gas leakage from background. At the same time, a feature map of gas leakage diffusion is generated through the plume model simulation module and fused with the detection results of YOLOv8. S3 feature extraction and classification: The gas injection and diffusion process is modeled through the plume model to generate a gas diffusion feature map. The output feature map of the plume model is spliced ​​with the gas region segmentation mask image generated by YOLOv8, and the classification branch is further processed to output a binary classification probability value to determine whether the image area is a gas leak. S4 loss function optimization and model adjustment: Multiple loss terms are added to the loss function to design the loss function and improve the gas leak detection accuracy and robustness of the network. S5 detection and classification result output: The trained and optimized YOLOv8 model can output the detection box and classification results of the gas leakage area.

2. The combustible gas cloud imaging method based on off-site sample learning of a refrigerated infrared camera according to claim 1 is characterized in that : In step S2, the improved YOLOv8 network includes: This scheme adds a classification branch to YOLOv8. It classifies the candidate boxes and segmentation masks generated by the target detection and image segmentation branches in combination with the physical features extracted by the plume model to distinguish gas leakage from the background. At the same time, the target boxes and segmentation masks generated by the detection branch of YOLOv8 are fused with the gas diffusion feature map generated by the plume model to provide richer input information, thereby improving the accuracy of gas leakage judgment. In addition, the YOLOv8 backbone network adopts an improved convolutional neural network architecture to enhance feature extraction capabilities and ensure the robustness of the network at multiple scales.

3. The combustible gas cloud imaging method based on off-site sample learning of a refrigerated infrared camera according to claim 1 is characterized in that : In step S2, the gas jet diffusion characteristics and plume model include: Assume that the gas concentration distribution C(x, y) can be represented by the following physical model: in: C(x, y, t) is the concentration of the gas at the position (x, y) at time t; Q is the gas leakage rate; σ x and σ y are the standard deviations of gas diffusion along the x and y directions, respectively, reflecting the directionality and range of gas diffusion; x and y represent spatial coordinates. Assume that a detection box is output through the YOLOv8 network, and a region R = {(x i ,y i )}, we can determine whether the object is a gas by calculating the symmetry of the gas concentration distribution in the area. Here, (x i ,y i ) are the pixel coordinates in the image. Use the gradient operator to calculate the concentration gradient in the target area: Gas leakage usually manifests itself as diffusion in one direction being more obvious than in another direction, resulting in a concentration distribution diagram showing certain directional differences. Directional characteristics of the target area concentration diagram: Where D asymmetry is a value that indicates a difference in symmetry or directionality.

4. The combustible gas cloud imaging method based on off-site sample learning of a refrigerated infrared camera according to claim 1 is characterized in that : In step S3, the sub-step of plume model feature extraction includes: The gas concentration distribution maps generated by the plume model are used as auxiliary feature inputs of the network. These feature maps contain the directionality and spatial distribution of gas diffusion. Subsequently, these feature maps are processed using a convolutional neural network (CNN) to extract feature vectors containing gas diffusion patterns and concatenate them with the visual features extracted by the network.

5. The combustible gas cloud imaging method based on off-site sample learning of a refrigerated infrared camera according to claim 1 is characterized in that In step S4, the sub-steps of loss function design include: The above symmetry evaluation and concentration gradient calculation are introduced into the loss function, and a jet diffusion consistency loss is designed. in: D asymmetry It is a value that evaluates the asymmetry of the target area and reflects the injection characteristics of the gas; It is the gradient of the gas concentration map, reflecting the directionality of gas diffusion; α and β are weight coefficients used to balance the contribution of symmetry and gradient terms; τ is a threshold value used to determine whether the gas has significant jet diffusion characteristics. If it exceeds this value, it is considered to be gas. To ensure the consistency between network features and plume model features, a plume model feature consistency loss is designed. The jet diffusion consistency loss L jet and plume model characteristics consistency L align Introduced into the total loss function, the loss function becomes: L total =λ det L det +λ seg L seg +λ cls L cls +λ align L align +λ jet L jet Among them, L det , L seg , L cls They are target detection loss, segmentation loss, classification loss and,λ det ,λ seg ,λ cls ,λ align ,λ jet is the corresponding weight coefficient.

Images