Underground pipe gallery data real-time processing system based on edge calculation
Through edge computing technology, multi-source data association relationships and cross-node knowledge migration are dynamically adjusted to form an self-organized collaborative computing cluster, solving the real-time and robustness of the underground pipeline monitoring system in a dynamic environment, and achieving efficient and accurate fault detection and decision-making.
Patent Information
- Application Number
- CN202510594327.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-09
- Publication Date
- 2025-08-15
AI Technical Summary
Traditional underground pipeline monitoring systems have insufficient adaptability to dynamic environments, low data processing efficiency and lack of system robustness, resulting in high false alarm rate, slow response speed and low computing power utilization efficiency, which cannot meet the real-time and reliability requirements under complex working conditions.
The dynamic perception primitive library based on edge computing, edge collaborative learning network, fault-driven dynamic task allocation module, multi-modal data channel optimization module, adaptive anomaly detection module, edge computing load balancing module, cross-modal data fusion decision-making module and global robustness enhancement module are adopted to realize dynamic perception, efficient collaborative processing and system-level fault tolerance of multi-source data.
It realizes high-precision real-time monitoring of underground pipe corridors under complex working conditions, reduces the false alarm rate, improves data processing efficiency and system robustness, and ensures rapid response and decision-making accuracy of fault detection.
Smart Images

Figure CN120495766A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a real-time processing system for underground pipeline gallery data based on edge computing, and belongs to the technical field of intelligent monitoring of underground pipeline galleries. Background Art
[0002] As an important part of urban infrastructure, the operation status of underground pipeline corridors directly affects public safety. It is necessary to conduct real-time monitoring and analysis of multi-source data such as pipeline pressure, temperature and humidity, and equipment vibration. Traditional monitoring systems have the following technical problems: 1. Insufficient adaptability to dynamic environments: Traditional solutions rely on fixed thresholds or single parameters to trigger abnormal alarms, and cannot effectively perceive the coupled changes of multiple parameters such as pipeline pressure, temperature and humidity. For example, when rainfall increases suddenly, the correlation between pipeline leakage risk and pressure and video texture features is significantly enhanced, but fixed thresholds cannot dynamically adjust data weights, resulting in a high false alarm rate and poor response to gradual faults (such as pipeline corrosion). )'s response is seriously delayed; 2. Bottleneck in data processing efficiency: unstructured data (such as video streams) need to rely on the cloud for centralized processing, resulting in edge nodes only assuming simple forwarding functions. For example, the full transmission of video streams to the cloud not only occupies bandwidth, but the cloud processing delay may exceed 10 seconds, which makes it difficult to meet the real-time decision-making needs of emergencies such as leakage; 3. Lack of system robustness: There is a lack of coordination mechanism between existing edge nodes, and local failures can easily cause data processing chain breaks. The industry usually improves reliability by adding redundant nodes, but the long-term idleness of redundant nodes leads to a waste of computing power, and manual switching takes a long time, making it difficult to restore key monitoring functions within 5 seconds.
[0003] To alleviate these issues, current technologies primarily employ the following improvements: 1. Predicting pipeline status through machine learning models, but these models rely on historical data training, resulting in insufficient generalization capabilities when faced with unknown environmental parameters (such as extreme weather), and thus unable to avoid false alarms. 2. Some computing tasks are decentralized to edge nodes, but the computing power of a single node is limited and cannot support real-time fusion analysis of multimodal data. 3. Data backup by master nodes improves reliability, but manual intervention is required to switch master nodes when they fail, resulting in recovery times exceeding 30 seconds and failing to meet the requirements for continuous monitoring of underground pipeline corridors. While these improvements have alleviated some of the issues at a local level, they have failed to address the following core contradictions: Traditional thresholds or models struggle to adapt to environmental changes, making it difficult for the system to balance detection accuracy and response speed under complex working conditions. A lack of knowledge sharing mechanisms between nodes prevents the dynamic reconstruction of computing clusters based on local anomalies, resulting in inefficient computing power utilization. The full transmission of video streams and independent processing of sensor data results in a loss of spatiotemporal correlation between multimodal data, impacting fault location accuracy. Therefore, how to achieve dynamic perception, efficient collaborative processing and system-level fault tolerance of multi-source data in underground pipeline corridors while meeting the real-time and reliability requirements under complex working conditions has become the technical problem to be solved by the present invention. Summary of the Invention
[0004] The present invention provides a real-time processing system for underground pipeline corridor data based on edge computing, the main purpose of which is to solve the problems of insufficient adaptability to dynamic environments, low data processing efficiency and lack of system robustness.
[0005] To achieve the above objectives, the present invention provides a real-time processing system for underground pipeline corridor data based on edge computing, comprising:
[0006] A dynamic perception primitive library is used to abstract multi-source heterogeneous data from underground pipeline corridors, including pipeline pressure, temperature and humidity, equipment vibration frequency, current waveform, and video surveillance texture changes and motion trajectories, into dynamic perception primitives. Each dynamic perception primitive contains at least data source identification, feature extraction rules, and initial association weight information.
[0007] The primitive reorganization module is configured to dynamically adjust the association rules and association weights between at least some of the dynamic perception primitives in the dynamic perception primitive library based on the real-time acquired environmental context information including rainfall, so as to form a dynamic feature combination adapted to the current environmental state, wherein the adjustment of the association weights is at least partially based on the function W new =W old +α×ΔC,W new is the new association weight, W old is the old association weight, ΔC is the change in environmental context, and α is the preset sensitivity coefficient;
[0008] The edge collaborative learning network consists of a lightweight federated learning framework deployed on multiple edge computing nodes. It is used to share the association rule parameters of dynamic perception primitives among multiple edge computing nodes to achieve cross-node knowledge transfer and model collaborative update;
[0009] The fault-driven dynamic task allocation module is configured to automatically trigger adjacent edge computing nodes to start high-precision data sampling mode when any edge computing node detects an abnormal dynamic feature combination that meets preset conditions, forming a local collaborative computing cluster for collaborative fault diagnosis and analysis;
[0010] The multimodal data channel optimization module is used to logically slice unstructured video stream data according to the physical structure of the underground pipeline corridor and upload only the video slices containing abnormal information to the cloud. It also performs adaptive Huffman coding compression on time series sensor data based on the periodic characteristics of pipeline pressure fluctuations to optimize data transmission efficiency.
[0011] An adaptive anomaly detection module is configured to detect anomalies, including gradual faults, in underground utility corridors in real time based on dynamically adjusted primitive association rules and weights, and output anomaly alarm information. The detection response time for gradual faults is less than or equal to five seconds.
[0012] The edge computing load balancing module is used to dynamically allocate computing tasks based on the computing load of each edge computing node in the edge collaborative learning network to improve the computing power utilization and balance of edge nodes;
[0013] A cross-modal data fusion decision module is configured to receive dynamic perception primitives and their associated weight information from different data modalities and adaptively assign fusion weights to each modality based on the context to improve the accuracy of the integrated decision;
[0014] A global robustness enhancement module, which is used to maintain the basic functional operation of the system through the edge collaborative learning network and the fault-driven dynamic task allocation module when a single edge computing node fails;
[0015] The visualization map generation module is used to generate visualization maps including heat maps of abnormal dynamic feature combinations to intuitively display fault location logic and system operation status.
[0016] Preferably, the primitive recombining module uses a sliding time window to analyze the frequency of the dynamic perception primitive combination. When the frequency of a specific abnormal dynamic perception primitive combination F abnormal Exceeds the preset threshold T f When , the adjacent edge computing nodes are triggered to start high-precision sampling mode.
[0017] Preferably, in the multimodal data channel optimization module, the video stream is logically sliced according to the physical structure partitions of the underground pipeline corridor such as valve areas and connections, and the video logical slices are bound to the spatial coordinate information of the corresponding sensor data.
[0018] Preferably, the parameter sharing process of the edge collaborative learning network adopts differential privacy encryption technology to ensure the data security of the primitive association rule parameters shared across nodes.
[0019] Preferably, it also includes a primitive conflict resolution mechanism for constructing a dynamic priority matrix based on the ant colony pheromone principle. When multiple dynamic perception primitive association rules conflict, the optimal weight distribution scheme is automatically selected based on the pheromone concentration of historical successful decision paths.
[0020] Preferably, the conflict types defined by the primitive conflict resolution mechanism include data source conflict, spatiotemporal conflict, weight conflict, confidence conflict and business logic conflict, and conflict type matching and pheromone concentration comparison are performed through a lightweight rule engine to achieve real-time response of less than fifty milliseconds.
[0021] Preferably, it also includes an energy consumption-aware task migration mechanism, which is configured to reuse existing temperature sensor data to build a thermodynamic model of the edge computing node, calculate the heat dissipation efficiency attenuation coefficient, and automatically migrate the computing task to the low-temperature area of the adjacent node when the node temperature exceeds a preset threshold.
[0022] Preferably, the energy-aware task migration mechanism gives priority to selecting adjacent nodes with copies of the same data shard for task migration to avoid data retransmission overhead.
[0023] Preferably, it also includes a cross-modal data timing alignment engine for enhancing the hardware clock synchronization accuracy, using key frame interpolation technology to compensate for transmission delays in video streams, and eliminating clock drift through sliding window mean filtering of sensor data streams, so that the timestamp alignment accuracy of multimodal data reaches within plus or minus one millisecond.
[0024] Preferably, it also includes an emergency decision topology protocol, which is used to dynamically reconstruct the data stream transmission priority according to the leakage alarm level of the underground pipeline corridor area. When the leakage level in a certain area exceeds the preset level, the video stream transmission path of the area is automatically switched to the user datagram protocol high-speed channel, and the triple sampling frequency compensation mechanism of the sensor data stream is triggered.
[0025] Compared with the background technology problems, the beneficial effects of the present invention are:
[0026] 1. By abstracting multi-source heterogeneous data into dynamic perception primitives and assigning associated weights, the system can automatically adjust the coupling relationship between data based on real-time environmental context (such as changes in rainfall). When environmental parameters suddenly change, the primitive reconstruction module drives weight updates through functions to dynamically enhance the correlation between video texture features and pressure sensor data, effectively identifying leakage precursors that are difficult to capture using traditional fixed rules. This dynamic coupling mechanism based on physical environment feedback avoids the traditional monitoring system's reliance on preset thresholds, makes the anomaly detection logic adaptive, and maintains high-precision decision-making under complex working conditions.
[0027] 2. The edge collaborative network based on lightweight federated learning realizes cross-node knowledge transfer while ensuring data privacy. When a node detects a combination of abnormal vibration and current waveform characteristics, the fault-driven module automatically activates the high-precision sampling mode of the adjacent node to form a local diagnostic cluster. This self-organizing collaborative mechanism not only avoids the delay of centralized processing in the cloud, but also resolves the bottleneck of single-point computing power through dynamic task allocation, so that the system can still maintain complete monitoring capabilities through topology reconstruction in the event of local failures, significantly improving the anti-destruction capability of the underground pipeline monitoring network.
[0028] 3. For unstructured video streams, the system logically slices the video according to physical partitions such as the pipeline corridor valve area and connection points, and only uploads video clips containing abnormal motion. At the same time, the sensor data stream is compressed through adaptive coding, and the spatial alignment transmission of multimodal data is achieved by binding the video slices with the pressure sensor coordinates of the corresponding position. This mechanism makes full use of the structural characteristics of the pipeline corridor, retaining the spatiotemporal correlation of key data while reducing bandwidth occupancy, providing accurate input for three-dimensional digital twins, and avoiding the loss of details caused by full transmission of traditional solutions.
[0029] 4. When spatiotemporal conflicts occur among multi-source primitive rules, the system constructs a dynamic priority matrix based on the ant colony pheromone principle, automatically avoiding conflicts by comparing the concentrations of historical successful paths. At the same time, it reuses temperature sensor data to construct a node thermodynamic model, migrating computing tasks to low-temperature areas before the heat dissipation efficiency decays. Two types of mechanisms form closed loops at the data decision layer and the physical resource layer, respectively: conflict resolution ensures logical consistency, and thermodynamic migration optimizes hardware load, jointly avoiding the common problems of logical oscillation and device overheating in traditional edge computing. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 This is a framework diagram of the real-time processing system for underground pipeline corridor data based on edge computing of the present invention;
[0031] Figure 2 This is a diagram of the lightweight federated learning and self-organizing diagnostic cluster framework in the edge computing network of the present invention.
[0032] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION
[0033] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0034] The present application provides a real-time processing system for underground pipe gallery data based on edge computing, including:
[0035] A dynamic perception primitive library is used to abstract multi-source heterogeneous data from underground pipeline corridors, including pipeline pressure, temperature and humidity, equipment vibration frequency, current waveform, and video surveillance texture changes and motion trajectories, into dynamic perception primitives. Each dynamic perception primitive contains at least data source identification, feature extraction rules, and initial association weight information.
[0036] The primitive reorganization module is configured to dynamically adjust the association rules and association weights between at least some of the dynamic perception primitives in the dynamic perception primitive library based on the real-time acquired environmental context information including rainfall, so as to form a dynamic feature combination adapted to the current environmental state, wherein the adjustment of the association weights is at least partially based on the function Wnew =W old +α×ΔC,W new is the new association weight, W old is the old association weight, ΔC is the change in environmental context, and α is the preset sensitivity coefficient;
[0037] The edge collaborative learning network consists of a lightweight federated learning framework deployed on multiple edge computing nodes. It is used to share the association rule parameters of dynamic perception primitives among multiple edge computing nodes to achieve cross-node knowledge transfer and model collaborative update;
[0038] The fault-driven dynamic task allocation module is configured to automatically trigger adjacent edge computing nodes to start high-precision data sampling mode when any edge computing node detects an abnormal dynamic feature combination that meets preset conditions, forming a local collaborative computing cluster for collaborative fault diagnosis and analysis;
[0039] The multimodal data channel optimization module is used to logically slice unstructured video stream data according to the physical structure of the underground pipeline corridor and upload only the video slices containing abnormal information to the cloud. It also performs adaptive Huffman coding compression on time series sensor data based on the periodic characteristics of pipeline pressure fluctuations to optimize data transmission efficiency.
[0040] An adaptive anomaly detection module is configured to detect anomalies, including gradual faults, in underground utility corridors in real time based on dynamically adjusted primitive association rules and weights, and output anomaly alarm information. The detection response time for gradual faults is less than or equal to five seconds.
[0041] The edge computing load balancing module is used to dynamically allocate computing tasks based on the computing load of each edge computing node in the edge collaborative learning network to improve the computing power utilization and balance of edge nodes;
[0042] A cross-modal data fusion decision module is configured to receive dynamic perception primitives and their associated weight information from different data modalities and adaptively assign fusion weights to each modality based on the context to improve the accuracy of the integrated decision;
[0043] A global robustness enhancement module, which is used to maintain the basic functional operation of the system through the edge collaborative learning network and the fault-driven dynamic task allocation module when a single edge computing node fails;
[0044] The visualization map generation module is used to generate visualization maps including heat maps of abnormal dynamic feature combinations to intuitively display fault location logic and system operation status.
[0045] Preferably, the primitive recombining module uses a sliding time window to analyze the frequency of the dynamic perception primitive combination. When the frequency of a specific abnormal dynamic perception primitive combination F abnormal Exceeds the preset threshold T f When , the adjacent edge computing nodes are triggered to start high-precision sampling mode.
[0046] Preferably, in the multimodal data channel optimization module, the video stream is logically sliced according to the physical structure partitions of the underground pipeline corridor such as valve areas and connections, and the video logical slices are bound to the spatial coordinate information of the corresponding sensor data.
[0047] Preferably, the parameter sharing process of the edge collaborative learning network adopts differential privacy encryption technology to ensure the data security of the primitive association rule parameters shared across nodes.
[0048] Preferably, it also includes a primitive conflict resolution mechanism for constructing a dynamic priority matrix based on the ant colony pheromone principle. When multiple dynamic perception primitive association rules conflict, the optimal weight distribution scheme is automatically selected based on the pheromone concentration of historical successful decision paths.
[0049] Preferably, the conflict types defined by the primitive conflict resolution mechanism include data source conflict, spatiotemporal conflict, weight conflict, confidence conflict and business logic conflict, and conflict type matching and pheromone concentration comparison are performed through a lightweight rule engine to achieve real-time response of less than fifty milliseconds.
[0050] Preferably, it also includes an energy consumption-aware task migration mechanism, which is configured to reuse existing temperature sensor data to build a thermodynamic model of the edge computing node, calculate the heat dissipation efficiency attenuation coefficient, and automatically migrate the computing task to the low-temperature area of the adjacent node when the node temperature exceeds a preset threshold.
[0051] Preferably, the energy-aware task migration mechanism gives priority to selecting adjacent nodes with copies of the same data shard for task migration to avoid data retransmission overhead.
[0052] Preferably, it also includes a cross-modal data timing alignment engine for enhancing the hardware clock synchronization accuracy, using key frame interpolation technology to compensate for transmission delays in video streams, and eliminating clock drift through sliding window mean filtering of sensor data streams, so that the timestamp alignment accuracy of multimodal data reaches within plus or minus one millisecond.
[0053] Preferably, it also includes an emergency decision topology protocol, which is used to dynamically reconstruct the data stream transmission priority according to the leakage alarm level of the underground pipeline corridor area. When the leakage level in a certain area exceeds the preset level, the video stream transmission path of the area is automatically switched to the user datagram protocol high-speed channel, and the triple sampling frequency compensation mechanism of the sensor data stream is triggered.
[0054] Example 1: The dynamic perception primitive library is responsible for abstracting multi-source data from underground pipe corridors (such as pipeline pressure, temperature and humidity, equipment vibration, current waveform, video monitoring, etc.) into dynamic perception primitives. Each dynamic perception primitive includes at least a data source identifier, feature extraction rules, and initial association weight information; in this embodiment, all sensor data are preliminarily preprocessed by the acquisition module, including steps such as signal denoising, normalization, and feature extraction. The feature extraction rules of each primitive adopt a standard algorithm based on physical principles to ensure that all data can be accurately mapped and processed according to the physical quantity category from which it originates. The primitive reorganization module dynamically adjusts the association rules and weights between primitives in the dynamic perception primitive library by acquiring environmental context information including rainfall in real time. Specifically, when the environmental parameters mutate, such as an increase in rainfall, the system will automatically adjust the association rules and weights between primitives according to the preset function model (for example: the association weight W new =W old +α×ΔC) to update the weight of the primitive, where α is the preset sensitivity coefficient, ΔC is the amount of environmental change, and W new is the new association weight, W old This dynamic adjustment process is completed in real time through edge computing nodes for the old association weights, ensuring that the system can automatically adapt to changing environmental conditions without relying on traditional fixed thresholds.
[0055] The edge collaborative learning network in this embodiment is based on a lightweight federated learning framework. Multiple edge computing nodes achieve cross-node knowledge transfer by sharing the association rule parameters of dynamic perception primitives. Specifically, each edge node trains a model with local data and transmits it to the central node in an encrypted manner. The central node then synchronizes the parameter updates to other nodes. To ensure data security, the parameter sharing process of federated learning uses differential privacy encryption technology to avoid leaking sensitive data during the sharing process. Through this collaborative learning mechanism, the system can achieve knowledge sharing between various nodes and trigger the high-precision sampling mode of adjacent nodes when local anomalies occur. When the system detects an abnormal dynamic feature combination, the fault driving module automatically triggers the adjacent edge computing nodes to start the high-precision data sampling mode, forming a local collaborative computing cluster for collaborative fault diagnosis and analysis. At this time, the dynamic task allocation module based on the sliding time window will adjust the allocation strategy of computing tasks according to the computing power load of the current node to avoid overloading a single node and ensure the efficiency and real-time performance of data processing. The task allocation process uses an optimization algorithm (such as a polling algorithm) to balance the computing load of each node, ensuring that the system can still maintain stable operation in the event of a local failure. The function of this module is to optimize the transmission efficiency of unstructured video streams. In this embodiment, the video stream data will be logically sliced according to the physical structure of the pipeline corridor, and each video slice will be bound to the pressure sensor data at the corresponding position through spatial coordinate information. In order to reduce the amount of data transmission, the system only uploads the video slices containing abnormal information to the cloud; this method effectively reduces bandwidth occupancy and avoids the loss of details caused by full transmission in traditional solutions. In addition, based on the periodic characteristics of pipeline pressure fluctuations, the system also uses an adaptive Huffman coding algorithm to compress the time series sensor data to improve data transmission efficiency.
[0056] The cross-modal data timing alignment engine is used to improve the timing accuracy of multimodal data. In this embodiment, the timestamps of the video stream and the sensor data stream are aligned using a sliding window mean filtering technique. This technique avoids the time difference caused by hardware clock drift and ensures that the time synchronization accuracy of different data sources is within plus or minus one millisecond. In addition, for the video stream, the system uses key frame interpolation technology to compensate for the transmission delay, further improving the synchronization between the video stream and the sensor data. This embodiment also includes an energy-aware task migration mechanism. By reusing existing temperature sensor data, a thermodynamic model of each edge computing node is constructed, and the heat dissipation efficiency attenuation coefficient is calculated. When the temperature of the node exceeds a preset threshold, the task will automatically migrate to an adjacent node with a lower temperature. This mechanism effectively avoids performance degradation caused by device overheating, and by selecting adjacent nodes with copies of the same data shard for task migration, the overhead of data retransmission is reduced, thereby improving the energy efficiency of the system. At the same time, in order to enhance the system's fault diagnosis capabilities, the embodiment also includes a visual map generation module, which intuitively displays the fault location logic and system operation status by generating a heat map containing a combination of abnormal dynamic features. Users can use this map to view the monitoring status and fault distribution of each area in real time, so as to quickly take countermeasures.
[0057] Example 2: This example combines Figures 1 to 2 , the implementation of the real-time processing system of underground pipeline corridor data based on edge computing is explained. Figure 1 As shown, Figure 1 It is mainly divided into three levels: dynamic perception layer, edge computing layer and decision output layer. In the dynamic perception layer, multi-source heterogeneous data are input into the dynamic perception primitive library. These data include pipeline pressure, video surveillance texture and equipment vibration frequency. The dynamic perception primitive library abstracts and integrates these multi-source data to form dynamic perception primitives. Subsequently, these primitive data are passed to the edge computing layer. In the edge computing layer, the primitive reorganization module is responsible for processing the primitive data from the dynamic perception layer to further analyze and process data such as pipeline pressure, video surveillance texture and equipment vibration frequency. The edge collaborative learning network is used to realize information sharing and collaborative work between edge nodes. The fault-driven task allocation module is responsible for dynamically allocating computing tasks when an anomaly is detected to achieve rapid response and processing. Finally, in the decision output layer, the adaptive anomaly detection module uses the data processed by the aforementioned layers to detect anomalies and generate corresponding alarm information. The visual map generation module displays the detection results in a graphical manner to facilitate users to understand and analyze the operating status of the system.
[0058] like Figure 2As shown in the figure, the process starts from the edge collaborative network at the top and first enters the judgment link of lightweight federated learning. According to the execution of federated learning, the process is divided into two branches: one branch is cross-node knowledge transfer, which includes the step of sharing model parameters; the other branch is the self-organizing diagnosis cluster. After a local anomaly is triggered, this branch executes the sampling instructions of the adjacent nodes, followed by high-precision sampling, and finally returns to the self-organizing diagnosis cluster through the step of enhancing the data, forming a closed-loop process.
[0059] Example 3: In this embodiment, in terms of the construction of the dynamic perception primitive library and the concretization of feature extraction rules, for pipeline pressure data, feature extraction rules may include calculating its mean, variance, peak-to-peak value and pressure change rate within a specific time window. These statistical features can reflect the stability and abnormal fluctuation of the pipeline; for temperature and humidity data, feature extraction rules may include obtaining the current value and calculating its short-term change trend and deviation from the safety threshold of the corridor design in combination with historical data; for equipment vibration frequency data, feature extraction rules may use fast Fourier transform to process the original vibration signal, extract the amplitude and energy distribution of the main vibration frequency, and judge whether the equipment operation status is abnormal; for current waveform data, feature extraction rules may include obtaining the current value and calculating the ... It can include analyzing the effective value, harmonic content and transient characteristics of the current signal, which are helpful in identifying potential faults of electrical equipment; for texture changes in video surveillance, the feature extraction rules can first pre-process the video frames, such as grayscale and noise reduction, and then use algorithms such as local binary patterns or gray-level co-occurrence matrices to extract the texture feature descriptors of the image, and judge whether there are abnormal textures such as water seepage and deformation by comparing the significance of changes in texture feature descriptors between consecutive frames; for motion trajectories, the moving targets in the monitoring screen can be identified and their motion trajectories can be tracked by combining background modeling with foreground extraction, and the length, direction, speed and other features of the trajectory can be extracted to analyze whether there are abnormal intrusions or object displacements. The initial association weight information of each dynamic perception primitive can be preliminarily set based on the statistical analysis of historical fault data of underground pipeline corridors and expert experience. For example, in areas prone to leakage, the initial association weight between pipeline pressure and video surveillance water seepage texture features can be set relatively high, and then dynamically adjusted by the primitive reconstruction module. These feature extraction rules and initial weight settings fully consider the requirements for data accuracy and real-time performance in industrial data processing, ensuring the validity and reliability of perception layer data.
[0060] About the correlation weight adjustment formula W in the primitive reorganization module new =W oldThe core variables and parameter details of +α×ΔC include the specific calculation method of the environmental context change. Taking rainfall as an example, the rainfall data per unit time (e.g., hourly) can be normalized and compared with the average rainfall at the previous moment or over a period of time. The difference or the absolute value of the rate of change is calculated to obtain a quantitative rainfall change indicator. If other environmental factors such as geological subsidence monitoring data need to be considered, their changes can be similarly quantified and weighted averaged or integrated into a unified environmental context change through a specific function. This integration method is based on prior knowledge of the impact of different environmental factors on pipeline corridor safety. As for the preset sensitivity coefficient, it is not a fixed global parameter, but can be set to different numerical ranges for different types of dynamic perception primitives or different types of environmental context changes. For example, for the adjustment of the correlation weight of pipeline pressure and video seepage texture due to rainfall changes, the preset sensitivity coefficient may be set larger to reflect its strong correlation, while for the adjustment of the correlation weight of equipment vibration due to temperature changes, the preset sensitivity coefficient may be relatively small; the specific values of these coefficients can be obtained through offline model training, using historical environmental data, sensor data and confirmed fault events for parameter optimization, or they can be set by experienced engineers according to the actual working conditions of the pipeline corridor in the early stage of system operation, and dynamically fine-tuned through the learning mechanism in subsequent operation to ensure the sensitivity and accuracy of the weight adjustment.
[0061] In terms of the lightweight implementation and parameter sharing mechanism of edge collaborative learning networks, the lightweight federated learning framework is mainly reflected in the model structure running on the edge computing nodes. For example, simplified neural network models or gradient boosting decision trees with low computational overhead are used. When uploading model parameters, technologies such as gradient sparsification and quantization compression are used to reduce communication overhead. The shared association rule parameters here can be specific to the model parameters used to judge the importance of different primitive combination patterns in the primitive reconstruction module, or the thresholds or model parameters used to distinguish between normal and abnormal states in the adaptive anomaly detection module. The specific mechanism of parameter sharing is that after each edge node uses its collected data to train the model or update the parameters locally, only the model updates or key parameters, rather than the original data, are sent via an encrypted channel to a logical coordination server (this server can also be temporarily served by an edge node, or a decentralized parameter aggregation method such as the gossip protocol is used). The coordination server securely aggregates the parameter updates from different nodes (for example, using homomorphic encryption aggregation or secure multi-party computation) and redistributes the aggregated global model parameters or updated rule parameters to each edge node to achieve knowledge sharing and model collaborative evolution.
[0062] In terms of the triggering and collaborative logic of the fault-driven dynamic task allocation module, the specific definition of the preset conditions and abnormal dynamic feature combinations, for example, when the value of the pipeline pressure primitive is lower than 50% of the normal threshold for three consecutive times, and at the same time the video monitoring texture change primitive detects a specific texture pattern related to leakage (such as the continuous expansion of the water stain area), then this combination is regarded as an abnormal dynamic feature combination, which meets the triggering preset conditions; once triggered, the high-precision data sampling mode can be specifically manifested as increasing the sampling frequency of the sensor related to the anomaly to three to five times the usual frequency, and at the same time instructing the video surveillance system to record a specific area with higher resolution or enable enhanced detection methods such as infrared imaging; the collaborative diagnosis and analysis of faults between adjacent nodes, its workflow can be: first, the node that triggers the anomaly sends a message to its edge coordination node A collaborative request is sent to several predefined neighboring nodes in the network (these neighboring relationships can be set based on the proximity of physical locations or the convenience of data transmission). The request contains a brief description of the abnormal feature combination and a timestamp. The neighboring node that receives the request immediately starts high-precision sampling and sends the collected relevant data (primitive information after preliminary processing and feature extraction) back to the requesting node or the designated aggregation node. Then, these nodes can jointly execute a preset distributed diagnosis algorithm. For example, by comparing the data differences and correlations of multiple nodes under the influence of the same fault event, they can cross-verify the authenticity of the fault, locate the source of the fault, and evaluate the impact range of the fault, forming a local, self-organizing diagnosis cluster to jointly complete the rapid response and analysis of the fault, thereby improving the accuracy and robustness of the diagnosis.
[0063] The logical slicing strategy of the video stream in the multimodal data channel optimization module can divide the underground corridor into several logical segments based on the three-dimensional geographic information system (GIS) data of the underground corridor. Each logical segment corresponds to a video monitoring range. In particular, key structural points such as valve wells, pipeline interfaces, and bends, or areas with a high incidence of historical failures, can be divided into more detailed slicing units. The basis for judging whether the video slice contains abnormal information can be running lightweight real-time video analysis algorithms on edge nodes, such as motion change detection based on optical flow method, simple target recognition based on pre-trained models (such as identifying abnormal water accumulation, human intrusion, etc.). When the detected changes or targets exceed the preset normal range, the video slice is considered to contain abnormal information, and it and the associated sensor data are uploaded; for the adaptive Huffman coding compression of time-series sensor data, its adaptability is reflected in the fact that the system will dynamically adjust the structure of the Huffman coding tree according to the distribution characteristics and change patterns of sensor values such as recent pipeline pressure, so that data values or change patterns with higher frequency can obtain shorter coding lengths, thereby improving the compression rate while ensuring lossless or near-lossless information. This dynamic adjustment can be performed periodically or triggered when the statistical characteristics of the data change significantly.
[0064] The adaptive anomaly detection module uses a combination of time series analysis and the changing trends of dynamic primitive weights to detect gradual faults. Specifically, the system continuously tracks the historical trends of dynamic sensing primitives related to pipeline corrosion and material aging (e.g., pipe wall thickness sensor data at specific locations, micro-vibration characteristics, localized temperature anomalies, etc.) and predicts future short-term changes in these primitives using methods such as exponential smoothing, ARIMA models, or trend prediction algorithms in machine learning. Simultaneously, the primitive reorganization module dynamically adjusts the weights of these gradual fault-related primitives based on environmental changes (e.g., humidity, soil pH, and other factors that may accelerate corrosion), making the system more sensitive to these potential gradual characteristics. When the deviation between the actual monitored primitive change trend and the predicted trend continues to accumulate and exceeds a dynamically adjusted threshold, or when multiple weakly correlated gradual characteristics, after dynamically enhanced weights, form a statistically significant abnormal combination pattern, the system determines that a gradual fault has occurred and can promptly issue an alarm before it escalates into a sudden, severe failure, thereby achieving early identification of gradual faults and a response within five seconds.
[0065] In the dynamic allocation strategy of the edge computing load balancing module, the measurement of computing load conditions can comprehensively consider the central processing unit utilization rate, memory remaining amount, network exit bandwidth occupancy rate and the length of the task queue currently being processed of the edge node; the specific strategy for dynamically allocating computing tasks can adopt a scheduling algorithm based on reinforcement learning. This algorithm continuously learns the historical load status, task processing completion time and resource consumption characteristics of each edge node to predict which node to assign new tasks to in order to obtain the best overall performance (such as the lowest average delay or the highest throughput). When the load of a node is too high, the algorithm will automatically direct new computing tasks (such as an anomaly analysis task of a video slice, or a complex feature extraction task of sensor data) to nodes with lower load and corresponding processing capabilities, thereby avoiding overload of a single node and improving the resource utilization and processing balance of the entire edge computing network.
[0066] The specific definition of context in the cross-modal data fusion decision module can include current environmental context information (such as the aforementioned rainfall and geological activity level), the current operating status of the system (such as whether it is in emergency response mode for a specific fault), and quality assessment information of the input data (for example, the image clarity score of a video sensor, or the data confidence of a pressure sensor); the adaptive allocation of fusion weights for each modal data can be based on a preset rule base, which is formulated by domain experts based on the contribution of each data modality to specific decisions (such as leakage judgment and structural health assessment) under different contextual conditions. For example, in heavy rain, the fusion weight of water-related texture features in video surveillance will be significantly increased; alternatively, an online learning model (such as a neural network based on an attention mechanism) can be used. This model can automatically learn and adjust the fusion weights of each modal data based on real-time context information and the correlation between each modal data and the final decision result, thereby making the fusion decision more intelligent and contextualized.
[0067] The specific maintenance mechanism of the global robustness enhancement module when a single edge computing node fails can be expressed as follows: first, the faulty node is discovered in time through the heartbeat detection mechanism of the edge collaborative learning network or the communication status monitoring between neighboring nodes; once a node failure is confirmed, the fault-driven dynamic task allocation module will immediately reallocate the computing tasks originally undertaken by it to one or more healthy adjacent nodes based on the type and priority of the task, as well as the current computing power load and data availability of other nodes (especially whether there are shard copies of relevant data); at the same time, the edge collaborative learning network can use its saved global model parameters or rules to help the node that newly takes over the task quickly restore or initialize the relevant computing model, ensuring the continuity of monitoring and data processing tasks, thereby maintaining the system's core monitoring functions and data processing capabilities without human intervention.
[0068] Specifically, the heat map generation logic in the visualization map generation module uses a two-dimensional or three-dimensional electronic map of the underground pipeline corridor as the base map. When the adaptive anomaly detection module outputs the abnormal dynamic feature combination and its location information, the system will superimpose marks of different colors or transparencies at the corresponding positions on the electronic map based on the severity level of the abnormal combination (the level can be obtained by a comprehensive assessment of the amplitude, duration, and primitive weight of the abnormal feature) and the frequency of occurrence. For example, areas with high severity levels or high occurrence frequencies are displayed in dark red and opaque, while areas with lower severity levels or lower frequencies are displayed in light yellow and translucent, thus forming a heat map that intuitively reflects the risk status of each area in the pipeline corridor, making it easier for operation and maintenance personnel to quickly locate high-risk areas and understand the distribution of faults.
[0069] In the primitive conflict resolution mechanism, the specific implementation of the dynamic priority matrix of the ant colony pheromone principle can be: when the association rules of multiple dynamic perception primitives (for example, rule A indicates to increase the weight of a certain video feature under certain conditions, while rule B indicates to reduce the weight of the feature at the same time) produce a logical conflict, the system regards each rule as a potential path in the ant colony algorithm, and the historical decision results (for example, a certain rule combination successfully predicted a fault or avoided a false alarm) will leave pheromones on this path; the dynamic priority matrix records the historical pheromone concentrations of different rules or rule combinations when solving similar conflict scenarios. When a new conflict occurs, the system will select the current pheromone. The rule or rule combination with the highest concentration is used as the optimal weight distribution or decision-making solution; pheromones evaporate over time, and successful decisions enhance them, enabling the system to dynamically adapt to environmental changes and accumulate experience; the lightweight rule engine predefines matching patterns for the five conflict types mentioned above. For example, a data source conflict may be defined as the large difference in similar data (such as two pressure sensor readings) from different sensors at the same monitoring point at the same time. The rule engine quickly matches the current primitive combination to see if it meets these predefined conflict patterns. Once a match is successful, the pheromone-based priority decision-making mechanism is called for processing to achieve the purpose of rapid response.
[0070] In the energy-aware task migration mechanism, the thermodynamic model of the edge computing node can be a simplified empirical model. For example, by recording the temperature rise rate and stable temperature of the node under different load conditions, the relationship function between temperature, computing load, and ambient temperature can be fitted; the heat dissipation efficiency attenuation coefficient can be dynamically evaluated based on the node fan speed, the dust accumulation inside the chassis (which can be obtained through indirect inference or regular maintenance records), and the deviation between the model-predicted temperature and the actual temperature; when the node core temperature predicted by the model (or the expected temperature adjusted according to the heat dissipation efficiency attenuation) exceeds the preset safety threshold (for example, 70 degrees Celsius), the system will trigger task migration, and prioritize migrating computing-intensive tasks to adjacent nodes with lower current temperatures and good network connections. If the target node stores a sharded copy of the data required for the task (this can be obtained through data distribution information in the edge collaborative network), the task is restored directly on the node, avoiding the overhead of data retransmission, thereby extending the equipment life and reducing energy consumption while ensuring stable system operation.
[0071] To enhance hardware clock synchronization accuracy, the cross-modal data timing alignment engine deploys lightweight Network Time Protocol (NTP) clients between edge nodes and regularly synchronizes with a high-precision time server. It also records the deviation of each node from the standard time and compensates for it when timestamping the data. For video stream keyframe interpolation, the receiver can use the timestamps and sequence numbers of the arriving keyframes to estimate the occurrence times of lost or delayed intermediate frames using linear or motion vector-based nonlinear interpolation to restore the original frame sequence timing as much as possible. When applying sliding window mean filtering to sensor data streams to avoid clock drift, the window size should be determined based on the specific sensor type and expected drift characteristics. For example, a larger window can be used for relatively slowly changing temperature data, while a smaller window can be used for rapidly changing pressure data. By averaging the timestamps of the data within the window, short-term jitter and long-term drift caused by clock hardware instability are smoothed out, ensuring millisecond-level time base consistency when fusion data from different sensors and videos.
[0072] In the emergency decision-making topology protocol, the leakage alarm level can be divided into multiple levels according to the magnitude of the leakage, the speed of development, and the potential degree of harm to the corridor structure and internal facilities, such as level one (minor leakage, need to be paid attention to), level two (moderate leakage, need to be dealt with as soon as possible), and level three (serious leakage, need immediate emergency response); when the leakage alarm level of a certain area reaches the preset highest level (such as level three), the specific operation of dynamically reconstructing the data stream transmission priority can be that the system sets the transmission priority of all video stream data in the area to the highest by controlling the network interface configuration of the edge node or calling the underlying network management interface, and may allocate a dedicated virtual LAN or bandwidth channel for it, and at the same time change its transmission protocol from guaranteed The reliable transmission control protocol is switched to the user datagram protocol high-speed channel with lower latency, sacrificing a certain degree of transmission reliability in exchange for extreme real-time performance, ensuring that the emergency command center can obtain on-site images in the first time; the triple sampling frequency compensation mechanism is that in this emergency, the key sensors such as pressure and flow in the instruction-related area are collected and reported at three times the usual frequency, providing more intensive data support for more accurate fault analysis and decision-making. At this time, the edge computing power load balancing module and the fault-driven dynamic task allocation module will work together to ensure that the increased data processing requirements can be effectively carried and avoid the system crash due to local data peaks. These are all extended implementation methods known to ordinary technicians in this field.
[0073] Example 4: In the application scenario of underground integrated pipeline corridors in key urban flood control areas, in order to achieve dynamic response to multi-source abnormal characteristics and collaborative processing of edge nodes under sudden events of heavy rainfall, this embodiment carries out systematic deployment and detailed technical implementation around the leakage monitoring and linkage diagnosis tasks within the pipeline corridor section. During the system startup phase, the edge node A located in the cross-intersection area of the pipeline corridor first completes the initial registration of the sensor, data standardization and feature extraction through its internal dynamic perception primitive library. Taking the locally installed pressure sensor as an example, the system collects the pressure change value in the pipeline at a frequency of one frame per second, and calculates its short-term variance and change rate characteristics within the sliding time window. Synchronously, the video frames collected by the camera are grayscale processed and texture feature encoded to extract local binary pattern distribution features, and the corresponding spatial coordinates are recorded at the same time. All perception results are structured and encapsulated into dynamic perception primitive units and assigned initial weights. At the same time, in order to cope with the increased probability of pipeline corridor leakage under sudden heavy rainfall conditions, the system obtains hourly rainfall data through the meteorological access module, and performs difference analysis with the historical baseline for the same period, and inputs the change rate as the environmental context change into the primitive reconstruction module. In this embodiment, the change is calculated by normalizing the hourly cumulative rainfall and subtracting the three-hour sliding average. By setting the sensitivity coefficient to 30% and combining the current context change value of 0.8, the adjustment of the association weights between dynamic perception primitives is completed; the calculation logic aims to enhance the connection strength between the characteristic primitives of water-related texture variation and the pressure fluctuation primitives, thereby improving the system's perception of potential leakage conditions.
[0074] During operation, if the system detects pressure variance exceeding the warning threshold for three consecutive time periods and simultaneously experiences a sudden change in the similarity of the water stain extension descriptor in the texture features (i.e., a drop of more than 50% in similarity between consecutive frames), the dynamic task scheduling module determines that the preset anomaly combination conditions have been met and initiates a high-precision sampling process. The sampling strategy is reflected in the following three steps: first, the pressure sensor sampling frequency is increased from one frame per second to five frames per second; second, the camera enters enhanced mode, enabling infrared and high-resolution dual-channel imaging; and third, simultaneous acquisition of humidity and vibration sensors is initiated to enhance the dimensional coverage of fault features. Furthermore, to achieve cross-node collaborative diagnosis capabilities, node A, which triggers the anomaly, broadcasts a diagnostic request to its physically adjacent nodes B and C via the edge collaborative learning network. The broadcast content includes the trigger time, primitive type, spatial coordinates, and feature summary. The recipient triggers the corresponding data collection process based on the request and returns a structured primitive vector to the diagnostic center node. In this embodiment, node A temporarily acts as the coordination center, integrates the primitives collected by multiple nodes, and then performs pattern matching analysis based on the preset feature aggregation rules. The specific strategy is that within the same time period, if the primitives of the same type at multiple nodes all show a trend of deviation in the same direction (such as a synchronous decrease in pressure and a sharp decrease in texture similarity), the system determines that the regional leakage risk is significant, and then triggers the emergency sampling topology reconstruction mechanism. In order to avoid the impact of data transmission bottlenecks on real-time response under abnormal conditions, the system implements logical slicing of the video stream, retaining only the frame segments that coincide with the current abnormal feature coordinates. The video logical slicing strategy is constructed based on the results of the three-dimensional modeling of the underground pipeline corridor. The spatial boundary of each slice is based on physical components (such as valves, corners, interfaces), and the spatial resolution accuracy is not less than one meter. Each video segment is bound to the corresponding spatial coordinates, timestamp and pressure sensor ID when it is generated, so as to achieve alignment of video and sensor data in time and space dimensions.
[0075] At the decision fusion layer, in order to avoid the error in the contribution of different modal data from amplifying the error in abnormal judgment, the system performs context-aware adjustment on the fusion weight. In this embodiment, when the meteorological module is marked as a heavy rainfall state, the system automatically increases the fusion weight of the video texture primitive and the humidity primitive to twice the original value, and compresses the vibration primitive weight to half of the original value to reflect the higher credibility of the leakage-related signal in this scenario. This adjustment is based on the context-feature correlation matrix constructed during the system's offline training for rapid retrieval to avoid additional delays caused by online calculations. In order to ensure the continuous operation capability of the entire system in the event of node-level failures, this embodiment also deploys a node-level computing power perception mechanism. Specifically, the system periodically collects the core temperature and load factor of the node equipment, and calculates the heat dissipation efficiency through a simple thermal function model. When it detects that the core temperature of node A is higher than 60 degrees Celsius for five consecutive minutes and the CPU load exceeds 80%, the system automatically migrates non-critical tasks on the node to node C; the migration task gives priority to the target node with a copy of the original data to avoid migration failure due to data missing. In the judgment of the data copy, the system uses the global data distribution map maintained in the edge collaborative network for rapid comparison to ensure that the target node can seamlessly continue the task status. At the same time, in order to achieve precise alignment of timestamps across modal data and ensure the accuracy of multimodal fusion, the system deploys a lightweight time synchronization module in the edge node and eliminates sensor clock jitter through sliding window mean filtering technology. In terms of video streaming, the receiving end uses a key frame interpolation algorithm to compensate for frame loss caused by network fluctuations, maintain the continuity of key image sequences, ensure that the timing alignment error of video and pressure data is controlled within plus or minus one millisecond, and improve the accuracy of event tracing and analysis. These are all extended implementation methods that can be known to ordinary technicians in this field.
[0076] Example 5: In this example, to achieve real-time adjustment of the association weights between dynamic perception primitives, the system constructs a parameter adjustment mechanism driven by environmental context information. Environmental context information refers to a set of external environmental parameters that reflect the operating environment of the underground pipeline corridor. Specifically, it may include the total rainfall per unit time, the amplitude of temperature fluctuation, the surface subsidence rate, and the amplitude of soil moisture change. To quantify the amount of environmental context change, the system runs a change detection module on each edge node to compare the difference between the values of various environmental parameters in the current time period and the historical statistical baseline, where the historical baseline is obtained through the sliding average method. For example, for the total rainfall, the system calculates the total rainfall in the current hour and subtracts it from the sliding average of the past three hours. The resulting difference is normalized and used as the rainfall change indicator at the current time point. The change indicators of multiple environmental parameters are weighted and integrated, and the weights are preset by domain experts based on the degree of influence of each parameter on the pipeline corridor leakage risk, ultimately forming a comprehensive environmental context change.
[0077] To guide the update of association weights, the system presets several types of sensitivity coefficients during the deployment phase based on the sensitivity of the correlation between different types of data sources. The sensitivity coefficient reflects the degree of response of the coupling relationship between a certain type of data sources to environmental changes. Its setting is based on the covariation trend analysis in historical fault data and the response sensitivity verification during the trial operation of the edge node. For example, in actual deployment, it was found that rainfall changes have a strong impact on the synergistic relationship between pressure data and video texture features. Therefore, the sensitivity coefficient of this type of data combination is set to a higher value; while temperature and humidity have a weaker effect on the coupling between equipment vibrations, and their sensitivity coefficient is correspondingly set to a lower level. The system supports fine-tuning of the sensitivity coefficient during operation by dynamically optimizing the parameter values based on feedback performance indicators (such as false alarm rate and warning accuracy).
[0078] When updating weights, the system multiplies the current change in the integrated environmental context by the preset sensitivity coefficient and adds the result to the original weight of the corresponding data combination to obtain the latest weight value at the current time. This update operation is performed independently within the primitive reassembly module of each edge computing node. The update frequency is centrally managed by the system's task scheduler, typically set to once every five minutes, to ensure a balance between real-time weight adjustment and system stability.
[0079] In addition to weight updates, the system also builds trigger judgment logic for abnormal dynamic feature combinations, and uses this to drive collaborative computing between edge nodes. During system operation, each edge node maintains a set of registered dynamic perception primitive combination templates. This template is formed based on pattern mining of historical fault events and contains typical fault symptom data combinations, such as the simultaneous occurrence of continuous pressure fluctuation anomalies and texture feature variations. During the actual monitoring process, the node matches the real-time primitive sequence with the template. When the matching degree exceeds the preset threshold and the number of consecutive occurrences exceeds a specific frequency threshold, the system determines that an abnormal dynamic feature combination exists, thereby triggering the high-precision sampling process.
[0080] The sampling process includes: first, temporarily increasing the sampling frequency of sensors in the current area by two to three times from the base frequency; second, instructing the video acquisition module to enable enhanced mode (such as high resolution or infrared assistance); and third, starting the synchronous acquisition of associated sensor groups (such as pressure, humidity, and vibration) and generating a data packet bound in time and space. To ensure the rational use of system resources, the node will evaluate the current computing load and temperature status before starting sampling. If the current node finds that the resources are tight, it will prioritize sending the task to a neighboring node with a copy of the relevant data for execution.
[0081] In addition, this embodiment specifically describes the context-aware weight adjustment mechanism of the fusion decision module. This module will continuously receive contextual information from the environmental perception module, and in combination with the confidence assessment results of the primitive sources, update the weight values of each modal data during fusion in real time. The weight adjustment rules are derived from the expert knowledge base before the system is deployed, covering the decision priority settings under different environmental conditions, such as increasing the data fusion weight related to humidity and video texture features in heavy rainfall scenes, and enhancing the weight distribution related to equipment vibration features in high temperature scenes. This mechanism can take effect quickly through the rule matching engine, and also supports further weight fine-tuning through online learning models.
[0082] In terms of module collaboration logic, each functional module interacts with data and control flows through a unified data channel protocol. For example, after completing the update of the associated weights, the primitive reorganization module pushes the results to the anomaly detection module. Once the latter detects a potential abnormal combination, it calls the task scheduling module to trigger the sampling mechanism. If the sampling task exceeds the capacity of the node, the scheduling module interacts with the edge collaborative learning network interface, which is responsible for selecting a node with sufficient computing resources and complete data copies as the task migration target. After the task migration is completed, its collection results are synchronously transmitted back to the original node and the central node for subsequent visualization and model optimization. The final results are then presented to the operation interface by the visualization map generation module to achieve closed-loop control. These are all extended implementation methods known to ordinary technicians in this field.
[0083] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.
[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention.
Claims
1. A real-time processing system for underground pipeline corridor data based on edge computing, characterized in that: The system comprises: A dynamic perception primitive library is used to abstract multi-source heterogeneous data from underground pipeline corridors, including pipeline pressure, temperature and humidity, equipment vibration frequency, current waveform, and video surveillance texture changes and motion trajectories, into dynamic perception primitives. Each dynamic perception primitive contains at least data source identification, feature extraction rules, and initial association weight information. The primitive reorganization module is configured to dynamically adjust the association rules and association weights between at least some of the dynamic perception primitives in the dynamic perception primitive library based on the real-time acquired environmental context information including rainfall, so as to form a dynamic feature combination adapted to the current environmental state, wherein the adjustment of the association weights is at least partially based on the function W new =W old +α×ΔC,W new is the new association weight, W old is the old association weight, ΔC is the change in environmental context, and α is the preset sensitivity coefficient; The edge collaborative learning network consists of a lightweight federated learning framework deployed on multiple edge computing nodes, which is used to share the association rule parameters of dynamic perception primitives among multiple edge computing nodes. The fault-driven dynamic task allocation module is configured to automatically trigger adjacent edge computing nodes to start high-precision data sampling mode when any edge computing node detects an abnormal dynamic feature combination that meets preset conditions, forming a local collaborative computing cluster for collaborative fault diagnosis and analysis. The multimodal data channel optimization module is used to logically slice unstructured video stream data according to the physical structure of the underground pipeline corridor and upload only the video slices containing abnormal information to the cloud. Based on the periodic characteristics of pipeline pressure fluctuations, the module performs adaptive Huffman coding compression on time series sensor data to optimize data transmission efficiency. The adaptive anomaly detection module is configured to perform real-time detection of abnormal situations, including gradual faults, occurring in underground pipelines based on dynamically adjusted primitive association rules and weights, and output abnormal alarm information, wherein the detection response time for gradual faults is less than or equal to five seconds; the edge computing power load balancing module is used to dynamically allocate computing tasks according to the computing power load of each edge computing node in the edge collaborative learning network; the cross-modal data fusion decision module is configured to receive dynamic perception primitives and their associated weight information from different data modalities, and adaptively allocate the fusion weights of each modal data based on the context; the global robustness enhancement module is used to maintain the basic functional operation of the system through the edge collaborative learning network and the fault-driven dynamic task allocation module when a single edge computing node fails; the visual map generation module is used to generate a visual map including a heat map containing abnormal dynamic feature combinations to intuitively display the fault location logic and system operation status.
2. The real-time processing system for underground pipeline gallery data based on edge computing according to claim 1 is characterized in that: The primitive reorganization module uses a sliding time window to analyze the frequency of dynamic perception primitive combinations. When the frequency of a specific abnormal dynamic perception primitive combination F abnormal Exceeds the preset threshold T f When , the adjacent edge computing nodes are triggered to start high-precision sampling mode.
3. The real-time processing system for underground pipeline gallery data based on edge computing according to claim 1 is characterized in that: In the multimodal data channel optimization module, the video stream is logically sliced according to the physical structure partitions of the valve area and the underground pipeline corridor at the connection, and the video logical slices are bound to the spatial coordinate information of the corresponding sensor data.
4. The real-time processing system for underground pipeline gallery data based on edge computing according to claim 1 is characterized in that: The parameter sharing process of the edge collaborative learning network adopts differential privacy encryption technology.
5. The real-time processing system for underground pipeline gallery data based on edge computing according to claim 1 is characterized in that: It also includes a primitive conflict resolution mechanism for constructing a dynamic priority matrix based on the ant colony pheromone principle. When multiple dynamic perception primitive association rules conflict, the optimal weight distribution scheme is automatically selected based on the pheromone concentration of historical successful decision paths.
6. The real-time processing system for underground pipeline gallery data based on edge computing according to claim 1 is characterized in that: It also includes an energy-aware task migration mechanism, which is configured to reuse existing temperature sensor data to build a thermodynamic model of the edge computing node, calculate the heat dissipation efficiency attenuation coefficient, and automatically migrate the computing task to the low-temperature area of the adjacent node when the node temperature exceeds the preset threshold.
7. The real-time processing system for underground pipeline gallery data based on edge computing according to claim 6 is characterized in that: The energy-aware task migration mechanism prioritizes adjacent nodes with copies of the same data shard for task migration.
8. The real-time processing system for underground pipeline gallery data based on edge computing according to claim 1 is characterized in that: It also includes a cross-modal data timing alignment engine to enhance the accuracy of hardware clock synchronization, use key frame interpolation technology to compensate for transmission delays in video streams, and eliminate clock drift in sensor data streams through sliding window mean filtering, so that the timestamp alignment accuracy of multimodal data can reach within plus or minus one millisecond.
9. The real-time processing system for underground pipeline gallery data based on edge computing according to claim 1 is characterized in that: It also includes an emergency decision-making topology protocol, which is used to dynamically reconstruct the data stream transmission priority based on the leakage alarm level of the underground pipeline corridor area. When the leakage level in a certain area exceeds the preset level, the video stream transmission path in that area is automatically switched to the User Datagram Protocol high-speed channel, and the triple sampling frequency compensation mechanism of the sensor data stream is triggered.
Citation Information
Cited By
LNG receiving station data co-processing system and method based on edge calculation
CN120751294A
LNG receiving station data collaborative processing system and method based on edge computing
CN120751294B
Multimedia unified management and control method and system based on edge computing nodes
CN120785712A
Industrial internet-of-things monitoring system for normal slag heat exchange process of inferior oil device
CN121028722A
Integrated geographic information acquisition method and system
CN121117941A
