A patrol management system and method for long-distance pipelines

Abstract

This invention relates to the field of pipeline inspection and management technology, specifically a long-distance pipeline inspection and management system and method, including a centralized control center, edge processing equipment, cameras, a cloud platform, and manual inspection terminals; the cameras are installed at designated locations for real-time monitoring of target objects. The advantages of this invention are: by introducing a data governance module and adding three-dimensional tags containing timestamps, location coordinates, and associated object IDs to multi-source data, and simultaneously establishing a core index master ledger and associated ledgers bound together by unique data IDs, this system achieves automatic association and fusion of unstructured, structured, and semi-structured data generated during inspections. This solves the "data silo" problem caused by inconsistent data formats and scattered storage in existing technologies, providing a complete and traceable data chain for each potential hazard or equipment status, supporting full-process traceability and analysis.

Description

Technical Field

[0001] This invention relates to the field of pipeline patrol and management technology, and in particular to a patrol and management system and method for long-distance pipelines. Background Technology

[0002] As a core infrastructure for aviation energy transportation, long-distance aviation fuel pipelines are characterized by long laying distances, complex environments, and high operational risks. Their safety inspections directly impact the stability and security of the aviation energy supply chain. Currently, aviation fuel pipeline inspections primarily rely on a combination of traditional manual inspections and rudimentary intelligent auxiliary methods. However, this approach has revealed numerous technical bottlenecks in practical applications, making it difficult to meet the demands of modern safety management. Existing aviation fuel pipeline inspections still predominantly employ the traditional "manual inspection" model, requiring inspectors to collect images and videos of pipeline equipment on-site and manually record inspection information. This model suffers from three core problems: First, extremely low data processing efficiency. The massive amounts of unstructured data generated daily require manual review, which is not only time-consuming and labor-intensive but also prone to missed hazards due to personnel fatigue and experience differences. Second, insufficient flexibility and real-time capability. It cannot monitor and record dynamic scenarios such as third-party construction in real time, and lacks complete process data to support post-incident traceability in the event of a safety accident. Third, low data correlation. Manually recorded structured information and unstructured data are disconnected, forming "data silos" that hinder full-process traceability and analysis.

[0003] With the development of computer vision technology, some general-purpose target detection models, such as the YOLO series, have begun to be applied to the field of pipeline inspection. However, due to the special characteristics of aviation fuel scenarios, these technologies have significant adaptability defects: existing technologies have not established a unified access and fusion standard for multi-source data in aviation fuel scenarios. Camera video streams, manual inspection terminal data, equipment ledgers, construction reporting information, etc., are stored in a scattered manner, lacking a standardized labeling system and association mechanism, resulting in unstructured data not being accurately mapped to structured data, making it difficult to extract data value; general-purpose target detection models have not been customized and optimized for aviation fuel inspection scenarios. On the one hand, they are difficult to adapt to complex environments such as backlight, rain, and night, and lack robustness to scenarios such as pipeline metal reflection and open backgrounds in storage areas; on the other hand, for aviation fuel-specific hidden danger targets, such as small pipeline corrosion, large construction machinery, and low-altitude smoke and fire, general-purpose models have problems with low detection rates for small targets and difficulty in identifying occluded targets. Furthermore, they have not designed industry-specific target classification systems and anchor frame structures, resulting in low recognition accuracy. The existing models rely solely on single visual features for detection, failing to consider the business logic and context of aviation fuel inspection. This leads to misidentification of clouds as smoke, birds as intrusion targets, and compliant construction machinery as illegal intrusions. The high false alarm rate increases the workload of maintenance personnel, severely impacting the practical application of intelligent inspection. Furthermore, the existing technology can only achieve preliminary hazard identification, lacking a full-process trajectory tracking mechanism for third-party construction machinery. It cannot record the complete "entry-operation-departure" process, making it difficult to trace illegal construction activities. Simultaneously, the identification results are disconnected from safety management processes, failing to form a closed-loop "identification-early warning-processing-review" chain, thus hindering the effective transformation of intelligent detection results into management value. Existing target detection models often suffer from high computational requirements and large model sizes, making real-time inference difficult on edge devices such as inspection cameras and mobile patrol terminals. Moreover, the lack of collaborative deployment solutions between the edge and cloud limits the ability to simultaneously meet real-time detection needs and the need for in-depth analysis of full-volume data, restricting large-scale application.

[0004] The aviation fuel industry has extremely stringent requirements for safe operation. The shortcomings of traditional inspection models and existing intelligent technologies have led to multiple dilemmas in aviation fuel pipeline inspection, such as "inaccurate identification of hidden dangers, lack of standardized data governance, lack of basis for construction traceability, and inflexible deployment and application," which seriously restrict the improvement of inspection efficiency and safety management level. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a patrol management system and method for long-distance pipelines, which effectively solves the deficiencies of the prior art.

[0006] To achieve the above objectives, one embodiment of the present invention provides a patrol management system and method for long-distance pipelines, including a centralized control center, an edge processing device, a camera, a cloud platform, and a manual inspection terminal; the camera is set at a designated location for real-time monitoring of target objects; the edge processing device is set at a designated location on the side of the inspection area and is electrically connected to the camera at the corresponding location; the edge processing device, the camera, and the manual inspection terminal are all electrically connected to the centralized control center; the cloud platform is communicatively connected to the edge processing device.

[0007] Preferably, in any of the above schemes, the edge processing device is equipped with an image recognition model, which is trained based on an improved YOLOX architecture, wherein the improved YOLOX architecture includes inserting a TransformerEncoder layer after the YOLOX CSPDarknet53 backbone network.

[0008] Preferably, the architecture of the image recognition model, as described in any of the above schemes, includes, in sequence: an input layer for receiving standardized preprocessed image data, a YOLOX-CSPDarknet53 backbone network, a feature dimensionality reduction and location coding layer, a TransformerEncoder enhancement layer, an FPN-PAN multi-scale fusion network, and a dedicated aviation fuel decoupling detection head.

[0009] Preferably, in any of the above schemes, the centralized control center includes a data governance module, which is used to perform fusion governance on multi-source inspection data. The data governance module adds three-dimensional tags containing timestamps, location coordinates, and associated object IDs to the data and establishes a core index master ledger and associated ledgers to associate the data.

[0010] Preferably, in any of the above schemes, the edge processing device further includes a false alarm suppression module, which integrates a strong quantization CNN model. The strong quantization CNN model is used to perform secondary verification on the recognition results that the image recognition model initially determines to be confused targets.

[0011] Preferably, in any of the above schemes, the system has a pre-set target risk level classification table, which defines different target categories and their corresponding risk levels and detection priorities.

[0012] A method for patrol and management of long-distance pipelines, used in the aforementioned system, includes the following steps: S1. Obtain image information sent by the camera; S2. Recognize the image information to obtain an image recognition result containing the target category and target location; S3. Determine the corresponding detection priority based on the target category; S4. Perform fault identification processing on the image information according to the detection priority to obtain the fault identification result; S5. Based on the fault identification results, generate a hidden fault package or retain image information, and send it to the centralized control center or cloud accordingly.

[0013] Preferably, in any of the above schemes, the process of recognizing image information is performed by an image recognition model based on an improved YOLOX architecture, including: 1) Extract image features using the YOLOX-CSPDarknet53 backbone network; 2) Reduce the dimensionality of the extracted features and add two-dimensional positional encoding; 3) Capture global context features through the TransformerEncoder layer; 4) Multi-scale feature fusion using FPN-PAN network; 5) Output the identification results through the aviation fuel-specific decoupling detection head.

[0014] Preferably, when the image recognition result belongs to a confused target, the method further includes a false alarm suppression step, which includes: 1) Extract local features of the obfuscated target object; 2) Input the local features into the strongly quantized CNN model for rapid determination to obtain the secondary verification result; 3) Based on the consistency between the secondary verification results and the initial identification results, decide whether to continue the subsequent process or report to the centralized control center.

[0015] Preferably, in any of the above schemes, the method further includes a data fusion governance step, which is executed by the centralized control center and includes: 1) Classify data into unstructured data, structured data, semi-structured data, and intelligent analysis results; 2) Standardize all types of data and add 3D labels; 3) Establish a core index master ledger and related ledgers, and bind and associate them through a unique data ID.

[0016] The present invention has the following advantages: 1. The patrol management system and method for this long-distance pipeline introduces a data governance module and adds three-dimensional labels containing timestamps, location coordinates, and associated object IDs to multi-source data. At the same time, it establishes a core index master ledger and associated ledgers and binds them through unique data IDs. This system realizes the automatic association and fusion of unstructured, structured, and semi-structured data generated during inspections, solving the "data silo" problem caused by inconsistent data formats and scattered storage in existing technologies. It provides a complete and traceable data chain for each hidden danger or equipment status, supporting the traceability and analysis of the entire process. 2. The patrol management system and method for this long-distance pipeline adopts an image recognition model based on the improved YOLOX architecture. This model inserts a TransformerEncoder layer after the YOLOX-CSPDarknet53 backbone network to capture global contextual features, and combines the FPN-PAN multi-scale fusion network with a decoupled detection head for aviation fuel. This improves the detection capability of targets with large scale differences in aviation fuel inspection scenarios, such as corrosion of small pipelines and large construction machinery. In particular, it enhances the recognition accuracy of small targets and occluded targets, thereby solving the technical problem of poor adaptability and low recognition accuracy of general target detection models in special aviation fuel scenarios. 3. The patrol management system and method for this long-distance pipeline, through the false alarm suppression module and its integrated strong quantization CNN model set in the edge processing equipment, performs secondary verification on the confusing target objects initially identified by the image recognition model, such as smoke and clouds. Combined with the mechanism of continuous optimization of the model by the centralized control center based on human feedback, a closed-loop process from identification, verification to correction is formed. This technical solution effectively reduces the high false alarm rate caused by environmental interference or model limitations, reduces the ineffective workload of operation and maintenance personnel, and improves the practicality and reliability of intelligent inspection. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited thereto.

[0019] Example 1: As Figure 1As shown, a patrol management system and method for long-distance pipelines includes a centralized control center, edge processing devices, cameras, a cloud platform, and manual inspection terminals. Cameras are deployed at designated locations near the monitored targets. The edge processing devices are deployed at designated locations on the patrol side and form data acquisition nodes with the cameras in their respective areas via electrical connections. The edge processing devices, cameras, and manual inspection terminals are all communicatively connected to the centralized control center. The cloud platform acts as a server, communicating with the edge processing devices and the centralized control center for data storage.

[0020] The edge processing device is equipped with an image recognition model, which is trained based on an improved YOLOX architecture. The improved YOLOX architecture includes inserting a TransformerEncoder layer after the YOLOX CSPDarknet53 backbone network.

[0021] The architecture of the image recognition model includes, in sequence: an input layer for receiving standardized preprocessed image data, a YOLOX-CSPDarknet53 backbone network, a feature dimensionality reduction and location coding layer, a TransformerEncoder enhancement layer, an FPN-PAN multi-scale fusion network, and a dedicated aviation fuel decoupling detection head.

[0022] The centralized control center includes a data governance module, which is used to integrate and govern multi-source inspection data. The data governance module adds three-dimensional tags containing timestamps, location coordinates and associated object IDs to the data, and establishes a core index master ledger and associated ledgers to associate the data.

[0023] The edge processing device also includes a false alarm suppression module, which integrates a strong quantization CNN model. The strong quantization CNN model is used to perform secondary verification on the recognition results that the image recognition model initially determined to be confusing targets.

[0024] The system has a pre-set target risk level classification table, as shown in the table below. The target risk level classification table defines different target categories and their corresponding risk levels and detection priorities.

[0025]

[0026] A method for patrol and management of long-distance pipelines, used in the aforementioned system, includes the following steps: S1. Obtain image information sent by the camera; S2. Recognize the image information to obtain an image recognition result containing the target category and target location; S3. Determine the corresponding detection priority based on the target category; S4. Perform fault identification processing on the image information according to the detection priority to obtain the fault identification result; S5. Based on the fault identification results, generate a hidden fault package or retain image information, and send it to the centralized control center or cloud accordingly.

[0027] The process of recognizing image information is performed by an image recognition model based on an improved YOLOX architecture, including: 1) Extract image features using the YOLOX-CSPDarknet53 backbone network; 2) Reduce the dimensionality of the extracted features and add two-dimensional positional encoding; 3) Capture global context features through the TransformerEncoder layer; 4) Multi-scale feature fusion using FPN-PAN network; 5) Output the identification results through the aviation fuel-specific decoupling detection head.

[0028] When the image recognition result belongs to a confused target, the method also includes a false alarm suppression step, which includes: 1) Extract local features of the obfuscated target object; 2) Input the local features into the strongly quantized CNN model for rapid judgment and obtain the secondary verification results; 3) Based on the consistency between the secondary verification results and the initial identification results, decide whether to continue the subsequent process or report to the centralized control center.

[0029] The method also includes a data fusion governance step, which is executed by a centralized control center and includes: 1) Classify data into unstructured data, structured data, semi-structured data, and intelligent analysis results; 2) Standardize all types of data and add 3D labels; 3) Establish a core index master ledger and related ledgers, and bind and associate them through a unique data ID.

[0030] The core of this embodiment lies in the image recognition model configured within the edge processing device, which is trained based on an improved YOLOX architecture. The improvement strategies include retaining the YOLOX CSPDarknet53 backbone network to ensure real-time performance, inserting a TransformerEncoder layer with an eight-head self-attention mechanism to extract global contextual features, thereby solving the detection challenges of small targets such as minor corrosion on pipes and occluded targets such as mechanical obstructions to pipes. Addressing the large scale differences of targets in aviation fuel scenarios, an adaptive anchor box generation algorithm is designed, generating six sets of anchor boxes based on the aviation fuel dataset through clustering, improving the accuracy of multi-scale target detection. The domain adaptation module uses transfer learning combined with a small amount of labeled aviation fuel data for fine-tuning, solving the adaptation problem of the general model in aviation fuel scenarios such as metal reflections on pipes.

[0031] The image recognition model's architecture consists of an input layer, a preprocessing module, a YOLOX-CSPDarknet53 backbone network, feature dimensionality reduction and positional encoding, and a TransformerEncoder enhancement layer. The output is then processed by an FPN-PAN multi-scale fusion network and a dedicated aviation fuel decoupled detection head. Input image information uses a standardized format with a uniform size of 640x640 pixels to balance edge computing power and detection accuracy. For different environments, Retinex-Net brightening and Gamma correction are used for backlighting or nighttime conditions, while dark channel prior dehazing is used for rainy or foggy days. Noise processing is achieved through median filtering, significantly improving the model's robustness in outdoor aviation fuel inspection environments. During training, sample images are enhanced with Mosaic by stitching four images together, random horizontal flipping, and color jittering to vary brightness and saturation, thus enhancing the model's generalization ability.

[0032] Further model optimizations include preserving the CSP structure to reduce redundant computation and adjusting the number of convolutional kernels for aviation fuel targets. Specifically, the Focus module layer undergoes a four-fold downsampling, the CSP1_1 layer extracts low-level textures, the SPPF module enhances multi-scale target adaptability, and the CSP2_1 and CSP2_2 layers strengthen mid-to-high-level feature extraction capabilities, ultimately outputting a multi-scale feature map. In the feature dimensionality reduction and positional encoding stages, the backbone network output features are uniformly reduced to 256 channels, and sinusoidal positional encoding is added to preserve spatial information. The TransformerEncoder enhancement layer uses a lightweight, stacked two-layer structure, capturing global features through multi-head self-attention and a feedforward network. The FPN-PAN fusion network improves detection accuracy by fusing global and local features in a top-down and bottom-up manner.

[0033] The false alarm suppression mechanism is activated when identifying confusing targets such as smoke or people. It uses a strongly quantized CNN model to perform secondary verification of local features, and determines the process direction based on the consistency of the results. At the same time, the spatiotemporal context verification requires that the target's detection confidence level meets the standard for three consecutive frames to take effect, and it also combines construction reporting information to reduce false alarms.

[0034] Example 2: This example focuses on describing the data fusion governance process, aiming to eliminate data silos and achieve full-process traceability, specifically as follows: The first step is to divide the inspection data into four categories: unstructured data such as videos and pictures, structured data such as equipment ledgers, semi-structured data such as manual records, and intelligent analysis results. The second step is to unify the data format to MP4 or JSON, with a resolution of 1080P and a time format accurate to milliseconds. The third step is to add 3D labels to all data, including timestamp location coordinates and associated object IDs, to achieve precise mapping. The fourth step is to define the required fields for each type of data, such as Device ID and Target Type, to ensure data standardization; The fifth step is to establish a core index master ledger that stores only the global index and three-dimensional labels as navigation indexes; The sixth step is to create related ledgers according to data types, including unstructured ledgers storing video paths, structured ledgers storing device information, and AI ledgers storing recognition results, etc. Step 7: Achieve one-click traceability by binding the main ledger and related ledgers with unique data IDs; Step 8: Based on 3D labels, automatically match data when the time difference is less than three seconds, the position difference is less than ten meters, and the IDs are consistent; automatically associate multiple types of data, such as illegal machinery with corresponding videos. The ninth step involves handling conflicting data by prioritizing structured data and high-confidence AI results, followed by manual review to ensure the integrity of the data chain.

[0035] The data governance process also supports real-time updates, synchronizing data to the ledger within one second of data collection and providing multi-dimensional query functions to retrieve data by device ID, time region, or hazard type. At the business level, ledger data is used to generate visual maps, such as machinery trajectories, automatically dispatching hazard work orders and tracking processing status to form a closed-loop management system. Database selection includes MySQL for managing relational data, MinIO for storing unstructured data, and InfluxDB for optimizing time-based query interfaces, adapting to HTTP and HTTPS protocols, supporting edge and cloud collaboration, and ensuring data consistency and system stability.

[0036] In summary, by introducing a data governance module and adding 3D labels containing timestamps, location coordinates, and associated object IDs to multi-source data, and by establishing a core index master ledger and associated ledgers and binding them with unique data IDs, this system achieves automatic association and fusion of unstructured, structured, and semi-structured data generated during inspections. This solves the "data silo" problem caused by inconsistent data formats and scattered storage in existing technologies, providing a complete and traceable data chain for each potential hazard or equipment status, supporting full-process traceability and analysis. Furthermore, by employing an image recognition model based on an improved YOLOX architecture, which inserts a TransformerEncoder layer after the YOLOX-CSPDarknet53 backbone network to capture global contextual features, and combines FPN-PAN multi-scale fusion... The network-decoupled detection head for aviation fuel enhances the detection capability for targets with large scale differences in aviation fuel inspection scenarios, such as corrosion of small pipelines and large construction machinery. In particular, it enhances the recognition accuracy of small and occluded targets, thus solving the technical problem of poor adaptability and low recognition accuracy of general target detection models in special aviation fuel scenarios. Through the false alarm suppression module set in the edge processing device and its integrated strong quantization CNN model, the initially identified confusing targets, such as smoke and clouds, are further verified. Combined with the mechanism of continuous model optimization based on human feedback in the centralized control center, a closed-loop process from recognition, verification to correction is formed. This technical solution effectively reduces the high false alarm rate caused by environmental interference or model limitations, reduces the ineffective workload of maintenance personnel, and improves the practicality and reliability of intelligent inspection.

[0037] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A patrol and management system for long-distance pipelines, characterized in that: It includes a centralized control center, edge processing equipment, cameras, cloud computing, and manual inspection terminals; the cameras are set at designated locations for real-time monitoring of target objects; the edge processing equipment is set at designated locations on the side of the inspection area and is electrically connected to the cameras in the corresponding area. The edge processing device, camera, and manual inspection terminal are all electrically connected to the centralized control center; the cloud is communicatively connected to the edge processing device.

2. The patrol and management system for long-distance pipelines according to claim 1, characterized in that: The edge processing device is equipped with an image recognition model, which is trained based on an improved YOLOX architecture. The improved YOLOX architecture includes inserting a TransformerEncoder layer after the YOLOX CSPDarknet53 backbone network.

3. The patrol and management system for long-distance pipelines according to claim 2, characterized in that: The architecture of the image recognition model includes, in sequence: an input layer for receiving standardized preprocessed image data, a YOLOX-CSPDarknet53 backbone network, a feature dimensionality reduction and location coding layer, a TransformerEncoder enhancement layer, an FPN-PAN multi-scale fusion network, and a dedicated aviation fuel decoupling detection head.

4. The patrol and management system for long-distance pipelines according to claim 3, characterized in that: The centralized control center includes a data governance module, which is used to integrate and govern multi-source inspection data. The data governance module adds three-dimensional tags containing timestamps, location coordinates and associated object IDs to the data, and establishes a core index master ledger and associated ledgers to associate the data.

5. The patrol and management system for long-distance pipelines according to claim 4, characterized in that: The edge processing device also includes a false alarm suppression module, which integrates a strong quantization CNN model. The strong quantization CNN model is used to perform secondary verification on the recognition results that the image recognition model initially determined to be confusing targets.

6. The patrol and management system for long-distance pipelines according to claim 5, characterized in that: The system has a pre-set target risk level classification table, which defines different target categories and their corresponding risk levels and detection priorities.

7. A method for patrol and management of long-distance pipelines, used in the system described in any one of claims 1-6, characterized in that: Includes the following steps: S1. Obtain image information sent by the camera; S2. Recognize the image information to obtain an image recognition result containing the target category and target location; S3. Determine the corresponding detection priority based on the target category; S4. Perform fault identification processing on the image information according to the detection priority to obtain the fault identification result; S5. Based on the fault identification results, generate a hidden fault package or retain image information, and send it to the centralized control center or cloud accordingly.

8. The patrol and management method for long-distance pipelines according to claim 7, characterized in that: The process of recognizing image information is performed by an image recognition model based on an improved YOLOX architecture, including: 1) Extract image features using the YOLOX-CSPDarknet53 backbone network; 2) Reduce the dimensionality of the extracted features and add two-dimensional positional encoding; 3) Capture global context features through the TransformerEncoder layer; 4) Multi-scale feature fusion using FPN-PAN network; 5) Output the identification results through the aviation fuel-specific decoupling detection head.

9. The patrol and management method for long-distance pipelines according to claim 7, characterized in that: When the image recognition result belongs to a confused target, the method further includes a false alarm suppression step, which includes: 1) Extract local features of the obfuscated target object; 2) Input the local features into the strongly quantized CNN model for rapid determination to obtain the secondary verification result; 3) Based on the consistency between the secondary verification results and the initial identification results, decide whether to continue the subsequent process or report to the centralized control center.

10. The patrol and management method for long-distance pipelines according to claim 7, characterized in that: The method further includes a data fusion governance step, which is executed by the centralized control center and includes: 1) Classify data into unstructured data, structured data, semi-structured data, and intelligent analysis results; 2) Standardize all types of data and add 3D labels; 3) Establish a core index master ledger and related ledgers, and bind and associate them through a unique data ID.

Images