A gas leakage intelligent detection and operation and maintenance early warning method and system based on an internet of things

Abstract

The application discloses a kind of based on Internet of Things gas leakage intelligent detection and operation and maintenance early warning method and system, belong to Internet of Things technical field, including through the multi-dimensional sensor cluster of deployment and synchronous edge gateway, to the synchronous perception and injection of multi-source asynchronous Internet of Things data stream, the present application constructs based on Internet of Things whole link intelligent detection and operation and maintenance early warning system, significantly improve the safety monitoring precision of gas pipe network, the intelligent level of risk research and judge and the timeliness and accuracy of operation and maintenance disposal, substantially reduce the probability of gas accident occurrence;By constructing dynamic graph including physical topology connection and fluid dynamics influence side, and combining time series graph neural network, the system can deeply learn the evolution law of node characteristics and the interaction mode between nodes under normal working conditions, real-time dynamic extraction of pre-leakage characteristics in online operation stage, acutely identify subtle abnormal changes such as pressure waveform, voiceprint spectrum, significantly enhance the sensitivity to early micro-leakage.

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things (IoT) technology, specifically referring to an IoT-based intelligent detection and maintenance early warning method and system for gas leaks. Background Technology

[0002] With the development of IoT technology, solutions have emerged that deploy multiple sensors such as pressure, flow, and vibration sensors on pipelines for multi-parameter monitoring to improve detection reliability. However, these solutions mostly remain at the level of simple data collection and independent threshold judgment, failing to effectively mine the deep spatiotemporal correlation information between multi-source asynchronous data, and unable to identify subtle abnormal precursors from the perspective of the overall system health status. At the same time, existing detection systems are severely disconnected from subsequent operation and maintenance processes. Alarm, dispatch, on-site handling, and result verification processes rely on manual coordination and experience-based judgment, lacking intelligent decision support and automated closed-loop verification, making it difficult to achieve rapid, accurate, and efficient emergency response.

[0003] However, existing intelligent detection and maintenance early warning methods for gas leaks still have certain shortcomings. Current technologies rely on threshold alarm stages of gas concentration sensors, which is essentially a passive response mode that treats existing problems. It can only issue alarms after a physical leak occurs and cannot warn of potential risks before a leak occurs. This isolated detection method is susceptible to environmental interference, has a high false alarm rate, and only provides isolated location information after an alarm, failing to pinpoint the root cause of the risk. There is a serious disconnect between alarms and maintenance, with the entire process from alarm to investigation and repair relying on manual experience and scheduling, resulting in low efficiency. Each link forms an information silo, lacking an intelligent closed loop from risk perception to handling verification, and failing to achieve proactive, accurate, and efficient management of gas system safety. Therefore, this paper proposes an intelligent detection and maintenance early warning method and system for gas leaks based on the Internet of Things. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for intelligent detection and early warning of gas leaks based on the Internet of Things, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method and system for intelligent detection and early warning of gas leaks based on the Internet of Things, comprising the following steps: S1. By deploying a multi-dimensional sensor cluster and synchronous edge gateway, synchronous sensing and injection of multi-source asynchronous IoT data streams are performed; S2. Dynamically extract pre-leakage features based on synchronous multi-source time-series data; S3. Based on the extracted pre-leakage features, risk tracing and vulnerability assessment are performed to locate initial vulnerability points and calculate risk entropy values; S4. Based on the output risk location and level, coupled with real-time resource status, generate adaptive operation and maintenance plans; S5. Receive and execute the generated maintenance plan through the AR maintenance terminal, and actively verify the on-site data; S6. Feed the verification results back to the time-series graph neural network model for online fine-tuning of the model; S7. By constructing a digital twin model, the data and status of all the above steps are integrated and mapped to achieve the fusion of full-link information and the visualization of the system health cockpit.

[0006] Preferably, in step S1, a multi-dimensional sensor cluster consisting of micro-pressure sensors, flow meters, acoustic vibration sensors, and temperature and humidity sensors is deployed at key nodes of the gas pipeline network and user terminal equipment. All sensors are connected to a local edge gateway via wired or low-power wireless protocols to form a monitoring subnet. A precise time synchronization protocol runs between the edge gateway and all connected sensor nodes. The edge gateway acts as the master clock and periodically broadcasts synchronization beacons. Each sensor node corrects its local clock. The edge gateway listens to and receives the raw data packets asynchronously reported by each sensor node in real time. The edge gateway's data processing extracts data packets from each buffer queue. Based on the established global synchronization clock, it uses the gateway's unified high-precision timestamp to overwrite or calibrate the original preliminary timestamp in the data packet, completing the timestamp injection. According to the injected timestamp, all multi-source data streams are aligned on the time axis according to a preset time window. The aligned multi-dimensional time series data blocks undergo preliminary quality checks. The synchronized multi-dimensional time series data blocks that pass the quality check are encapsulated according to a predefined format.

[0007] Preferably, in step S2, the synchronous multi-source time-series data block is mapped to a dynamic graph structure, and the sensor cluster of each monitoring point is defined as a graph node, with edges established between the nodes in the graph. In the initial deployment phase of the system or during the known leak-free healthy phase, the dynamic graph sequences accumulated over a long period under normal operating conditions are input into the time-series graph neural network for training. During the online operation phase, real-time uploaded synchronous data blocks are continuously constructed into real-time dynamic graph sequences, which are then input into the trained time-series graph neural network to dynamically extract pre-leakage features, achieving the following: , In the formula, This represents the pre-leakage characteristic value of node v in the time t plot. Represents a nonlinear transformation function. This indicates the length of the time window, and i represents the index of the time step. Indicates the time decay weight. This represents the actual state embedding vector of graph node v at time step ti. This represents the expected state embedding vector of graph node v at time step ti. Indicates feature fusion operator, Represents a set The number of elements in the table, i.e., the number of neighbors of node v, where u represents the index of a neighbor node. Indicates the correlation deviation measure. This represents the actual association strength on the edge connecting nodes v and u at time step ti. Indicates that at time step ti, the edge The expected strength of the association.

[0008] Preferably, in step S3, the pre-leakage characteristics are obtained, and a risk subgraph is dynamically constructed in conjunction with the pipeline network physical topology. Specifically, a feature threshold is preset, nodes whose pre-leakage characteristics exceed the threshold are marked as abnormal nodes, and a connected subgraph containing all abnormal nodes and the connecting edges between them is extracted. Within the dynamic risk subgraph, the propagation pattern of the pre-leakage characteristics in time and space is analyzed. In the time dimension, the time points when the pre-leakage characteristics of each node in the subgraph first exceed the threshold are compared, and the node with the earliest time is marked as the initial abnormal node. In the spatial dimension, the information propagation model on the graph is used, with the initial abnormal node as the hypothetical source, to simulate the propagation of abnormal characteristics in the subgraph topology.

[0009] Preferably, in step S3, the node that best matches the spatiotemporal analysis results is assumed to be the initial vulnerability point. A comprehensive risk entropy value is calculated for the initial vulnerability point and its corresponding risk subgraph, implemented as follows: , In the formula, Represents the risk entropy value. This represents the average of the current feature values ​​of all nodes within the risk subgraph. The standard deviation of the eigenvalues This indicates the duration from the first detection of the anomaly to the present. This represents the average topological importance score of nodes within the risk subgraph. This represents an adjustable weighting coefficient. The larger the R value, the higher the overall risk. The calculated overall risk entropy value is mapped to a preset discrete risk level, and a vulnerability assessment report is output.

[0010] Preferably, in step S4, a vulnerability assessment report is received, dynamic resource status is obtained and standardized in real time from the operation and maintenance management platform, all risk points to be dealt with are sorted according to risk level and risk entropy value, and the priority of disposal is determined. Under the premise of satisfying all constraints, the optimal operation and maintenance team and equipment that are closest, have the same skills, and have an available time window are dynamically matched for each risk point, appropriate spare parts are allocated to the task, and the inventory status is updated. For the generated scheduling task, fine-grained operation instructions are matched and generated from the contingency plan knowledge base. The scheduling plan, detailed operation instructions, safety procedures, and electronic documents are packaged to generate a structured adaptive operation and maintenance contingency plan package and pushed out.

[0011] Preferably, in step S5, the AR maintenance terminal worn by the maintenance personnel receives the adaptive maintenance contingency plan package in real time via the network. The terminal parses the contingency plan package and matches it with its own positioning system based on the target location information within it to confirm that the personnel have arrived at or correctly located the target work area, thus completing the on-site scenario binding. On the AR display interface of the terminal, the detailed operation steps, equipment identification information, and safety procedures in the contingency plan are guided to the operators in a virtual-real overlay manner. The operators perform the operation step by step according to the AR guidance. During this process, the terminal actively controls its integrated or external multiple sensors to synchronously collect on-site verification data according to the preset procedures. The terminal performs real-time analysis on the collected multimodal verification data and compares it with the expected standards in the contingency plan or the normal state benchmark stored in the system. The re-measured gas concentration value is compared with the safety threshold. Based on the analysis results, the system automatically generates a preliminary verification conclusion and encapsulates the verification conclusion, all collected raw data, operation completion time, and operator identification into a structured on-site verification report.

[0012] Preferably, in step S6, the system receives the on-site verification report, extracts the core information from the report, including the verified initial vulnerability location, the type of operation and maintenance, the on-site verification conclusion, and the corresponding time range; the system automatically aligns this information with the historical database, extracts the pre-leaking feature time series data and associated original multi-source sensor data from a period of time before the current operation and maintenance, and this aligned feature-result data pair constitutes a labeled fine-tuning sample; based on the on-site verification conclusion, the fine-tuning sample is assigned a supervisory signal label. The newly constructed labeled fine-tuning samples are input into the temporal graph neural network for real-time fine-tuning using a mini-batch incremental learning approach.

[0013] Preferably, in step S7, a high-fidelity three-dimensional digital twin model is constructed based on the GIS, BIM, and equipment asset data of the physical gas pipeline network. In the data platform of the digital twin, the heterogeneous data of different sources, formats, and time scales imported in step S1 are fused and processed. Through a unified spatiotemporal index, events such as early warning features, risk assessment, operation and maintenance actions, and verification feedback are accurately associated and bound with specific equipment and pipe segment objects in the twin, forming a complete event context of the entire life cycle of each entity object across multiple steps. Based on the fused full-link data, a set of hierarchical system health indicators are calculated in the twin. The digital twin 3D scene and health indicators are rendered collaboratively in multiple views on the command center's large screen or professional terminal through graphics, charts, heat maps, and other forms. Intelligent alarm rules based on health indicators are set to provide proactive visual alarms for global risk level jumps and high-risk situations that have not been handled within the time limit.

[0014] Preferably, it includes a synchronous perception injection module, a dynamic feature extraction module, a risk assessment module, a contingency plan generation module, a contingency plan execution and verification module, a model optimization module, and a digital twin visualization module.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. By constructing an IoT-based end-to-end intelligent detection and operation and maintenance early warning system, this invention significantly improves the safety monitoring accuracy, risk assessment intelligence level, and operation and maintenance timeliness and accuracy of gas pipeline networks, realizing the transformation from passive emergency repair to proactive prevention and greatly reducing the probability of gas accidents. 2. By constructing a dynamic graph containing physical topological connections and fluid dynamic influence edges, and combining it with a time-series graph neural network, the system can deeply learn the evolution law of node features and the interaction mode between nodes under normal operating conditions. During the online operation phase, it can dynamically extract pre-leakage features in real time and keenly identify subtle abnormal changes such as pressure waveform and acoustic spectrum. Even in environments with high noise interference, it can effectively separate the real pre-leakage signals, significantly enhancing the sensitivity to early micro-leakage. 3. This invention constructs a dynamic risk subgraph and calculates a comprehensive risk entropy value. Combined with spatiotemporal propagation mode analysis, it can accurately locate the initial vulnerability point, i.e. the most likely physical origin of the risk. By quantitatively evaluating the feature dispersion, duration and topological importance within the risk subgraph, the generated risk entropy value can objectively reflect the comprehensive risk level and output a vulnerability assessment report containing precise location markers and key supporting information, which greatly improves the accuracy of risk tracing. 4. This invention generates adaptive operation and maintenance plans by coupling real-time resource status, automatically sorts the priority of handling according to risk level, and dynamically matches the optimal operation and maintenance team and equipment that are closest, have the same skills, and have an available time window. At the same time, it automatically associates spare parts inventory and safety procedures. The generated structured plan package contains detailed operation instructions, historical maintenance records, and electronic documents, ensuring that operation and maintenance personnel can obtain comprehensive operation guidance before departure, maximizing the utilization efficiency of operation and maintenance resources, shortening the time from risk discovery to personnel arrival on site, standardizing the operation process, effectively reducing the risk of human operation error, and realizing the high efficiency and standardization of operation and maintenance handling. Attached Figure Description

[0016] Figure 1 This invention provides the operational flow of an IoT-based intelligent gas leak detection and maintenance early warning method. Figure 1 ; Figure 2 This invention provides the operational flow of an IoT-based intelligent gas leak detection and maintenance early warning method. Figure 2 ; Figure 3 This invention provides the operational flow of an IoT-based intelligent gas leak detection and maintenance early warning method. Figure 3 ; Figure 4 This invention provides the operational flow of an IoT-based intelligent gas leak detection and maintenance early warning method. Figure 4 ; Figure 5 This invention provides the operational flow of an IoT-based intelligent gas leak detection and maintenance early warning method. Figure 5 ; Figure 6 This is a schematic diagram of the structure of an intelligent gas leak detection and maintenance early warning system based on the Internet of Things according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example Please see Figures 1-6 As shown, the present invention provides a technical solution comprising the following steps: S1. By deploying a multi-dimensional sensor cluster and synchronous edge gateway, synchronous sensing and injection of multi-source asynchronous IoT data streams are performed; S2. Dynamically extract pre-leakage features based on synchronous multi-source time-series data; S3. Based on the extracted pre-leakage features, risk tracing and vulnerability assessment are performed to locate initial vulnerability points and calculate risk entropy values; S4. Based on the output risk location and level, coupled with real-time resource status, generate adaptive operation and maintenance plans; S5. Receive and execute the generated maintenance plan through the AR maintenance terminal, and actively verify the on-site data; S6. Feed the verification results back to the time-series graph neural network model for online fine-tuning of the model; S7. By constructing a digital twin model, the data and status of all the above steps are integrated and mapped to achieve the fusion of full-link information and the visualization of the system health cockpit.

[0019] In this embodiment, in step S1, a multi-dimensional sensor cluster consisting of micro-pressure sensors, flow meters, acoustic vibration sensors, and temperature and humidity sensors is deployed at key nodes of the gas pipeline network, such as valves, interfaces, branches, and user terminal equipment, such as gas meters and stove connection points. All sensors are connected to the local edge gateway via wired or low-power wireless protocols, such as LoRa and NB-IoT, to form a monitoring subnet.

[0020] Specifically, a precise time synchronization protocol runs between the edge gateway and all connected sensor nodes. The edge gateway acts as the master clock, periodically broadcasting synchronization beacons. Each sensor node corrects its local clock. The edge gateway listens for and receives raw data packets asynchronously reported by each sensor node in real time. Each data packet contains the sensor ID, the raw reading, and the local preliminary timestamp. The gateway opens an independent first-in-first-out buffer queue for each data type to temporarily store the raw data. The edge gateway's data processing extracts data packets from each buffer queue.

[0021] Specifically, based on the established global synchronization clock, the original preliminary timestamp in the data packet is overwritten or calibrated using the gateway's unified high-precision timestamp to complete the time stamp injection. According to the injected timestamp, all multi-source data streams are aligned on the time axis according to a preset time window. For a certain type of data missing within the same time window, linear interpolation is performed to fill it in based on the data before and after it, forming a strictly synchronized multi-dimensional time series data block. The aligned multi-dimensional time series data block undergoes a preliminary quality check. The synchronized and aligned multi-dimensional time series data blocks that pass the quality check are encapsulated according to a predefined format, such as a structured data packet containing fields such as synchronization timestamp, sensor ID array, various sensor value arrays, and data quality flag bits.

[0022] In this embodiment, in step S2, the synchronous multi-source time-series data block is mapped to a dynamic graph structure. The sensor cluster of each monitoring point is defined as a graph node. The static attributes of the node include type and geographical location. The dynamic feature vector of the node is formed by splicing together multi-dimensional sensor time-series data aligned within the current time window, such as pressure waveform segments, flow sequences, acoustic spectrum features, and temperature and humidity values, after standardization and preliminary feature engineering, such as calculating short-time Fourier transform coefficients. Specifically, edges are established between nodes in the graph. These edges are categorized as follows: structural edges based on physical network topology connections; and influence edges derived from fluid dynamics simulations or correlation analysis of historical data. Each edge is assigned an initial weight and includes a learnable time delay parameter to model the time difference of signal or influence propagation between nodes. In the early stages of system deployment or during a known healthy phase with no leaks, dynamic graph sequences accumulated over a long period under normal operating conditions are input into a temporal graph neural network for training. Temporal graph neural networks typically combine graph convolutional networks and recurrent neural networks. Their core objective is to learn the evolution of node features over time under normal conditions and the spatiotemporal correlation patterns between nodes through edge interactions. Training is conducted in an unsupervised or self-supervised manner, and the model learns the normal patterns by reconstructing the input or predicting the short-term node states in the future.

[0023] Specifically, during the online operation phase, the real-time uploaded synchronous data blocks are continuously constructed into a real-time dynamic graph sequence, and input into a pre-trained time-series graph neural network to dynamically extract pre-leakage features, achieving the following: , In the formula, This represents the pre-leakage characteristic value of node v in the time t plot. Represents a nonlinear transformation function. This indicates the length of the time window, and i represents the index of the time step. Indicates the time decay weight. This represents the actual state embedding vector of graph node v at time step ti. This represents the expected state embedding vector of graph node v at time step ti. Indicates feature fusion operator, Represents a set The number of elements in the table, i.e., the number of neighbors of node v, where u represents the index of a neighbor node. Indicates the correlation deviation measure. This represents the actual association strength on the edge connecting nodes v and u at time step ti. Indicates that at time step ti, the edge The expected strength of the association.

[0024] In this embodiment, in step S3, pre-leakage features are acquired, and a risk subgraph is dynamically constructed in conjunction with the pipeline network physical topology. Specifically, a feature threshold is preset, nodes whose pre-leakage features exceed the threshold are marked as abnormal nodes, and a connected subgraph containing all abnormal nodes and the connecting edges between them is extracted. Within the dynamic risk subgraph, the propagation pattern of pre-leakage features in time and space is analyzed. In the time dimension, the time points when the pre-leakage features of each node in the subgraph first exceed the threshold are compared, and the 1-2 nodes with the earliest time are marked as initial abnormal nodes. In the spatial dimension, the information propagation model on the graph is used, with the initial abnormal nodes as the hypothetical source, to simulate the propagation of abnormal features in the subgraph topology and calculate the matching degree with the time series of the actual feature values ​​of each node.

[0025] In this embodiment, in step S3, the node that best matches the spatiotemporal analysis results is assumed to be the initial vulnerability point, i.e., the most likely physical origin of the risk, such as a tiny crack or loose interface. A comprehensive risk entropy value is calculated for the initial vulnerability point and its corresponding risk subgraph, implemented as follows: , In the formula, Represents the risk entropy value. This represents the average of the current feature values ​​of all nodes within the risk subgraph. The standard deviation of the eigenvalues This indicates the duration from the first detection of the anomaly to the present. This represents the average topological importance score of nodes within the risk subgraph. This represents an adjustable weighting coefficient; the larger the R value, the higher the overall risk.

[0026] Specifically, the calculated comprehensive risk entropy value is mapped to a preset discrete risk level, and a vulnerability assessment report is output. The report clearly includes: the precise location identifier of the initial vulnerability point, such as device ID, geographical coordinates, comprehensive risk entropy value, corresponding risk level, and key supporting information, such as the dominant risk factor and the scope of the risk submap.

[0027] In this embodiment, in step S4, a vulnerability assessment report is received, and the dynamic resource status is obtained and standardized in real time from the operation and maintenance management platform, including: the real-time location of available inspection personnel / teams, skill qualification tags, current task queue and estimated completion time, available vehicle location, status and location of special inspection equipment, and related spare parts inventory.

[0028] All potential risks are sorted based on risk level and risk entropy value to determine priority. Under the premise of meeting all constraints, the optimal maintenance team and equipment that are closest, have the right skills, and have an available time window are dynamically matched for each risk. At the same time, appropriate spare parts are allocated to the task and the inventory status is updated. For the generated scheduling task, fine-grained operation instructions are matched and generated from the contingency plan knowledge base. The scheduling plan, detailed operation instructions and safety procedures, as well as electronic documents such as work orders, links to equipment operation manuals, and historical maintenance records are packaged to generate a structured adaptive maintenance contingency plan package and pushed out.

[0029] In this embodiment, in step S5, the AR maintenance terminal worn by the maintenance personnel receives the adaptive maintenance contingency plan package in real time via the network. The terminal parses the contingency plan package and matches it with its own positioning system based on the target location information in the package to confirm that the personnel have arrived at or correctly located the target work area, thus completing the on-site scenario binding. On the AR display interface of the terminal, the detailed operation steps, equipment identification information, and safety procedures in the contingency plan are guided to the operators in a virtual-real overlay manner. The operators then follow the AR guidance to perform the operations step by step.

[0030] During the process, the terminal actively controls its integrated or external sensors to synchronously collect on-site verification data according to preset procedures. For example, when inspecting interfaces, it automatically takes high-definition close-up photos and videos; after maintenance is completed, it automatically starts the connected portable high-precision gas detector to retest trace concentrations and record readings; at the same time, it records key voice notes of the operators throughout the process; these data are automatically tagged with metadata such as timestamps, geographical locations, and operation step IDs. The terminal performs real-time analysis on the collected multimodal verification data and compares it with the expected standards in the plan or the normal state benchmark stored in the system. For example, it performs visual difference analysis on the captured interface images and historical normal images to determine whether corrosion or displacement has been addressed. The system compares the retested gas concentration value with the safety threshold. Based on the analysis results, the system automatically generates a preliminary verification conclusion. For example, if the visual condition of the sealing surface returns to normal, the retested concentration returns to zero, the hidden danger is eliminated, or a new crack is found, and an upgraded treatment is required, the verification conclusion, all collected raw data, operation completion time, and operator identification are packaged together into a structured on-site verification report.

[0031] In this embodiment, in step S6, a field verification report is received, and the core information in the report is extracted, including the verified initial vulnerability location, operation and maintenance handling type, field verification conclusion and corresponding time range; the system automatically aligns this information with the historical database and extracts the information from a period of time before the current operation and maintenance.

[0032] Specifically, the aligned feature-result data pairs, consisting of pre-leakage feature time-series data and associated raw multi-source sensor data, constitute a labeled fine-tuning sample. Based on the field verification results, a supervisory signal label is assigned to the fine-tuning sample: If the conclusion is that the hidden danger has been eliminated, that is, the operation and maintenance has solved the real problem, then the pre-leakage features within this sample time period are marked as true positive early warning signals and used as positive samples to enhance the model's ability to identify this type of feature pattern.

[0033] If the conclusion is a false alarm, meaning no hidden danger was found on site, the sample is marked as a false positive warning signal and used as a negative sample to suppress the model from producing similar false judgments.

[0034] If the conclusion is that a new hidden danger has been discovered, that is, a problem found on site that is not exactly the same as the warning but is related, then the sample is marked as a warning signal that needs to be generalized, and is used to adjust the sensitivity of the model's feature extraction. The newly constructed labeled fine-tuning samples are input into the temporal graph neural network for real-time fine-tuning using a mini-batch incremental learning approach.

[0035] In this embodiment, in step S7, a high-fidelity three-dimensional digital twin model is constructed based on the GIS, BIM, and equipment asset data of the physical gas pipeline network. This model includes static pipeline topology and equipment attributes. By establishing a data channel for continuous two-way communication with the physical world, the core data and status of the aforementioned steps are synchronized and mapped in real time.

[0036] In the data platform of the digital twin, the heterogeneous data from different sources, formats and time scales imported into S1 are integrated and processed. Through a unified spatiotemporal index, events such as early warning features, risk assessment, operation and maintenance actions and verification feedback are accurately associated and bound with specific equipment and pipeline objects in the twin, forming a complete event context for each entity object throughout its entire life cycle and across multiple steps.

[0037] Specifically, based on the fused end-to-end data, a set of hierarchical system health indicators are calculated in the twin, including: Global Health Score: A comprehensive score calculated based on the number, level, and distribution range of currently active alerts; Regional risk heat map: generated by spatial aggregation and interpolation of risk entropy values; Model performance metrics: Showing the latest accuracy, recall, etc. of the S6 model after fine-tuning; Operation and maintenance efficiency indicators: such as the timeliness rate of emergency response and the rate of elimination of hidden dangers.

[0038] Specifically, the digital twin 3D scene and health indicators are rendered collaboratively in multiple views on the command center's large screen or professional terminal through graphics, charts, heat maps, and other forms. Intelligent alarm rules based on health indicators are set to proactively and visually alert on situations such as a jump in the overall risk level or high-risk situations that have not been handled within the specified time.

[0039] This embodiment includes a synchronous perception injection module, a feature dynamic extraction module, a risk assessment module, a contingency plan generation module, a contingency plan execution and verification module, a model optimization module, and a digital twin visualization module.

[0040] Working Principle: A multi-dimensional sensor cluster, including sensors for micro-pressure, flow, acoustic vibration, and temperature and humidity, is deployed at key nodes of the gas pipeline network. An edge gateway acts as the master clock, running a precise time synchronization protocol and periodically broadcasting beacons to correct the local clocks of each sensor node. After receiving asynchronously reported raw data packets, the gateway overwrites the original preliminary timestamp with a unified, high-precision global timestamp, completing the time stamp injection. Subsequently, it aligns the multi-source data streams on the time axis according to a preset time window, filling missing data with linear interpolation to form strictly synchronized multi-dimensional time-series data blocks. These blocks undergo quality inspection and structured encapsulation to eliminate time deviations in the heterogeneous multi-source data. The synchronized multi-source time-series data is mapped into a dynamic graph structure, displaying the sensor data at each monitoring point. Clusters are defined as graph nodes, and static attributes such as type and geographical location are assigned to them. Dynamic data such as pressure waveforms, flow sequences, and acoustic spectrum within the current time window are standardized and used as node feature vectors. Structural edges based on physical topology and influence edges based on fluid dynamics or historical correlation are established between nodes, and learnable time delay parameters are assigned to model signal propagation time difference. Using a time-series graph neural network trained through unsupervised learning in the health phase, dynamic graph sequences are input online in real time, and the evolution law of node features and spatiotemporal correlation patterns between nodes under normal operating conditions are deeply learned. By comparing the deviation between the actual state and the expected state and the changes in the correlation strength between neighboring nodes, pre-leakage features that characterize the precursors of minor leaks are dynamically extracted. Based on the extracted pre-leakage features and the dynamic construction of the pipeline network physical topology, a risk subgraph is constructed. Anomaly nodes are screened using preset thresholds, and their connected subgraphs are extracted. The propagation patterns of features are then analyzed from both temporal and spatial dimensions. In the temporal dimension, the order in which the features of each node first exceed the threshold is compared to pinpoint the initial anomaly node. In the spatial dimension, the propagation process of the anomaly features along the topology is simulated using the initial node as the hypothetical source, and the matching degree with the actual observation sequence is calculated, thereby accurately locating the most likely physical origin of the risk, i.e., the initial vulnerability point. Subsequently, considering the average and dispersion of node features within the risk subgraph, the duration of the anomaly, and the topological importance score of the node, a comprehensive risk entropy value is calculated to quantify the risk level. Finally, a vulnerability assessment report containing precise location, risk level, and key supporting factors is generated. After receiving the vulnerability assessment report, the dynamic resource status in the operation and maintenance management platform is acquired and standardized in real time, including the location and skills of inspection personnel, vehicle status, distribution of testing equipment, and spare parts inventory. The system prioritizes the risk points to be addressed based on risk level and risk entropy value, while meeting time windows, skill matching, and distance constraints. The algorithm dynamically matches the optimal combination of maintenance team and equipment, and automatically allocates the necessary spare parts to update inventory. Then, it retrieves and generates fine-grained operation instructions from the contingency plan knowledge base, packaging scheduling plans, detailed operation steps, safety procedures, and related electronic documents into a structured adaptive maintenance contingency plan package. This enables intelligent matching of resources and tasks and automated delivery of work instructions. AR terminals worn by maintenance personnel receive and parse the adaptive maintenance contingency plan package in real time, using the positioning system to match and confirm arrival at the work area and complete context binding. The terminal displays operation steps, equipment information, and safety procedures to the personnel through an AR interface using a virtual-real overlay method, guiding them to execute tasks correctly. Simultaneously, it actively controls integrated or external sensors during operation, automatically collecting high-definition images, videos, gas concentration retest data, and voice annotations, among other multimodal verification data, and automatically attaches spatiotemporal metadata tags. The terminal compares and analyzes the collected data in real time with the contingency plan's expected standards or historical normal benchmarks, performing visual difference judgments and concentration threshold verifications. It automatically generates verification conclusions including hazard elimination status or newly discovered issues, and packages them into a structured on-site verification report. The system receives on-site verification reports and extracts core information. This information is then aligned with pre-leakage feature time-series data and raw multi-source sensor data in the historical database to construct fine-tuned samples labeled with supervisory signals. Based on the on-site verification conclusions, the system automatically labels the sample properties: if the hazard is eliminated, it is marked as a positive sample to enhance the model's recognition ability; if it is confirmed as a false alarm, it is marked as a negative sample to suppress misjudgments; if a new hazard is discovered, it is marked as a sample requiring generalization to adjust sensitivity. These labeled new samples are then input into the time-series graph neural network using a mini-batch incremental learning approach for real-time fine-tuning. This allows the model to continuously absorb actual on-site feedback, dynamically optimize parameter weights, adapt to environmental changes, and continuously improve the accuracy of identifying real leak features and the handling of false alarms. Filtering capabilities; Building information model and equipment asset data are used to construct a high-fidelity 3D digital twin model, establishing a two-way data channel with the physical world, and real-time synchronously mapping the core data and status of the entire chain from perception to verification; In the digital twin data platform, a unified spatiotemporal index is used to fuse heterogeneous data, and events such as early warning, assessment, operation and maintenance and feedback are precisely bound to specific pipeline equipment objects to form a full life cycle event context; Based on the fused data, hierarchical indicators such as global health score, regional risk heat map, model performance and operation and maintenance efficiency are calculated, and the 3D scene and health chart are intuitively presented on the command terminal through multi-view collaborative rendering technology. At the same time, according to preset rules, active visual alarms are triggered for abnormal situations such as risk level jumps or failure to handle within time.

[0041] 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 likenesses.

[0042] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An Internet of Things-based gas leakage intelligent detection and operation and maintenance early warning method, characterized in that, Includes the following steps: S1. By deploying a multi-dimensional sensor cluster and synchronous edge gateway, synchronous sensing and injection of multi-source asynchronous IoT data streams are performed; S2. Dynamically extract pre-leakage features based on synchronous multi-source time-series data; S3. Based on the extracted pre-leakage features, risk tracing and vulnerability assessment are performed to locate initial vulnerability points and calculate risk entropy values; S4. Based on the output risk location and level, coupled with real-time resource status, generate adaptive operation and maintenance plans; S5. Receive and execute the generated maintenance plan through the AR maintenance terminal, and actively verify the on-site data; S6. Feed the verification results back to the time-series graph neural network model for online fine-tuning of the model; S7. By constructing a digital twin model, the data and status of all the above steps are integrated and mapped to achieve the fusion of full-link information and the visualization of the system health cockpit. 2.The gas leakage intelligent detection and operation and maintenance early warning method based on the Internet of Things according to claim 1, characterized in that: In step S1, multi-dimensional sensor clusters are deployed at key nodes of the gas pipeline network and user terminal equipment. All sensors are connected to a local edge gateway via wires to form a monitoring subnet. A precise time synchronization protocol runs between the edge gateway and all connected sensor nodes. The edge gateway acts as the master clock and periodically broadcasts synchronization beacons. Each sensor node corrects its local clock. The edge gateway listens for and receives the raw data packets asynchronously reported by each sensor node in real time. The edge gateway's data processing extracts data packets from each buffer queue. Based on the established global synchronization clock, it uses the gateway's unified high-precision timestamp to overwrite or calibrate the original preliminary timestamp in the data packet, completing the timestamp injection. According to the injected timestamp, all multi-source data streams are aligned on the time axis according to a preset time window. The aligned multi-dimensional time series data blocks undergo preliminary quality checks. The synchronized and aligned multi-dimensional time series data blocks that pass the quality check are encapsulated according to a predefined format.

3. The gas leakage intelligent detection and operation and maintenance early warning method based on the Internet of Things according to claim 1, characterized in that: In S2, the synchronous multi-source time-series data block is mapped to a dynamic graph structure, and the sensor cluster of each monitoring point is defined as a graph node, and edges are established between the nodes in the graph. In the early stages of system deployment or during a known healthy phase with no leaks, the dynamic graph sequences accumulated over a long period under normal operating conditions are input into the time-series graph neural network for training. During the online operation phase, the real-time uploaded synchronous data blocks are continuously constructed into a real-time dynamic graph sequence, and input into a pre-trained time-series graph neural network to dynamically extract pre-leakage features, achieving the following: ， In the formula, Indicates pre-leakage characteristics, Represents a nonlinear transformation function. This indicates the length of the time window, and i represents the index of the time step. Indicates the time decay weight. This represents the actual state embedding vector of graph node v at time step ti. This represents the expected state embedding vector of graph node v at time step ti. Indicates feature fusion operator, Represents a set The number of elements in the array, where u represents the index of the neighboring node. Indicates the correlation deviation measure. This represents the actual association strength on the edge connecting nodes v and u at time step ti. Indicates that at time step ti, the edge The expected strength of the association.

4. The gas leakage intelligent detection and operation and maintenance early warning method based on the Internet of Things according to claim 1, characterized in that: In step S3, pre-leakage characteristics are acquired, and a risk subgraph is dynamically constructed in conjunction with the pipeline network physical topology. Specifically, a feature threshold is preset, and nodes whose pre-leakage characteristics exceed the threshold are marked as abnormal nodes. A connected subgraph containing all abnormal nodes and the connecting edges between them is extracted. Within the dynamic risk subgraph, the propagation pattern of pre-leakage characteristics in time and space is analyzed. In the time dimension, the time points when the pre-leakage characteristics of each node in the subgraph first exceed the threshold are compared, and the node with the earliest time is marked as the initial abnormal node. In the spatial dimension, an information propagation model on the graph is used, with the initial abnormal node as the hypothetical source, to simulate the propagation of abnormal characteristics in the subgraph topology.

5. The gas leakage intelligent detection and operation and maintenance early warning method based on the Internet of Things according to claim 4, characterized in that: In step S3, the node that best matches the spatiotemporal analysis results is assumed to be the initial vulnerability point. A comprehensive risk entropy value is calculated for the initial vulnerability point and its corresponding risk subgraph, implemented as follows: ， In the formula, Represents the risk entropy value. This represents the average of the current feature values ​​of all nodes within the risk subgraph. The standard deviation of the eigenvalues This indicates the duration from the first detection of the anomaly to the present. This represents the average topological importance score of nodes within the risk subgraph. The calculated comprehensive risk entropy value is mapped to a preset discrete risk level, and a vulnerability assessment report is output.

6. The method for intelligent detection and early warning of gas leaks based on the Internet of Things according to claim 5, characterized in that: In step S4, a vulnerability assessment report is received, dynamic resource status is obtained and standardized in real time from the operation and maintenance management platform, all risk points to be dealt with are sorted according to risk level and risk entropy value, and the priority of handling is determined. Under the premise of meeting all constraints, the optimal operation and maintenance team and equipment that are closest, have the same skills, and have an available time window are dynamically matched for each risk point. For the generated scheduling task, fine-grained operation instructions are matched from the contingency plan knowledge base. The scheduling plan, detailed operation instructions, safety procedures, and electronic documents are packaged to generate a structured adaptive operation and maintenance contingency plan package and pushed out.

7. The gas leakage intelligent detection and operation and maintenance early warning method based on the Internet of Things according to claim 1, characterized in that: In S5, the AR maintenance terminal worn by maintenance personnel receives adaptive maintenance plan packages in real time through the network. The terminal parses the plan package and matches it with its own positioning system based on the target location information in it to confirm that the personnel have arrived at or correctly located the target work area, thus completing the on-site scenario binding. On the terminal's AR display interface, the detailed operating steps, equipment identification information, and safety procedures in the contingency plan are guided to the operators in a virtual-real overlay manner. The operators follow the AR guidance step by step to perform the operation. The terminal actively controls its integrated or external multiple sensors to synchronously collect on-site verification data according to the preset procedures. The terminal performs real-time analysis on the collected multimodal verification data and compares it with the normal state benchmark of the expected standard in the contingency plan. The re-measured gas concentration value is compared with the safety threshold. Based on the analysis results, the system automatically generates a preliminary verification conclusion and encapsulates the verification conclusion, all collected raw data, operation completion time, and operator identification into a structured on-site verification report.

8. The gas leakage intelligent detection and operation and maintenance early warning method based on the Internet of Things according to claim 7, characterized in that: In step S6, the system receives the field verification report and extracts the core information from the report. The system automatically aligns this information with the historical database and extracts a labeled fine-tuning sample from the pre-leaked feature time series data and associated original multi-source sensor data from a period of time before the current operation and maintenance. The aligned feature-result data pairs constitute a fine-tuning sample. Based on the field verification conclusion, the fine-tuning sample is assigned a supervisory signal label. The newly constructed labeled fine-tuning sample is then input into the time series graph neural network for real-time fine-tuning. 9.The gas leakage intelligent detection and operation and maintenance early warning method based on the Internet of Things according to claim 1, characterized in that: In S7, a three-dimensional digital twin model is constructed based on the GIS, BIM and equipment asset data of the physical gas pipeline network. In the data platform of the digital twin, the heterogeneous data of different sources, formats and time scales of the whole link imported in S1 are fused and processed. Through a unified spatiotemporal index, events are accurately associated and bound with specific equipment and pipe segment objects in the twin, forming a complete event context of the entire life cycle of each entity object across multiple steps. Based on the fused whole link data, a set of hierarchical system health indicators are calculated in the twin. The digital twin 3D scene and health indicators are rendered collaboratively in multiple views on the command center's large screen or professional terminal through graphics, charts, heat maps, and other forms. Intelligent alarm rules based on health indicators are set to provide proactive visual alarms for global risk level jumps and high-risk situations that have not been handled within the time limit.

10. The Internet of Things-based gas leakage intelligent detection and operation and maintenance early warning system realized by the Internet of Things-based gas leakage intelligent detection and operation and maintenance early warning method according to claim 1, characterized in that: It includes a synchronous perception injection module, a dynamic feature extraction module, a risk assessment module, a contingency plan generation module, a contingency plan execution and verification module, a model optimization module, and a digital twin visualization module.

Images