Multi-head attention mechanism and field self-adaption integrated industrial pipeline water leakage detection method and system
By integrating multi-head attention mechanisms with domain-adaptive methods, the problems of data scarcity and environmental interference in industrial pipeline leak detection are solved, achieving high-precision and robust cross-domain detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-03
- Publication Date
- 2026-03-13
AI Technical Summary
Existing technologies for detecting leaks in industrial pipelines suffer from problems such as data scarcity, strong environmental interference, and insufficient cross-domain performance, leading to frequent misjudgments and missed detections.
We employ a method that integrates multi-head attention mechanism and domain adaptation. By constructing a multi-head attention feature enhancement module and domain-adaptive adversarial training, and combining synthetic and real data, we enhance the model's cross-domain generalization ability and robustness.
It significantly improves the detection capability of small targets and blurred edges, reduces the false detection rate, and achieves high-precision leak detection under data-scarce conditions, making it suitable for complex industrial scenarios.
Smart Images
Figure CN121661319A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of computer vision and intelligent detection technology, specifically to an industrial pipeline leakage detection method and system that integrates multi-head attention mechanism and domain adaptation. Background Technology
[0002] In the industrial manufacturing sector, pipeline systems carry high-temperature and high-pressure media, and their safety is directly related to production stability and the safety of personnel. A leak can not only cause significant economic losses but also potentially trigger a serious safety accident.
[0003] Traditional leak detection relies on manual inspections or acoustic sensors, but real leak events are rare and data confidentiality and labeling costs are high, resulting in an extreme scarcity of training samples. At the same time, interference from strong steam, metal reflection, and corrosion at the scene can cause frequent misjudgments and missed detections in visual detection.
[0004] Furthermore, the synthetic leaked images generated by software differ significantly from real images in terms of texture, lighting, and background. Existing general-purpose object detection algorithms suffer from high false negative rates for small targets, poor edge blurring recognition, and insufficient cross-domain performance when directly applied to industrial scenarios. Therefore, a solution combining attention mechanisms and domain adaptation is needed to improve the robustness and accuracy of detection. Summary of the Invention
[0005] The purpose of this invention is to provide an industrial pipeline leakage detection method and system that integrates multi-head attention mechanism and domain adaptation.
[0006] This invention relates to the fields of computer vision and intelligent detection technology, and to an intelligent detection method for industrial pipeline leaks based on deep learning, particularly a cross-domain leak detection method that integrates multi-head attention mechanism and domain adaptation technology, which is suitable for industrial leak detection tasks such as blast furnace pipelines in steel plants where data is scarce and environmental interference is strong.
[0007] To achieve the objectives of this invention, the technical solution provided by this invention is as follows:
[0008] First aspect
[0009] This application provides a method for detecting leaks in industrial pipelines that integrates multi-head attention mechanisms and domain adaptation, comprising the following steps:
[0010] Step S1: Construction of industrial pipeline image dataset;
[0011] Step S2: Design of multi-head attention feature enhancement module;
[0012] Step S3: Domain-Adaptive Adversarial Training;
[0013] Step S4: Model training and inference application.
[0014] Step S1 includes the following:
[0015] (1) Training dataset construction:
[0016] First, a publicly available leaked image dataset is introduced as an open-source source domain image to provide the model with basic knowledge of leak patterns;
[0017] Secondly, use 3D simulation software to generate defective images of the target domain with precise leak annotations;
[0018] Finally, defect-free images of the target domain in a real industrial setting are acquired.
[0019] The three parts of data are combined to form a hybrid dataset containing both defective and defect-free samples for adversarial training.
[0020] (2) Evaluation test set construction: In order to objectively evaluate the generalization ability of the model, an evaluation test set independent of the training set is constructed. The evaluation test set is constructed by fusing the leak foreground generated by 3D software with the real leak-free pipeline background, and is reviewed and labeled by industry experts to ensure its visual realism and labeling accuracy.
[0021] In step S2, MDANet, as a feature enhancement method, is integrated between the feature extraction layers of the backbone network to form a feature extractor; the processing flow specifically includes the following:
[0022] (1) Feature input and distribution: The feature map of the upstream network is used as input and copied first to provide a data basis for subsequent multi-branch parallel processing;
[0023] (2) Branch 1: Global context modeling. This branch is responsible for capturing long-distance dependencies in the feature maps. Its core is to use a multi-head attention mechanism to perform weighted modeling of the input features using the following formula:
[0024]
[0025] Where Q, K, and V are the query, key, and value obtained from the input feature mapping, respectively, and d is the channel dimension. Through this mechanism, this branch can effectively understand the overall layout and global information of the screen.
[0026] (3) Branch 2: Feature compression and receptive field expansion. This branch expands the perception range while reducing the amount of computation by reducing the spatial resolution of the feature map, and focuses on integrating more macroscopic regional information.
[0027] (4) Branch 3: Local detail refinement. This branch focuses on processing the local fine texture and edge information of the feature map. To highlight key feature regions, this branch integrates a channel attention enhancement mechanism, which is calculated as follows:
[0028]
[0029] Among them, X c w represents the feature map of the c-th channel. c The corresponding attention weight is σ, which is the Sigmoid function. In this way, background interference can be adaptively suppressed and the area with leakage signs can be highlighted.
[0030] (5) Feature integration and abstraction and output: The processing results of branch 1 and branch 2 are initially fused. Then, the fusion result and the output of branch 3 enter the global information integration stage to carry out the final cross-dimensional information extraction, and finally generate a highly abstract downstream feature map. This feature map integrates information from three dimensions: global, macro and local, and is passed to the subsequent layers of the network for use.
[0031] The backbone network is YOLOv8.
[0032] Step S3 includes the following:
[0033] (1) Data input and feature extraction: The source domain data and target domain mixed data constructed in step S1 are input into the MDANet-YOLOv8 feature extractor to extract high-dimensional features;
[0034] (2) Gradient inversion and domain discrimination: The extracted features first pass through a gradient inversion layer GRL. This layer does not change the features during forward propagation, but inverts the gradient sign during backward propagation. Subsequently, the features are fed into a domain discriminator D, which outputs the discrimination result of the domain to which it belongs.
[0035] (3) Adversarial Loss and Backward Optimization: The output of the domain discriminator is used to calculate the domain adversarial loss, which aligns the feature distributions of the source and target domains. Its core idea can be expressed by the following formula:
[0036]
[0037] Among them, D s With D tLet f(·) represent the sample distributions of the source and target domains, respectively. f(·) is the feature extractor, and D(·) is the domain discriminator. The loss is updated iteratively through backpropagation. Its gradient is inverted by GRL and then applied unilaterally to the feature extractor. The goal of this optimization is to make the features generated by the feature extractor confuse the domain discriminator to the greatest extent possible, and finally achieve domain invariance.
[0038] The feature extractor is the MDANet-YOLOv8 feature extractor.
[0039] Step S4 includes the following:
[0040] (1) Model adversarial training: The MDANet-YOLOv8 framework, which integrates MDANet, is trained end-to-end using the domain adversarial optimization method;
[0041] (2) Performance verification and deployment: Using the independent evaluation test set built in step S1, the cross-domain generalization performance of the trained detection model is evaluated; after the evaluation is passed, the model is advanced to the industrial inspection inference deployment stage, and it is deployed on the field server or edge device to receive real-time monitoring images and output detection results, so as to realize the final implementation of industrial scenario applications.
[0042] Second aspect
[0043] Corresponding to the above method, this application provides an industrial pipeline leakage detection system that integrates multi-head attention mechanism and domain adaptation, including the following units: dataset construction unit, module design unit, model training and inference application unit, and domain adaptive adversarial training unit;
[0044] The dataset construction unit is used for constructing industrial pipeline image datasets;
[0045] The module design unit is used for the design of the multi-head attention feature enhancement module;
[0046] The domain-adaptive adversarial training unit is used for domain-adaptive adversarial training;
[0047] The model training and inference application unit is used for model training and inference applications.
[0048] Specifically, the dataset construction unit is used to perform the following:
[0049] (1) Training dataset construction:
[0050] First, a publicly available leaked image dataset is introduced as an open-source source domain image to provide the model with basic knowledge of leak patterns;
[0051] Secondly, use 3D simulation software to generate defective images of the target domain with precise leak annotations;
[0052] Finally, defect-free images of the target domain in a real industrial setting are acquired.
[0053] The three parts of data are combined to form a hybrid dataset containing both defective and defect-free samples for adversarial training.
[0054] (2) Evaluation test set construction: In order to objectively evaluate the generalization ability of the model, an evaluation test set independent of the training set is constructed. The evaluation test set is constructed by fusing the leak foreground generated by 3D software with the real leak-free pipeline background, and is reviewed and labeled by industry experts to ensure its visual realism and labeling accuracy.
[0055] In the module design unit execution, MDANet, as a feature enhancement method, is integrated between the feature extraction layers of the backbone network to form a feature extractor; the processing flow specifically includes the following:
[0056] (1) Feature input and distribution: The feature map of the upstream network is used as input and copied first to provide a data basis for subsequent multi-branch parallel processing;
[0057] (2) Branch 1: Global context modeling. This branch is responsible for capturing long-distance dependencies in the feature maps. Its core is to use a multi-head attention mechanism to perform weighted modeling of the input features using the following formula:
[0058]
[0059] Where Q, K, and V are the query, key, and value obtained from the input feature mapping, respectively, and d is the channel dimension. Through this mechanism, this branch can effectively understand the overall layout and global information of the screen.
[0060] (3) Branch 2: Feature compression and receptive field expansion. This branch expands the perception range while reducing the amount of computation by reducing the spatial resolution of the feature map, and focuses on integrating more macroscopic regional information.
[0061] (4) Branch 3: Local detail refinement. This branch focuses on processing the local fine texture and edge information of the feature map. To highlight key feature regions, this branch integrates a channel attention enhancement mechanism, which is calculated as follows:
[0062]
[0063] Among them, X c w represents the feature map of the c-th channel. c The corresponding attention weight is σ, which is the Sigmoid function. In this way, background interference can be adaptively suppressed and the area with leakage signs can be highlighted.
[0064] (5) Feature integration and abstraction and output: The processing results of branch 1 and branch 2 are initially fused. Then, the fusion result and the output of branch 3 enter the global information integration stage to carry out the final cross-dimensional information extraction, and finally generate a highly abstract downstream feature map. This feature map integrates information from three dimensions: global, macro and local, and is passed to the subsequent layers of the network for use.
[0065] This invention discloses an industrial pipeline leak detection method integrating multi-head attention mechanism and domain adaptation, comprising: constructing a cross-domain dataset consisting of synthetic leak images, real background images, and pseudo-leak images; designing a multi-head attention module (MDANet) integrating dynamic multi-kernel convolution and multi-path fusion mechanisms to enhance the feature representation of blurred edges and small target water traces; introducing an improved domain adversarial network (DANN), combining high-gradient region focusing and dynamic loss weighting strategies to reduce the feature distribution difference between the synthetic and real domains; and finally training a leak detection model with strong cross-domain generalization ability. This invention can still achieve accurate detection of industrial pipeline leaks even in the absence of real leak samples, and effectively suppresses false detections caused by interference from steam, reflections, etc., making it suitable for complex industrial scenarios such as blast furnace pipelines in steel plants.
[0066] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0067] (1) Enhance the detection capability of small targets and blurred edges: Through dynamic multi-kernel dilated convolution and multi-path fusion mechanism, the detection rate of weak watermarks and small leaks is significantly improved.
[0068] (2) Strong cross-domain generalization ability: Through the improved DANN structure and adversarial focus mechanism, the distribution gap between synthetic and real data is effectively narrowed, solving the domain transfer problem.
[0069] (3) High data utilization: By combining synthetic and real data, high-precision detection results can still be obtained even when the number of real leakage samples is limited.
[0070] (4) Robustness improvement: The dynamic loss weight and background noise injection mechanism effectively suppress interference factors such as steam, metal reflection, and corrosion, which significantly reduces the false detection rate.
[0071] (5) High practical application value: This method only requires a small number of real images to achieve high-performance cross-domain detection, and is suitable for real-time leakage monitoring in industrial environments such as steel plants. Attached Figure Description
[0072] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention;
[0073] Figure 2This is a schematic diagram of the design process of the multi-head attention feature enhancement module provided in an embodiment of the present invention;
[0074] Figure 3 The flowchart of the domain-adaptive adversarial training method provided in the embodiments of the present invention. Detailed Implementation
[0075] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0076] The technical solution of this invention aims to provide a systematic, end-to-end method for detecting cross-domain leaks in industrial pipelines. The overall process of this method is as follows: Figure 1 As shown, it covers three core modules: data construction, model adversarial training, and performance verification and deployment. The specific technical steps are broken down as follows:
[0077] Step S1: Construction of Industrial Pipeline Image Dataset
[0078] This step corresponds to Figure 1 The data building module in the system is designed to provide an independent, high-quality data foundation for subsequent model training and performance validation.
[0079] (1) Training dataset construction:
[0080] First, a publicly available leaked image dataset is introduced as an open-source source domain image to provide the model with basic knowledge of leak patterns.
[0081] Secondly, 3D simulation software is used to generate defective images of the target domain with precise leakage annotations.
[0082] Finally, defect-free images of the target domain from real-world industrial scenarios are collected. These three data sets are then combined to form a hybrid dataset containing both defective and defect-free samples for adversarial training.
[0083] (2) Evaluation Test Set Construction: To objectively evaluate the generalization ability of the model, a test set independent of the training set needs to be constructed. This test set is constructed by fusing the leak foreground generated by 3D software with the background of a real leak-free pipeline, and is reviewed and annotated by industry experts to ensure its visual realism and annotation accuracy.
[0084] Step S2: Design of Multi-Head Attention Feature Enhancement Module (MDANet)
[0085] This step aims to enhance the model's ability to perceive and represent difficult-to-detect features such as blurred edges and small objects. MDANet, as a feature enhancement method, is integrated between the feature extraction layers of the backbone network (such as YOLOv8). Its processing flow is as follows: Figure 2 As shown, it specifically includes:
[0086] (1) Feature input and distribution: The method takes the upstream feature map of the network as input and first copies it to provide a data basis for subsequent multi-branch parallel processing.
[0087] (2) Branch 1: Global Context Modeling. This branch is responsible for capturing long-distance dependencies in the feature maps. Its core is the use of a multi-head attention mechanism, which models the input features using the following weighted formula:
[0088]
[0089] Where Q, K, and V represent the query, key, and value obtained from the input feature mapping, respectively, and d is the channel dimension. Through this mechanism, this branch can effectively understand the overall layout and global information of the screen.
[0090] (3) Branch 2: Feature compression and receptive field expansion. This branch expands the perception range while reducing the computational load by reducing the spatial resolution of the feature map, focusing on integrating more macroscopic regional information.
[0091] (4) Branch 3: Local Detail Refinement. This branch focuses on processing local fine texture and edge information in the feature map. To highlight key feature regions, this branch integrates a channel attention enhancement mechanism, which is calculated as follows:
[0092]
[0093] Among them, X c w represents the feature map of the c-th channel. c The corresponding attention weight is σ, where σ is the Sigmoid function. In this way, the present invention can adaptively suppress background interference and highlight areas showing signs of leakage.
[0094] (5) Feature integration and abstraction and output: The processing results of branch 1 and branch 2 are initially fused. Then, the fusion result and the output of branch 3 enter the global information integration stage to carry out the final cross-dimensional information extraction, and finally generate a highly abstract downstream feature map. This feature map integrates information from three dimensions: global, macro and local, and is passed to the subsequent layers of the network for use.
[0095] Step S3: Domain Adaptive Adversarial Training
[0096] This step aims to reduce the difference in feature distribution between synthetic and real-domain data, thereby improving the model's cross-domain generalization performance. The method flow is as follows: Figure 3 As shown, the core lies in performing unidirectional adversarial optimization on the feature extractor:
[0097] (1) Data input and feature extraction: The source domain data and target domain mixed data constructed in step 1 are input into the MDANet-YOLOv8 feature extractor to extract high-dimensional features.
[0098] (2) Gradient inversion and domain discrimination: The extracted features first pass through a gradient inversion layer (GRL), which does not change the features during forward propagation, but inverts the sign of their gradients during backward propagation. Subsequently, the features are fed into a domain discriminator D, which outputs the discrimination result of the domain to which the features belong.
[0099] (3) Adversarial Loss and Backward Optimization: The output of the domain discriminator is used to calculate the domain adversarial loss, which aligns the feature distributions of the source and target domains. Its core idea can be expressed by the following formula:
[0100]
[0101] Among them, D s With D t Let f(·) represent the sample distributions of the source and target domains, respectively. Let f(·) be the feature extractor and D(·) be the domain discriminator. The loss is updated iteratively through backpropagation. Its gradient is inverted by GRL and then applied unilaterally to the feature extractor. The goal of this optimization is to maximize the obfuscation of the domain discriminator by the features generated by the feature extractor, ultimately achieving domain invariance.
[0102] Step S4: Model Training and Inference Application
[0103] This step corresponds to Figure 1 The last two modules are the process of implementing the above technical solutions.
[0104] (1) Model adversarial training: The MDANet-YOLOv8 framework, which integrates MDANet (step S2), is trained end-to-end using a domain adversarial optimization method (step S3). During training, strategies such as gradient clipping and online background noise injection can be used to further improve the training stability and environmental robustness of the model.
[0105] (2) Performance Verification and Deployment: Using the independent evaluation test set constructed in step S1, the trained detection model is evaluated for cross-domain generalization performance. After passing the evaluation, the model is advanced to the industrial inspection inference deployment stage, deployed on field servers or edge devices, receiving real-time monitoring images and outputting detection results, thus realizing the final implementation of industrial application scenarios.
[0106] Finally, it should be noted that the above embodiments are merely illustrative and explanatory of the present invention, and are not intended to limit the present invention to the scope of the described embodiments. Furthermore, those skilled in the art will understand that the present invention is not limited to the above embodiments, and many more variations and modifications can be made based on the teachings of the present invention, all of which fall within the scope of protection claimed by the present invention.
Claims
1. A method for detecting leaks in industrial pipelines that integrates multi-head attention mechanisms and domain adaptation, characterized in that, Includes the following steps: Step S1: Construction of industrial pipeline image dataset; Step S2: Design of multi-head attention feature enhancement module; Step S3: Domain-Adaptive Adversarial Training; Step S4: Model training and inference application.
2. The industrial pipeline leakage detection method integrating multi-head attention mechanism and domain adaptation as described in claim 1, characterized in that, Step S1 includes the following: (1) Training dataset construction: First, a publicly available leaked image dataset is introduced as an open-source source domain image to provide the model with basic knowledge of leak patterns; Secondly, use 3D simulation software to generate defective images of the target domain with precise leak annotations; Finally, defect-free images of the target domain in a real industrial setting are acquired. The three parts of data are combined to form a hybrid dataset containing both defective and defect-free samples for adversarial training. (2) Evaluation test set construction: In order to objectively evaluate the generalization ability of the model, an evaluation test set independent of the training set is constructed. The evaluation test set is constructed by fusing the leak foreground generated by 3D software with the real leak-free pipeline background, and is reviewed and labeled by industry experts to ensure its visual realism and labeling accuracy.
3. The industrial pipeline leakage detection method integrating multi-head attention mechanism and domain adaptation as described in claim 1, characterized in that, In step S2, MDANet, as a feature enhancement method, is integrated between the feature extraction layers of the backbone network to form a feature extractor; the processing flow specifically includes the following: (1) Feature input and distribution: The feature map of the upstream network is used as input and copied first to provide a data basis for subsequent multi-branch parallel processing; (2) Branch 1: Global context modeling. This branch is responsible for capturing long-distance dependencies in the feature maps. Its core is to use a multi-head attention mechanism to perform weighted modeling of the input features using the following formula: Where Q, K, and V are the query, key, and value obtained from the input feature mapping, respectively, and d is the channel dimension. Through this mechanism, this branch can effectively understand the overall layout and global information of the screen. (3) Branch 2: Feature compression and receptive field expansion. This branch expands the perception range while reducing the amount of computation by reducing the spatial resolution of the feature map, and focuses on integrating more macroscopic regional information. (4) Branch 3: Local detail refinement. This branch focuses on processing the local fine texture and edge information of the feature map. To highlight key feature regions, this branch integrates a channel attention enhancement mechanism, which is calculated as follows: Among them, X c w represents the feature map of the c-th channel. c The corresponding attention weight is σ, which is the Sigmoid function. In this way, background interference can be adaptively suppressed and the area with leakage signs can be highlighted. (5) Feature integration and abstraction and output: The processing results of branch 1 and branch 2 are initially fused. Then, the fusion result and the output of branch 3 enter the global information integration stage to carry out the final cross-dimensional information extraction, and finally generate a highly abstract downstream feature map. This feature map integrates information from three dimensions: global, macro and local, and is passed to the subsequent layers of the network for use.
4. The industrial pipeline leakage detection method integrating multi-head attention mechanism and domain adaptation as described in claim 3, characterized in that, The backbone network is YOLOv8.
5. The industrial pipeline leakage detection method integrating multi-head attention mechanism and domain adaptation according to claim 4, characterized in that, Step S3 includes the following: (1) Data input and feature extraction: The source domain data and target domain mixed data constructed in step S1 are input into the MDANet-YOLOv8 feature extractor to extract high-dimensional features; (2) Gradient inversion and domain discrimination: The extracted features first pass through a gradient inversion layer GRL. This layer does not change the features during forward propagation, but inverts the gradient sign during backward propagation. Subsequently, the features are fed into a domain discriminator D, which outputs the discrimination result of the domain to which it belongs. (3) Adversarial Loss and Backward Optimization: The output of the domain discriminator is used to calculate the domain adversarial loss, which aligns the feature distributions of the source and target domains. Its core idea can be expressed by the following formula: Among them, D s With D t Let f(·) represent the sample distributions of the source and target domains, respectively. f(·) is the feature extractor, and D(·) is the domain discriminator. The loss is updated iteratively through backpropagation. Its gradient is inverted by GRL and then applied unilaterally to the feature extractor. The goal of this optimization is to make the features generated by the feature extractor confuse the domain discriminator to the greatest extent possible, and finally achieve domain invariance.
6. The industrial pipeline leakage detection method integrating multi-head attention mechanism and domain adaptation as described in claim 4, characterized in that, The feature extractor is the MDANet-YOLOv8 feature extractor.
7. The industrial pipeline leakage detection method integrating multi-head attention mechanism and domain adaptation as described in claim 6, characterized in that, Step S4 includes the following: (1) Model adversarial training: The MDANet-YOLOv8 framework, which integrates MDANet, is trained end-to-end using the domain adversarial optimization method; (2) Performance verification and deployment: Using the independent evaluation test set built in step S1, the cross-domain generalization performance of the trained detection model is evaluated; after the evaluation is passed, the model is advanced to the industrial inspection inference deployment stage, and it is deployed on the field server or edge device to receive real-time monitoring images and output detection results, so as to realize the final implementation of industrial scenario applications.
8. An industrial pipeline leakage detection system integrating multi-head attention mechanism and domain adaptation, characterized in that, It includes the following units: dataset construction unit, module design unit, model training and inference application unit, and domain adaptive adversarial training unit; The dataset construction unit is used for constructing industrial pipeline image datasets; The module design unit is used for the design of the multi-head attention feature enhancement module; The domain-adaptive adversarial training unit is used for domain-adaptive adversarial training; The model training and inference application unit is used for model training and inference applications.
9. The industrial pipeline leakage detection system integrating multi-head attention mechanism and domain adaptation as described in claim 8, characterized in that, The dataset construction unit is specifically used to perform the following: (1) Training dataset construction: First, a publicly available leaked image dataset is introduced as an open-source source domain image to provide the model with basic knowledge of leak patterns; Secondly, use 3D simulation software to generate defective images of the target domain with precise leak annotations; Finally, defect-free images of the target domain in a real industrial setting are acquired. The three parts of data are combined to form a hybrid dataset containing both defective and defect-free samples for adversarial training. (2) Evaluation test set construction: In order to objectively evaluate the generalization ability of the model, an evaluation test set independent of the training set is constructed. The evaluation test set is constructed by fusing the leak foreground generated by 3D software with the real leak-free pipeline background, and is reviewed and labeled by industry experts to ensure its visual realism and labeling accuracy.
10. The industrial pipeline leakage detection system integrating multi-head attention mechanism and domain adaptation as described in claim 8, characterized in that, In the module design unit execution, MDANet, as a feature enhancement method, is integrated between the feature extraction layers of the backbone network to form a feature extractor; the processing flow specifically includes the following: (1) Feature input and distribution: The feature map of the upstream network is used as input and copied first to provide a data basis for subsequent multi-branch parallel processing; (2) Branch 1: Global context modeling. This branch is responsible for capturing long-distance dependencies in the feature maps. Its core is to use a multi-head attention mechanism to perform weighted modeling of the input features using the following formula: Where Q, K, and V are the query, key, and value obtained from the input feature mapping, respectively, and d is the channel dimension. Through this mechanism, this branch can effectively understand the overall layout and global information of the screen. (3) Branch 2: Feature compression and receptive field expansion. This branch expands the perception range while reducing the amount of computation by reducing the spatial resolution of the feature map, and focuses on integrating more macroscopic regional information. (4) Branch 3: Local detail refinement. This branch focuses on processing the local fine texture and edge information of the feature map. To highlight key feature regions, this branch integrates a channel attention enhancement mechanism, which is calculated as follows: Among them, X c w represents the feature map of the c-th channel. c The corresponding attention weight is σ, which is the Sigmoid function. In this way, background interference can be adaptively suppressed and the area with leakage signs can be highlighted. (5) Feature integration and abstraction and output: The processing results of branch 1 and branch 2 are initially fused. Then, the fusion result and the output of branch 3 enter the global information integration stage to carry out the final cross-dimensional information extraction, and finally generate a highly abstract downstream feature map. This feature map integrates information from three dimensions: global, macro and local, and is passed to the subsequent layers of the network for use.