Aviation oil pipeline unmanned aerial vehicle intelligent inspection method and system

Abstract

The invention discloses an aviation oil pipeline unmanned aerial vehicle intelligent inspection method and system, and the method comprises the steps: collecting visual image data along a pipeline through an inspection terminal carried by an unmanned aerial vehicle, covering a pipeline body and a surrounding environment, and recognizing an abnormal scene endangering the safety of the pipeline; inputting the image data into an abnormal scene recognition model deployed at an unmanned aerial vehicle end, and judging whether a preset type of abnormality is included; if at least one type of abnormity is identified, generating an alarm signal; alarm and related images are uploaded to a remote monitoring center server through wireless communication, and intelligent unmanned inspection of the running state of the pipeline is achieved. An image acquisition and recognition model is integrated at an unmanned aerial vehicle end, a pipeline and an environment are sensed in real time, and key abnormity is automatically recognized; the edge deployment model reduces invalid return, and only triggers alarm uploading when a risk is detected; and in combination with wireless communication return alarms, high-reliability monitoring is realized, and intelligent and refined guarantee is provided for safe operation of aviation oil pipelines.

Description

Technical Field

[0001] This invention belongs to the field of pipeline safety monitoring and intelligent inspection technology, specifically relating to an intelligent inspection method and system for aviation fuel pipelines based on an unmanned aerial vehicle (UAV) platform. Background Technology

[0002] As a critical infrastructure connecting airports and fuel storage facilities, aviation fuel pipelines bear the important responsibility of ensuring the safe and stable delivery of aviation fuel. Their operational safety directly impacts the normal order of civil aviation transportation and major public safety. Because aviation fuel pipelines typically span long distances and traverse complex geographical environments, including urban suburbs, farmland, mountains, and rivers, they are susceptible to threats from natural factors (such as geological subsidence and soil erosion) and human activities (such as construction and illegal encroachment), posing significant safety hazards such as leaks, ruptures, and even fires and explosions. Therefore, establishing an efficient and reliable pipeline inspection mechanism to promptly identify and address potential risks is a core element in ensuring the safe operation of the aviation fuel supply system.

[0003] Traditional aviation fuel pipeline inspections primarily rely on manual patrols. Inspectors periodically walk or drive along the pipeline route, visually assessing for any anomalies. This method is not only labor-intensive, time-consuming, and has limited coverage, but it is also heavily influenced by weather, terrain, and human experience, making it difficult to guarantee continuity and consistency. Especially in complex terrain or inclement weather conditions, manual inspections are inefficient and risky, failing to meet the urgent demands of modern pipeline safety management for real-time and precise monitoring. While information technology has led to the introduction of video surveillance and sensor monitoring, their fixed deployment points still prevent dynamic and flexible monitoring of long-distance pipelines.

[0004] In recent years, drone technology has been widely used in inspection operations in the power, petroleum, and transportation industries due to its advantages such as high mobility, wide field of vision, and flexible operation. By equipping drones with high-definition cameras or infrared sensors, visual image data along pipelines can be quickly acquired, significantly improving inspection efficiency and coverage density. Invention patent CN113203049B discloses a pipeline safety intelligent monitoring and early warning system and method. Through the combination of a fiber optic early warning module, a video intelligent recognition module, a drone patrol module, and an artificial intelligence patrol module, it achieves automatic monitoring of the entire pipeline, solving the problems of high intensity and inaccessible areas of manual inspection, and improving the real-time monitoring and processing capabilities of pipeline safety. Combining image recognition and artificial intelligence algorithms, drone inspection is developing from a "collection + playback" model to an intelligent "flying and recognizing" approach, providing a new technological path for proactive early warning and rapid response in pipeline safety.

[0005] However, most existing UAV inspection systems adopt a "collect first, analyze later" working mode, meaning that after completing the flight mission, a large amount of raw image data is transmitted back to the ground center for centralized processing. This results in a delay in the detection of abnormal events and makes it difficult to meet the needs of real-time alarms. In addition, the continuous transmission of high-definition image data puts enormous pressure on wireless communication bandwidth, especially in remote areas with limited communication conditions, which can easily lead to transmission delays or interruptions, affecting the reliability of the monitoring system. At the same time, most systems lack dedicated identification capabilities for typical risk scenarios of aviation fuel pipelines (such as open flames, illegal construction, and foundation collapse), resulting in high false alarm and false negative rates. They have not yet achieved true edge intelligence and closed-loop response, which restricts the in-depth application of UAV inspection in aviation fuel pipeline safety protection. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an intelligent inspection method and system for aviation fuel pipelines using unmanned aerial vehicles (UAVs). By realizing edge intelligent recognition and proactive alarm at the UAV end, the real-time performance, accuracy, and system response efficiency of the inspection are improved.

[0007] The objective of this invention is achieved through the following technical solution: A method for intelligent inspection of aviation fuel pipelines using unmanned aerial vehicles (UAVs) includes the following steps: Visual image data along the aviation fuel pipeline is collected using an inspection terminal mounted on a drone. The visual image data covers the pipeline itself and its surrounding geographical environment and is used to identify objects in abnormal scenarios that may endanger the safe operation of the pipeline. The visual image data is input into an abnormal scene recognition model deployed on the drone to identify whether the visual image data contains an abnormal scene of a preset category; the abnormal scene includes open flame or smoke in the first abnormal scene, illegal excavation operation in the second abnormal scene, large construction machinery approaching in the third abnormal scene, and exposed pipe or foundation collapse in the fourth abnormal scene. If the identification results indicate the existence of at least one preset abnormal scenario, a corresponding alarm signal will be generated. The alarm signals and associated visual image data are uploaded to the server system of the remote monitoring center via a wireless communication link, enabling intelligent unmanned inspection of the aviation fuel pipeline's operating status.

[0008] By equipping drones with inspection terminals, the system continuously collects images of the aviation fuel pipeline during flight using its visual sensors, covering the pipeline itself and its surrounding environment. The collected image data is then input in real time into a lightweight anomaly scene recognition model pre-deployed on the drone. This model analyzes the images based on deep learning algorithms and automatically determines whether there are pre-set high-risk anomaly scenes such as open flames, smoke, illegal excavation, large machinery approaching, or foundation collapse. Once any anomaly is detected, the system immediately generates an alarm signal locally and transmits the alarm information along with the corresponding image fragment back to the remote monitoring center in real time via a wireless communication link. The monitoring center can then quickly locate the risk and take appropriate measures, thus forming a closed-loop workflow of front-end perception, edge intelligent recognition, real-time alarm, and remote response, achieving efficient, autonomous, and intelligent inspection of the aviation fuel pipeline's safety status.

[0009] As a preferred method, the abnormal scene recognition model is trained in the following way: Acquire a raw image sample set containing various abnormal scenarios, the raw image sample set covering real-world images of aviation fuel pipelines under different meteorological conditions, light intensity and topography; Data cleaning is performed on the original image sample set to remove blurry, overexposed, or heavily background-interferenced images, resulting in a cleaned image sample set. The various abnormal scenes in the cleaned image sample set are categorized and labeled, and the bounding boxes of each detected object in the image are labeled to generate an annotated labeled image sample set. The labeled image sample set is divided into a training sample set and a test sample set, wherein the training sample set is used for model training and the test sample set is used for model performance verification. The training sample set is input into the initial deep neural network model, and the backpropagation algorithm is used to iteratively optimize the model parameters until the model's recognition accuracy on the test sample set reaches the preset standard, thus obtaining the trained abnormal scene recognition model.

[0010] As a preferred approach, the loss function optimization during the model training process includes: Input the labeled images from the training sample set into the current model to obtain the predicted bounding boxes, abnormal scene category labels, and confidence scores; Calculate the comprehensive loss value between the prediction result and the true label. The comprehensive loss consists of three parts: bounding box regression loss, category classification loss and confidence loss. The bounding box regression loss is measured using the generalized intersection-union loss function, defined as: ;in, Indicates the predicted bounding box. This represents the actual bounding box; GIoU is the generalized intersection-union ratio, and the GIoU value is between 0 and 1. The larger the value, the higher the degree of overlap. The model parameters are updated using gradient descent based on the total loss, and this process is repeated iteratively until the model converges.

[0011] As a preferred embodiment, after generating the labeled image sample set, the following steps are also included: An independent test sample set is divided from the labeled image sample set, and the test sample set includes labeled images that did not participate in the training and their corresponding abnormal scene annotation information; After the model training is completed, the test sample set is input into the trained model to obtain the test recognition results; Compare the test identification results with the real annotations, and calculate the key performance indicators of the model, including mean precision (mAP), recall, and false positive rate; When the average precision is ≥90%, the recall rate is ≥85%, and the false positive rate is ≤5%, the abnormal scene recognition model is deemed to meet the on-site deployment conditions.

[0012] As a preferred approach, after confirming that the abnormal scene recognition model meets the deployment conditions, the following is also included: Test the inference latency of the model on an embedded computing platform for drones; If the inference latency exceeds the preset real-time threshold, the model is subjected to lightweight processing, including channel pruning, weight quantization and knowledge distillation techniques, to obtain a lightweight abnormal scene recognition model. Evaluate the accuracy loss of the lightweight model on the test sample set; If the accuracy loss does not exceed 2 percentage points, the lightweight anomaly scene recognition model will be deployed to the drone for online real-time analysis.

[0013] As a preferred method, the process of inputting visual image data into an anomaly scene recognition model and outputting anomaly scene recognition results includes: The input image is divided into a regular grid structure, and multiple candidate region boxes with different scales and aspect ratios are preset within each grid. For candidate region boxes containing the detected object, predict the abnormal scene category to which it belongs, and adjust the bounding box position to fit the actual target contour; The nonmaximum suppression algorithm is applied to retain the detection result with the highest confidence from the overlapping prediction boxes, and outputs the target's location coordinates, abnormal scene category label and confidence score.

[0014] As a preferred approach, the process of generating candidate region boxes and predicting targets includes: The input image is convolved by the backbone feature extraction network of the abnormal scene recognition model to generate a first feature map with a first resolution, and the first feature map contains high-level semantic information. The first feature map is upsampled using a feature pyramid structure and then fused with a second feature map output from a shallow network. The second feature map has a higher resolution than the first feature map, thus preserving more texture and edge details. A third feature map, which has the highest resolution, is further introduced from the original image or the output of the primary convolutional layer to enhance the detection capability of small-sized objects. The fused multi-level features are fed into the detection head module, which outputs the probability of the abnormal scene category, the bounding box offset, and the confidence of the target presence for each candidate region.

[0015] As a preferred method, it also includes: Based on the temperature distribution sequence along the pipeline in thermal infrared images, a temperature variation curve along the pipeline is constructed. Calculate the local thermal anomaly index The formula is as follows: ;in, This is the set of temperature values ​​measured on the surface of the current pipe section. This is the highest temperature value in the set; This represents the average temperature of the corresponding pipe section under historical normal operating conditions. The standard deviation of historical temperatures; when At that time, among them To preset the risk assessment threshold, and The system determined that there was a risk of leakage in the area, triggering a Level 1 warning signal.

[0016] As a preferred approach, when multiple drones collaborate to perform inspection tasks, the following is also included: Real-time acquisition of remaining battery power for each drone and the urgency weight of the tasks to be performed ; Establish a task priority scoring function Defined as: ;in, This is the drone's current remaining battery power. This is the rated capacity of the battery, therefore This is the normalized ratio of remaining battery power. The urgency weight is the value of the task, which takes a value in the range of [0,1] after normalization. is the energy preference coefficient, with a value range of [0,1], used to adjust the relative weight of energy status and task urgency; The central dispatch system is based on the calculated The system dynamically assigns high-priority tasks to the drone with the highest score, achieving coordinated optimization of energy consumption and task requirements.

[0017] A drone-based intelligent inspection system for aviation fuel pipelines includes: The drone inspection device is configured to carry a multi-source sensor terminal to perform automatic patrols along the aviation fuel pipeline and collect visual image data. The edge intelligence analysis module is installed on the drone itself and is used to run a lightweight abnormal scene recognition model to perform local real-time analysis on the collected visual image data to identify whether there are preset abnormal scenes. The alarm triggering module is used to generate corresponding alarm signals when abnormal scenarios are detected; The communication transmission module is used to upload the alarm signal and associated image data to the remote monitoring center; The preset abnormal scenarios include open flames or smoke, illegal excavation operations, the approach of large construction machinery, exposed pipes, or foundation collapse.

[0018] The present invention has at least the following beneficial effects: By integrating visual image acquisition and abnormal scene recognition models at the UAV terminal, the present invention achieves real-time intelligent perception of the pipeline body and its surrounding environment. It can automatically identify key abnormal scenes such as open flames, smoke, illegal excavation, large machinery approaching, exposed pipelines, or foundation collapse during flight inspections, significantly improving the timeliness and accuracy of inspections. By deploying recognition models at the edge, invalid data backhaul is reduced, and alarms are triggered and associated images are uploaded only when preset risks are detected, effectively reducing communication bandwidth pressure and improving response efficiency. Combined with wireless communication links, alarm information is transmitted back to the monitoring center, achieving highly reliable pipeline safety monitoring, reducing the cost and safety risks of manual inspections, and providing intelligent and refined protection measures for the safe operation of aviation fuel pipelines. Attached Figure Description

[0019] To reveal the technical details of the embodiments of the present invention, the accompanying drawings involved in the embodiments will be briefly described below. It should be emphasized that these drawings only present several embodiments of the present invention and should not be considered as defining the scope of the invention. For those skilled in the art, other related drawings can still be derived based on these drawings without inventive effort.

[0020] Figure 1 This is a flowchart of an intelligent unmanned aerial vehicle (UAV) inspection method for aviation fuel pipelines according to the present invention; Figure 2 This is a diagram illustrating the training process of an abnormal scene recognition model. Figure 3This is a schematic diagram illustrating the working principle of an unmanned aerial vehicle (UAV) intelligent inspection system for aviation fuel pipelines. Detailed Implementation

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0022] In the following description, embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. However, it should be understood that the present disclosure is not limited to the specific forms shown herein. Rather, it should be understood to encompass various variations, equivalents, and / or alternatives to the embodiments of the present disclosure. In illustrating the drawings, the same reference numerals will be used to denote similar components.

[0023] In the various embodiments of this disclosure, the terms "first," "second," "the first," or "the second" are intended to modify different components and not to indicate order and / or importance, nor do they constitute a limitation on the respective components. For example, a first user equipment and a second user equipment represent different user equipments, although they both fall under the category of user equipment. Similarly, a first component may be named a second component, and a second component may be named a first component, without changing their essential attributes within the scope of this disclosure.

[0024] In this disclosure, terminology is used to describe specific embodiments and does not constitute a limitation thereof. In this context, the use of the singular form also encompasses the plural form, unless otherwise expressly stated herein. In the course of description, terms such as “comprising” or “having” are intended to indicate the presence of features, quantities, steps, operations, structural components, parts, or combinations thereof, and do not preclude the possibility or addition of one or more other features, quantities, steps, operations, structural components, parts, or combinations thereof.

[0025] It should be clarified that while the following description provides detailed specific information to aid in a comprehensive understanding of the exemplary embodiments, those skilled in the art will recognize that the exemplary embodiments can be implemented even without these specific details. For example, the system may be illustrated using block diagrams to avoid excessive detail that could obscure the clarity of the example. In other cases, to maintain the clarity of the example, unnecessary details of well-known processes, structures, and techniques may be omitted.

[0026] like Figure 1 As shown, a method for intelligent inspection of aviation fuel pipelines using unmanned aerial vehicles (UAVs) includes the following steps: Visual image data along the aviation fuel pipeline is collected using an inspection terminal mounted on a drone. The visual image data covers the pipeline itself and its surrounding geographical environment and is used to identify objects in abnormal scenarios that may endanger the safe operation of the pipeline. The visual image data is input into an abnormal scene recognition model deployed on the drone to identify whether the visual image data contains an abnormal scene of a preset category; the abnormal scene includes open flame or smoke in the first abnormal scene, illegal excavation operation in the second abnormal scene, large construction machinery approaching in the third abnormal scene, and exposed pipe or foundation collapse in the fourth abnormal scene. If the identification results indicate the existence of at least one preset abnormal scenario, a corresponding alarm signal will be generated. The alarm signals and associated visual image data are uploaded to the server system of the remote monitoring center via a wireless communication link, enabling intelligent unmanned inspection of the aviation fuel pipeline's operating status.

[0027] This embodiment places AI recognition capabilities at the edge of the drone, constructing an integrated intelligent inspection mechanism encompassing perception, decision-making, and transmission. During autonomous flight along a pre-set route of the aviation fuel pipeline, the drone captures real-time visual images of the pipeline using its onboard camera. The system then runs an anomaly scene recognition model locally, rapidly inferring and judging four typical risks that may threaten pipeline safety: open flames and smoke, illegal excavation, approaching heavy machinery, exposed pipeline, or foundation collapse. Alarms are triggered only when an anomaly is confirmed, and key images and location information are selectively uploaded, avoiding the continuous transmission of massive amounts of useless data. This on-demand alarm and precise data transmission mode enables near real-time detection and remote synchronous monitoring of pipeline safety hazards, improving the response speed and operational efficiency of the entire inspection system.

[0028] In a preferred embodiment, see Figure 2 The abnormal scene recognition model is trained in the following way: Acquire a raw image sample set containing various abnormal scenarios, the raw image sample set covering real-world images of aviation fuel pipelines under different meteorological conditions, light intensity and topography; Data cleaning is performed on the original image sample set to remove blurry, overexposed, or heavily background-interferenced images, resulting in a cleaned image sample set. The various abnormal scenes in the cleaned image sample set are categorized and labeled, and the bounding boxes of each detected object in the image are labeled to generate an annotated labeled image sample set. The labeled image sample set is divided into a training sample set and a test sample set, wherein the training sample set is used for model training and the test sample set is used for model performance verification. The training sample set is input into the initial deep neural network model, and the backpropagation algorithm is used to iteratively optimize the model parameters until the model's recognition accuracy on the test sample set reaches the preset standard, thus obtaining the trained abnormal scene recognition model.

[0029] The training process for the anomaly scene recognition model is one of learning and refining judgment capabilities from real-world scene data. First, the system collects a large number of real-world images of the aviation fuel pipeline route taken under various weather, lighting, and terrain conditions. These images encompass a variety of potential risk scenarios, including open flames, smoke, excavation operations, heavy machinery, and exposed pipelines, ensuring the model can adapt to complex and changing on-site environments. Next, these raw images are cleaned, removing blurry, overexposed, or heavily interfered-with low-quality samples to ensure the reliability of subsequent training data. Then, the retained images are meticulously labeled. Professionals assign category labels to each type of anomaly scene and precisely mark the target's location in the image with bounding boxes, enabling the model not only to determine "whether" an anomaly exists but also "where" it is. Subsequently, these labeled images are divided into training and test sets. The former is used to "teach" the model to recognize various anomalies, while the latter is used to "test" the model's recognition capabilities. During training, the system inputs the training set data into an initial deep neural network. The model continuously compares the differences between its own recognition results and the standard answer, and automatically adjusts its internal parameters using the backpropagation mechanism to gradually improve the recognition accuracy until it reaches the preset accuracy requirement on the test set. Finally, a lightweight and high-precision anomaly recognition model that can run efficiently on the drone is formed, providing intelligent support for real-time risk assessment on site.

[0030] In a preferred embodiment, the loss function optimization during model training includes: Input the labeled images from the training sample set into the current model to obtain the predicted bounding boxes, abnormal scene category labels, and confidence scores; Calculate the comprehensive loss value between the prediction result and the true label. The comprehensive loss consists of three parts: bounding box regression loss, category classification loss and confidence loss. The bounding box regression loss is measured using the generalized intersection-union ratio (GIoULoss) loss function, defined as: ;in, Indicates the predicted bounding box. This represents the actual bounding box, expressed in pixel coordinates; the GIoU value ranges from 0 to 1, with a higher value indicating a higher degree of overlap. The model parameters are updated using gradient descent based on the total loss, and this process is repeated iteratively until the model converges.

[0031] During model training, to enable the system to more accurately identify abnormal scenes and precisely locate target positions, a loss function is needed to measure the gap between the model's current predictions and the actual situation, thus guiding continuous model improvement. Specifically, when a labeled image is input into the training model, the model outputs its predictions for anomaly type, bounding box, and confidence level. The system then compares these predictions with manually labeled standard answers, calculating a comprehensive "error score." This score consists of three parts: first, whether the bounding box accurately matches the target object's position; second, whether the anomaly category is correctly classified; and third, whether the confidence level given by the model is reasonable. To more scientifically evaluate the matching degree of the bounding boxes, the Generalized Intersection over Union (GIoU) ​​evaluation method is used. It not only considers the overlap area between the predicted and ground truth bounding boxes but also their relative positional relationship, avoiding misjudgments caused by excessive distance or directional deviation. The closer the GIoU value is to 1, the more accurate the prediction, and the smaller the corresponding loss. Ultimately, the system weights and combines these three losses into a single total loss value, and uses the gradient descent algorithm to adjust the model's internal parameters in reverse, making the next prediction closer to the true result. This process is iterated repeatedly on the training set until the model's recognition accuracy stabilizes and meets the requirements, thus ensuring that it can quickly and accurately detect safety hazards around aviation fuel pipelines during actual inspections.

[0032] In a preferred embodiment, after generating the labeled image sample set, the method further includes: An independent test sample set is divided from the labeled image sample set, and the test sample set includes labeled images that did not participate in the training and their corresponding abnormal scene annotation information; After the model training is completed, the test sample set is input into the trained model to obtain the test recognition results; Compare the test results with the actual annotations, and calculate the key performance indicators of the model, including mean precision, recall, and false positive rate. When the average precision is ≥90% (or 0.9), the recall rate is ≥85%, and the false positive rate is ≤5%, the abnormal scene recognition model is deemed to meet the on-site deployment conditions.

[0033] After image annotation and the generation of annotated sample sets, rigorous performance verification is required to objectively evaluate whether the trained anomaly scene recognition model truly possesses practical capabilities. To this end, the system reserves a portion of the annotated data as a test sample set. These images are never seen by the model during the entire training process, ensuring the fairness and reliability of the test results. Once the model is trained, it is applied to these independent test images, allowing the model to automatically identify anomalies and output detection results. Subsequently, the system compares the model's output with manually annotated standard answers, statistically analyzing the model's recognition performance across various anomalies, and calculating several key performance indicators: mean precision (mAP) measures the overall accuracy of the model; recall reflects the model's ability to detect as many real risks as possible; and false positive rate assesses whether the model frequently "falsely" identifies normal situations as anomalies. Only when the model simultaneously achieves high standards across these indicators—namely, an accuracy rate exceeding 90%, a risk detection rate exceeding 85%, and no more than 5 false alarms per 100 judgments—is the model considered capable of practical deployment on drones. This verification mechanism ensures that the model not only performs well on training data, but also runs stably and reliably in real inspection environments, avoiding monitoring failures or frequent false alarms that interfere with normal work due to inaccurate identification.

[0034] In a preferred embodiment, after confirming that the abnormal scene recognition model meets the deployment conditions, the method further includes: Test the inference latency of the model on an embedded computing platform for drones; If the inference latency exceeds the preset real-time threshold, the model is subjected to lightweight processing, including channel pruning, weight quantization and knowledge distillation techniques, to obtain a lightweight abnormal scene recognition model. Evaluate the accuracy loss of the lightweight model on the test sample set; If the accuracy loss does not exceed 2 percentage points, the lightweight anomaly scene recognition model will be deployed to the drone for online real-time analysis.

[0035] After the anomaly scene recognition model passes performance verification, it's crucial to ensure its real-time operation within the limited hardware resources of the drone. Therefore, inference speed testing on the actual embedded computing platform is necessary, measuring the time required for the model to process each image frame. If the test reveals that the model's processing speed is too slow, causing latency exceeding the system's set real-time requirements, it will fail to meet the "simultaneous data acquisition and recognition" requirement during flight, easily leading to missed detections or delayed alarms. To address this, the original model needs lightweight optimization. This involves removing redundant computational units in the network through channel pruning, reducing the storage precision of model parameters using weight quantization, and combining this with knowledge distillation techniques to allow the smaller model to learn the discrimination logic of the larger model, thereby significantly compressing the model size and improving operational efficiency. After lightweighting, the model's recognition accuracy is re-evaluated on the test set to ensure that performance degradation is controlled within an acceptable range, i.e., accuracy loss does not exceed 2 percentage points. Only when both accuracy and speed standards are met is the lightweight model deployed to the drone, truly achieving efficient, low-latency online real-time analysis capabilities and ensuring the smoothness and reliability of inspection tasks.

[0036] In a preferred embodiment, the process of inputting visual image data into an abnormal scene recognition model and outputting abnormal scene recognition results includes: The input image is divided into a regular grid structure, and multiple candidate region boxes with different scales and aspect ratios are preset within each grid. For candidate region boxes containing the detected object, predict the abnormal scene category to which it belongs, and adjust the bounding box position to fit the actual target contour; The nonmaximum suppression algorithm is applied to retain the detection result with the highest confidence from the overlapping prediction boxes, and outputs the target's location coordinates, abnormal scene category label and confidence score.

[0037] After the drone acquires visual images along the pipeline, the system divides the image into small, regular grids. Each grid contains multiple candidate detection boxes of varying sizes and shapes to accommodate the size differences of various abnormal targets, such as open flames, machinery, and collapses, based on distance and angle. The model first determines which candidate boxes might cover the abnormal target and performs detailed analysis on these areas: predicting the risk scenario of the area, such as smoke or an excavator; and automatically adjusting the position and size of the candidate boxes to more closely match the actual contour of the target, improving positioning accuracy. Since the same target may be detected repeatedly by multiple adjacent candidate boxes, the system further employs non-maximum suppression technology. By comparing the confidence scores of each overlapping box, it automatically retains the most reliable result and eliminates other redundant or duplicate predictions. The final output is a concise and accurate detection result, including the specific location coordinates of the abnormal target, the corresponding category label, and the credibility of the system's judgment. This provides a clear and reliable basis for subsequent alarm generation and information uploading, ensuring that the entire recognition process is comprehensive yet non-redundant.

[0038] In a preferred embodiment, the process of generating candidate region boxes and predicting targets includes: The input image is convolved by the backbone feature extraction network of the abnormal scene recognition model to generate a first feature map with a first resolution, and the first feature map contains high-level semantic information. The first feature map is upsampled using a feature pyramid structure and then fused with a second feature map output from a shallow network. The second feature map has a higher resolution than the first feature map, thus preserving more texture and edge details. A third feature map, which has the highest resolution, is further introduced from the original image or the output of the primary convolutional layer to enhance the detection capability of small-sized objects. The fused multi-level features are fed into the detection head module, which outputs the probability of the abnormal scene category, the bounding box offset, and the confidence of the target presence for each candidate region.

[0039] In image recognition, the model first performs layer-by-layer convolution operations on the input image along the aviation fuel pipeline through the backbone network, extracting a low-resolution feature map containing rich high-level semantic information. This map helps the model understand the overall category and structure of objects in the image, such as determining whether a certain area is machinery or flames. However, relying solely on high-level features can easily lose details, especially for distant or small-sized anomalous targets (such as distant excavators or initial smoke). To address this, the system introduces a feature pyramid structure, upsampling the high-level feature map and fusing it with higher-resolution feature maps from shallower layers of the network. The shallower feature maps retain more texture, edge, and detail information from the original image, which helps to accurately locate target contours. To further improve sensitivity to minor anomalies, the system also introduces higher-resolution primary feature maps to enhance the model's ability to perceive small-sized targets. Ultimately, these features from different levels, possessing both semantic understanding and detail restoration capabilities, are integrated and fed into the detection module. The module analyzes each preset candidate region box to determine whether there are any anomalies, what type of risk it belongs to, and outputs the corresponding category probability, bounding box adjustment parameters, and the reliability of the detection result, thereby achieving comprehensive and accurate identification of various safety hazards around the pipeline.

[0040] In a preferred embodiment, the inspection terminal integrates a visible light camera, an infrared thermal imager, and a lidar sensor; The process of acquiring visual image data simultaneously obtains visible light images, thermal infrared images, and three-dimensional point cloud data. The abnormal scene recognition model is a multimodal fusion model, whose inputs include visible light images, thermal infrared images, and point cloud projection images, and whose output is the fusion recognition result; Among them, thermal infrared images are used to identify areas with abnormal temperatures on the pipe surface, while point cloud data is used to detect surface deformation or terrain disturbances caused by human construction.

[0041] The inspection terminal is equipped with a visible light camera, an infrared thermal imager, and a lidar sensor, enabling it to simultaneously collect multiple types of images and data. During missions, the terminal can not only capture conventional visible light photographs but also acquire thermal infrared images and 3D point cloud data of the target area. This data is crucial for a comprehensive understanding of the pipeline's condition. Thermal infrared images are primarily used to identify temperature anomalies, such as localized overheating caused by leaks or blockages; while 3D point cloud data helps detect surface deformation or any terrain disturbance caused by human construction, providing critical information about ground subsidence or excavation activities.

[0042] The multimodal fusion model integrates data from these different sources for analysis. It considers not only the intuitive visual information provided by visible light images but also the sensitivity of thermal infrared images to temperature changes and the spatial structural features reflected in point cloud data. Through this comprehensive processing approach, the model can more accurately identify potential anomalies. For example, it can detect leak risks by comparing expected temperatures under normal operating conditions with actual measurements, or monitor land movement that may affect pipeline safety using changes in terrain data. The final fusion identification result integrates information from three data sources, providing a more comprehensive and accurate basis for assessing pipeline health. This method significantly improves the reliability and accuracy of anomaly detection, making maintenance work more efficient.

[0043] In a preferred embodiment, it further includes: Based on the temperature distribution sequence along the pipeline in thermal infrared images, a temperature variation curve along the pipeline is constructed. Calculate the local thermal anomaly index The formula is as follows: ;in, This is the set of temperature values ​​measured on the surface of the current pipe section. This is the highest temperature value in the set; This represents the average temperature of the corresponding pipe section under historical normal operating conditions. The standard deviation of historical temperatures; when At that time, among them To preset the risk assessment threshold, and The system determined that there was a risk of leakage in the area, triggering a Level 1 warning signal.

[0044] During inspections, the system not only acquires visible light images but also simultaneously obtains thermal infrared images along the pipeline, utilizing their temperature-sensitive characteristics to analyze the heat distribution on the pipeline surface. By extracting temperature data along the pipeline's direction from the infrared images, a continuous temperature variation curve is constructed, visually reflecting the thermal state of the entire pipeline section. To identify abnormal heat areas, the system focuses on analyzing the highest temperature on a specific section of the pipeline and compares it with the average temperature and temperature fluctuation range at that location under historical normal operating conditions. This comparison is not simply based on the size of the temperature difference but incorporates statistical patterns from long-term data to determine whether the current high temperature significantly deviates from the normal range. When a temperature significantly higher than the daily level is detected, exceeding the normal fluctuation range by several standard deviations, a localized thermal anomaly is considered to have occurred. At this point, the thermal anomaly index calculated by the system will exceed a preset safety threshold (usually set between 2 and 4), thus determining that there may be fuel leaks, friction overheating, or other potential risks causing temperature increases in that area. A Level 1 warning signal is immediately triggered, alerting the monitoring center to pay close attention and take timely action, achieving early detection and proactive prevention of potential safety incidents.

[0045] In a preferred embodiment, it further includes: Pipe surface temperature field acquired by infrared thermal imager Surface micro-deformation field extracted by lidar A thermal-mechanical coupling leakage risk assessment model was constructed; a comprehensive leakage risk index per unit area was defined. The calculation formula is as follows: ;in, This indicates the overall leakage risk index; This represents the two-dimensional temperature distribution of the pipe region in an infrared image, with units of °C. The characteristic length reference value is set at 1m; The temperature Laplace operator represents the second-order spatial rate of change of the temperature gradient, reflecting the degree of local heat concentration, and is expressed in °C / m². 2 ; This represents the reference temperature baseline value, which is the average temperature of the pipeline under historical normal operating conditions, and the unit is °C. This represents the vertical displacement field of the Earth's surface reconstructed from laser point clouds, in mm. The modulus represents the surface displacement gradient, which represents the rate of change of surface height per unit distance, and is expressed in mm / m (i.e., ‰ slope). This represents the reference value for surface deformation sensitivity, set to 1 mm / m, used for normalization processing, with the unit being mm / m; , Denotes the weighting coefficients, satisfying ,and The adjustments are made dynamically based on the pipe section material, burial depth, and geological conditions.

[0046] when At that time, among them To set a preset risk threshold, if a potential leakage risk is determined in the area, a level-two early warning signal is triggered and a high-frequency re-inspection task is initiated.

[0047] The aforementioned model, through joint analysis of "thermal anomaly gradient" and "surface strain gradient," captures the dual precursors of heat diffusion and soil loosening caused by leaks. Compared to single-index detection, it can effectively distinguish between environmental disturbances (such as sunlight causing temperature rise without deformation) and actual leak events, significantly reducing the false alarm rate. For example, direct sunlight may cause... Increased, but not causing significant surface deformation. ,therefore It is still within a safe range; however, an actual leak would simultaneously cause local high temperatures and soil softening, leading to a simultaneous increase in both indicators, thus making it accurately identifiable.

[0048] In a preferred embodiment, after generating the alarm signal, a tiered response mechanism is further included: For a Level 1 alarm (corresponding to open flame or explosion risk), immediately control the drone to return to base and automatically push alarm information to the emergency command system; For Level 2 alarms (corresponding to illegal excavation or the approach of large machinery), control the drone to hover and record a high-definition video stream, which is then uploaded to the monitoring center for manual review. For Level 3 alarms (corresponding to vegetation cover or slight subsidence), record the alarm location and include it in the periodic maintenance plan.

[0049] After the system identifies an anomaly and generates an alarm signal, it automatically activates a tiered response mechanism based on different risk levels to ensure that the response measures are both timely and appropriate. For Level 1 alarms, such as the detection of open flames or serious threats like explosion risks, the system determines that the on-site environment may endanger the drone and surrounding safety. It immediately interrupts the inspection mission, controls the drone to autonomously return to a safe area, and automatically pushes the alarm information to the emergency command system to quickly activate the emergency plan and gain valuable time for response. For Level 2 alarms, such as the discovery of illegal excavation or large construction machinery approaching a pipeline, although not immediately damaging but posing a high risk, the drone will hover at a safe altitude, continuously record high-definition video, and transmit it back to the monitoring center in real time. This provides clear on-site evidence for manual review and decision-making, avoiding misjudgments or omissions. For Level 3 alarms, such as minor issues like excessive vegetation cover or slight ground subsidence, the system records them as potential hazards, marks their specific locations, and incorporates them into subsequent regular maintenance plans, achieving reasonable resource allocation and refined management. This tiered response mechanism enables the system to adopt matching response strategies based on different levels of risk, which not only improves the intelligence level of emergency response but also enhances the safety and management efficiency of the inspection system.

[0050] In a preferred embodiment, when multiple drones cooperate to perform inspection tasks, the method further includes: Real-time acquisition of remaining battery power for each drone and the urgency weight of the tasks to be performed ; Establish a task priority scoring function Defined as: ;in, This is the drone's current remaining battery power (unit: mAh). This refers to the battery's rated capacity (unit: mAh). This is the normalized ratio of remaining battery power. The urgency weight is the value of the task, which takes a value in the range of [0,1] after normalization. is the energy preference coefficient, with a value range of [0,1], used to adjust the relative weight of energy status and task urgency; The central dispatch system is based on the calculated The system dynamically assigns high-priority tasks to the drone with the highest score, achieving coordinated optimization of energy consumption and task requirements.

[0051] When multiple drones collaborate on fuel pipeline inspection missions, a dynamic scheduling mechanism based on battery power and mission urgency is introduced to ensure efficient and safe mission completion. The current remaining battery power of each drone, as well as the importance and urgency of its upcoming tasks, are collected and evaluated in real time. The system uses a comprehensive scoring function to measure the suitability of each drone for a new task. This score considers both the drone's own energy status (more sufficient power means stronger execution capability) and the urgency of the task itself (more critical tasks have higher priority). By adjusting the weights of these two factors, the system can flexibly balance power priority and task priority. For example, it prioritizes ensuring continued operation when battery power is low, or dispatches the nearest drone to respond in case of sudden risks. The central scheduling system makes unified decisions based on the comprehensive scores of each drone, automatically assigning high-priority tasks to the most suitable drone, avoiding situations where low-battery drones are assigned heavy tasks or high-urgency tasks are delayed. This mechanism effectively achieves rational resource allocation in multi-drone collaboration, improving overall inspection efficiency and system stability.

[0052] In a preferred embodiment, it further includes: Risk heat map generated based on Geographic Information System (GIS) Current drone location Construct a risk coverage benefit function within a unit flight distance. A total energy consumption model is established based on battery discharge characteristics. Define the path optimization objective function Its expression is: ; It is a comprehensive benefit index for routes, used to evaluate the merits of routes; Plan the flight path curve for the drone; The path element length; This is a GIS risk heat map. The higher the value, the higher the frequency of anomalies or the more serious the consequences in the area in history. It is normalized to [0, 1]. The benchmark value for unit risk coverage is set to 1; The empirical formula for the power consumption of a drone is: ,in Flight speed (m / s) An empirical coefficient related to the drone model. These are empirical coefficients calibrated through experiments or simulations based on the specific model, structure, motor efficiency, aerodynamic characteristics, and other physical properties of the drone. The rated power reference value is taken as the typical cruise power (e.g., 150W). Let be the weighting coefficient, satisfying and It can be adjusted according to the task mode (e.g., emphasizing routine inspections). Emergency response is biased. ).

[0053] The central scheduling system uses an improved A* algorithm or reinforcement learning strategy to solve the problem. Maximize the optimal path The command is then sent to the corresponding drone for execution.

[0054] The aforementioned model implements the intelligent navigation logic of "prioritizing areas with higher risk." When the battery is fully charged, the system tends to select routes with a high density of high-risk areas to improve safety and security efficiency; while when the battery is low, it automatically reduces... improve The system prioritizes the return route with the lowest energy consumption to ensure system safety. Furthermore, the integral form of the objective function is applicable to arbitrarily complex paths and possesses good engineering feasibility.

[0055] In a preferred embodiment, the UAV takes off autonomously and performs inspection tasks according to a preset flight route generated based on a geographic information system (GIS); During flight, it maintains two-way data interaction with the central server via 4G / 5G or satellite communication links; When the communication link is interrupted, the UAV switches to autonomous navigation mode, uses the onboard SLAM algorithm to build a local environment map and perform dynamic obstacle avoidance, and continues to fly along the original route after the communication is restored.

[0056] When performing inspection missions, the drone first automatically generates a flight path that follows the route of the aviation fuel pipeline based on pre-imported Geographic Information System (GIS) data. This includes information such as take-off and landing points, waypoint coordinates, and flight altitude, enabling one-click autonomous take-off and map-based flight, ensuring comprehensive inspection coverage and optimal path. During normal flight, the drone maintains real-time two-way communication with the ground monitoring center via 4G / 5G or satellite communication links, uploading identified anomaly data and receiving remote commands, achieving full control. However, in complex environments such as mountainous areas, tunnels, or areas with severe electromagnetic interference, communication signals may be temporarily interrupted. In this case, the drone will not blindly stop flying or become uncontrollable, but will automatically switch to autonomous navigation mode. Relying on onboard sensors and SLAM (Simultaneous Localization and Mapping) algorithms, it will perceive the surrounding environment in real time, build a local 3D map, and dynamically avoid obstacles such as birds, temporary buildings, or high-voltage power line towers. Throughout the period of communication loss, the drone continuously records its flight trajectory and environmental information. Once communication is restored, it automatically calibrates its position and reconnects to the original flight path to continue completing the remaining inspection tasks. This mechanism significantly improves the adaptability and mission continuity of drones in complex geographical environments, ensuring the integrity and safety of inspection operations.

[0057] A drone-based intelligent inspection system for aviation fuel pipelines, participating in... Figure 3 ,include: The drone inspection device is configured to carry a multi-source sensor terminal to perform automatic patrols along the aviation fuel pipeline and collect visual image data. The edge intelligence analysis module is installed on the drone itself and is used to run a lightweight abnormal scene recognition model to perform local real-time analysis on the collected visual image data to identify whether there are preset abnormal scenes. The alarm triggering module is used to generate corresponding alarm signals when abnormal scenarios are detected; The communication transmission module is used to upload the alarm signal and associated image data to the remote monitoring center; The preset abnormal scenarios include open flames or smoke, illegal excavation operations, the approach of large construction machinery, exposed pipes, or foundation collapse.

[0058] The aviation fuel pipeline drone intelligent inspection system integrates a flight platform with intelligent sensing technology to achieve automated and intelligent monitoring of the safety conditions along the pipeline. The system uses drones as the core carrier, equipped with multi-source sensor terminals, to autonomously fly along a preset route, continuously collecting high-definition visual image data of the pipeline and its surrounding environment. Crucially, the system integrates an edge intelligent analysis module on the drone, with a built-in trained lightweight deep learning model that can process image data in real time during flight. It can locally identify typical risk scenarios such as open flames, smoke, illegal excavation, the approach of large machinery, exposed pipelines, or foundation collapse without relying on a remote server. Once an anomaly is detected, the alarm triggering module immediately generates an alarm signal of the corresponding level and quickly transmits the alarm information and key images back to the remote monitoring center via 4G / 5G or satellite links through the communication transmission module. This "front-end identification, local decision-making, and on-demand transmission" working mode not only significantly reduces data transmission volume and improves response speed but also enhances the system's reliability in weak network or remote environments, truly achieving efficient and accurate unmanned intelligent inspection.

[0059] In a preferred embodiment, the remote monitoring center includes: The central processing server is used to receive alarm signals and inspection data from multiple drones; The data storage module is used to store historical inspection images, alarm records, and model version information; The human-computer interaction module is used to visualize alarm locations, drone flight trajectories, and on-site images on an electronic map; The scheduling and control module is used to send mission instructions to the drone, including route adjustment, emergency return to home, or fixed-point reshoot.

[0060] The remote monitoring center, acting as the "command brain" of the entire inspection system, is responsible for the centralized management and coordinated scheduling of the operational status and inspection data of multiple drones. The center's processing server receives alarm signals, image data, and flight information uploaded from each drone in real time, ensuring that all abnormal events are aggregated at the center immediately. This data is then stored in the data storage module, including historical inspection images, alarm records, and model version information, forming a complete data archive for subsequent traceability analysis and model optimization. Through the human-machine interface module, operators can intuitively view the real-time flight trajectory of each drone and the specific location of alarms on an electronic map, and retrieve on-site images or videos for verification, achieving "clear visibility, accurate location, and clear judgment" through visualized monitoring. Simultaneously, the scheduling and control module supports manual intervention, allowing remote commands to be sent to drones based on the situation, such as adjusting flight paths, executing emergency return-to-home orders, or conducting targeted re-photographing of suspicious areas, thus flexibly responding to emergencies. The entire monitoring center has built an efficient, visible, and manageable intelligent operation and maintenance platform by organically combining data aggregation, information display, and remote control, which has improved the overall response capability and decision-making level of aviation fuel pipeline safety management.

[0061] In a preferred embodiment, the edge intelligent analysis module includes: The feature extraction submodule is used to extract multi-scale feature maps from the input image; The feature fusion submodule is used to fuse deep semantic features with shallow detail features to improve the detection capability of small targets; The detection output submodule is used to generate detection results based on fused features, including target bounding boxes, abnormal scene category labels, and confidence scores. The model compression submodule is used to prune and quantize the original abnormal scene recognition model to adapt to the limited computing resources of the drone.

[0062] The edge intelligence analysis module is the core intelligent unit deployed on the drone, responsible for real-time processing and anomaly identification of acquired images during flight. This module first performs multi-layer convolution operations on the input image through a feature extraction submodule, extracting feature maps of different scales. These feature maps reflect rich information from the overall structure to local details of the image. To improve the ability to identify distant or small-sized anomalous targets (such as distant construction machinery or initial smoke), the feature fusion submodule effectively fuses semantic information from the deep network with edge and texture details preserved by the shallow network, enabling the model to not only understand what something is but also where it is located. Subsequently, the detection output submodule analyzes each candidate region based on the fused high-quality features, ultimately outputting a precise target location bounding box, the corresponding anomaly type (such as open flame, excavation, etc.), and the credibility of the system's judgment. Considering the limited computing resources and power consumption of drones, the module also integrates a model compression submodule. Through techniques such as pruning to remove redundant calculations and quantizing to reduce parameter precision, the original massive recognition model is compressed into a lightweight version. While ensuring recognition accuracy, the operating speed is significantly improved, enabling it to run efficiently and stably on airborne hardware and truly achieve real-time intelligent analysis capabilities of "recognizing while flying".

[0063] In a preferred embodiment, a multimodal sensing subsystem is further included, the subsystem comprising: The visible light imaging unit is used to acquire high-resolution visible light images along the aviation fuel pipeline. Infrared temperature measurement unit is used to acquire thermal radiation images of the pipe surface and identify areas with abnormal temperatures; The laser scanning unit is used to collect three-dimensional point cloud data of the terrain and identify surface subsidence or excavation behavior; The data synchronization unit is used to perform timestamp alignment and spatial registration of multi-source sensor data to ensure the consistency of multimodal information.

[0064] The multimodal perception subsystem integrates multiple sensors to achieve comprehensive, three-dimensional perception of aviation fuel pipelines and their surrounding environment. The system includes a visible light imaging unit capable of capturing clear, high-definition images to identify visually apparent anomalies such as pipeline surface damage, illegal construction, and vegetation cover; an infrared thermography unit that captures thermal radiation information to generate temperature distribution maps of the pipeline surface, particularly suitable for detecting localized overheating caused by fuel leaks or friction, compensating for the limitations of visible light in identifying hidden hazards at night; and a laser scanning unit that uses lidar technology to acquire three-dimensional point cloud data of the ground and features, accurately reflecting terrain undulations and effectively detecting surface disturbances such as foundation collapse, excavation, or earthwork accumulation. To ensure that information from these different sensors can work collaboratively, the system is also equipped with a data synchronization unit. By precisely timestamping each frame of image and point cloud data and aligning spatial coordinates, it unifies data from different perspectives and formats under the same spatiotemporal reference, avoiding misjudgments caused by time delays or positional deviations. This multi-source data fusion approach enables the system to not only "see clearly" but also "sense accurately" and "measure precisely," improving the comprehensiveness and reliability of pipeline safety hazard identification in complex environments.

[0065] In a preferred embodiment, multiple drones are deployed in a cluster, with each drone responsible for the inspection of different pipe sections. Each drone shares local environmental perception information through an edge collaboration mechanism to achieve dynamic obstacle avoidance and path optimization; The central server dynamically adjusts the inspection frequency and flight altitude of each drone based on meteorological data, airspace status, and pipeline risk level. The central server is also equipped with a task scheduling optimization engine for executing task priority scoring functions. This enables task allocation based on a synergy between energy state and task urgency.

[0066] The swarm deployment of multiple drones enables each drone to be dedicated to inspecting a specific section of the aviation fuel pipeline, ensuring comprehensive coverage and monitoring of the entire pipeline network. Through edge collaboration mechanisms, each drone can not only execute tasks independently but also share its perceived local environmental information in real time. This enables dynamic obstacle avoidance and optimizes flight paths to improve inspection efficiency. For example, when a drone detects an obstacle ahead, it transmits this information to other nearby drones, which then adjust their flight paths to avoid collisions.

[0067] The central server acts as the brain of this system. It not only collects and analyzes critical information such as meteorological data, airspace status, and pipeline risk levels, but also dynamically adjusts the operating modes of each drone based on this information, including inspection frequency and flight altitude. For example, in severe weather conditions or high-risk areas, the central server increases the drone's inspection frequency and adjusts its flight altitude to ensure safety and monitoring effectiveness. Furthermore, the task scheduling optimization engine configured on the central server further enhances the system's intelligence. By assessing task urgency and drone energy status, it assigns task priority scoring functions (S), thereby achieving an optimal task allocation strategy based on energy utilization efficiency and task urgency. This approach ensures that all drones operate at the highest efficiency at all times, meeting the continuous monitoring needs of pipeline safety while balancing equipment maintenance and energy consumption. In this way, the entire system maximizes resource utilization efficiency while ensuring safety, achieving intelligent management and operation.

[0068] In summary, this invention provides a complete, efficient, and intelligent UAV inspection method and system for aviation fuel pipelines. By deeply integrating multimodal perception, edge intelligent analysis, hierarchical response mechanisms, and cluster collaborative scheduling, it achieves automated and high-precision monitoring of safety hazards along aviation fuel pipelines. The system not only enables real-time anomaly identification in complex environments at the UAV end, effectively reducing data transmission pressure, but also significantly improves the comprehensive judgment capability for various risks such as open flames, illegal construction, surface deformation, and temperature anomalies through the fusion of multi-source information such as visible light, infrared, and lidar. Combined with thermal anomaly index analysis and hierarchical alarm strategies, the system possesses early warning and differentiated handling capabilities; while the multi-UAV collaborative mechanism and autonomous navigation function without network access further enhance the reliability and coverage of inspection operations. The entire solution forms a closed-loop system of front-end perception, local decision-making, accurate data transmission, and central control, which improves inspection efficiency and response speed while reducing labor costs and safety risks, providing strong technical support for the safe operation and maintenance of aviation fuel pipelines. It has good practical value and broad prospects for widespread application.

[0069] The above description of the present invention is merely an example of specific embodiments and is not intended to limit its scope of protection. It should be understood that those skilled in the art can make various modifications, equivalent substitutions, or improvements based on the disclosed technical solutions without departing from the core ideas and basic principles of the present invention, and all such modifications should be considered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims, covering equivalent structures of the described technical features and any substantial improvements and variations thereof.

Claims

1. A method for intelligent inspection of aviation fuel pipelines using unmanned aerial vehicles (UAVs), characterized in that, Includes the following steps: Visual image data along the aviation fuel pipeline is collected using an inspection terminal mounted on a drone. The visual image data covers the pipeline itself and its surrounding geographical environment and is used to identify objects in abnormal scenarios that may endanger the safe operation of the pipeline. The visual image data is input into an abnormal scene recognition model deployed on the drone to identify whether the visual image data contains an abnormal scene of a preset category; the abnormal scene includes open flame or smoke in the first abnormal scene, illegal excavation operation in the second abnormal scene, large construction machinery approaching in the third abnormal scene, and exposed pipe or foundation collapse in the fourth abnormal scene. If the identification results indicate the existence of at least one preset abnormal scenario, a corresponding alarm signal will be generated. The alarm signals and associated visual image data are uploaded to the server system of the remote monitoring center via a wireless communication link, enabling intelligent unmanned inspection of the aviation fuel pipeline's operating status.

2. The intelligent unmanned aerial vehicle (UAV) inspection method for aviation fuel pipelines as described in claim 1, characterized in that, The abnormal scene recognition model is trained in the following way: Obtain a raw image sample set containing various abnormal scenarios. The raw image sample set covers real-world images of aviation fuel pipelines under different meteorological conditions, light intensities, and topographical features. Data cleaning is performed on the original image sample set to remove blurry, overexposed, or heavily background-interferenced images, resulting in a cleaned image sample set. The various abnormal scenes in the cleaned image sample set are categorized and labeled, and the bounding boxes of the detected objects in the images are labeled to generate an annotated labeled image sample set. The labeled image sample set is divided into a training sample set and a test sample set, wherein the training sample set is used for model training and the test sample set is used for model performance verification. The training sample set is input into the initial deep neural network model, and the backpropagation algorithm is used to iteratively optimize the model parameters until the model's recognition accuracy on the test sample set reaches the preset standard, thus obtaining the trained abnormal scene recognition model.

3. The intelligent inspection method for aviation fuel pipelines by unmanned aerial vehicles as described in claim 2, characterized in that, The loss function optimization during model training includes: Input the labeled images from the training sample set into the current model to obtain the predicted bounding boxes, abnormal scene category labels, and confidence scores; Calculate the comprehensive loss value between the prediction result and the true label. The comprehensive loss consists of three parts: bounding box regression loss, category classification loss and confidence loss. The bounding box regression loss is measured using the generalized intersection-union loss function, defined as: ;in, Indicates the predicted bounding box. This represents the actual bounding box; the GIoU value is between 0 and 1, with a higher value indicating a higher degree of overlap. The model parameters are updated using gradient descent based on the total loss, and this process is repeated iteratively until the model converges.

4. The intelligent inspection method for aviation fuel pipelines by unmanned aerial vehicles as described in claim 2, characterized in that, After generating the labeled image sample set, the following is also included: An independent test sample set is divided from the labeled image sample set, and the test sample set includes labeled images that did not participate in the training and their corresponding abnormal scene annotation information; After the model training is completed, the test sample set is input into the trained model to obtain the test recognition results; Compare the test identification results with the real annotations, and calculate the key performance indicators of the model, including mean precision, recall, and false positive rate; When the average precision is ≥90%, the recall rate is ≥85%, and the false positive rate is ≤5%, the abnormal scene recognition model is deemed to meet the on-site deployment conditions.

5. The intelligent inspection method for aviation fuel pipelines by unmanned aerial vehicles as described in claim 4, characterized in that, After confirming that the abnormal scene recognition model meets the deployment conditions, the following steps are also included: Test the inference latency of the model on an embedded computing platform for drones; If the inference latency exceeds the preset real-time threshold, the model is subjected to lightweight processing, including channel pruning, weight quantization and knowledge distillation techniques, to obtain a lightweight abnormal scene recognition model. Evaluate the accuracy loss of the lightweight model on the test sample set; If the accuracy loss does not exceed 2 percentage points, the lightweight anomaly scene recognition model will be deployed to the drone for online real-time analysis.

6. The intelligent inspection method for aviation fuel pipelines by unmanned aerial vehicles as described in claim 1, characterized in that, The process of inputting visual image data into an anomaly scene recognition model and outputting anomaly scene recognition results includes: The input image is divided into a regular grid structure, and multiple candidate region boxes with different scales and aspect ratios are preset within each grid. For candidate region boxes containing the detected object, predict the abnormal scene category to which it belongs, and adjust the bounding box position to fit the actual target contour; The nonmaximum suppression algorithm is applied to retain the detection result with the highest confidence from the overlapping prediction boxes, and outputs the target's location coordinates, abnormal scene category label and confidence score.

7. The intelligent unmanned aerial vehicle (UAV) inspection method for aviation fuel pipelines as described in claim 6, characterized in that, The process of generating candidate region boxes and predicting targets includes: The input image is convolved by the backbone feature extraction network of the abnormal scene recognition model to generate a first feature map with a first resolution, and the first feature map contains high-level semantic information. The first feature map is upsampled using a feature pyramid structure and then fused with a second feature map output from a shallow network. The second feature map has a higher resolution than the first feature map, thus preserving more texture and edge details. A third feature map, which has the highest resolution, is further introduced from the original image or the output of the primary convolutional layer to enhance the detection capability of small-sized objects. The fused multi-level features are fed into the detection head module, which outputs the probability of the abnormal scene category, the bounding box offset, and the confidence of the target presence for each candidate region.

8. The intelligent inspection method for aviation fuel pipelines by unmanned aerial vehicles as described in claim 1, characterized in that, Also includes: Based on the temperature distribution sequence along the pipeline in thermal infrared images, a temperature variation curve along the pipeline is constructed. Calculate the local thermal anomaly index The formula is as follows: ;in, This is the set of temperature values ​​measured on the surface of the current pipe section. This is the highest temperature value in the set; This represents the average temperature of the corresponding pipe section under historical normal operating conditions. The standard deviation of historical temperatures; when At that time, among them To preset the risk assessment threshold, and The system determined that there was a risk of leakage in the area, triggering a Level 1 warning signal.

9. The intelligent inspection method for aviation fuel pipelines by unmanned aerial vehicles as described in claim 1, characterized in that, When multiple drones work together to perform inspection tasks, it also includes: Real-time acquisition of remaining battery power for each drone and the urgency weight of the tasks to be performed ; Establish a task priority scoring function Defined as: ;in, This is the drone's current remaining battery power. This is the battery's rated capacity, therefore This is the normalized ratio of remaining battery power. The urgency weight is the value of the task, which takes a value in the range of [0,1] after normalization. is the energy preference coefficient, with a value range of [0,1], used to adjust the relative weight of energy status and task urgency; The central dispatch system is based on the calculated The system dynamically assigns high-priority tasks to the drone with the highest score, achieving coordinated optimization of energy consumption and task requirements.

10. A drone-based intelligent inspection system for aviation fuel pipelines, characterized in that, include: The drone inspection device is configured to carry a multi-source sensor terminal to perform automatic patrols along the aviation fuel pipeline and collect visual image data. The edge intelligence analysis module is installed on the drone itself and is used to run a lightweight abnormal scene recognition model to perform local real-time analysis on the collected visual image data to identify whether there are preset abnormal scenes. The alarm triggering module is used to generate corresponding alarm signals when abnormal scenarios are detected; The communication transmission module is used to upload the alarm signal and associated image data to the remote monitoring center; The preset abnormal scenarios include open flames or smoke, illegal excavation operations, the approach of large construction machinery, exposed pipes, or foundation collapse.

Images