Mining pipeline liquid leakage positioning and intelligent valve closing system based on 5G + AI

Abstract

The invention discloses a mining pipeline leakage positioning and intelligent valve closing system based on 5G + AI. The mining pipeline leakage positioning and intelligent valve closing system comprises a multi-mode sensor deployed at a pipeline key node, a mining 5G communication network, a cloud edge collaborative analysis platform and an intelligent execution unit. The sensor collects vibration and pressure waveform data, and the vibration and pressure waveform data are synchronously transmitted through a 5G network in a low-delay manner; the edge analysis unit is internally provided with a lightweight AI model for leakage initial judgment and triggering; the cloud analysis unit performs fusion analysis on multi-node synchronous data by adopting a space-time convolutional neural network to realize sub-wavelength level accurate positioning and intensity evaluation of a leakage point; and the decision engine generates an intelligent instruction containing a dynamic valve closing strategy according to the positioning and evaluation result, and issues the intelligent instruction to the intelligent valve for execution through the 5G network. The problems that a traditional method is poor in positioning precision, lagged in response and high in false alarm rate are solved, rapid sensing, accurate positioning, intelligent decision making and self-adaptive optimization of leakage events are achieved, and the safety and reliability of the mining liquid supply pipeline are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of mine safety and intelligent equipment technology, specifically relating to an intelligent monitoring and safety handling system for mine high-pressure fluid transport pipelines that integrates advanced communication and artificial intelligence technologies. Background Technology

[0002] The underground fluid supply network in coal mines is responsible for supplying power to critical equipment such as hydraulic supports and emulsion pump stations. It is characterized by high pressure (31.5 MPa), long distances (over 3 km per roadway), and numerous branches (dozens per working face). Due to the combined effects of mining stress, humid air, coal dust erosion, and hydrogen sulfide corrosion, pipeline joints, welds, and valve bodies are highly susceptible to fatigue cracks, leading to sudden leaks of emulsions or water-based media. Statistics show that in the past five years, key state-owned coal mines have recorded over 1,200 fluid supply system leaks. Approximately 38% of these leaks were not detected in time, resulting in water accumulation in roadways, insufficient equipment pressure, and ultimately, working face shutdowns, with an average economic loss exceeding 3 million yuan per incident. About 7% of the leaks, under high-pressure atomization, formed flammable droplets, inducing secondary coal dust explosions and posing significant safety hazards.

[0003] Currently, the common underground management model is an open-loop system: sensor monitoring – underground ring network transmission – manual interpretation in the surface monitoring room – telephone notification to maintenance workers. Sensors are mostly deployed at single points, and the data collected is limited to pressure, flow, or negative pressure wave signals. Communication relies on industrial Ethernet and multi-level fieldbus relays, with end-to-end latency typically exceeding 500ms, increasing exponentially with tunnel extension. Decision-making depends on the experience of the shift operator, with an average time of 3–5 minutes from alarm to dispatch. Execution still requires maintenance personnel to enter the mine with manual hydraulic pliers or gate valve wrenches to close the valve, resulting in a shutdown delay of over 10 minutes. During this period, the leaked medium spreads along the tunnel at approximately 0.3m / s, often exceeding 50m in contamination range, potentially flooding cables, switches, and other electrical equipment, escalating the severity of the accident. Furthermore, existing monitoring methods generally have a location error greater than 50m in branching pipe networks, making it difficult to distinguish between leak signals and interference from mining machinery vibrations, resulting in a false alarm rate consistently above 30%, leading to a vicious cycle of inaccurate alarms and failure to report potential hazards. Although some mines have tried to introduce digital twin visualization platforms, the low frequency of model updates and the lack of on-site deformation data mean that they remain in the stage of being able to see but not control, and cannot replace the automatic shutdown function.

[0004] Foreign remote liquid supply systems are represented by RMI (UK), KAMAT (Germany), and Hauhinco, whose emulsion pumps and pump station control systems have consistently been industry leaders. RMI's system, based on a CAN bus, can centrally or locally control four pumps and one water tank. Its five-plunger pump has a maximum flow rate of 1080 L / min and a pressure of 100 MPa, with power consumption reduced by 40%. KAMAT's system, based on a BUS bus, controls four main pumps, two pre-boost pumps, and two water tanks. It uses electric unloading valves and variable frequency speed control, achieving a flow rate of up to 1360 L / min, and includes online monitoring of pH and conductivity. Hauhinco employs distributed control, with all monitoring and control points centrally connected, making local operation impossible. Its HyPres pipeline system can extend the high-pressure liquid supply distance to 4000 m. These systems focus on improving mechanical performance, with slow improvements in control functions, and none have achieved millisecond-level leak detection or remote automatic shutdown.

[0005] There are numerous domestic fluid supply manufacturers, and pump parameters have been upgraded from 500L / min and 40MPa to 630L / min and 40MPa. They generally employ variable frequency multi-drive systems and electronically controlled precision fluid dispensing, and have added water treatment and low-flow pressure-maintaining pumps, resulting in more comprehensive control functions than their foreign counterparts. However, both domestic and international systems remain at the level of monitoring plus manual intervention. This leads to large leak location errors, long shutdown lags, and an inability to form closed-loop control, failing to meet the mandatory requirement of Article 443 of the "Coal Mine Safety Regulations" that underground fluid supply systems must achieve remote emergency shut-off. Furthermore, intelligent monitoring in coal mines faces common problems such as poor reliability of precursor information collection, insufficient integration depth of cloud platforms, weak database security, and immature artificial intelligence applications, severely hindering further improvements in the safety level of fluid supply systems. Summary of the Invention

[0006] This invention aims to overcome the technical defects of existing mining pipeline leakage monitoring methods, such as poor positioning accuracy, slow response, high false alarm rate, and coarse control. By constructing a closed-loop system that integrates ubiquitous sensing, high-speed computing, intelligent decision-making, and precise control, it achieves sub-minute-level accurate positioning of leakage events, second-level automatic isolation, and extremely low false alarm rate, thereby fundamentally improving the safety of mine fluid supply and emergency response capabilities.

[0007] The solution to the technical problem of this invention is as follows: a mining pipeline leakage location and intelligent valve shut-off system based on 5G and artificial intelligence, comprising: multiple sensor nodes deployed at key nodes of the pipeline for real-time acquisition of vibration and pressure waveform data of the pipeline; a mining 5G communication network for low-latency, high-reliability transmission of the waveform data acquired by the sensor nodes; a cloud-edge collaborative analysis platform, comprising: an edge analysis unit deployed underground for real-time processing and initial leakage judgment of data from a single sensor node; a cloud intelligent analysis unit deployed on the ground for receiving synchronous waveform data from multiple sensor nodes triggered by the edge analysis unit, and accurately locating and assessing the leakage point based on a spatiotemporal deep learning model; and an intelligent execution unit, comprising at least one electric valve installed on the pipeline for receiving control commands from the cloud intelligent analysis unit and executing valve shut-off operations; wherein, the cloud intelligent analysis unit is also used to generate intelligent control commands containing target valve identification and valve shut-off strategies based on the leakage location and assessment results, and to send them to the intelligent execution unit through the 5G communication network.

[0008] Preferably, at least a portion of the plurality of sensor nodes are multimodal sensor nodes, integrating vibration sensors and dynamic pressure sensors.

[0009] Preferably, the edge analysis unit has a built-in first artificial intelligence model for quickly analyzing single-point vibration signals to determine suspected leakage events; the cloud intelligent analysis unit has a built-in second artificial intelligence model, which is a spatiotemporal convolutional neural network model, for fusing and analyzing synchronous pressure and vibration waveform data of multiple nodes and outputting the spatial coordinates of the leakage point.

[0010] Preferably, the mining 5G communication network is used to ensure the time synchronization accuracy of the waveform data collected by the multiple sensor nodes, and to provide an end-to-end low-latency communication link to the edge analysis unit and the intelligent execution unit.

[0011] Preferably, the valve closing strategy is dynamically generated based on the leakage intensity assessment results, including the valve closing speed curve.

[0012] Preferably, the system further includes a case database and a model optimization module; the case database is used to store waveform data, location results and control feedback information related to leakage events; the model optimization module is used to incrementally train the first artificial intelligence model and / or the second artificial intelligence model using the data in the case database.

[0013] Preferably, after determining a suspected leakage event, the edge analysis unit triggers the associated sensor nodes to enter a high-frequency synchronous acquisition mode.

[0014] Preferably, the system is connected to a mining digital twin platform to display the leak location results, pipeline status, and valve action information in three dimensions.

[0015] Compared to existing technologies, this solution, through the comprehensive implementation of the above-mentioned inventive points, brings the following significant and interrelated beneficial technical effects: 1. A leapfrog improvement in positioning accuracy and response speed has been achieved. Thanks to 5G ensuring the synchronous and real-time acquisition of multi-source high-fidelity data, and the ability of cloud-based AI models to fully explore the spatiotemporal correlation features in these data, the system has improved the leak location accuracy from the hundreds of meters level of traditional methods to within ±10 meters (sub-minute level). At the same time, relying on the rapid initial judgment of edge AI and the low-latency command transmission of 5G, the entire response time from the occurrence of the leak to the activation of valve closure has been shortened to the second level, which is two orders of magnitude better than the manual handling mode.

[0016] 2. Significantly improves system reliability and availability, with an extremely low false alarm rate. The dual AI filtering mechanism of edge detection and cloud confirmation effectively distinguishes between real leaks and common strong disturbances such as pump and valve start-up / shutdown and material impact, reducing false alarms caused by non-real leaks by over 90%. This makes alarm information highly reliable, avoiding frequent false alarms and wasted maintenance resources.

[0017] 3. It achieves synergistic optimization of safety and equipment lifespan. Based on AI-assessed dynamic valve closure strategies, it can tailor closure actions according to the severity of the leak. For small leaks, it employs slow closure to avoid water hammer, while for large ruptures, it uses rapid closure for quick isolation. This refined control not only minimizes the damage from leaks but also significantly reduces the impact damage to the pipeline system caused by water hammer from emergency valve closures, extending the service life of critical equipment.

[0018] 4. An intelligent system paradigm that is evolvable and easy to operate and maintain has been constructed. The system's self-learning capability ensures long-term stability and continuous optimization of its performance, reducing the difficulty and frequency of later maintenance and debugging. At the same time, the integration with the digital twin platform provides an intuitive visual monitoring and simulation interface, making the status of complex underground pipeline systems clear at a glance, greatly improving operation and maintenance management efficiency and safety pre-control capabilities. Attached Figure Description

[0019] Figure 1 It is a system hierarchy architecture diagram; Figure 2 This is a schematic diagram illustrating the system workflow and the mechanism of AI model operation; Figure 3 This is a flowchart of cloud-edge collaborative AI data processing; Figure 4 This is a flowchart of the system's intelligent valve shut-off process; Figure 5This is a schematic diagram of the intelligent valve shut-off method of the system. Detailed Implementation

[0020] The technical solution of the present invention will be described in detail below with reference to the embodiments. It should be understood that these embodiments are intended to clearly illustrate the core concept and implementation of the present invention, and are not intended to limit the scope of protection of the present invention in any way.

[0021] Example 1: A 5G+AI-based system for locating and intelligently shutting off leaks in mining pipelines. This example addresses three major technical bottlenecks in existing high-pressure liquid supply pipeline leak monitoring systems for mines: low location accuracy, slow response speed, and high false alarm rate. Traditional techniques such as negative pressure wave methods are limited by sensor deployment density and the complex acoustic environment underground, often resulting in location errors of tens to hundreds of meters. From the occurrence of a leak to manual confirmation and remote valve closure, the time required can be as long as several minutes, which can easily lead to the escalation of the accident. Furthermore, conventional threshold alarms are easily interfered with by operating conditions such as pump and valve start-up and shutdown, generating a large number of invalid alarms. To address these issues, this invention constructs an active defense system that deeply integrates 5G communication and artificial intelligence (AI) analysis. Its core objective is to achieve a comprehensive performance breakthrough with sub-minute-level accurate location, second-level intelligent valve closure, and an extremely low false alarm rate.

[0022] The physical architecture of this system is as follows: Figure 1 As shown, the system is built around the concepts of ubiquitous sensing, high-speed computing, cloud-edge collaboration, and precise control. Its key lies in using 5G communication to reconstruct the information flow and using AI algorithms to endow the system with the ability to understand and make decisions.

[0023] The system's sensory nerve endings consist of multimodal sensor nodes deployed at pump stations, pipeline segment points, and key valve groups. Each node integrates a vibration acceleration sensor, a dynamic pressure sensor, and an intrinsically safe 5G communication module. The vibration and pressure sensors synchronously acquire raw waveform data at a frequency of at least 1kHz. The biggest difference from previous systems is that this high-frequency data is no longer aggregated layer by layer through traditional industrial buses with limited bandwidth and variable latency; instead, it is directly connected to the underground 5G mining private network via the 5G modules built into the nodes. To ensure full coverage, intrinsically safe 5G base stations and low-power 5G remote radio units (RRUs) are deployed as needed in the tunnels, forming a high-bandwidth, low-latency, and highly reliable underground IoT backbone network. This allows massive amounts of raw waveform data to be directly uploaded with a stable end-to-end latency of less than 50 milliseconds, providing an unprecedented data foundation for the real-time and accurate analysis of subsequent AI models.

[0024] This system innovatively adopts a cloud-edge collaborative AI computing architecture with lightweight edge inference and precise cloud training. Edge side (rapid response unit): A high-performance AI inference chip is built into the intrinsically safe edge computing gateway for mining, located near the pumping station or regional control center. This chip deploys a compressed and optimized leak detection model. This model is a lightweight convolutional neural network (CNN) responsible for real-time analysis of the vibration spectrum uploaded by a single sensor node, quickly determining whether a leak is suspected or not, with a computational latency controlled within 10 milliseconds. Once the edge model determines a suspected leak, the gateway is immediately activated and issues instructions via the 5G network to all sensor nodes in the relevant pipe section, triggering them to enter a high-frequency synchronous acquisition mode and uploading a high-precision synchronous waveform data block from the seconds before and after this critical moment. This mechanism ensures both low-power operation under normal conditions and the acquisition of the highest quality analysis data in accident situations. Cloud side (the brain for precise analysis): The AI ​​analysis main engine of the system is deployed in the ground data center, its core being a leak localization and assessment model based on a spatiotemporal convolutional neural network (ST-CNN). The network's input consists of time-series pressure and vibration waveforms from multiple sensors upstream and downstream of the suspected leaking pipe section, strictly synchronized via a 5G network. The network structure is specially designed: its temporal convolutional layers extract the transient characteristics of each sensor signal (e.g., wavefront characteristics of negative pressure waves); spatial convolutional layers learn the correlation and propagation patterns between sensor signals at different spatial locations (e.g., implicit patterns of time difference of arrival (TDOA)). Through end-to-end training, the model can directly map the precise location coordinates of the leak point and estimate the leak intensity from the raw data, avoiding the error accumulation problems caused by inaccurate manual feature extraction and inaccurate sound velocity settings in traditional methods. The CNN processes not natural images, but multi-channel time-series data with a defined spatiotemporal structure generated by the sensor network. We construct a multi-dimensional feature map from the one-dimensional vibration waveform, pressure waveform, and their distribution on the spatial sensor nodes, which serves as the input to the CNN.

[0025] The core of this system, the Spatiotemporal Convolutional Neural Network (ST-CNN), takes as input a three-dimensional tensor I ∈ R^(N×T×C). N (spatial dimension): represents the number of sensor nodes involved in the analysis. Each node is considered a "spatial location." T (temporal dimension): represents the length of the strictly time-synchronized waveform data points acquired from each node. This is equivalent to the length of a "time series." C (channel dimension): represents the number of signal types acquired at each time point. For example, C=2 indicates that one channel is vibration acceleration data and another channel is dynamic pressure data. This is similar to the R, G, and B channels of a color image.

[0026] The model training data comes from two sources: historical real-world leak event data: collecting synchronous waveform data recorded by multiple sensors during past leaks. Each data sample corresponds to a known real leak location label (coordinate X) and a leak intensity label (Q) assessed based on subsequent maintenance records. Simulation and data augmentation data: generating a large amount of simulated waveform data by simulating leaks at different locations, with different orifice sizes, and under different background pressures in the digital twin model. Simultaneously, noise addition, time shifting, and amplitude scaling are applied to existing real-world data to enhance the model's robustness to various disturbances and unknown operating conditions.

[0027] The preprocessing step is applied to each T×C slice (i.e., data from a single sensor). Standardization: z-score standardization is performed on the data from each channel to remove baseline offset and dimensional effects. Bandpass filtering: Digital filters are designed based on the main energy frequency band of the leakage signal (e.g., 100Hz-2kHz) to filter out low-frequency creep and high-frequency electromagnetic noise. Input tensor construction: The preprocessed T×C data blocks from N sensors are arranged according to the actual spatial order (or logical topological order) of the sensors in the pipeline network and stacked to form the final N×T×C three-dimensional tensor I. The technical effect of this step is that it transforms the spatiotemporal sensing information of the physical world into a structured data format with local correlation that can be efficiently processed by CNNs, laying the foundation for subsequent automatic feature extraction.

[0028] ST-CNN employs a dual-branch feature fusion structure to extract temporal local features and spatially dependent features separately and collaboratively. The temporal feature extraction branch mainly consists of a stack of multiple one-dimensional convolutional layers and pooling layers. The convolutional kernels slide along the T-dimensional (time axis), acting on the data from each sensor (a 1×T×C local region). The technical problem it addresses is that transient signals generated by leaks (such as the wavefront of negative pressure waves) have specific local waveform patterns. The one-dimensional convolutional layers, through different convolutional kernels, can automatically learn and identify these key local temporal patterns (such as pulses of specific shapes and oscillation decay), thereby effectively distinguishing leak signals from random noise or interference signals generated by the start-up and shutdown of other equipment. Its technical effect is: achieving high-sensitivity and interference-resistant preliminary perception of leak events at a "single observation point."

[0029] The spatiotemporal feature fusion and spatial relationship learning branch reshapes the feature map output from the time branch (whose shape has been changed to N×T'×F, where T' is the compressed length of the time dimension and F is the number of feature channels), merging the N and T' dimensions into a two-dimensional space to form a (N*T')×F grid. This is then fed into a two-dimensional convolutional layer (Conv2D). These two-dimensional convolutional kernels slide on this grid, their receptive fields simultaneously covering multiple sensor nodes (N dimension) and multiple consecutive time points (T' dimension). The key technical problem addressed is the propagation effect of the leakage signal. The time difference (TDOA) and amplitude attenuation of the signal arriving at different sensors contain the location information of the leakage point. The two-dimensional convolutional layer can automatically learn the complex correlations and dependencies between features at different spatial locations (between sensor nodes), i.e., learn the spatiotemporal propagation pattern of the signal. The technical effect is that the network can directly internalize the acoustic propagation characteristics of the pipeline network from the data without manually setting the wave velocity or calculating the time difference, thereby achieving precise location of the leakage source on the spatial network.

[0030] The fully connected layer and output are formed by flattening the high-level features extracted by the convolutional layers. The final output layer consists of two heads: a localization regression head, which outputs a two-dimensional or three-dimensional coordinate X representing the estimated location of the leak point; and an intensity assessment regression head, which outputs a scalar Q representing the estimated leak intensity. The loss function uses a composite loss function combining mean squared error (MSE): L = α * MSE(X, X_true) + β * MSE(Q, Q_true), which optimizes both localization and assessment accuracy.

[0031] During forward propagation (i.e., leakage analysis), the original waveform tensor I flows through the network. The temporal convolutional layer first "scans" each sensor node, extracting local transient features related to the leakage and filtering out irrelevant temporal patterns. The spatiotemporal convolutional layer then "scans" these extracted feature maps, inferring the source of signal propagation by comparing feature activation patterns at different locations. Finally, the fully connected layer maps these abstract features, which incorporate spatiotemporal context, to specific physical quantities (coordinates and intensity).

[0032] The actuator employs an intrinsically safe and explosion-proof intelligent electric ball valve for mining applications, and its controller also integrates a 5G communication module. Unlike traditional valves that only receive on / off commands, the intelligent valve in this system can receive structured valve-closing strategies from the cloud control platform. This strategy not only includes valve-closing commands but may also include a dynamically calculated closing speed curve based on AI-assessed leakage intensity and pipeline pressure (e.g., rapid shut-off for large leaks, slow closure for small leaks to prevent water hammer), as well as a status confirmation and reporting process after valve closure. The low latency of the 5G network allows the command transmission time from cloud analysis completion to valve activation to be compressed to the hundreds of milliseconds level.

[0033] like Figure 2 and Figure 3 As shown, the system enters intelligent monitoring mode upon power-up, and its workflow is a closed loop driven by data and AI. The first stage: routine monitoring and edge screening. All sensor nodes continuously collect data at the base frequency and upload compressed feature data (such as RMS and peak values) to the cloud platform via the 5G network for health status visualization. Simultaneously, vibration data from each node is analyzed in real-time by a lightweight edge CNN model. This model, trained on a large amount of historical data (including normal operating conditions, pump start-stop interference, and real leak cases), has learned to distinguish specific frequency band vibration patterns caused by leaks from background noise. The technical effect is that most meaningless background fluctuations are filtered directly at the edge, and only highly suspicious event trigger signals and associated high-precision data blocks are reported. This significantly reduces the data throughput pressure on the cloud and reduces system-level false alarms caused by common interference.

[0034] Phase Two: Leak Confirmation, Precise Location, and Intelligent Decision-Making. Upon receiving a suspected leak alarm and synchronization data block from the edge gateway, the cloud platform immediately initiates a cloud-based ST-CNN location model for deep analysis. The model's processing goes beyond simple threshold comparison; it involves deep feature abstraction and fusion reasoning of multi-channel spatiotemporal signals. For example, the model can understand the correlation between the steepness of the pressure signal's falling edge and the leak's orifice diameter, and it can also perceive the leak direction information implied by the time difference in the arrival time of vibration signals at different sensors. Crucially, the model internalizes experience gained from pipeline topology, media properties, and even historical leak data during training. Its output location result is a direct spatial coordinate with an accuracy within ±10 meters, and it has a certain degree of fault tolerance for single sensor failures. After location is established, the AI ​​decision engine calculates the optimal valve closure scheme (e.g., which valves to close, the closure order, and the speed) within milliseconds based on the leak location, intensity estimate, and pipeline hydraulic model, and generates the corresponding control command set.

[0035] Phase Three: Precise Control and Feedback Learning. Control commands are instantly transmitted to the target smart valve via the 5G network. After parsing the commands, the valve controller executes the closing action according to the preset dynamic performance curve and feeds back real-time valve position and torque information to the cloud platform. The platform automatically archives all data from this leakage event, including raw waveforms, AI positioning process data, and valve execution results, into a case library. This case library will be used periodically for incremental training and optimization of the cloud-based ST-CNN model and the edge CNN model, enabling the AI ​​model to continuously learn new operating conditions and leakage patterns as the system runs, achieving a self-evolving capability that becomes more accurate with use. This is the core mechanism for the system to maintain high reliability over the long term.

[0036] Through the aforementioned integrated hardware and software implementation, this embodiment achieves the following substantial technological advancements: 1. It resolves the contradiction between precise positioning and rapid response. The high bandwidth and low latency of 5G networks enable the bottleneck-free transmission of massive synchronous waveform data, providing an input foundation for complex AI models. AI models (especially ST-CNN) can directly extract deep spatiotemporal correlation features from this data, achieving sub-minute-level, high-precision positioning that traditional algorithms cannot reach. The connection between technical means and effects lies in this: without 5G, AI models are useless; without AI, 5G transmission is merely a data deluge. The combination of the two leads to a qualitative leap in precise positioning capabilities. 2. It achieves a leap from alarm to intelligent decision-making and control. The system utilizes AI not only for detection but also for evaluation and decision-making. AI-based leakage intensity assessment transforms valve closing strategies (such as closing speed) from fixed to dynamic optimization, effectively balancing the contradiction between isolating leaks and preventing water hammer. The connection between the technical means and the results lies in the fact that the AI ​​decision engine embeds hydraulic dynamics knowledge, and its output is a control strategy that considers the overall safety of the system. The low latency of 5G ensures that this strategy can be executed seamlessly and quickly, forming a complete intelligent control closed loop. 3. The system's self-evolving performance capability is built. By establishing a complete data closed loop from data acquisition and AI analysis to execution feedback, and using this closed-loop data to continuously train the model, the system can adapt to the slow changes in mine pipeline conditions and new leakage modes. The connection between the technical means and the results lies in the fact that the cloud-edge collaborative architecture makes model updates feasible, while 5G ensures that a large amount of field data can be efficiently transmitted back for training. Ultimately, this ensures that the system's reliability does not decrease over time but is expected to improve, solving the performance degradation problem caused by the rigid management logic of traditional systems.

[0037] Example 2: The method based on the system described in Example 1 is as follows Figure 4 and Figure 5 As shown.

[0038] Phase 1: System Initialization and Routine Monitoring. After power-on, the system first completes an initialization self-test, including confirming the connection status of sensor nodes, 5G communication links, edge gateways, and cloud services. After initialization, the system enters a 24 / 7 routine monitoring mode. Data Acquisition: All deployed sensor nodes continuously acquire vibration acceleration signals and dynamic pressure signals from the pipeline at a base frequency (e.g., 500Hz). Edge Light Processing and Uplink: The raw data acquired is preprocessed locally on the nodes (e.g., analog-to-digital conversion, preliminary filtering) and then connected to the underground 5G private network via an integrated 5G module for stable, low-latency transmission to the edge computing gateway.

[0039] Edge screening and cloud-based panoramic visualization: A lightweight AI model within the edge gateway analyzes the vibration signal flow of each node in real time, calculating a continuous anomaly index. At this stage, the model's primary function is to filter background noise, uploading only highly compressed statistical features (such as mean and variance) and the anomaly index to the cloud. After receiving data from all nodes, the cloud platform dynamically updates the global status view of the entire pipeline network on the digital twin interface, such as displaying pressure distribution using color temperature maps, achieving all-weather, panoramic visual monitoring. During this stage, the system operates in a low-power, wide-area monitoring state, providing a background baseline for subsequent precise analysis.

[0040] Phase Two: Leak Triggering and Confirmation. When a pipeline leaks or experiences strong abnormal vibrations: Edge Model Trigger: Vibration waves of a specific frequency band generated at the leak point propagate along the pipeline and are captured by nearby sensor nodes. The AI ​​model in the edge gateway analyzes the node and detects a sharp increase in its anomaly index exceeding a first threshold, immediately making a preliminary judgment of suspected leakage. The effect is: initial identification of potential threats is completed within seconds, and isolated interference such as single-person operation and equipment collisions is greatly filtered out.

[0041] Regarding the triggered synchronous acquisition: The edge gateway then sends a high-frequency synchronous acquisition command via the 5G network to all relevant sensor nodes within a certain range upstream and downstream of the trigger point. These nodes immediately switch to high-frequency mode (e.g., 2kHz) and utilize the high-precision clock synchronization function of the 5G network to begin acquiring raw pressure and vibration waveform data with strictly synchronized timestamps.

[0042] Data packet upload: After the high-precision acquisition window, which lasts for hundreds of milliseconds, ends, each node uploads a synchronous data packet containing data before and after the critical moment to the cloud at high speed via the 5G network. The core effect of this is to acquire the highest quality and most complete raw data for precise positioning only when absolutely necessary, balancing data quality with system energy consumption and bandwidth costs.

[0043] Phase Three: Precise Cloud Positioning and Intelligent Decision-Making. Upon receiving the synchronization data packet, the cloud platform initiates the core analysis process: spatiotemporal fusion AI analysis, activating the spatiotemporal convolutional neural network in the cloud. This model simultaneously receives synchronized time-series data uploaded from sensors at multiple spatial locations. The model's temporal convolutional layers specifically learn the transient characteristics of a single signal waveform (such as the steepness of the pressure drop edge), while the spatial convolutional layers learn the correlation and propagation delay patterns between signals at different locations. Through end-to-end training, the model can directly output the precise spatial coordinates (X) of the leak point and an estimated leak intensity (Q). The resulting technical benefits are: avoiding the cumulative errors caused by manually extracting features, calculating wave velocity, and solving equations in traditional methods, and achieving high-precision positioning within ±10 meters even in noisy environments.

[0044] Intelligent Strategy Generation: The coordinates and intensity information output by the AI ​​positioning engine are fed into the intelligent decision-making module. This module, combined with a pre-set pipeline topology map and hydraulic model, performs simulation within milliseconds to calculate the optimal response plan. This plan is a structured valve-closing strategy, explicitly including: the valve IDs to be closed, the order of closure, and the closing speed curve for each valve dynamically calculated based on the leakage intensity Q (e.g., slow closure for small leaks to prevent water hammer, and fast closure for large ruptures). The effect of this step is to upgrade the response action from simple valve closure to a systematic strategy execution that considers global safety.

[0045] Phase Four: Precise Execution, Effect Verification, and Model Evolution. Command Issuance and Execution: The generated intelligent valve-closing strategy is instantly transmitted to the target intelligent electric valve via the 5G network. The valve controller parses the command, drives the motor to execute the closing action according to the specified speed curve, and feeds back real-time valve position and torque information to the cloud.

[0046] Effect closed-loop and confirmation: The cloud platform monitors valve feedback signals and pressure changes from sensors near the leak point. Once it is confirmed that the pressure has returned to normal and the leak signal has disappeared, the leak alarm is marked as handled in the twin model, and an event handling report is generated.

[0047] Data archiving and model self-evolution: All data from the entire event, from raw synchronization waveforms, AI intermediate analysis results, location coordinates, decision-making strategies to valve execution feedback, is automatically packaged and stored in the case database. The system periodically uses this newly added real-world case data to incrementally train and fine-tune the lightweight edge model and the ST-CNN model in the cloud. The long-term effect of this closed-loop learning process is that the system's perception capabilities and decision-making accuracy can continuously improve with runtime, adapting to pipeline aging and changes in operating conditions, achieving system self-optimization and evolution.

[0048] The structures shown in the accompanying figures of this embodiment are merely simplified schematic diagrams illustrating the principle of the intelligent valve-closing system. They aim to clearly demonstrate the core architecture and data / control flow logic of the system, comprised of a 5G converged network, cloud-edge collaborative AI computing power, and an intelligent control unit. This does not constitute a limitation on the physical form of the equipment, specific installation methods, or all interface protocols. The technical solution described in this embodiment focuses on the key technical features that play a crucial role in achieving sub-minute-level precise positioning, intelligent decision-making, and second-level valve closure (such as 5G direct transmission synchronous waveforms, ST-CNN spatiotemporal fusion positioning model, AI-based dynamic valve closure strategy, and closed-loop learning mechanism). While auxiliary or conventional engineering details necessary for system implementation in underground coal mines, such as intrinsically safe and explosion-proof housing design, equipment heat dissipation and waterproof structure, specific 5G base station deployment spacing, sensor mounting fixtures, and backup power configuration, are not described in detail, this does not mean that this technical solution does not include or ignore such details, nor does it mean that such conventional design choices necessary for achieving system functions are not within the scope of this technical solution's feasibility. Based on the core concepts disclosed in this invention, and combined with common knowledge in mining safety standards and the fields of communication and automation, those skilled in the art can naturally and reasonably determine and implement these details, and they should all be considered to be covered within the scope of feasibility and protection of this technical solution.

Claims

1. A 5G and artificial intelligence-based system for locating and intelligently shutting off leaking liquids in mining pipelines, characterized in that: include: Multiple sensor nodes are deployed at key nodes of the pipeline to collect vibration and pressure waveform data of the pipeline in real time; A mining 5G communication network is used to transmit waveform data collected by the sensor nodes with low latency and high reliability. The cloud-edge collaborative analytics platform includes: Edge analysis units, deployed underground, are used for real-time processing of data from individual sensor nodes and initial leak detection; The cloud-based intelligent analysis unit, deployed on the ground, is used to receive synchronous waveform data from multiple sensor nodes triggered by the edge analysis unit, and to accurately locate and assess the intensity of the leak point based on a spatiotemporal deep learning model. The intelligent execution unit includes at least one electric valve installed on the pipeline, which is used to receive control commands from the cloud-based intelligent analysis unit and perform valve closing operations; The cloud-based intelligent analysis unit is also used to generate intelligent control commands containing target valve identification and valve closing strategies based on the leak location and assessment results, and to send them to the intelligent execution unit through the 5G communication network.

2. The system according to claim 1, characterized in that, At least some of the multiple sensor nodes are multimodal sensor nodes, integrating vibration sensors and dynamic pressure sensors.

3. The system according to claim 1 or 2, characterized in that, The edge analysis unit has a built-in first artificial intelligence model, which is used to quickly analyze single-point vibration signals to determine suspected leakage events; the cloud intelligent analysis unit has a built-in second artificial intelligence model, which is a spatiotemporal convolutional neural network model, used to fuse and analyze synchronous pressure and vibration waveform data of multiple nodes and output the spatial coordinates of the leakage point.

4. The system according to claim 3, characterized in that, The mining 5G communication network is used to ensure the time synchronization accuracy of the waveform data collected by the multiple sensor nodes, and to provide an end-to-end low-latency communication link to the edge analysis unit and the intelligent execution unit.

5. The system according to claim 1, characterized in that, The valve closing strategy is dynamically generated based on the leakage intensity assessment results, including the valve closing speed curve.

6. The system according to claim 1, characterized in that, The system also includes a case database and a model optimization module; the case database is used to store waveform data, location results and control feedback information related to leakage events; the model optimization module is used to incrementally train the first artificial intelligence model and / or the second artificial intelligence model using the data in the case database.

7. The system according to claim 1, characterized in that, After determining that a suspected leakage event has occurred, the edge analysis unit triggers the associated sensor nodes to enter a high-frequency synchronous acquisition mode.

8. The system according to claim 1, characterized in that, The system is connected to a mining digital twin platform to provide a three-dimensional visualization of leak location results, pipeline status, and valve action information.

Images