A pipeline leakage risk control method, system, medium and product

Abstract

This application provides a risk management method, system, medium, and product for pipeline leakage, relating to the field of electrical digital data processing technology. It utilizes the spatial topology of the pipeline network and the theoretical wave velocity of fluids to calculate the "physical propagation time difference," introducing the wave propagation laws of the physical world as a hard constraint into the algorithm, establishing objective causal relationships between multi-point data. Subsequently, a spatial attention mask generated based on this time difference is used for pre-filtering of multimodal sequences. This mechanism, before the data is input into a deep recurrent neural network, isolates macroscopic operational fluctuations and microscopic environmental vibrations that occur concurrently in absolute time but lack direct physical causal relationships in physical propagation space, from a spatiotemporal perspective. This effectively suppresses irrelevant environmental interference, reducing the false alarm rate.

Description

Technical Field

[0001] This application relates to the field of electrical digital data processing technology, and in particular to a method, system, medium, and product for risk management of pipeline leakage. Background Technology

[0002] Pipeline networks, as key infrastructure for ensuring the transportation of energy and materials, are widely used in the allocation and transportation of oil, natural gas, and chemical raw materials.

[0003] In early pipeline management systems, monitoring relied heavily on single-variable methods (such as pressure sensors alone) for information collection. This resulted in high false alarm and missed alarm rates when the system faced complex operating conditions. Furthermore, when an alarm was triggered, it typically required manual confirmation from the control room and manual remote valve closure. Consequently, the average response time from initial leak detection to final containment was over 20 minutes, during which a significant amount of media was lost, and the emergency response remained passive and delayed for an extended period.

[0004] Subsequent improvements typically rely on hybrid sensor arrays deployed at various nodes of the pipeline network to continuously collect physical sequences covering multiple dimensions, including internal pipeline pressure, external environmental vibration, and temperature. These sequences are then fed into deep learning recurrent neural networks (RNNs) or convolutional neural networks (CNNs). These models, based on their inherent multi-layered neuron mapping weights, concurrently extract cross-modal data combination and correlation features within the global receptive field.

[0005] Because main pipelines often stretch for hundreds or even thousands of kilometers, they inevitably cross various complex geological structures and areas with frequent human activity, such as heavy-duty traffic arteries and large-scale hub pumping stations and pressure regulating centers. The current technology, which mixes and processes global multimodal data within an absolute time window concurrently, suffers from a significantly increased false alarm rate. Specifically, the long and complex physical environment of pipeline networks constantly contains numerous independent and compliant interference events. For example, routine scheduling operations of remote hub pumps and valves objectively induce macroscopic pressure fluctuations propagating along the pipeline axis; while accidental crossings or construction work by heavy-duty machinery (such as heavy trucks) outside the pipeline can trigger strong but extremely localized microscopic pipe wall vibrations. When these two independent events, which have no causal relationship in three-dimensional physical space, accidentally overlap concurrently on the time axis, the existing models will passively capture the high-frequency characteristics of "macroscopic fluid pressure fluctuations" and "localized strong vibrations" without differentiation within the same global input matrix. Ultimately, these spatially unrelated interference events will be incorrectly classified as high-risk leaks by the algorithm, resulting in a high false alarm rate in complex pipeline network operation and maintenance scenarios. Summary of the Invention

[0006] This application provides a method, system, medium, and product for risk management of pipeline leaks, used to reduce false alarm rates.

[0007] In a first aspect, this application provides a risk management method for pipeline leakage, comprising: sampling a pipeline network using sensor nodes to obtain a multimodal physical sequence; performing time alignment and denoising processing on the multimodal physical sequence to obtain a multimodal time series sequence; determining the physical conduction time difference based on the multimodal time series sequence, the pipeline network spatial topology, and the theoretical wave velocity of the fluid; determining a spatial attention mask based on the physical conduction time difference; filtering the multimodal time series sequence using the spatial attention mask to obtain a sub-multimodal time series sequence; inputting the sub-multimodal time series sequence into a preset deep recurrent neural network to obtain a dynamic leakage confidence index; and generating and sending a closed-loop blocking control command when the dynamic leakage confidence index exceeds a preset warning threshold.

[0008] By employing the aforementioned technical solution, the "physical propagation time difference" is calculated using the spatial topology of the pipeline network and the theoretical wave velocity of fluids. This introduces the wave propagation laws of the physical world as a hard constraint into the algorithm, establishing an objective causal relationship between multi-point data. Subsequently, a spatial attention mask generated based on this time difference is used for pre-filtering of multimodal sequences. This mechanism, before the data is input into the deep recurrent neural network, isolates macroscopic operational fluctuations and microscopic environmental vibrations that occur concurrently in absolute time but lack direct physical causal connections in the physical propagation space. This effectively suppresses irrelevant environmental interference, reducing the false alarm rate.

[0009] In conjunction with some embodiments of the first aspect, in some embodiments, the steps of determining the physical conduction time difference based on the multimodal time series, the pipeline spatial topology, and the theoretical fluid wave velocity, and determining the spatial attention mask based on the physical conduction time difference, specifically include: extracting the fluid wave temporal components from the multimodal time series; calculating the ratio of the spatial distance to the preset theoretical fluid wave velocity to obtain the physical conduction time difference of the fluid wave temporal components transmitted between sensing nodes; using the physical conduction time difference as a sliding alignment time window, performing time axis compensation splicing operations on the fluid wave temporal components collected by upstream and downstream sensing nodes to generate a wave spatiotemporal alignment matrix; performing differential derivation operations on the wave spatiotemporal alignment matrix to obtain the leakage source coordinate distribution tensor where the upstream and downstream wave features intersect and overlap; and transforming the leakage source coordinate distribution tensor through a feature mapping function to obtain the spatial attention mask.

[0010] By employing the above technical solution, and using the pre-calculated physical propagation time difference as a sliding alignment time window, time-axis compensation splicing is performed on the wave sequences belonging to upstream and downstream nodes. This reverse compensation cancels the inherent time lag effect caused by the long-distance propagation of physical pressure waves in the medium, aligning the spatially discrete waveforms with the same frequency. Based on this alignment matrix, differential operations can be performed to objectively reveal the physical origin calibration (leakage source coordinate tensor) of the intersection and overlap of upstream and downstream waves.

[0011] In conjunction with some embodiments of the first aspect, in some embodiments, the step of inputting the sub-multimodal time series sequence into a preset deep recurrent neural network to obtain a dynamic leakage confidence index specifically includes: performing dimensionality reduction on the sub-multimodal time series sequence to obtain several single-modal time series sequences; inputting the single-modal time series sequence and the corresponding historical state background data into the deep recurrent neural network to obtain future trend prediction data; calculating the difference between the future trend prediction data and the single-modal time series sequence to obtain a multidimensional feature residual matrix; and mapping and compressing the multidimensional feature residual matrix to a predetermined interval to obtain the dynamic leakage confidence index.

[0012] By employing the aforementioned technical solution, the single-modal time series is fused with corresponding historical background data within the deep network. This enables the algorithm to not only possess memory inertia but also to deduce future trend predictions under normal pipeline operation conditions. Based on this, the difference between the actual occurrence sequence and the predicted trend features is calculated to generate a multi-dimensional feature residual matrix. This allows the discovery of subtle leakage events that were originally hidden beneath macroscopic fluctuations and exhibited slow curves, thereby improving the model's sensitivity to capturing early subtle leakage events.

[0013] In conjunction with some embodiments of the first aspect, in some embodiments, after the step of inputting the sub-multimodal time series sequence into a preset deep time recurrent neural network to obtain the dynamic leakage confidence index, the method further includes: training local historical data using edge computing nodes to obtain edge local network gradients; calculating and allocating fusion weight coefficients based on the data signal-to-noise ratio corresponding to each edge computing node; using the fusion weight coefficients to perform homomorphic encryption and parameter aggregation accumulation on multiple edge local network gradients to generate a new feature calibration benchmark; and distributing the new feature calibration benchmark to different edge computing nodes.

[0014] By adopting the above technical solution, edge computing terminals in various locations independently extract gradients based on locally evolving physical parameters, ensuring continuous model upgrades. When converging to the cloud, the system does not blindly equate the parameters of each node, but dynamically allocates fusion weight coefficients based on the data signal-to-noise ratio, ensuring that only high-quality, high-confidence physical evolution experience can guide the upgrade of the global model.

[0015] In conjunction with some embodiments of the first aspect, in some embodiments, the step of generating and sending a closed-loop blocking control command when the dynamic leakage confidence index exceeds a preset warning threshold specifically includes: when the dynamic leakage confidence index exceeds a preset warning threshold, calculating the topological spatial correlation coefficient between the sensor node and the upstream and downstream sensor nodes based on the pipeline network topology; determining whether the topological spatial correlation coefficient conforms to a preset overall translation feature rule; if it does not conform to the overall translation feature rule, generating and sending a closed-loop blocking control command; if it conforms to the overall translation feature rule, determining the multimodal physical sequence as local historical data, and executing the step of training the local historical data using edge computing nodes to obtain the edge local network gradient.

[0016] By employing the above technical solution, the system calculates the topological spatial correlation coefficient between upstream and downstream sensor nodes before triggering the closed-loop interruption control command. If the correlation analysis shows that the pipeline network characteristics exhibit an overall translational pattern, the anomaly is determined to be a global baseline drift caused by environmental factors, rather than physical damage. Based on this logic, the system intercepts erroneously triggered closed-loop interruption control commands and automatically converts the drift data for that period into high-value negative samples for local model retraining. This closed-loop mechanism, which uses macroscopic physical topological laws as algorithmic constraints, endows the system with the ability to adapt to environmental changes and intercepts large-scale false positive shutdown false alarms induced by pipeline aging and environmental changes at the decision-making end.

[0017] In conjunction with some embodiments of the first aspect, in some embodiments, the step of generating and sending a closed-loop blocking control command when the dynamic leakage confidence index exceeds a preset warning threshold specifically includes: when the dynamic leakage confidence index falls within a first warning interval, generating a silent patrol work order and sending the silent patrol work order containing the coordinate anchor point information of the alarm pipe section to the inspection terminal device; when the dynamic leakage confidence index falls within a second verification interval, generating a multimodal evidence collection cross command, using the multimodal evidence collection cross command to drive the peripheral detection equipment to the target coordinate to perform acoustic and optical evidence collection and data feedback comparison; when the dynamic leakage confidence index is greater than the blocking defense threshold, generating a closed-loop blocking control command, driving the upstream and downstream valves of the pipeline to perform electric shut-off displacement, and simultaneously sending a high-level signal to start the pressure relief pump group configured by the system; the blocking defense threshold is greater than the second verification interval and greater than the first warning interval.

[0018] By adopting the above technical solution and establishing a three-tiered progressive response interval, silent patrols are triggered in the primary warning interval, or external equipment is dispatched for multimodal cross-verification in the intermediate review interval. These two pre-buffer layers construct fault tolerance and eliminate the risk of erroneous production shutdowns caused by occasional interference signals through low-cost external verification. When the confidence index exceeds the highest blocking defense threshold, the mechanical valve position is forcibly locked and the pressure relief pump group is started through the underlying electrical command, achieving the dual goals of minimizing the false alarm rate and maximizing disaster control efficiency.

[0019] In conjunction with some embodiments of the first aspect, in some embodiments, before the step of sampling the pipeline network using sensor nodes to obtain a multimodal physical sequence, the method further includes: adjusting the adhesion force angle of the sensor node on the pipe wall using a laser collimation and alignment device; confirming the physical calibration reference surface when the deviation of the adhesion force angle is determined to be within a preset tolerance range; injecting a test fluid medium into the pipeline network and applying an unloading oscillation step signal via a bypass; calculating the overall transmission time from the moment the unloading oscillation step signal is generated to the moment the closed-loop blocking control command is calculated and output; when it is determined that the overall transmission time meets the preset safety delay constraint and the local data sample size is empty, calling a preset domain adaptive mapping algorithm to extract the fusion weight parameters of the historical pipeline operation library; injecting the fusion weight parameters into a deep time recurrent neural network to generate a priori network weight base, and performing the step of sampling the pipeline network using sensor nodes to obtain a multimodal physical sequence.

[0020] By adopting the above technical solutions, before the system is put into use, the sensor attachment angle is physically calibrated using a laser collimation device, eliminating signal scattering and distortion caused by off-axis effects at the source of perception. A step signal is applied to measure the end-to-end time delay, matching the communication link with the algorithm processing. Addressing the "cold start" dilemma of newly built pipelines failing to converge due to a lack of historical data, a domain adaptive algorithm is used to inject prior weight parameters from older pipelines with similar operating conditions. This allows newly deployed systems to possess fault diagnosis capabilities without a lengthy training and adjustment period.

[0021] In a second aspect, this application provides a risk management system for pipeline leakage, which includes: one or more processors and a memory; the memory is coupled to one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to cause the risk management system for pipeline leakage to perform the method described in the first aspect and any possible implementation thereof.

[0022] Thirdly, this application provides a computer program product containing instructions that, when run on a pipeline leakage risk management system, cause the pipeline leakage risk management system to perform the method described in the first aspect and any possible implementation thereof.

[0023] Fourthly, this application provides a computer-readable storage medium including instructions that, when executed on a pipeline leakage risk management system, cause the pipeline leakage risk management system to perform the method described in the first aspect and any possible implementation thereof.

[0024] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0025] 1. By calculating the "physical propagation time difference" using the spatial topology of the pipeline network and the wave velocity in fluid theory, the wave propagation laws of the physical world are introduced into the algorithm as a hard constraint, establishing an objective causal relationship between multi-point data. Subsequently, a spatial attention mask generated based on this time difference is used for pre-filtering of multimodal sequences. This mechanism, before the data is input into the deep recurrent neural network, isolates macroscopic operational fluctuations and microscopic environmental vibrations that occur concurrently in absolute time but lack a direct physical causal relationship in the physical propagation space. This effectively suppresses irrelevant environmental interference and reduces the false alarm rate. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a risk management method for pipeline leakage in an embodiment of this application;

[0027] Figure 2 This is another flowchart illustrating the risk management method for pipeline leakage in the embodiments of this application;

[0028] Figure 3 This is another flowchart illustrating the risk management method for pipeline leakage in the embodiments of this application;

[0029] Figure 4 This is an exemplary hardware structure diagram of a pipeline leakage risk management system in the embodiments of this application. Detailed Implementation

[0030] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.

[0031] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0032] Please see Figure 1 , Figure 1 This is a flowchart illustrating a risk management method for pipeline leakage in an embodiment of this application;

[0033] A risk management method for pipeline leakage includes:

[0034] S101. Use sensor nodes to sample the pipeline network and obtain multimodal physical sequences;

[0035] Among them, a sensing node refers to a data acquisition terminal that is deployed on the outside of the pipe wall or inside the valve chamber and integrates multiple sensing units and communication modules; a multimodal physical sequence refers to a set of continuous analog or digital data that reflects the operating status of the pipeline, collected by different types of sensors such as pressure, vibration, and temperature at the same time reference.

[0036] In some embodiments, the sensing node is attached to the outer wall of the pipe by magnetic attraction.

[0037] S102. Perform time alignment and denoising on the multimodal physical sequence to obtain a multimodal time series sequence;

[0038] Among them, time alignment refers to the process of correcting the timestamps of scattered data frames based on a unified clock protocol to keep them synchronized on the time axis; denoising processing is used to represent the operation of removing environmental background noise and non-stationary random interference through time domain filtering and frequency domain analysis; multimodal time series refers to a high-quality data stream with a unified spatiotemporal reference after time synchronization and signal enhancement processing.

[0039] S103. Determine the physical conduction time difference based on the multimodal time series, pipeline spatial topology, and theoretical fluid wave velocity;

[0040] Among them, the pipeline spatial topology represents a vector data structure that digitally describes the geographical location, connection relationship and length attribute of the pipeline; the fluid theoretical wave velocity refers to the theoretical propagation speed of pressure waves under specific media, pipe materials and operating conditions; and the physical transmission time difference refers to the theoretical time difference generated by the propagation of the wave generated by the same physical disturbance source to different sensor nodes.

[0041] Specifically, the negative pressure wave generated by the pipeline leak propagates bidirectionally upstream and downstream along the pipeline, causing sensors at different locations to receive signals at different times. The algorithm combines Geographic Information System (GIS) data to calculate the pipeline distance between each node and, based on the wave velocity estimated from the real-time operating conditions of the medium, calculates the theoretical time delay of signals from the same wave source reaching different nodes, providing a physical benchmark for subsequent spatiotemporal correlation analysis.

[0042] In some specific embodiments, the path length is calculated by extracting the mileage coordinates of the pipeline segment from the database, and combined with the medium wave velocity constant under standard operating conditions, the static theoretical time difference is directly calculated using the linear ratio between the path length and the wave velocity. This is not limited here.

[0043] In actual use, in complex terrains (such as mountains and urban pipe networks), the actual pipeline laying path contains a large number of bends and undulations. Directly using the straight-line distance on the ground will lead to errors in time difference estimation.

[0044] In some preferred embodiments, a three-dimensional digital twin model or BIM model containing pipe fitting details is introduced. The algorithm performs calculus path accumulation along the pipeline centerline trajectory to obtain the fluid transmission distance, which is used to replace the straight-line distance to ensure the physical accuracy of time difference calculation.

[0045] In some embodiments, step S103 specifically includes:

[0046] S1031. Extract the fluid wave time series components from the multimodal time series;

[0047] Among them, the fluid fluctuation time series component is used to represent a subset of data extracted from a dataset containing multiple physical properties, which specifically characterizes the pressure pulsation and macroscopic oscillation characteristics of the medium inside the pipe caused by forced pressure.

[0048] Specifically, after collecting a complete dataset including environmental background noise, local thermal disturbances, and mechanical vibrations, the system utilizes signal separation technology to remove thermal hysteresis changes and local high-frequency friction signals from the pipe wall, targeting the negative pressure wave suction effect and fluid dynamic characteristics caused by the leakage accident. This step aims to accurately extract the core sequence reflecting the energy transfer and pressure transmission of the fluid within the pipe from the mixed signal, providing a high signal-to-noise ratio input for subsequent spatiotemporal correlation analysis.

[0049] S1032. Based on the spatial distance between adjacent sensing nodes in the pipeline network topology, and combined with the preset theoretical fluid wave velocity calculation ratio, the physical transmission time difference of the fluid wave time sequence component between sensing nodes is obtained.

[0050] Among them, the spatial distance between adjacent sensing nodes refers to the physical pipeline length between two acquisition terminals obtained by mapping based on the actual pipeline laying path; the preset fluid theoretical wave velocity represents the theoretical propagation speed of pressure waves in the pipe derived from the physical properties of the current transported medium (such as crude oil and natural gas) and the pipeline operating conditions; the physical transmission time difference is used to represent the theoretical time difference that objectively exists when the wave generated by the same physical disturbance source propagates along the pipeline axis to two sensors at different locations.

[0051] Specifically, given that the negative pressure wave generated by pipeline leakage propagates bidirectionally upstream and downstream along the pipeline, there is a time lag in the signal received by sensors at different locations. The algorithm extracts three-dimensional mileage data from the pipeline geographic information system (GIS), combines it with the medium wave velocity corrected for temperature and pressure effects, and uses the ratio of transmission distance to propagation speed to calculate the theoretical time difference for the same wave source to reach different sensor nodes, providing a physical benchmark for spatiotemporal alignment.

[0052] S104. Determine the spatial attention mask based on the physical conduction time difference;

[0053] Spatial attention mask refers to a matrix or vector generated based on physical time difference, used to mark suspected leakage areas in global data and suppress the weights of irrelevant areas.

[0054] Specifically, to avoid interference from globally irrelevant data, the system uses physical transmission time difference as an alignment benchmark to perform time axis compensation (shifting) on ​​upstream and downstream data sequences. When the compensated upstream and downstream signal characteristics (such as voltage drop and waveform correlation) exhibit maximum correlation, their corresponding spatial location is the coordinate of the suspected leak point. Based on this coordinate, the system generates a spatially weighted mask, highlighting data in the suspected area of ​​interest and suppressing the weight of data in distant areas.

[0055] In some specific embodiments, the source coordinates are locked using the signal reciprocity peak after time difference compensation, and a rectangular window function mask is constructed with these coordinates as the center.

[0056] In some other specific embodiments, a soft attention mask is generated with the positioning point as the expectation and the positioning error as the variance, which has the highest center weight and decays smoothly outwards. This is not limited here.

[0057] In actual use, the unidirectional pressure wave generated by the operation of remote pump valves may cause false correlation peaks, resulting in the mask pointing to the wrong location.

[0058] In some preferred embodiments, bidirectional propagation verification logic is introduced before generating the mask. The system extracts the signal waveforms received by the upstream sensor and the downstream sensor respectively, and calculates the peak value of their cross-correlation function after time axis alignment. If the upstream waveform lags behind the downstream waveform and has a high correlation, or the downstream waveform lags behind the upstream waveform and has a high correlation (one-way delay), it is determined to be unidirectional interference. Only when both the upstream and downstream signals exhibit lag characteristics relative to a certain intermediate point (bidirectional divergence), i.e., confirming that the waveforms have the characteristics of bidirectional diffusion from the same source, is the generation of a leakage concern mask allowed. If the waveforms only exhibit unidirectional transmission characteristics (consistent with pump and valve operation rules), it is determined to be interference, and the generation of a leakage concern mask is prohibited.

[0059] In some specific embodiments, step S104 specifically includes:

[0060] S1041. Using the physical conduction time difference as a sliding alignment time window, perform time axis compensation splicing operation on the fluid fluctuation time series components collected by upstream and downstream sensing nodes to generate a fluctuation spatiotemporal alignment matrix.

[0061] Among them, the sliding alignment time window refers to the adjustment window constructed based on the physical transmission time difference, which is used to translate data on the time axis to achieve window overlap; the time axis compensation splicing operation represents the process of time displacement compensation for multiple sets of data sequences with physical time delays to offset transmission delays and achieve feature synchronization; the fluctuation spatiotemporal alignment matrix is ​​used to represent the data matrix in which the upstream and downstream sequences achieve waveform side by side alignment on the logical time axis after time difference compensation.

[0062] Specifically, due to the physical time difference between the leakage waveforms received by the upstream and downstream sensors, the system uses the calculated physical propagation time difference as a compensation. The algorithm shifts the downstream waveform data, which is received later, forward on the time axis (or delays the upstream data) until the two waveforms are aligned at the node on the compensated logical time plane. This process eliminates the phase difference caused by physical propagation and generates an alignment matrix for the neural network to analyze causal relationships.

[0063] In some specific embodiments, a timestamp backwards method is used to construct a virtual buffer queue based on a standard time axis. The timestamp tail of the downstream fluctuation sequence is deducted (subtracting the time difference offset), and the compensated valid data frames are sequentially filled into the corresponding upstream time slots to generate a synchronization matrix.

[0064] In other specific embodiments, the upstream waveform is used as the anchoring reference, and a coarse sliding window is set using the physical propagation time difference. Within the window, a sliding cross-correlation calculation is performed on the downstream waveform to find the extreme point of the correlation coefficient as a fine alignment anchor point, thus completing the matrix stitching. This is not limited here. It should be noted that if the sampling frequencies at both ends are inconsistent, resampling or interpolation processing needs to be performed in advance to ensure data density matching.

[0065] S1042. Perform differential derivation on the wave spatiotemporal alignment matrix to obtain the leakage source coordinate distribution tensor where the upstream and downstream wave characteristics intersect and overlap.

[0066] Among them, differential inference operation is used to represent the algorithm of extracting features (such as slope and phase) along the opposite propagation direction of the aligned wave matrix and performing differential comparison to locate the physical intersection point; the leakage source coordinate distribution tensor is a multidimensional array containing pipeline distance coordinate probabilities and their confidence weights, and its extreme points indicate the specific location of physical damage.

[0067] Specifically, after completing the spatiotemporal alignment of the wave data, the algorithm uses an inference engine to simulate the backward propagation and attenuation process of the wave within the pipe wall. By calculating the difference between the forward distortion of the upstream wave and the backward distortion of the downstream wave, the algorithm finds the extreme point (minimum difference point) of the fitted feature vector at the virtual spatial node. This node is then identified as the common physical origin of the two waves, and the system transforms it into a coordinate distribution tensor carrying spatial location probabilities.

[0068] In some specific embodiments, the system adopts the wave velocity gradient iterative approximation method, establishes a virtual source point traversal table along the pipeline in the alignment matrix, calculates the attenuation compensation coefficient from each source point to the sensors at both ends, compares the sum of differences of the compensated waveforms, and determines the coordinates of the source point with the smallest sum of differences as the leakage source distribution tensor, which is not limited here.

[0069] S1043. The spatial attention mask is obtained by transforming the coordinate distribution tensor of the leakage source through the feature mapping function.

[0070] Among them, the feature mapping function refers to the nonlinear mathematical transformation formula that converts spatial coordinate probabilities into neural network input weights; the spatial attention mask refers to a matrix or filter generated based on the mapping, used to highlight suspected leakage areas and suppress the weights of irrelevant areas.

[0071] Specifically, after calculating the wave source coordinate distribution, the system transforms it into a spatial attention mask overlaying the original data stream using a feature mapping function. The mask assigns a high weight close to 1 at the center of the wave source coordinates, and the weight decreases with increasing distance (e.g., a cliff-like or bell-shaped curve). This mask gives the system the ability to "spatial focus," guiding subsequent models to focus on data features in high-probability leakage areas.

[0072] As can be seen, by using the pre-calculated physical propagation time difference as a sliding alignment time window, time-axis compensation splicing is performed on the wave sequences belonging to upstream and downstream nodes. This reverse compensation cancels the inherent time lag effect caused by the long-distance propagation of physical pressure waves in the medium, aligning the spatially discrete waveforms with the same frequency. Based on this alignment matrix, differential operations can be performed to objectively reveal the physical origin calibration (leakage source coordinate tensor) of the intersection and overlap of upstream and downstream waves.

[0073] S105. Use spatial attention masking to filter multimodal time series sequences to obtain sub-multimodal time series sequences;

[0074] Among them, filtering refers to the operation of using mask weights and the original multimodal sequence to perform weighted operations, retain the key regional features and remove background noise; the sub-multimodal time series is used to represent the data subset that, after spatial domain filtering, mainly contains the physical features (such as local vibration and temperature changes) around the suspected leak point.

[0075] Specifically, the system performs element-wise multiplication or weighted operations on the global multimodal sequence and the spatial attention mask. Data within the high-weight region of the mask (i.e., data near suspected leakage points) is preserved, while data within the low-weight region (such as distant environmental noise) is suppressed or set to zero. This processing effectively separates local leakage features from global background interference, improving the input signal-to-noise ratio of subsequent models.

[0076] In some specific embodiments, a hard threshold mask is used to retain only the original sensor data within the spatial / temporal window covered by the mask, while discarding the data outside the window, thereby achieving data truncation. This is not limited here.

[0077] In other specific embodiments, a soft-weight mask is used to attenuate the data amplitude based on the distance from the suspected point, while retaining some surrounding background information as a reference to achieve smooth filtering. This is not limited here.

[0078] S106. Input the sub-multimodal time series into a preset deep time recurrent neural network to obtain the dynamic leakage confidence index;

[0079] Among them, the deep time recurrent neural network represents a deep learning model (such as LSTM and GRU) that can process sequential data and capture long-short-term dependencies; the dynamic leakage confidence index refers to the probability value (0-1 or 0-100%) output by the model that represents the current pipeline state as a leakage state.

[0080] Specifically, the system inputs spatially filtered and frequency-enhanced subsequences (containing dimensions such as pressure, vibration, and temperature) into a pre-trained deep recurrent neural network. The network not only analyzes the signal characteristics at the current moment but also performs temporal reasoning by combining historical states with its internal memory units. The model comprehensively evaluates the degree of deviation of the current signal from the historical baseline and outputs the confidence level of pipeline leakage through nonlinear mapping.

[0081] In some embodiments, step S105 specifically includes:

[0082] S1061. Dimensional reduction and deconstruction of the sub-multimodal time series sequence yields several single-modal time series sequences;

[0083] Among them, the sub-multimodal time series is used to represent the high signal-to-noise ratio data block around the suspected leak point, which contains multiple physical attributes (pressure, vibration, temperature), after being filtered by spatial attention mask; dimensionality reduction refers to the action of unbinding and separating multi-dimensional coupled data according to the physical detection source using matrix decomposition technology; the single-modal time series represents the independent data sequence that records the evolution law of a single physical field variable after decomposition.

[0084] Specifically, to avoid network gradient problems caused by differences in dimensions and physical properties (such as pressure in MPa and temperature in °C) when mixed inputs, the system performs a deconstruction operation on the sub-multimodal sequences. Based on the channel identifier of the data frame, the intertwined pressure, vibration, and temperature data are peeled off layer by layer and sorted into a parallel sequence array with a single physical meaning and mutual independence, in preparation for subsequent refined feature analysis.

[0085] S1062. Input the single-modal time series and the corresponding historical state background data into a deep time recurrent neural network to obtain future trend prediction data;

[0086] Among them, historical background data is used to represent historical benchmark data collected during normal operation cycles that reflect the standard physical fluctuation patterns of the pipe section under specific seasons and operating conditions; deep time recurrent neural network refers to a deep learning architecture with gated memory mechanism (such as LSTM, GRU) that can handle long and short-term time series dependencies and perform evolutionary inference; future trend prediction data refers to the theoretical waveform curve that the pipe section should present in future time steps under normal operating conditions, which is deduced by the network based on historical benchmarks and current inputs.

[0087] Specifically, the system utilizes a deep recurrent neural network to process single-modal sequences and simultaneously loads historical health background data for the pipeline segment. The network uses gating units to combine historical memory (long-term patterns) with current input (short-term state) to predict the expected physical trajectory under current operating conditions without any anomalies (i.e., the prediction baseline). This prediction data serves as an objective reference for determining whether the current input deviates from the normal state.

[0088] This step leverages the nonlinear mapping capabilities of deep networks, essentially performing a high-dimensional phase space reconstruction. The network maps one-dimensional single-modal time series data into a high-dimensional feature space to find the trajectory of the sequence in the phase space (i.e., the dynamic evolution manifold of the fluid).

[0089] Historical background data provides the constraints and boundaries for this high-dimensional phase space (i.e., limits the range of trajectories for normal operation). When a new sequence is input, the network actually calculates the next landing point of the trajectory line within this limited phase space, based on the current trajectory point position and along the tangent direction of the manifold. This landing point is projected back into one-dimensional space, which is the future trend prediction data.

[0090] S1063. Calculate the difference between the future trend prediction data and the single-modality time series to obtain the multidimensional feature residual matrix;

[0091] Here, the difference refers to the absolute deviation vector between the predicted data under normal operating conditions and the currently collected data; the multidimensional feature residual matrix is ​​used to represent the tensor structure containing cross-dimensional abnormal mutation signals that cannot be explained by safety experience after the normal scheduling fluctuation background is removed by subtraction operation.

[0092] Specifically, the algorithm performs a subtraction operation of the same dimension, subtracting the "future trend prediction data" generated by the network from the currently measured single-modal sequence. This process effectively filters out legitimate and predictable scheduling fluctuations (such as pressurization), and the remaining residuals are the abnormal components that deviate from expectations (such as small pressure drops and high-frequency vibrations). This matrix is ​​composed of residual vectors of multiple physical dimensions, intuitively reflecting the degree of variation in the pipeline state.

[0093] S1064. The multidimensional feature residual matrix is ​​mapped and compressed to a predetermined interval to obtain the dynamic leakage confidence index.

[0094] Among them, mapping compression refers to using the nonlinear activation function at the end of the neural network to force the scalar value of the residual matrix, whose value range may diverge, to be mapped to a standardized numerical range; the predetermined range is used to represent the normalized decision space (such as 0-1 or 0-100%), where 0 represents absolute safety and 1 represents deterministic leakage; the dynamic leakage confidence index is a floating-point index calculated by the model based on the residual matrix, which reflects the probability of pipeline leakage in real time.

[0095] Specifically, the system uses a fully connected layer to aggregate the multidimensional feature residual matrix into a scalar score, and then maps it to the 0-1 range using sigmoid and other sigmoid functions. Under this mapping mechanism, small background noise residuals are compressed to near 0, while significant and cross-modal consistent anomalous residuals (such as voltage drop combined with acoustic emission) will exceed the threshold and approach 1, intuitively quantifying the current leakage risk.

[0096] It is evident that within the deep network, the fusion of single-modal time series with accompanying historical background data not only endows the algorithm with memory inertia but also enables it to deduce future trend predictions under normal pipeline operation conditions. Based on this, the difference between the actual occurrence sequence and the predicted trend features is calculated to generate a multi-dimensional feature residual matrix. This allows the discovery of subtle leakage events that were originally hidden beneath macroscopic fluctuations and exhibited slow curves, thereby improving the model's sensitivity to capturing early subtle leakage events.

[0097] S107. If the dynamic leakage confidence index exceeds the preset warning limit, generate and send a closed-loop blocking control command.

[0098] Among them, the closed-loop blocking control command refers to the control signal automatically generated by the system to drive field actuators (such as valves and pumps) to perform physical isolation and pressure relief; the preset warning limit represents the confidence threshold for triggering automatic shutdown protection.

[0099] Specifically, the system monitors the leakage confidence index in real time. When the index is in the low-risk range, it only records or alerts; when the index exceeds the extreme danger threshold (e.g., 85%), the system determines it as a deterministic leakage event. At this point, the algorithm logic bypasses the manual confirmation process and directly generates high-priority control commands, which are sent to the field PLC via the industrial bus to drive the upstream and downstream shut-off valves to close and initiate the pressure relief process, achieving automatic response.

[0100] It is evident that by calculating the "physical propagation time difference" using the spatial topology of the pipeline network and the wave velocity in fluid theory, the wave propagation laws of the physical world are introduced into the algorithm as a hard constraint, establishing an objective causal relationship between multi-point data. Subsequently, a spatial attention mask generated based on this time difference is used for pre-filtering of multimodal sequences. This mechanism, before the data is input into the deep recurrent neural network, isolates macroscopic operational fluctuations and microscopic environmental vibrations that occur concurrently in absolute time but lack direct physical causal relationships in the physical propagation space. This effectively suppresses irrelevant environmental interference and reduces the false alarm rate.

[0101] In actual use, the pipe wall thickness gradually decreases due to the continuous scouring and abrasion caused by solid particles carried by the fluid inside the pipe. This leads to physical variations in its inherent acoustic resonance frequency and vibration propagation mode. This scenario causes a deviation between the actual operating state of the pipeline network and the original training baseline of the model, resulting in a decrease in the model's ability to recognize normal operating conditions, a phenomenon known as "model drift."

[0102] Please see Figure 2 , Figure 2 This is another flowchart illustrating the risk management method for pipeline leakage in the embodiments of this application;

[0103] Therefore, in some embodiments, after step S106, the method further includes:

[0104] S201. Use edge computing nodes to train local historical data to obtain the edge local network gradient;

[0105] Among them, edge computing nodes refer to industrial-grade gateways or servers deployed in valve chambers or pump stations, possessing independent computing power and storage capabilities, and responsible for performing localized data processing and model training; local historical data refers to the historical operating data stored by the edge nodes, reflecting the specific environment and operating conditions of the pipe section where the node is located; edge local network gradient refers to the partial derivative direction vector calculated by the edge nodes after backpropagating the model using local data, which is used to guide the updating of model parameters.

[0106] Specifically, to address the decline in generalization ability of the standard model caused by environmental differences and seasonal variations across different sections of the pipeline network, edge nodes utilize locally accumulated historical data (such as temperature and pressure curves for specific seasons) for self-training. By fine-tuning the model parameters locally, the nodes calculate the network gradient update direction adapted to local characteristics, achieving personalized model calibration.

[0107] S202. Calculate and allocate fusion weight coefficients based on the data signal-to-noise ratio of each edge computing node;

[0108] Among them, the data signal-to-noise ratio refers to the ratio of the effective signal energy to the background noise energy of the data collected by the edge nodes, reflecting the purity and value of the data; the fusion weight coefficient is used to represent the contribution or confidence weight of the gradient update package assigned to each edge node when the global model is aggregated.

[0109] Specifically, to prevent low-quality data (such as high-noise environmental data) from dragging down the performance of the global model, the system dynamically allocates fusion weights based on the signal-to-noise ratio (SNR) of each node. High-quality data nodes receive high weights, while low-quality data nodes have their weights reduced, ensuring that the evolution of the global model is driven by high-quality data.

[0110] S203. Homomorphic encryption and parameter aggregation accumulation are performed on the gradients of multiple edge local networks using fusion weight coefficients to generate a new feature calibration benchmark.

[0111] Among them, homomorphic encryption is used to represent an encryption technology that allows direct arithmetic operations (such as weighted summation) in the ciphertext state; the new feature calibration benchmark refers to the global model parameter update matrix that integrates the latest operating features of the entire network after weighted aggregation and decryption verification.

[0112] Specifically, the cloud server acts as the aggregation center, using assigned weight coefficients to weight and accumulate the encrypted gradients uploaded by each edge node (FedAvg algorithm) without decryption (to protect privacy). This process integrates the experience of all nodes across the network to generate an updated global model baseline.

[0113] S204. Distribute the new feature calibration benchmark to different edge computing nodes.

[0114] Among them, the new feature calibration benchmark is the latest global model parameters; distribution refers to the action of synchronizing the parameter matrix to each edge node through the communication network.

[0115] Specifically, the cloud will broadcast the latest global model parameters (new baseline) to all edge nodes via fiber optic or wireless network. Upon receiving the data, the edge nodes will update their local models, thus achieving a synchronous improvement in the network-wide monitoring capabilities.

[0116] As can be seen, the independent extraction of gradients by edge computing terminals in various locations based on locally evolving physical parameters ensures the continuous upgrading of the model. When converging to the cloud, the system does not blindly equate the parameters of each node, but dynamically allocates fusion weight coefficients based on the data signal-to-noise ratio, ensuring that only high-quality, high-confidence physical evolution experience can guide the upgrading of the global model.

[0117] Step S107 specifically includes:

[0118] S1071. When the dynamic leakage confidence index exceeds the preset warning limit, calculate the topological spatial correlation coefficient between the sensing node and the upstream and downstream sensing nodes based on the pipeline spatial topology.

[0119] Among them, the pipeline spatial topology refers to the vector data structure constructed based on geographic information that describes the physical location relationship and connectivity of sensor nodes; the topology spatial correlation coefficient is used to represent the degree of statistical correlation between the signal waveform collected by the current alarm sensor node and the signal waveform of its physically connected upstream and downstream adjacent nodes within a specific time window. This coefficient usually takes a value range of -1 to 1, reflecting the synchronicity and causality of waveform changes.

[0120] Specifically, when the dynamic leakage confidence index calculated by the system through S106 exceeds the safety threshold (e.g., 90%), the algorithm initiates secondary verification logic before triggering irreversible physical blocking actions. The system extracts the pressure change sequence of the current alarm node and retrieves the upstream and downstream sensor nodes directly connected to it based on the pipeline network topology. Subsequently, the algorithm calculates the Pearson correlation coefficient or cross-correlation function of the pressure sequences of the alarm node and its neighboring nodes within the same time period. This step aims to verify whether the current high confidence level possesses the spatial propagation characteristics unique to fluid mechanics (such as negative correlation or delayed correlation caused by the bidirectional propagation of negative pressure waves) through a macroscopic spatial dimension physical correlation check, thereby distinguishing between local single-point anomalies and systemic changes across the entire pipeline.

[0121] S1072. Determine whether the correlation coefficient of the topological space conforms to the preset overall translation feature rule;

[0122] Among them, the preset overall translation feature rule is used to represent a specific spatial correlation pattern, that is, the physical parameters (such as pressure and temperature) of all nodes (or nodes in a large area) along the entire line show a highly positive correlation and a synchronous rise and fall trend with almost no time difference, which is usually caused by environmental factors (such as a sudden drop in temperature or system pressure regulation) rather than a single point leakage.

[0123] Specifically, the correlation coefficient obtained from system analysis S1071 is examined to determine if it approaches 1 (highly positive correlation). If the fluctuations of the alarm node and the upstream and downstream nodes are found to be highly consistent not only in morphology but also almost synchronous in time (without fluid propagation delay) or exhibit a general baseline drift, then it is determined to meet the "overall translational characteristics." This means that the current pressure anomaly does not originate from the energy release at a specific rupture point (leakage would cause the pressure to be lowest at the source and attenuate to both sides), but rather that the entire pipeline system is uniformly affected by some macroscopic external condition (such as the change in medium density caused by diurnal temperature differences).

[0124] S1073. If the overall translation characteristic rule is not met, a closed-loop blocking control command will be generated and sent down.

[0125] Among them, the closed-loop interruption control command refers to the high-priority electrical signal generated by the central control unit that directly drives the field actuators (such as shut-off valves and vent valves).

[0126] Specifically, when the system confirms that the current confidence index is extremely high (S107 judgment result), and the spatial topology verification in S1072 eliminates the possibility of overall environmental drift (i.e., confirming that the abnormal waveform possesses the spatial heterogeneity and single-point divergence characteristic of leakage), the algorithm determines that this is a physical damage event. At this point, the safety logic bypasses all manual verification steps and immediately generates a sequence of blocking instructions for the pipe segment. The instructions are sent to the field PLC via industrial Ethernet or a dedicated safety bus, forcibly triggering the shut-off actions of upstream and downstream valves to achieve physical isolation.

[0127] S1074. If the overall translation feature rule is met, the multimodal physical sequence will be determined as local historical data, and S201 will be executed.

[0128] Local historical data refers to sample data that has been verified by the system as non-leakage operating conditions and reflects the normal physical characteristics of the pipeline network under specific environments (such as extreme temperatures).

[0129] Specifically, if S1072 determines that the current abnormally high confidence level is due to an overall shift (false positive) caused by environmental factors, the system not only intercepts and blocks the command, but also marks the data sequence that caused the false alarm as a highly valuable "negative sample". This sequence is automatically archived as "local historical data" and immediately triggers step S201 in Example 4. The edge node uses this batch of newly identified environmental drift data to retrain the local model (fine-tuning), enabling the model to learn to recognize and adapt to this specific environmental change, so that it will no longer falsely report when encountering similar working conditions in the future, achieving the system's self-immunity and evolution.

[0130] As can be seen, before triggering the closed-loop interruption control command, the system calculates the topological spatial correlation coefficient between upstream and downstream sensor nodes. If the correlation analysis shows that the pipeline network characteristics exhibit an overall translational pattern, the anomaly is determined to be a global baseline drift caused by environmental factors, rather than physical damage. Based on this logic, the system intercepts erroneously triggered closed-loop interruption control commands and automatically converts the drift data for that period into high-value negative samples for local model retraining. This closed-loop mechanism, which uses macroscopic physical topological laws as algorithmic constraints, endows the system with the ability to adapt to environmental changes and intercepts large-scale false positive shutdown false alarms induced by pipeline aging and environmental changes at the decision-making end.

[0131] In actual use, when encountering occasional electromagnetic interference or transient signal crosstalk, it is very easy to accidentally trigger an emergency shutdown command for the entire line. Such unnecessary system shutdowns not only lead to a surge in energy consumption and equipment damage during pipeline restart, but also cause economic losses from unplanned production stoppages. Therefore, it is difficult to strike a balance between ensuring safety and maintaining production continuity.

[0132] Therefore, another way to implement S107 is:

[0133] S1076. If the dynamic leakage confidence index falls within the first warning range, generate a silent patrol work order and send the silent patrol work order containing the coordinate anchor point information of the alarm pipe section to the patrol terminal equipment.

[0134] The first warning interval refers to the range where the confidence index is in a low-risk range (e.g., 60%-75%), indicating that there may be a slight anomaly but it has not yet reached the standard for a definitive leak. The silent patrol work order is used to represent a verification task order that does not trigger an audible or visual alarm, does not interfere with normal production, and is only pushed to specific maintenance personnel through the background. The coordinate anchor information refers to the geographical coordinates (latitude, longitude, and pipeline station number) of the suspected anomaly point.

[0135] Specifically, when the confidence index output by the AI ​​model fluctuates within the low-risk concern zone, the system determines that the current stage is "suspected early leakage." To avoid frequently disturbing the control room operators, the system adopts a "silent monitoring" strategy. The algorithm automatically generates an electronic work order containing the location of the abnormal pipe section, the type of abnormality (such as a slight pressure drop trend), and the time of occurrence, and silently pushes it to the handheld terminal APP of the area inspector via a dedicated mobile network, guiding them to focus on checking this area during the next routine inspection, thus achieving proactive risk screening.

[0136] S1077. If the dynamic leakage confidence index falls within the second review interval, generate a multimodal evidence collection cross command and use the multimodal evidence collection cross command to drive the peripheral detection equipment to the target coordinates to perform acoustic and optical evidence collection and data feedback comparison.

[0137] The second verification interval refers to the range where the confidence index is in the medium risk range (e.g., 75%-85%), indicating that the abnormal characteristics are significant but occasional interference still needs to be excluded; the multimodal forensics cross command is used to represent the control command for scheduling third-party independent detection resources (such as drones and robots) to perform audiovisual verification of specific targets.

[0138] Specifically, when the confidence index climbs to the highly suspected area, the system determines that data deduction alone is insufficient for a definitive conclusion, and empirical evidence from the physical world must be introduced. The algorithm generates scheduling instructions to wake up nearby automated peripheral equipment (such as inspection robots or drone airports along the route). After receiving the instructions, the equipment automatically plans a path to the alarm coordinates. Upon arrival at the site, the equipment activates its onboard high-sensitivity microphone array (to listen for leak sounds) and gas laser telemetry instrument (to detect methane concentration), and transmits the collected audio-visual data back to the cloud in real time. The cloud performs a "cross-comparison" between the transmitted real-time data and the data from the sensors inside the pipe (internal waveform and external sound and light) to determine whether to escalate to the highest alarm level.

[0139] S1078. When the dynamic leakage confidence index is determined to be greater than the blocking defense threshold, a closed-loop blocking control command is generated to drive the upstream and downstream valves of the pipeline to perform electric shut-off displacement, and a high-level signal is sent simultaneously to start the pressure relief pump group configured by the system; the blocking defense threshold is greater than the second review interval and greater than the first warning interval.

[0140] Among them, the blocking defense threshold refers to the critical value at which the confidence index reaches an extremely high risk standard (such as >85%), which almost eliminates all possible false alarms in probability; the motor shut-off displacement represents the physical stroke of the electric valve actuator from the fully open state to the fully closed state.

[0141] It is easy to understand that the threshold of the blocking defense line is greater than the upper limit of the second review interval, and the lower limit of the second review interval is greater than the upper limit of the first warning interval.

[0142] Specifically, when the confidence index exceeds the highest defense line, it indicates that a leak is inevitable and the situation is critical. The system automatically blocks all intermediate confirmation steps. The algorithm immediately generates a sequence of highest-priority ESD (Emergency Stop) commands. On one hand, the commands drive the electric actuators of the line shut-off valves upstream and downstream of the leak point to cut off the medium source; on the other hand, a high-level hardwired signal is simultaneously output to the starter cabinet of the pressure relief pump in the venting zone, forcibly starting the pump group to suck up the residual medium in the suction pipe and reduce the leakage. This set of combined actions is executed concurrently within seconds, achieving physical disaster isolation.

[0143] It is evident that by establishing a three-tiered progressive response interval, triggering silent patrols in the primary warning interval, or scheduling external equipment for multimodal cross-verification in the intermediate review interval, these two pre-buffer layers construct fault tolerance and eliminate the risk of erroneous production shutdowns caused by occasional interference signals through low-cost external verification. Furthermore, when the confidence index exceeds the highest blocking threshold, the mechanical valve position is forcibly locked and the pressure relief pump unit is activated via underlying electrical commands, achieving the dual objectives of minimizing false alarm rates and maximizing disaster control efficiency.

[0144] In practical applications, pipeline monitoring systems face a cold start problem during the transition from design and planning to physical deployment and initial operation. At the hardware level, the actual installation angle and coupling state of field sensors often exhibit uncontrollable deviations in degrees of freedom, leading to off-axis distortion in the acquired signals compared to the predetermined design parameters. At the algorithm level, newly built systems lack historical sample data accumulation, making it difficult for deep learning models to achieve effective initial feature convergence.

[0145] Please see Figure 3 , Figure 3 This is another flowchart illustrating the risk management method for pipeline leakage in the embodiments of this application;

[0146] Therefore, in some embodiments, before step S102, the method further includes:

[0147] S301. Use a laser collimation and alignment device to adjust the adhesion force angle of the sensing node on the pipe wall. If the deviation of the adhesion force angle is reduced to within the preset tolerance range, confirm the physical calibration reference surface.

[0148] Among them, the laser collimation and alignment device is used to represent a high-precision optical auxiliary installation tool; the adhesion force angle refers to the angle between the sensitive axis of the sensor (especially the vibration sensor) and the normal of the tube wall tangent; the physical calibration reference plane refers to the fixed state after confirming that the sensor installation posture fully meets the design requirements (such as orthogonal 90 degrees).

[0149] Specifically, during the installation phase before system commissioning, to eliminate signal distortion caused by human installation errors (such as off-axis effects leading to loss of shear wave components), construction personnel use a laser collimator to project a baseline. The sensor base is adjusted until the sensor axis coincides with the laser baseline within a preset error range, ensuring the angular deviation is less than a preset tolerance (e.g., ±2°). At this point, the clamps are locked, and the system records the current installation posture as the "physical calibration reference plane," which serves as the geometric reference zero point for all subsequent signal acquisition.

[0150] S302. Inject test fluid medium into the pipeline network and apply unloading oscillation step signal via bypass;

[0151] The test fluid medium is usually water or inert gas, used to simulate real working conditions during the commissioning phase; the unloading oscillation step signal refers to the negative pressure wave and fluid excitation signal artificially created by quickly opening the bypass pressure relief valve to simulate the instantaneous rupture of the pipeline.

[0152] Specifically, after physical installation, the system needs to undergo a full-link practical exercise. After filling the pipe with the test medium and pressurizing it to the working pressure, a standard "leakage step" is artificially created by rapidly opening the solenoid valve on the bypass (millisecond-level action). This signal serves as an excitation source, generating negative pressure waves and vibration waves in the pipeline network to test the system's sensing sensitivity and response speed.

[0153] S303. Calculate the total transmission time from the moment the unloading oscillation step signal is generated to the moment the closed-loop blocking control command is calculated and output;

[0154] The overall transmission time refers to the time difference between the physical stimulus (T0) and the control command (T1), covering the total delay of physical propagation, sensor acquisition, network transmission, cloud computing and command issuance.

[0155] Specifically, the system records the timestamp of the bypass valve opening (as its true value) and monitors the timestamp of the final blocking command generated in the cloud. The difference between the two yields the total response time of the system. This metric is the core criterion for evaluating whether the system has "real-time emergency response capability" and must meet preset safety constraints (such as <8 seconds); otherwise, the system cannot go online.

[0156] S304. If the overall transmission time meets the preset safety delay constraint and the local data sample size is empty, call the preset domain adaptive mapping algorithm to extract the fusion weight parameters of the historical pipeline operation library.

[0157] Among them, the preset safety delay constraint refers to the maximum allowable response time required by industry standards; Domain Adaptation is used to represent a machine learning technique that transfers knowledge from the source domain (old pipeline) to the target domain (new pipeline).

[0158] Specifically, once the response speed test meets the standards, to address the "cold start" challenge caused by the lack of historical data for the new pipeline (empty local sample size), the system activates the transfer learning engine. Instead of training from scratch, the algorithm extracts model weights from a cloud database of "mature old pipelines" similar in diameter, material, and transport medium to the current pipeline. Using a domain adaptation algorithm, the feature distribution of the old pipeline is mapped and aligned to the feature space of the new pipeline, enabling the new system to inherit the discriminative capabilities of the old system.

[0159] S305. Inject the fused weight parameters into the deep time recurrent neural network to generate the prior network weight base, and execute S101.

[0160] Among them, the prior network weight base refers to the neural network parameter state that has initial intelligence after transfer injection.

[0161] Specifically, the system formally loads the adjusted migration weights into the newly deployed deep recurrent neural network. At this point, the deep recurrent neural network possesses mature leak detection logic. The system then activates step S101 to initiate the real-time data acquisition and monitoring process.

[0162] As can be seen, before the system is put into use, the sensor attachment angle is physically calibrated using a laser collimation device, eliminating signal scattering and distortion caused by off-axis effects at the source of perception. By applying a step signal to measure the end-to-end latency, the communication link and algorithm processing are matched. Addressing the "cold start" dilemma of newly built pipelines failing to converge due to a lack of historical data, a domain adaptive algorithm is used to inject prior weight parameters from older pipelines with similar operating conditions. This allows newly deployed systems to possess fault diagnosis capabilities without a lengthy training and adjustment period.

[0163] The following describes an exemplary pipeline leakage risk management system 400 provided in an embodiment of this application. Figure 4 This is an exemplary hardware structure diagram of the pipeline leakage risk management system 400 provided in this application embodiment.

[0164] In some embodiments, the pipeline leakage risk management system 400 is a computer device or includes a computer device. The computer device includes a processor, memory, and a network interface connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device stores data. The network interface of the computer device is used to communicate with other external terminals or servers via a network connection. In some embodiments, the network interface can be a wired network interface; in some embodiments, the network interface can also be a wireless network interface. When the computer program is executed by the processor, it implements the methods in the embodiments of this application.

[0165] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0166] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0167] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as meaning "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0168] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0169] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for risk management of pipeline leakage, characterized in that, include: The pipeline network is sampled using sensor nodes to obtain multimodal physical sequences; The multimodal physical sequence is time-aligned and denoised to obtain a multimodal time sequence; The physical conduction time difference is determined based on the multimodal time series, the spatial topology of the pipeline network, and the theoretical wave velocity of the fluid. Determine the spatial attention mask based on the physical conduction time difference; The spatial attention mask is used to filter the multimodal time series to obtain a sub-multimodal time series; The sub-multimodal time series is input into a preset deep time recurrent neural network to obtain a dynamic leakage confidence index; If the dynamic leakage confidence index exceeds the preset warning limit, a closed-loop blocking control command is generated and sent down.

2. The method according to claim 1, characterized in that, The physical conduction time difference is determined based on the multimodal time sequence, the pipeline network spatial topology, and the theoretical wave velocity of the fluid. The steps for determining the spatial attention mask based on the physical transit time difference specifically include: Extract the fluid fluctuation time series components from the multimodal time series; Based on the spatial distance between adjacent sensing nodes in the pipeline network topology, and combined with the preset theoretical wave velocity ratio of the fluid, the physical transmission time difference of the fluid wave temporal component between sensing nodes is obtained. Using the physical conduction time difference as a sliding alignment time window, time axis compensation splicing operation is performed on the fluid fluctuation time series components collected by upstream and downstream sensing nodes to generate a fluctuation spatiotemporal alignment matrix. The difference derivation operation of the aforementioned spatiotemporal alignment matrix is ​​used to obtain the leakage source coordinate distribution tensor where the upstream and downstream wave characteristics intersect and overlap; The spatial attention mask is obtained by transforming the coordinate distribution tensor of the leakage source through a feature mapping function.

3. The method according to claim 1 or 2, characterized in that, The step of inputting the sub-multimodal time series into a preset deep recurrent neural network to obtain the dynamic leakage confidence index specifically includes: The sub-multimodal time series sequences are dimensionality reduced and deconstructed to obtain several single-modal time series sequences; The single-modal time series and the corresponding historical state background data are input into the deep time recurrent neural network to obtain future trend prediction data; The difference between the future trend prediction data and the single-modal time series is calculated to obtain a multidimensional feature residual matrix; The multidimensional feature residual matrix is ​​mapped and compressed to a predetermined interval to obtain the dynamic leakage confidence index.

4. The method according to claim 1, characterized in that, After the step of inputting the sub-multimodal time series into a preset deep time recurrent neural network to obtain the dynamic leakage confidence index, the method further includes: The local edge network gradient is obtained by training on local historical data using edge computing nodes. Calculate and assign fusion weight coefficients based on the data signal-to-noise ratio of each edge computing node; The fusion weight coefficients are used to homomorphically encrypt and aggregate the parameters of multiple edge local network gradients to generate a new feature calibration benchmark. The new feature calibration benchmark is distributed to different edge computing nodes.

5. The method according to claim 4, characterized in that, The step of generating and sending a closed-loop blocking control command when the dynamic leakage confidence index exceeds a preset warning threshold specifically includes: When the dynamic leakage confidence index exceeds the preset warning limit, the topological spatial correlation coefficient between the sensing node and the upstream and downstream sensing nodes is calculated based on the pipeline spatial topology. Determine whether the correlation coefficient of the topological space conforms to the preset overall translation feature rule; If the overall translation feature rules are not met, a closed-loop blocking control command is generated and sent downwards. If the overall translation feature rule is met, then the multimodal physical sequence is determined as local historical data, and the step of training the local historical data using edge computing nodes to obtain the edge local network gradient is executed.

6. The method according to claim 1, characterized in that, The step of generating and sending a closed-loop blocking control command when the dynamic leakage confidence index exceeds a preset warning threshold specifically includes: If the dynamic leakage confidence index falls within the first warning range, a silent patrol work order is generated and the silent patrol work order containing the coordinate anchor point information of the alarm pipe section is sent to the patrol terminal device. If the dynamic leakage confidence index falls within the second verification interval, a multimodal evidence collection cross-instruction is generated. The multimodal evidence collection cross-instruction is used to drive the peripheral detection equipment to the target coordinates to perform acoustic and optical evidence collection and data feedback comparison. If the dynamic leakage confidence index is determined to be greater than the blocking defense threshold, the closed-loop blocking control command is generated to drive the upstream and downstream valves of the pipeline to perform electric shut-off displacement, and a high-level signal is sent simultaneously to start the pressure relief pump group configured in the system; the blocking defense threshold is greater than the second review interval and greater than the first warning interval.

7. The method according to claim 1, characterized in that, Before the step of sampling the pipeline network using sensor nodes to obtain multimodal physical sequences, the method further includes: The adhesion angle of the sensing node on the pipe wall is adjusted using a laser collimation and alignment device. When the deviation of the adhesion angle is determined to be within the preset tolerance range, the physical calibration reference surface is confirmed. Test fluid medium is injected into the pipeline network, and an unloading oscillation step signal is applied via a bypass. Calculate the total transmission time from the moment the unloading oscillation step signal is generated to the moment the closed-loop blocking control command is calculated and output; If it is determined that the overall transmission time meets the preset safety delay constraint and the local data sample size is empty, the preset domain adaptive mapping algorithm is called to extract the fusion weight parameters of the historical pipeline operation library. The fusion weight parameters are injected into the deep time recurrent neural network to generate a priori network weight base, and the step of sampling the pipeline network using sensor nodes to obtain multimodal physical sequences is performed.

8. A risk management system for pipeline leakage, characterized in that, The pipeline leakage risk management system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the pipeline leakage risk management system to perform the method as described in any one of claims 1-7.

9. A computer program product containing instructions, characterized in that, When the computer program product is run on a pipeline leakage risk management system, the pipeline leakage risk management system performs the method as described in any one of claims 1-7.

10. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is executed on the pipeline leakage risk management system, the pipeline leakage risk management system performs the method as described in any one of claims 1-7.

Images