A method for monitoring leakage in pressure pipelines based on multi-sensor fusion

Abstract

This invention discloses a pressure pipeline leakage monitoring method based on multi-sensor fusion, comprising the following steps: acquiring pressure, flow rate, vibration, acoustic emission, temperature, and valve status data of the target pressure pipeline and performing time calibration, anomaly removal, normalization, and pipe segment mapping to obtain multi-sensor aligned data; extracting fluctuation, offset, impact, and consistency features to generate operational fusion features; inputting the operational fusion features into an improved Koopman pipeline state evolution model to obtain predicted state features; generating residual evidence based on the predicted state features and obtaining leakage candidate fusion evidence through a D-S evidence theory fusion model; determining the leakage state and inverting the location based on the fusion evidence, and outputting leakage monitoring results and leakage location estimation results. This invention achieves fusion identification and location of pressure pipeline leaks under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of pressure pipeline safety monitoring technology, specifically to a pressure pipeline leakage monitoring method based on multi-sensor fusion. Background Technology

[0002] Pressure pipelines are widely used in oil, natural gas, chemical, heating, water supply, and industrial transmission systems, undertaking the task of transporting high-pressure media or continuous fluids. Because pressure pipelines are constantly exposed to the combined effects of pressure fluctuations, temperature changes, media erosion, valve regulation, external vibration, and material aging, pipe wall corrosion, loosening of connections, seal failure, weld cracks, and localized mechanical damage can easily develop into leakage hazards. Once a leak occurs, it can not only cause media loss and system shutdown but also potentially lead to combustion, explosion, pollution, equipment damage, and personnel safety accidents. Therefore, continuous and accurate leakage monitoring of the operating status of pressure pipelines is necessary.

[0003] Existing methods for monitoring leaks in pressure pipelines mainly include pressure threshold judgment, flow balance analysis, acoustic emission detection, vibration detection, and manual inspection. Pressure threshold and flow balance methods are relatively simple to implement, but they are prone to false alarms during pump start-up and shutdown, valve opening and closing, load fluctuations, and normal pressure regulation. Acoustic emission and vibration detection are sensitive to minor leaks, but the results are easily affected by installation location, pipeline materials, background noise, and signal propagation attenuation. Manual inspection relies on personnel experience and lacks real-time and continuous capabilities, making it difficult to meet the online monitoring needs of long-distance pipeline networks and complex industrial scenarios.

[0004] As the number of sensors increases, multi-source data such as pressure, flow, vibration, acoustic emission, temperature, and valve status can reflect pipeline operation changes from different perspectives. However, different sensor data suffer from inconsistent sampling frequencies, significant differences in dimensions, spatial dispersion, and varying noise levels. Simply stitching together or superimposing thresholds from multiple sensor data sets makes it difficult to accurately distinguish between actual leaks, operating condition changes, valve disturbances, and sensor anomalies. Especially under complex operating conditions, the normal state itself undergoes dynamic changes, and traditional methods often lack modeling of the evolution of pipeline operating states, resulting in insufficient stability in leak identification and inadequate accuracy in leak location estimation.

[0005] Therefore, how to provide a pressure pipeline leakage monitoring method based on multi-sensor fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a pressure pipeline leakage monitoring method based on multi-sensor fusion. This invention fully utilizes multi-sensor data fusion, pipeline state evolution modeling, and evidence fusion analysis techniques. It details the complete process of synchronously aligning, extracting features, predicting states, generating residual evidence, determining leakage states, and estimating leakage locations for pressure, flow, vibration, acoustic emission, temperature, and valve state data during the operation of a pressure pipeline. It has the advantages of fully utilizing multi-source data, strong adaptability to complex operating conditions, high stability in leakage identification, and accurate leakage location determination.

[0007] A pressure pipeline leakage monitoring method based on multi-sensor fusion according to an embodiment of the present invention includes the following steps: S1. Obtain multi-sensor raw data of the target pressure pipeline during the monitoring period, and perform time calibration, anomaly removal, normalization and pipe segment mapping on the multi-sensor raw data to obtain multi-sensor aligned data. S2. Based on the pressure sequence, flow sequence, vibration sequence, acoustic emission sequence, temperature sequence and valve state sequence in the multi-sensor aligned data, extract the fluctuation feature, offset feature, impact feature and consistency feature to obtain the operation fusion feature; S3. Input the fusion features into the improved Koopman pipeline state evolution model, select the state evolution matrix according to the operating condition code, valve state code and pipe segment pressure level, map and predict the current operating state, and obtain the predicted state features. S4. Based on the operational fusion characteristics and predicted state characteristics, perform residual calculation, standardization, and evidence splitting to obtain pressure residual evidence, flow residual evidence, vibration residual evidence, and acoustic emission residual evidence. S5. Input the residual evidence into the DS evidence theory fusion model, and perform probability allocation based on sensor health status, distance weight and historical noise level to obtain leakage candidate fusion evidence. S6. Based on the leakage probability allocation, non-leakage probability allocation, and uncertain probability allocation in the leakage candidate fusion evidence, the state is determined and the leakage monitoring results are obtained. S7. Based on the abnormal sensor nodes, disturbance arrival time difference and propagation speed corresponding to the leakage monitoring results, the leakage location estimation result is obtained by inversion.

[0008] Optionally, step S1 includes the following steps: S11. Acquire pressure, flow, vibration, acoustic emission, temperature sampling data and valve status data collected by sensor nodes along the target pressure pipeline, and obtain pipeline basic parameters, sensor configuration parameters and historical operating parameters to obtain multi-sensor raw dataset; S12. Based on a unified clock reference and sensor sampling frequency, timestamp correction, resampling, and monitoring window segmentation are performed on the multi-sensor raw dataset to obtain synchronized sensing data. S13. Based on the sensor range, historical drift range and abrupt change amplitude limitations, perform anomaly removal, missing data filling, drift correction and normalization on the synchronous sensing data to obtain clean sensing data. S14. Based on the sensor node installation location, pipe segment connection relationship, medium flow direction and valve position, perform spatial mapping and pipe segment attribution marking on the cleaning sensor data to obtain multi-sensor alignment data.

[0009] Optionally, step S2 includes the following steps: S21. Calculate the sampling point difference, window mean offset, peak-valley variation amplitude and pressure-flow matching difference based on the pressure sequence and flow sequence in the multi-sensor aligned data to obtain the pressure-flow fluctuation offset characteristics. S22. Based on the vibration sequence and acoustic emission sequence in the multi-sensor aligned data, perform frequency band decomposition, energy statistics, impact peak extraction and duration statistics to obtain the mechanical disturbance impact characteristics. S23. Based on the valve state sequence, the temperature change sequence calculated from the temperature sequence, the pressure and flow fluctuation offset characteristics, and the mechanical disturbance and impact characteristics, the working condition segments are divided to obtain the steady-state delivery segment, the valve regulation segment, and the abnormal candidate segment. S24. Based on the change direction, change amplitude, occurrence order and propagation interval of each sensor node within the same working condition segment, perform consistency calculation to obtain the operation fusion characteristics.

[0010] Optionally, step S3 includes the following steps: S31. Based on the pressure-flow combination characteristics, valve status characteristics, temperature compensation characteristics, and cross-sensor consistency characteristics in the operation fusion characteristics, and combined with the pipe section pressure level, medium flow velocity range, and pipe section connection relationship, the current monitoring window is coded to obtain the operating condition vector. S32. Input the running fusion features and operating condition vectors into the improved Koopman pipeline state evolution model's up-dimensional mapping network, and perform nonlinear embedding, operating condition splicing, and state normalization on the low-dimensional sensing features to obtain the high-dimensional state vector. The up-dimensional mapping network includes a multi-branch up-dimensional mapping structure, a working condition coding splicing layer, and a state normalization layer; The multi-branch up-dimensional mapping structure includes a pressure-flow mapping branch and a vibration-acoustic emission mapping branch; S33. Based on the operating condition vector, valve opening change, medium flow rate and pipe pressure level, select the target state evolution matrix from the candidate state evolution matrix to obtain the operating condition matching state evolution parameters. S34. Perform state recursion based on the high-dimensional state vector and the working condition matching state evolution parameters, and map the recursive state back to the original feature space through a dimensionality reduction mapping network to obtain the predicted state features.

[0011] Optionally, step S32 includes the following steps: S321. Establish a pressure and flow input vector based on the pressure and flow combination characteristics in the operation fusion characteristics, and establish a vibration and acoustic emission input vector based on the vibration energy increment and acoustic emission peak intensity in the mechanical disturbance and impact characteristics to obtain the branch input characteristics; S322. Input the branch input features into the multi-branch up-dimensional mapping structure of the improved Koopman pipeline state evolution model, perform nonlinear embedding processing on the pressure and flow input vector through the pressure and flow mapping branch to obtain the pressure and flow embedding vector, and perform nonlinear embedding processing on the vibration and acoustic emission input vector through the vibration and acoustic emission mapping branch to obtain the vibration and acoustic emission embedding vector. S323. Generate a working condition embedding vector based on the working condition vector, and input the working condition embedding vector, pressure-flow embedding vector, and vibration-acoustic emission embedding vector into the working condition encoding splicing layer of the improved Koopman pipeline state evolution model to obtain a joint state input vector. S324. Input the joint state input vector into the state normalization layer of the improved Koopman pipeline state evolution model to perform scale unification and state distribution correction on the feature amplitudes formed by different sensing branches, and obtain a high-dimensional state vector.

[0012] Optionally, step S33 includes the following steps: S331. Based on the conveying load level, valve action range, temperature change range and pipe section pressure level in the working condition vector, the working condition index is encoded to obtain the working condition index value. The operating condition index value is the matrix selection identifier for the operating condition type corresponding to the current monitoring window; S332. Based on the operating condition index value, call the candidate library of operating condition grouping matrices in the improved Koopman pipeline state evolution model to obtain the initial evolution matrix; The candidate library of working condition grouping matrices is a set of state evolution matrices stored according to steady-state delivery, valve regulation, load fluctuation and pressure transition working conditions respectively; S333. The initial evolution matrix is ​​scaled according to the pipe section length, medium flow velocity and pipe section pressure level to obtain the pressure-velocity correction matrix and scale correction parameters. The scale correction parameter is a matrix scaling parameter calculated from the physical properties of the pipe section and the operating state of the medium. S334. Calculate the feature difference data based on the operation fusion characteristics of the current monitoring window and the previous monitoring window, and calculate the valve state change based on the valve state data of the two monitoring windows. Input the feature difference data and the valve state change into the gating parameter generation unit in the improved Koopman pipeline state evolution model to obtain the gating coefficient. S335. Adjust the local evolution parameters in the pressure-velocity correction matrix according to the gating coefficient to obtain the target state evolution matrix; S336. Based on the target state evolution matrix, scale correction parameters and gating coefficients, the parameters are sorted to obtain the working condition matching state evolution parameters.

[0013] Optionally, step S4 includes the following steps: S41. Based on the difference between the operational fusion characteristics and the predicted state characteristics in the same dimension, residual calculation is performed to obtain the pressure residual sequence, flow residual sequence, vibration residual sequence and acoustic emission residual sequence; S42. Based on the pipe section pressure level, medium flow rate, distance between adjacent sensing nodes and historical fluctuation range, each residual sequence is segmented and standardized to obtain a standardized residual sequence. S43. Based on the residual amplitude, duration, mutation direction, cross-node propagation order and sensor node location in the standardized residual sequence, the evidence is split to obtain the initial pressure residual evidence, initial flow residual evidence, initial vibration residual evidence and initial acoustic emission residual evidence. S44. Based on the matching relationship between each initial residual evidence and the valve state change, the operating condition disturbance component caused by valve regulation is eliminated to obtain pressure residual evidence, flow residual evidence, vibration residual evidence and acoustic emission residual evidence.

[0014] Optionally, step S5 includes the following steps: S51. Calculate the leakage support, non-leakage support, and uncertainty support based on the pressure residual evidence, flow residual evidence, vibration residual evidence, and acoustic emission residual evidence, respectively, to obtain single-source evidence support data; S52. Based on the sensor health status, drift range, noise level, number of intermittent samplings, and communication delay, perform credibility discounting on the single-source evidence support data to obtain discounted evidence data. S53. Based on the abnormal sensing nodes and pipe segment connection relationships corresponding to the residual evidence, candidate pipe segments are determined, and the discounted evidence data is spatially weighted according to the candidate pipe segments, the direction of medium propagation, and the degree of signal attenuation to obtain weighted evidence data. S54. Input the weighted evidence data into the DS evidence theory fusion model, perform combination calculations on the basic probability allocation of multiple evidence sources, and obtain leaked candidate fusion evidence.

[0015] Optionally, step S6 includes the following steps: S61. Calculate the state difference based on the leakage probability allocation and non-leakage probability allocation in the leakage candidate fusion evidence, and make an initial judgment based on the changing trend of the leakage probability allocation in adjacent monitoring windows to obtain the leakage state judgment value. S62. Based on the uncertainty probability distribution in the leakage candidate fusion evidence, the sensor health status, and the statistical results of the leakage candidate fusion evidence within the continuous monitoring window, the leakage status judgment value is corrected for the confidence interval to obtain the corrected judgment value. The statistical results include the number of abnormal persistence windows and the number of consecutive increases in leakage probability; S63. Based on the comparison between the correction judgment value and the suspected leakage threshold and leakage confirmation threshold in the preset state threshold library, and combined with the number of abnormal evidence sources and the consistency of cross-sensor evidence, the target pressure pipeline is classified into states to obtain state judgment data. S64. Based on the state determination data, the fusion probability allocation in the leakage candidate fusion evidence, the source of abnormal evidence corresponding to the residual evidence, the number of abnormal persistence windows and the corresponding monitoring window number, the results are encapsulated to obtain the leakage monitoring results.

[0016] Optionally, step S7 includes the following steps: S71. Extract the residual peak time based on the abnormal sensor nodes and residual evidence corresponding to the leakage monitoring results, and process the disturbance arrival time of the residual peak time to obtain the multi-node disturbance arrival time series. S72. Based on the arrival time sequence of multi-node disturbances, pipe segment connection relationship and medium flow direction, and combined with the pipe segment propagation velocity determined by medium type, pipe material and pipe segment pressure level, calculate the propagation time difference from candidate leak point to adjacent sensing node to obtain candidate location error data. S73. Select the candidate pipe segment position with the smallest error based on the candidate position error data, and correct the position by combining the source nodes of pressure residual evidence, acoustic emission residual evidence and vibration residual evidence to obtain corrected candidate position data. S74. Based on the corrected candidate location data, candidate pipe segment number and the length ratio from the upstream sensor node, the location results are sorted to obtain the leakage location estimation result.

[0017] The beneficial effects of this invention are: by performing time calibration, anomaly removal, normalization, and pipe segment mapping on multi-source data such as pressure, flow rate, vibration, acoustic emission, temperature, and valve status, this invention enables sensor data with different sampling frequencies, different dimensions, and different installation locations to enter a unified analysis framework, reducing the interference caused by fluctuations in single sensor data on leak judgment and improving the completeness of the description of the operating status of pressure pipelines.

[0018] This invention utilizes an improved Koopman pipeline state evolution model with fused feature inputs. By mapping low-dimensional sensing features to a high-dimensional evolution space through an upscaling mapping network, and combining operating condition vectors, a candidate library of operating condition grouping matrices, scale correction parameters, and gating parameter generation units, it adapts and models the pipeline state evolution process under different transport loads, valve actions, pressure levels, and media flow rates. This allows the predicted state features to more closely reflect the normal operating patterns under different operating conditions, thus providing a more stable reference basis for subsequent residual evidence generation.

[0019] This invention generates pressure residual evidence, flow residual evidence, vibration residual evidence, and acoustic emission residual evidence based on the differences between operational fusion characteristics and predicted state characteristics. It transforms leakage judgment from a single threshold comparison into multi-source residual evidence analysis, which can more effectively distinguish between real leakage, valve regulation disturbance, load change, and sensor anomaly, and reduce the probability of misjudgment under complex operating conditions.

[0020] This invention further fuses multiple types of residual evidence through the DS evidence theory fusion model, and generates leak candidate fusion evidence by combining sensor health status, distance weight, and historical noise level, so that the conflict and uncertainty between evidence from different sources can be uniformly handled; then, the location is inverted based on the abnormal sensor node, the time difference of arrival of the disturbance, and the propagation speed, which can provide the leak location estimation result while identifying the leak status, thereby improving the continuity, stability and positioning accuracy of pressure pipeline leak monitoring. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a pressure pipeline leakage monitoring method based on multi-sensor fusion proposed in this invention; Figure 2 This is a schematic diagram illustrating the multi-sensor raw data processing, runtime fusion feature extraction, and operational condition segmentation process in this invention. Figure 3 This is a schematic diagram of the improved Koopman pipeline state evolution model, residual evidence fusion, and leak location estimation process in this invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0023] refer to Figures 1-3 A method for monitoring leaks in pressure pipelines based on multi-sensor fusion includes the following steps: S1. Obtain multi-sensor raw data of the target pressure pipeline during the monitoring period, and perform time calibration, anomaly removal, normalization and pipe segment mapping on the multi-sensor raw data to obtain multi-sensor aligned data. S2. Based on the pressure sequence, flow sequence, vibration sequence, acoustic emission sequence, temperature sequence and valve state sequence in the multi-sensor aligned data, extract the fluctuation feature, offset feature, impact feature and consistency feature to obtain the operation fusion feature; S3. Input the fusion features into the improved Koopman pipeline state evolution model, select the state evolution matrix according to the operating condition code, valve state code and pipe segment pressure level, map and predict the current operating state, and obtain the predicted state features. S4. Based on the operational fusion characteristics and predicted state characteristics, perform residual calculation, standardization, and evidence splitting to obtain pressure residual evidence, flow residual evidence, vibration residual evidence, and acoustic emission residual evidence. S5. Input the residual evidence into the DS evidence theory fusion model, and perform probability allocation based on sensor health status, distance weight and historical noise level to obtain leakage candidate fusion evidence. S6. Based on the leakage probability allocation, non-leakage probability allocation, and uncertain probability allocation in the leakage candidate fusion evidence, the state is determined and the leakage monitoring results are obtained. S7. Based on the abnormal sensor nodes, disturbance arrival time difference and propagation speed corresponding to the leakage monitoring results, the leakage location estimation result is obtained by inversion.

[0024] refer to Figure 1 and Figure 2 In this embodiment, S11 acquires pressure, flow, vibration, acoustic emission, temperature sampling data and valve status data collected by sensor nodes along the target pressure pipeline, and acquires pipeline basic parameters, sensor configuration parameters and historical operating parameters to obtain multi-sensor raw dataset.

[0025] Among them, pressure, flow, vibration, acoustic emission and temperature sampling data are collected by sensor nodes installed at different pipe sections of the target pressure pipeline, valve status data are collected by valve control unit or valve status acquisition device, pipeline basic parameters are from pipeline basic information table, sensor configuration parameters are from equipment configuration table, and historical operating parameters are from historical database and operation and maintenance records.

[0026] The basic pipeline parameters include pipe segment number, pipe segment length, pipe diameter, pipe wall thickness, pipe material, medium type, medium flow direction, upstream node number, downstream node number, and valve installation location. The sensor configuration parameters include sensor type, installation location, sampling frequency, measurement range, communication address, calibration coefficient, and maintenance status. The historical operating parameters include pressure operating range, flow operating range, temperature operating range, valve action records, historical noise level, and historical maintenance records.

[0027] The system arranges the above data into a matrix according to the monitoring window. The matrix rows correspond to the sampling time, and the matrix columns correspond to the sensor node, sensor type, pipe section number, and valve status fields, thus forming a multi-sensor raw dataset that can be read and calculated uniformly.

[0028] In this embodiment, S12 performs timestamp correction, resampling, and monitoring window segmentation on the multi-sensor raw dataset based on a unified clock reference and sensor sampling frequency to obtain synchronized sensing data. The unified clock reference is a time reference formed by jointly verifying the monitoring platform server clock, the edge acquisition unit synchronization clock, and the timestamp uploaded by the sensor node. The resampling frequency is a unified sampling frequency determined based on the pressure and flow sampling frequency, the density of vibration and acoustic emission events, and the time resolution of leakage disturbance propagation.

[0029] The system corrects the time offset in the data uploaded by different sensor nodes, resamples data with different sampling frequencies, compresses high-frequency vibration and acoustic emission data into a window sequence that matches the pressure and flow data according to the window statistics method, and maps low-frequency valve status data to the corresponding monitoring window according to the status holding method.

[0030] For high-frequency vibration data, the system calculates the root mean square value, peak value, kurtosis, dominant frequency energy, and high-frequency energy ratio within each monitoring window to form vibration window statistics. For acoustic emission data, the system extracts event count, peak intensity, duration, rise time, and cumulative energy value within each monitoring window to form an acoustic emission window sequence. For pressure and flow data, the system retains the original sampling trend and calculates the window mean, maximum value, minimum value, and slope. For valve status data, the system generates valve action markers based on the valve opening change time and valve action direction to obtain synchronous sensing data.

[0031] In this embodiment, S13 performs anomaly removal, missing data filling, drift correction and normalization on the synchronous sensing data according to the sensor range, historical drift range and abrupt change amplitude limitations, to obtain clean sensing data.

[0032] The historical drift range is the data offset range obtained by statistically analyzing the zero-point drift and range drift of the sensor during the leak-free and stable operation phase. The mutation amplitude limit is the maximum allowable variation amplitude determined based on the variation range of adjacent sampling points during the normal operation of the target pressure pipeline, the allowable fluctuation range caused by valve action, and the upper limit of the sensor range. The missing length limit is the upper limit of data missing set based on the number of consecutive missing sampling points, the stability of the communication link, and the length of the monitoring window.

[0033] The system removes isolated spikes that are short-lived and do not conform to the physical propagation laws of pipelines. For short-term missing data, it uses adjacent window interpolation or adjacent nodes in the same pipe segment to fill in the missing data. For long-term missing nodes, it marks them as low health status. Anomaly removal does not directly delete all mutation signals. Instead, it first determines whether the mutation signal conforms to the propagation order of leakage disturbance between adjacent sensing nodes. If the mutation only occurs in a single sensing node and there is no response from adjacent nodes, the mutation is treated as an isolated anomaly. If the mutation forms a sequential response on multiple nodes according to the direction of medium flow, the mutation is retained and enters the subsequent residual evidence generation process.

[0034] In this embodiment, S13 also performs type-based normalization processing on different types of data.

[0035] Pressure data is normalized based on historical pressure ranges for pipe sections of the same pressure level; flow data is normalized based on transport load ranges and pipe diameter parameters; vibration and acoustic emission data are normalized based on sensor installation location, pipe material, and historical noise levels; temperature data is normalized based on medium temperature ranges and ambient temperature variation ranges; and historical noise levels are noise baseline data obtained from background fluctuation statistics of the same sensor node during a leak-free stable operation phase.

[0036] Through the above processing, the system can reduce the interference of sensor glitches, communication jitter and single-point acquisition errors on leakage judgment, while avoiding the accidental deletion of early signals of real leakage.

[0037] In this embodiment, S14 performs spatial mapping and pipe segment attribution marking on the cleaning sensor data based on the sensor node installation location, pipe segment connection relationship, medium flow direction and valve location, to obtain multi-sensor alignment data.

[0038] Specifically, the system maps cleaning sensor data to corresponding pipe segment numbers, upstream sensor node numbers, downstream sensor node numbers, valve influence range, and media flow direction markers, forming multi-sensor aligned data with time index, node index, pipe segment index, and sensor type index.

[0039] After the pipe segment mapping is completed, the pressure, flow rate, vibration, acoustic emission, temperature and valve status within the same pipe segment can be called by subsequent steps as a unified object. The cross-pipe segment propagation relationship can also be tracked through the upstream and downstream node numbers. For pipe networks with branches, diameter-changing pipe segments or multiple valves, the system uses the branch connection point, diameter-changing position and valve position as topology boundary points, so that subsequent working condition division and position inversion can be performed based on the real pipeline structure.

[0040] refer to Figure 2 In this embodiment, S21 calculates the sampling point difference, window mean offset, peak-valley variation amplitude, and pressure-flow matching difference based on the pressure sequence and flow sequence in the multi-sensor alignment data to obtain the pressure-flow fluctuation offset characteristics.

[0041] Among them, the pressure sequence and flow sequence are calculated by differential calculation, mean calculation, peak-valley amplitude calculation and pressure-flow matching difference calculation according to the monitoring window. The pressure-flow matching difference is the degree of synchronous deviation between pressure change and flow change within the same pipe section. Under normal transportation conditions, pressure change and flow change usually remain matched within a certain delay range. Under leakage conditions, the pressure may drop continuously or drop suddenly in some areas, and the flow may be inconsistent between upstream and downstream flow. Therefore, the pressure-flow matching difference can reflect whether the fluid transportation relationship deviates from the normal pattern. The pressure-flow fluctuation deviation characteristics include pressure window slope, flow window slope, pressure mean deviation, flow mean deviation, pressure peak-valley difference, flow peak-valley difference, pressure-flow direction consistency and pressure-flow amplitude matching difference.

[0042] In this embodiment, S22 performs frequency band decomposition, energy statistics, impact peak extraction, and duration statistics based on the vibration sequence and acoustic emission sequence in the multi-sensor aligned data to obtain mechanical disturbance impact characteristics. The frequency band boundary of the frequency band decomposition is a segmented frequency range determined based on the sensor response range, the vibration frequency distribution of the pipeline material, and the high-frequency energy concentration interval in the historical leakage samples. The duration statistics rule is a continuous statistical method formed based on the number of windows in which the abnormal amplitude continuously exceeds the historical fluctuation interval. The historical fluctuation interval is the normal variation range of pressure, flow rate, vibration, and acoustic emission obtained by statistically analyzing the target pipeline's leak-free operation data. The mechanical disturbance impact characteristics include the sudden increase in vibration energy, the peak intensity of acoustic emission, the number of impact duration windows, and the proportion of high-frequency energy.

[0043] Since leakage jets, pipe wall micro-vibrations, and media disturbances may manifest as short-term impact enhancement or continuous high-frequency energy increases in the vibration and acoustic emission channels, the system uses the mechanical disturbance impact characteristics and pressure and flow fluctuation offset characteristics together to form the basis data for subsequent leakage judgment, rather than relying solely on pressure drop or flow change for judgment.

[0044] In this embodiment, S23 divides the operating conditions into segments based on the valve state sequence, the temperature change sequence calculated from the temperature sequence, the pressure and flow fluctuation offset characteristics, and the mechanical disturbance impact characteristics, resulting in steady-state delivery segments, valve regulation segments, and abnormal candidate segments. The segment division rules are based on the operating state division rules determined by the magnitude of the delivery load change, the frequency of valve opening change, the medium flow velocity change range, the temperature change magnitude, and the pressure change trend. The steady-state delivery segment corresponds to the monitoring segment where the pressure, flow rate, and valve state are all within a stable range. The valve regulation segment corresponds to the monitoring segment where the pressure and flow rate fluctuate synchronously due to the change in valve opening. The abnormal candidate segment corresponds to the monitoring segment where the pressure and flow rate are mismatched, the vibration and acoustic emission impact is enhanced, or the cross-node propagation sequence is abnormal.

[0045] The significance of dividing the operating conditions into segments is that the system can use different state evolution matrices under valve regulation and load fluctuation conditions, avoiding mistaking normal regulation processes for leakage anomalies. In anomaly candidate segments, the system pays more attention to the consistency of multi-source evidence, enabling early leakage signals to be continuously tracked.

[0046] In this embodiment, S24 performs consistency calculations based on the change direction, change magnitude, occurrence order, and propagation interval of each sensing node within the same operating condition segment to obtain the operation fusion characteristics.

[0047] Cross-node consistency characteristics include consistency of upstream and downstream pressure change direction in the same pipe segment, arrival sequence of pressure disturbances in adjacent pipe segments, occurrence interval of vibration shocks between adjacent nodes, intensity attenuation relationship of acoustic emission events among multiple nodes, and time matching relationship between pressure and flow rate changes and valve actions.

[0048] The system splices together pressure and flow fluctuation offset features, mechanical disturbance impact features, temperature compensation features, and cross-node features according to pipe segment numbers to obtain operational fusion features. Operational fusion features can be represented as a three-dimensional tensor, with the first dimension being the monitoring window, the second dimension being the pipe segment or node, and the third dimension being the feature channel, which facilitates the subsequent reading of the improved Koopman pipeline state evolution model.

[0049] refer to Figure 3 In this embodiment, S31 encodes the operating condition of the current monitoring window based on the pressure-flow combination characteristics, valve status characteristics, temperature compensation characteristics, and cross-sensor consistency characteristics in the operation fusion characteristics, and combines the pipe section pressure level, medium flow velocity range, and pipe section connection relationship to obtain the operating condition vector.

[0050] The operating condition vector includes the conveying load level, valve actuation range, valve actuation direction, temperature change range, pipe section pressure level, medium flow velocity range, and pipe section topology marker.

[0051] The purpose of operating condition coding is to write the pipeline operating state and the physical conditions of the pipe section into the model input, so that the improved Koopman pipeline state evolution model can distinguish different operating conditions such as steady-state delivery, valve regulation, load fluctuation and pressure transition when predicting the current operating state.

[0052] In this embodiment, S32 will run the improved Koopman pipeline state evolution model with fused features and operating condition vector inputs, and perform nonlinear embedding, operating condition splicing and state normalization processing on the low-dimensional sensing features to obtain a high-dimensional state vector. The improved mapping network includes a multi-branch improved mapping structure, an operating condition encoding splicing layer and a state normalization layer. The multi-branch improved mapping structure includes a pressure flow mapping branch and a vibration acoustic emission mapping branch.

[0053] The setting of the up-dimensional mapping network is not simply to increase the number of neural network layers, but to process sensing feature branches with different physical meanings. The pressure and flow mapping branch mainly learns the slowly changing state relationship in the medium transportation process, while the vibration and acoustic emission mapping branch mainly learns the high-frequency change relationship caused by leakage impact, mechanical disturbance and local anomaly. The operating condition encoding splicing layer adds valve action, pressure level, medium flow rate and pipe section connection information into the high-dimensional state expression, so that the high-dimensional state vector contains both sensor observation features and pipeline operating conditions and spatial position constraints.

[0054] In this embodiment, S321 establishes a pressure-flow input vector based on the pressure-flow combination feature in the operation fusion feature, and establishes a vibration-acoustic emission input vector based on the vibration energy surge and acoustic emission peak intensity in the mechanical disturbance impact feature, thus obtaining the branch input feature.

[0055] The pressure-flow input vector includes the pressure window slope, flow window slope, pressure mean offset, flow mean offset, and pressure-flow matching difference. The vibration-acoustic emission input vector includes the vibration energy burst, acoustic emission peak intensity, number of impact duration windows, and high-frequency energy ratio. The branch input features are arranged according to the sensor type and pipe segment number, so that each branch can read the data channel with the corresponding physical meaning.

[0056] In this embodiment, S322 inputs the branch input features into the multi-branch up-dimensional mapping structure in the improved Koopman pipeline state evolution model, performs nonlinear embedding processing on the pressure and flow input vector through the pressure and flow mapping branch to obtain the pressure and flow embedding vector, and performs nonlinear embedding processing on the vibration and acoustic emission input vector through the vibration and acoustic emission mapping branch to obtain the vibration and acoustic emission embedding vector.

[0057] The pressure-flow mapping branch and the vibration-acoustic emission mapping branch each contain an input layer, a nonlinear transformation layer, and an embedded output layer. The output dimensions of the two branches can be set according to the number of feature channels and pipe segments. The pressure-flow mapping branch is biased towards expressing slow-changing transport states, while the vibration-acoustic emission mapping branch is biased towards expressing rapid impact states, thereby avoiding mutual interference between different types of features in the same mapping layer.

[0058] In this embodiment, S323 generates a working condition embedding vector based on the working condition vector, and inputs the working condition embedding vector, pressure flow embedding vector, and vibration acoustic emission embedding vector into the working condition encoding splicing layer of the improved Koopman pipeline state evolution model to obtain a joint state input vector.

[0059] The operating condition embedding vector is obtained by converting the pipe section pressure level, medium flow velocity range, valve status code, temperature change range, and pipe section connection relationship. The operating condition encoding splicing layer splices the operating condition embedding vector with the two sensing branch embedding vectors in the feature channel dimension, so that the joint state input vector simultaneously contains operating observation information, pipe section physical conditions, and operating condition category information.

[0060] In this embodiment, S324 inputs the joint state input vector into the state normalization layer of the improved Koopman pipeline state evolution model, performs scale unification and state distribution correction on the feature amplitudes formed by different sensing branches, and obtains a high-dimensional state vector.

[0061] The state normalization layer unifies the scale of different branch characteristics, preventing the pressure and flow channels from suppressing the vibration and acoustic emission channels due to their large amplitude, and also preventing the high-frequency sensing channels from affecting the overall state recursion stability due to excessive local impact.

[0062] Therefore, the improved Koopman pipeline state evolution model can represent the normal state evolution law of pressure pipelines in a unified high-dimensional space and provide predicted state characteristics for subsequent residual calculation.

[0063] In this embodiment, S33 selects the target state evolution matrix from the candidate state evolution matrix based on the operating condition vector, valve opening change, medium flow rate, and pipe pressure level, and obtains the operating condition matching state evolution parameters.

[0064] Traditional state evolution models typically use a fixed evolution matrix to uniformly extrapolate all operating conditions, making it difficult to distinguish the differences in state changes under steady-state delivery, valve regulation, load fluctuation, and pressure transition conditions.

[0065] In this implementation, multiple candidate state evolution matrices are established according to the operating condition type, and the pipe section pressure level, valve action range, medium flow rate and temperature change range are converted into operating condition index values. The operating condition index values ​​are used as matrix selection identifiers to call the initial evolution matrix from the operating condition group matrix candidate library.

[0066] The candidate matrix for operating condition grouping is a set of matrices trained based on historical operating data, operating condition labels, and pipe section pressure levels. The number of candidate matrices is determined by the number of operating condition categories, the number of pipe section pressure levels, and the upper limit of model complexity.

[0067] In this embodiment, S331 performs working condition index encoding based on the conveying load level, valve action range, temperature change range, and pipe section pressure level in the working condition vector to obtain the working condition index value.

[0068] The operating condition index value is the matrix selection identifier for the operating condition type corresponding to the current monitoring window.

[0069] Steady-state delivery, valve regulation, load fluctuation, and pressure transition conditions each correspond to different index ranges, enabling the model to select an evolution matrix that better reflects the current state change pattern based on the index value of the condition.

[0070] In this embodiment, S332 calls the candidate library of the working condition grouping matrix in the improved Koopman pipeline state evolution model according to the working condition index value to obtain the initial evolution matrix.

[0071] The candidate library of operating condition grouping matrices is a set of state evolution matrices stored separately for steady-state delivery, valve regulation, load fluctuation, and pressure transition operating conditions.

[0072] The state evolution matrix corresponding to steady-state transportation conditions tends to maintain the steady recursion of pressure and flow. The state evolution matrix corresponding to valve regulation conditions tends to characterize short-term pressure fluctuations caused by valve action. The state evolution matrix corresponding to load fluctuation conditions tends to characterize the coupling relationship between flow rate changes and pressure response. The state evolution matrix corresponding to pressure transition conditions tends to characterize the state changes during start-up, shutdown, or pipeline segment switching.

[0073] In this embodiment, S333 performs a scaling correction on the initial evolution matrix based on the pipe segment length, medium flow velocity, and pipe segment pressure level to obtain a pressure-velocity correction matrix and a scaling correction parameter; the scaling correction parameter is a matrix scaling parameter calculated from the pipe segment physical properties and medium operating state.

[0074] Since different pipe lengths, medium flow velocities, and pressure levels correspond to different state propagation speeds and response amplitudes, the initial evolution matrix under the same operating condition still needs to be corrected in combination with the physical properties of the pipe section in order to obtain a pressure-velocity correction matrix that is suitable for the current pipe section.

[0075] In this embodiment, S334 calculates feature difference data based on the operational fusion characteristics of the current monitoring window and the previous monitoring window, and calculates the valve state change based on the valve state data of the two monitoring windows. The feature difference data and the valve state change are then input into the gating parameter generation unit in the improved Koopman pipeline state evolution model to obtain the gating coefficient.

[0076] The hidden layer width of the gate parameter generation unit is determined by the network parameter width based on the dimension of the running fusion feature, the dimension of the working condition vector, and the size of the training samples. After receiving the feature difference data and the valve state change, the gate parameter generation unit outputs the adjustment coefficient corresponding to the local parameters of the target state evolution matrix.

[0077] When the changes between adjacent monitoring windows mainly come from valve actions, the gating coefficient increases the weight of evolution parameters related to valve regulation; when the changes between adjacent monitoring windows mainly come from pressure-flow mismatch or enhanced vibration-acoustic emission impact, the gating coefficient increases the weight of evolution parameters related to abnormal propagation; when the operating state remains stable, the gating coefficient maintains a smooth change in matrix parameters.

[0078] In this embodiment, S335 adjusts the local evolution parameters in the pressure-velocity correction matrix according to the gating coefficient to obtain the target state evolution matrix.

[0079] Specifically, the system applies the gating coefficient to the local parameters in the pressure-velocity correction matrix that correspond to the pressure-flow state, vibration-acoustic emission state, and operating condition state, so that the target state evolution matrix can be finely adjusted to follow the state changes of the current monitoring window.

[0080] By combining the candidate library of working condition grouping matrices with the gating parameter generation unit, the model no longer uses the same matrix to explain all state changes, but selects and adjusts more suitable state evolution parameters according to the actual operating scenario.

[0081] In this embodiment, S336 organizes the parameters according to the target state evolution matrix, scale correction parameters and gating coefficients to obtain the working condition matching state evolution parameters.

[0082] The working condition matching state evolution parameters include the target state evolution matrix, the corresponding working condition index value, the scale correction parameter, and the gating coefficient. After reading the working condition matching state evolution parameters, the state recursion unit can perform the recursion from the current state to the predicted state in the high-dimensional state space.

[0083] In this embodiment, S34 performs state recursion based on the high-dimensional state vector and the working condition matching state evolution parameters, and maps the recursive state back to the original feature space through a dimensionality reduction mapping network to obtain the predicted state features.

[0084] The state recursion unit receives the high-dimensional state vector and the target state evolution matrix, and completes the mapping prediction from the current state to the next state. The dimensionality reduction mapping network maps the recursive state back to the original feature space and outputs the predicted state features. The predicted state features include predicted pressure and flow combination features, predicted mechanical disturbance and impact features, predicted temperature compensation features, and predicted cross-node consistency features, which are used for subsequent residual calculation by fusing features with actual operation features.

[0085] In this embodiment, the training data for the improved Koopman pipeline state evolution model comes from the target pressure pipeline historical operation database, maintenance records, manually marked leakage records, valve action records, and leakage-free steady-state operation records.

[0086] The training samples are organized according to the monitoring window. Each sample includes the operation fusion features, the operating condition vector, the real operating features of the next window, the valve status change, and the pipe section status label.

[0087] Pipeline segment status labels include normal transport, valve regulation, load fluctuation, suspected leakage, and confirmed leakage. For scenarios with few actual leakage events, historical maintenance confirmation records, pressure test records, and leakage disturbance data generated by the simulation platform can be included in the training set, and the leaking pipeline segment and disturbance arrival time can be marked by manual verification.

[0088] The manual verification method can be jointly confirmed by operation records, inspection records, shutdown and maintenance records, and on-site handling records, so that the training samples not only include abnormal waveforms of sensors, but also leakage events corresponding to the actual pipe section locations.

[0089] In this embodiment, the improved Koopman pipeline state evolution model is trained by jointly optimizing the up-dimensional mapping network, the working condition encoding splicing layer, the gating parameter generation unit, the state recursion unit, and the down-dimensional mapping network.

[0090] The training objectives include the error between the predicted state features and the actual operating features in the next window, the smoothness constraint of the state evolution matrix, and the difference constraint between the working condition grouping matrices.

[0091] The loss function can be taken in the following simple form: ; in, ; To predict the error between state characteristics and actual operating characteristics, ; These are the constraint terms of the state evolution matrix. ; For constraint weights.

[0092] During training, the monitoring window length, batch size, learning rate, number of matrix candidates, and width of the gated hidden layer are set according to the pipeline size, number of sensor nodes, and number of training samples.

[0093] Training stops when the model's prediction error on the validation set no longer decreases after several consecutive training epochs, or when the maximum number of training epochs is reached; the maximum number of training epochs is the number of training epochs to terminate training, which is set based on the size of the training set, the convergence speed of the validation set error, and the stability of the model parameters.

[0094] The validation set prediction error no longer decreases when the decrease in the validation set error over multiple consecutive training rounds is less than the error change limit; the error change limit is a convergence judgment parameter determined based on historical training records and model output stability.

[0095] refer to Figure 3 In this embodiment, S41 performs residual calculation based on the difference between the operation fusion characteristics and the predicted state characteristics in the same dimension to obtain the pressure residual sequence, flow residual sequence, vibration residual sequence and acoustic emission residual sequence.

[0096] Specifically, the system performs a dimension-by-dimensional difference calculation between the actual observed operational fusion characteristics and the predicted state characteristics output by the improved Koopman pipeline state evolution model. The pressure channel forms a pressure residual sequence, the flow channel forms a flow residual sequence, the vibration channel forms a vibration residual sequence, and the acoustic emission channel forms an acoustic emission residual sequence. The residual sequence is used to represent the deviation of the actual operating state from the predicted normal evolution state within the current monitoring window.

[0097] In this embodiment, S42 performs segmented standardization processing on each residual sequence based on the pipe section pressure level, medium flow rate, distance between adjacent sensing nodes, and historical fluctuation range to obtain a standardized residual sequence.

[0098] Segmented standardization can be based on pipe segments, or on pressure levels and medium flow velocity ranges.

[0099] For high-pressure pipe sections, the system standardizes them according to the historical residual distribution of the high-pressure operating range; for low-pressure pipe sections, the system standardizes them according to the historical residual distribution of the low-pressure operating range; for pipe sections near valves, the system combines valve action records to separately statistically analyze the residual distribution, avoiding normal fluctuations near valves from being mistaken for abnormal evidence.

[0100] Through segmented standardization, the differences in dimensions, operational intensity, and sensor sensitivity between different pipe sections are compressed into a unified analytical scale.

[0101] In this embodiment, S43 performs evidence splitting based on the residual amplitude, duration, mutation direction, cross-node propagation order, and sensor node location in the standardized residual sequence to obtain initial pressure residual evidence, initial flow residual evidence, initial vibration residual evidence, and initial acoustic emission residual evidence.

[0102] Pressure residual evidence mainly records the magnitude of pressure drop, the number of pressure drop duration windows, the location of pressure anomaly nodes, and the propagation sequence of pressure anomalies; flow residual evidence mainly records the degree of deviation between upstream and downstream flow, the direction of flow change, and the duration of flow anomalies; vibration residual evidence mainly records the sudden increase in vibration energy, peak impact value, abnormal frequency band, and the location of abnormal nodes; acoustic emission residual evidence mainly records the intensity of acoustic emission events, event count, duration, and the arrival sequence of multiple nodes.

[0103] In this embodiment, S44 eliminates the operating condition disturbance component caused by valve adjustment based on the matching relationship between each initial residual evidence and the valve state change, and obtains pressure residual evidence, flow residual evidence, vibration residual evidence and acoustic emission residual evidence.

[0104] The system combines valve action time, valve action direction, valve opening change range, and affected pipe section to identify normal operating disturbances caused by valve regulation. It removes residual components that match the valve action time, valve action direction, and affected pipe section, and retains residual evidence that matches the leakage disturbance propagation law.

[0105] The affected pipe section is the range of upstream and downstream pipe sections affected by valve regulation, determined based on the valve installation location, pipe connection relationship, medium flow direction, and valve action records.

[0106] In this embodiment, evidence splitting does not simply divide residuals into two categories: abnormal and normal. Instead, it forms structured evidence that can be fused based on the source of residuals, the propagation relationship of residuals, and the reliability of sensors.

[0107] For pressure residuals, if the pressure anomaly only occurs after the valve is activated and the pressure and flow rates change in the same direction, the system will classify the pressure residual as a non-leakage supporting factor; if the pressure anomaly is accompanied by a decrease in downstream pressure, an increase in upstream flow compensation, and a consistent propagation sequence between adjacent nodes, the system will classify the pressure residual as a leakage supporting factor.

[0108] For vibration and acoustic emission residuals, if the anomaly is concentrated in a single sensor and there is no response from adjacent nodes, the system will classify the anomaly as an uncertainty factor; if the anomaly exhibits intensity attenuation and time delay along the pipe direction, the system will classify the anomaly as a leakage supporting factor.

[0109] refer to Figure 3In this embodiment, S51 calculates the leakage support, non-leakage support, and uncertain support based on the pressure residual evidence, flow residual evidence, vibration residual evidence, and acoustic emission residual evidence, respectively, to obtain single-source evidence support data.

[0110] The leakage support is determined based on whether the residual amplitude exceeds the standardized fluctuation range of the corresponding pipe section, whether the anomaly propagates along the medium flow direction, and whether the anomaly forms a reasonable arrival time difference between adjacent nodes.

[0111] Non-leakage support is determined based on the degree of matching between residuals and valve action, load changes, and temperature changes; uncertainty support is determined based on sensor health status, data missingness, and noise level.

[0112] Leakage support, non-leakage support, and uncertain support together form the basic probability allocation of single-source evidence, with pressure evidence, flow evidence, vibration evidence, and acoustic emission evidence each corresponding to an independent source of evidence.

[0113] In this embodiment, S52 performs credibility discounting on the single-source evidence support data based on the sensor health status, drift range, noise level, number of intermittent samplings, and communication delay to obtain discounted evidence data.

[0114] The sensor health status is calculated based on sensor drift range, data loss rate, number of communication delays, noise level, and historical maintenance records.

[0115] Historical noise levels are based on noise baseline data obtained from the residual distribution statistics of the same pipe section under leak-free operating conditions.

[0116] Sensor health status can be categorized into high health, medium health, and low health. High health corresponds to sensor nodes with continuous data, stable noise, and normal maintenance records. Medium health corresponds to sensor nodes with slight drift or occasional communication delays. Low health corresponds to sensor nodes with continuous data loss, abnormally high noise, or abnormal maintenance records. For anomalous evidence generated by low-health nodes, the system increases uncertainty support and decreases leakage support. For anomalous evidence generated by high-health nodes and consistent with the propagation relationship of neighboring nodes, the system retains a higher leakage support.

[0117] In this embodiment, S53 determines candidate pipe segments based on the abnormal sensing nodes and pipe segment connection relationships corresponding to the residual evidence, and performs spatial weight correction on the discounted evidence data based on the candidate pipe segments, the direction of medium propagation, and the degree of signal attenuation to obtain weighted evidence data.

[0118] The distance weight is a spatial correction weight determined based on the topological distance between the abnormal sensing node and the candidate pipe segment, the direction of medium propagation, the node installation spacing, and the degree of signal attenuation. Abnormal sensing nodes closer to the candidate pipe segment correspond to higher spatial weights, while abnormal sensing nodes farther from the candidate pipe segment and not conforming to the direction of medium propagation correspond to lower spatial weights.

[0119] For acoustic emission and vibration evidence, the system also adjusts the distance weight by incorporating the signal intensity attenuation relationship; for pressure and flow evidence, the system adjusts the distance weight by incorporating the upstream and downstream fluid transport relationship.

[0120] In this embodiment, S54 inputs the weighted evidence data into the DS evidence theory fusion model, performs a combination operation on the basic probability allocation of multiple evidence sources, and obtains leaked candidate fusion evidence.

[0121] The DS evidence theory fusion model combines the basic probability assignments of pressure evidence, flow evidence, vibration evidence, and acoustic emission evidence to output leakage probability assignment, non-leakage probability assignment, and uncertainty probability assignment.

[0122] The data connection between the model and the improved Koopman pipeline state evolution model is as follows: the improved Koopman pipeline state evolution model outputs predicted state features, the residual evidence generation unit forms multiple types of residual evidence based on the predicted state features and the operational fusion features, and the DS evidence theory fusion model receives residual evidence and auxiliary data such as sensor health status, distance weight, and historical noise level, and outputs leakage candidate fusion evidence. Through this connection method, leakage judgment does not directly rely on a single sensor threshold, but integrates the anomaly intensity, spatial propagation relationship and data credibility of multiple evidence sources based on the predicted residuals.

[0123] refer to Figure 1 and Figure 3 In this embodiment, S61 calculates the state difference based on the leakage probability allocation and non-leakage probability allocation in the leakage candidate fusion evidence, and makes an initial judgment based on the changing trend of the leakage probability allocation in adjacent monitoring windows to obtain the leakage state judgment value.

[0124] Specifically, the system reads the leakage probability allocation, non-leakage probability allocation, and uncertain probability allocation of the current monitoring window, calculates the difference between the leakage probability allocation and the non-leakage probability allocation, and combines the continuous changing trend of the leakage probability allocation in multiple adjacent monitoring windows to form a leakage status judgment value.

[0125] If the leakage probability distribution continues to increase while the non-leakage probability distribution continues to decrease, the leakage status determination value increases; if the leakage probability distribution only increases briefly within a single monitoring window and does not form a continuous trend in subsequent windows, the leakage status determination value decreases.

[0126] In this embodiment, S62 performs a confidence interval correction on the leakage status judgment value based on the uncertainty probability allocation in the leakage candidate fusion evidence, the sensor health status, and the statistical results of the leakage candidate fusion evidence within the continuous monitoring window, to obtain the corrected judgment value; the statistical results include the number of abnormal continuous windows and the number of times the leakage probability has increased continuously.

[0127] The confidence interval correction rule is a judgment value correction rule established based on the sensor health status level, uncertainty probability allocation, number of consecutive abnormal windows, and statistical results of historical false alarm samples. The role of confidence interval correction is that when the leakage probability allocation is high but the uncertainty probability allocation is also high, the system will not immediately output a confirmation of leakage, but will make corrections based on the trend of consecutive windows and the sensor health status.

[0128] When the probability of leakage increases continuously and the evidence from multiple sources is consistent, the system increases the credibility of the judgment value; when the anomaly comes only from nodes in a low health state, the system decreases the credibility of the judgment value.

[0129] In this embodiment, S63 compares the correction judgment value with the suspected leakage threshold and leakage confirmation threshold in the state threshold library, and combines the number of abnormal evidence sources and the consistency of cross-sensor evidence to classify the target pressure pipeline into states and obtain state judgment data.

[0130] The status threshold library is a set of hierarchical thresholds formed by statistical analysis of different pipe section pressure levels, medium types, historical operating fluctuation ranges, sensor node densities, and manual review records. The suspected leakage threshold and the leakage confirmation threshold correspond to the status judgment boundaries under different risk levels.

[0131] For pipe sections with high pressure, high risk, or low sensor node density, the status threshold library can use a more sensitive suspected leakage threshold to detect anomalies as early as possible; for pipe sections with frequent valve operation or high historical noise, the status threshold library can increase the leakage confirmation threshold and require more evidence sources to participate in the confirmation.

[0132] Based on the comparison results, the system classifies the status of the target pressure pipeline into normal, suspected leakage, confirmed leakage, and uncertain data states.

[0133] In this embodiment, S64 encapsulates the results based on the state determination data, the fusion probability allocation in the leakage candidate fusion evidence, the abnormal evidence source corresponding to the residual evidence, the number of abnormal persistence windows, and the corresponding monitoring window number to obtain the leakage monitoring results.

[0134] Leakage monitoring results include status category, leakage probability allocation, non-leakage probability allocation, uncertain probability allocation, source of abnormal evidence, abnormal sensor node, number of abnormal persistence windows, corresponding monitoring window number, and evidence consistency marker.

[0135] The status category can be normal, suspected leak, confirmed leak, and data uncertainty; the source of abnormal evidence can be one or more of pressure residual evidence, flow residual evidence, vibration residual evidence, and acoustic emission residual evidence; the evidence consistency marker is used to record whether different evidence sources are consistent in time sequence, spatial location, and direction of change.

[0136] refer to Figure 3 In this embodiment, S71 extracts the residual peak time based on the abnormal sensing nodes and residual evidence corresponding to the leakage monitoring results, and performs disturbance arrival time processing on the residual peak time to obtain a multi-node disturbance arrival time series.

[0137] The system reads abnormal sensor nodes and residual evidence from the leakage monitoring results, and extracts the residual peak times from the pressure residual, acoustic emission residual and vibration residual. The residual peak times of multiple nodes are sorted and organized. The multi-node disturbance arrival time series records the time order of abnormal disturbances at different sensor nodes, which is the time basis for subsequent location inversion.

[0138] In this embodiment, S72 calculates the propagation time difference from the candidate leak point to the adjacent sensing node based on the arrival time sequence of multi-node disturbances, the connection relationship of pipe segments, and the direction of medium flow, and in combination with the pipe segment propagation velocity determined by the medium type, pipe material, and pipe segment pressure level, to obtain candidate location error data.

[0139] The propagation velocity of a pipe segment is a disturbance propagation parameter determined based on the medium type, pipe material, pipe diameter, pipe wall thickness, pipe segment pressure level, and historical disturbance propagation records. It can be calculated from the basic pipe parameters or calibrated from historical valve action or pressure test data.

[0140] For the same pipe segment, different media types and different pressure levels may correspond to different disturbance propagation velocities. Therefore, when the system performs location inversion, it reads the corresponding propagation velocity according to the candidate pipe segment, rather than using a uniform propagation velocity for the entire pipe network.

[0141] In this embodiment, S73 selects the candidate pipe segment position with the smallest error based on the candidate position error data, and performs position correction by combining the source nodes of pressure residual evidence, acoustic emission residual evidence and vibration residual evidence to obtain corrected candidate position data.

[0142] If the pressure residual evidence, acoustic emission residual evidence, and vibration residual evidence all point to the same candidate pipe segment, the system increases the credibility of the candidate pipe segment location.

[0143] If multiple sources of evidence point to different candidate pipe segments, the system corrects the candidate positions based on sensor health status, residual strength, arrival time consistency, and spatial distance weights, retaining candidate positions with smaller errors and higher levels of evidence support.

[0144] In this embodiment, S74 organizes the location results based on the corrected candidate location data, candidate pipe segment number and the length ratio from the upstream sensing node to obtain the leakage location estimation result.

[0145] The leak location estimation results may include candidate pipe segment number, length ratio from upstream sensor node, corresponding upstream node number, corresponding downstream node number, number of abnormal sensor nodes participating in the location, and location confidence level; the location confidence level is the degree of confidence of the location determined based on the number of abnormal nodes participating in the location, consistency of arrival time series, magnitude of candidate location error, and sensor health status.

[0146] When multiple sensor nodes participate in the localization and the candidate location error is small, the localization reliability level is high; when the number of abnormal nodes is small or there is obvious conflict in the time series, the localization reliability level is low, the system outputs the candidate location range and prompts that it needs to be combined with inspection and verification.

[0147] Example 1: To verify the feasibility of this invention in pressure pipeline leakage monitoring, it was applied to a pressure medium transmission pipeline network in an industrial park. This network comprises 12 continuous pipe sections, 18 sensing nodes, and 4 electrically operated regulating valves, with a total pipeline length of approximately 6.8 km. The normal transmission pressure range is 1.60 MPa to 2.35 MPa, and the normal flow rate range is 420 m³ / h to 680 m³ / h. Each sensing node collects pressure, flow rate, vibration, acoustic emission, and temperature data. Pressure and flow rate data are sampled at the second level, vibration and acoustic emission data are converted into window statistics by an edge acquisition unit, and valve status data is synchronously uploaded by the valve control unit.

[0148] In this embodiment, the monitoring platform sets the monitoring window to 30 seconds and the sliding step size to 10 seconds. Taking the pipe segment containing adjacent sensing nodes numbered P07 to P09 as an example, within a certain continuous monitoring segment, the pressure at node P07 decreases from 2.18 MPa to 2.06 MPa, the pressure at node P08 decreases from 2.15 MPa to 1.98 MPa, and the pressure at node P09 decreases from 2.11 MPa to 1.94 MPa. During the same period, the upstream flow rate increases from 552 m³ / h to 571 m³ / h, and the downstream flow rate decreases from 548 m³ / h to 521 m³ / h, forming a significant pressure-flow mismatch. In the vibration channel, the vibration energy statistics of the pipe segment containing node P08 increase from 0.42 g² to 0.86 g², the acoustic emission event count increases from 18 times per window to 47 times, and the peak acoustic emission intensity increases from 62 dB to 81 dB. The valve status log shows that the opening of adjacent valves remained between 68% and 70% within this monitoring segment without any significant movement. Therefore, the platform did not classify this segment as a valve regulation segment, but rather as an abnormal candidate segment.

[0149] After performing time calibration, missing data completion, and normalization on the aforementioned data, the platform generates multi-sensor aligned data. Subsequently, the platform extracts features such as pressure window slope, flow window slope, pressure-flow matching difference, sudden increase in vibration energy, peak acoustic emission intensity, acoustic emission event count, and cross-node propagation interval to obtain operational fusion features. In this monitoring segment, the pressure-flow matching difference increases from 0.08 during normal operation to 0.31, and the cross-node anomaly arrival time difference is 12s and 15s respectively between P07, P08, and P09, consistent with the abnormal change characteristics of propagation along the medium flow direction.

[0150] During the model analysis phase, the platform runs a Koopman pipeline state evolution model improved by incorporating feature inputs. The model selects the state evolution matrix corresponding to load fluctuations and abnormal candidate operating conditions based on the current pressure level, medium flow velocity range, valve status, and pipe segment connection relationships, and adjusts local evolution parameters through a gating parameter generation unit. The model's predicted state output shows that, under the condition that the current valve opening is basically stable, the predicted pressure value at node P08 should remain between 2.12 MPa and 2.16 MPa, the predicted pressure value at node P09 should remain between 2.08 MPa and 2.13 MPa, and the predicted acoustic emission event count should be less than 25 times per window. Significant deviations exist between the actual monitored values ​​and the predicted values, with the pressure residual at node P08 reaching 0.17 MPa, the pressure residual at node P09 reaching 0.16 MPa, the acoustic emission event count residual at node P08 reaching 22 times, and the vibration energy residual reaching 0.39 g².

[0151] The platform generates pressure residual evidence, flow residual evidence, vibration residual evidence, and acoustic emission residual evidence based on the residual results. The support for leakage is 0.72 for pressure residual evidence, 0.68 for flow residual evidence, 0.63 for vibration residual evidence, and 0.76 for acoustic emission residual evidence. Since the sensor health status of node P08 is 0.94 and that of node P09 is 0.91, and the communication latency is less than 1 second, the platform only applies a slight discount to the above evidence. After distance weighting and signal attenuation correction, the weighted overall support of candidate pipe segment K08 is higher than that of adjacent pipe segments K07 and K09.

[0152] Subsequently, the platform inputs the weighted multi-source evidence into the DS evidence theory fusion model. The fusion results show that candidate pipe segment K08 has a leakage probability of 0.82, a non-leakage probability of 0.09, and an uncertain probability of 0.09; adjacent pipe segment K07 has a leakage probability of 0.41, a non-leakage probability of 0.34, and an uncertain probability of 0.25; adjacent pipe segment K09 has a leakage probability of 0.46, a non-leakage probability of 0.29, and an uncertain probability of 0.25. Based on the status threshold library, the suspected leakage threshold is 0.60 and the confirmed leakage threshold is 0.78 at the current pressure level. Therefore, the platform classifies pipe segment K08 as a confirmed leak and pipe segments K07 and K09 as segments of abnormal concern.

[0153] During the location inversion phase, the platform extracts the peak times of the residuals from nodes P07, P08, and P09. The abnormal peak value at node P07 occurs at 1260s, at node P08 at 1272s, and at node P09 at 1287s. Combining the pipe segment connection relationships, media flow direction, and pipe segment propagation velocity, the platform calculates the propagation time difference from the candidate leak point to the upstream and downstream sensor nodes and selects the candidate location by minimizing the error. The calculation results show that the candidate leak point is located in pipe segment K08, approximately 0.43km from the upstream node P08, accounting for 38% of the length of pipe segment K08. Since the pressure residual, vibration residual, and acoustic emission residual all point to pipe segment K08, and the arrival time series matches the media flow direction, the platform marks the location confidence level as high.

[0154] For comparative verification, this embodiment also compares the traditional pressure threshold alarm method with the present invention. In this segment, the traditional pressure threshold method only triggers an alarm after the pressure drop exceeds a fixed threshold, at 1320 seconds, and cannot distinguish between valve regulation disturbances and actual leakage disturbances. The present invention confirms the leakage at 1280 seconds, 40 seconds earlier than the traditional method, and simultaneously outputs the candidate leak pipe section, the source of abnormal evidence, and the location estimation result. Subsequent verification showed that the leak point was located in the middle-upstream position of pipe section K08, consistent with the location estimation result output by the present invention.

[0155] Through the above process, this embodiment can continuously monitor the leakage status and leakage location in pipeline operation environments where valve regulation and load fluctuations coexist.

[0156] Table 1 Comparison of pressure pipeline leakage monitoring data between the present invention and existing monitoring methods.

[0157] As shown in Table 1, compared with the traditional "pressure threshold + flow difference judgment" method, the present invention significantly improves the completeness, stability, and location capability of pressure pipeline leakage monitoring. The traditional method mainly relies on pressure and flow data for judgment, with limited data types and susceptibility to fluctuations from a single sensor, valve adjustments, and load changes. The present invention simultaneously introduces six types of data: pressure, flow, vibration, acoustic emission, temperature, and valve status. Through timestamp correction, resampling, and monitoring window segmentation, it achieves unified alignment, reducing the data time alignment error from 850ms to 120ms. The effective window percentage after missing data completion increases from 76.5% to 94.2%, demonstrating that the present invention provides a more complete and stable data foundation for subsequent status analysis.

[0158] In adapting to complex operating conditions, traditional methods easily misjudge normal operating condition changes such as valve opening and closing, and load fluctuations as leaks. Therefore, the number of false alarms for valve regulation and load fluctuation conditions reaches 17 and 14, respectively. This invention, through operating condition segmentation and an improved Koopman pipeline state evolution model, predicts the normal operating state under different transport loads, valve actions, pressure levels, and media flow rates. Then, it generates residual evidence based on the difference between the actual operating characteristics and the predicted state characteristics, reducing the number of false alarms for valve regulation to 5 and the number of false alarms for load fluctuations to 4. This demonstrates that this invention does not simply rely on threshold judgment, but first establishes a state evolution benchmark under normal operating conditions, and then identifies abnormal changes deviating from the benchmark.

[0159] Regarding leak detection capabilities, this invention demonstrates higher detection rates for both small-flow and sudden leaks. The detection rate for small-flow leaks increased from 68.4% to 89.7%, primarily because while small leaks may not show significant changes in pressure and flow, they can still generate identifiable signals through acoustic emission, vibration, and multi-node residual propagation characteristics. The detection rate for sudden leaks increased from 91.2% to 97.8%, indicating that the stability of leak status identification is further enhanced after multiple types of residual evidence are used in the judgment. Simultaneously, the average leak detection delay was reduced from 18.6 seconds to 7.4 seconds, demonstrating that this invention can capture multi-sensor anomalies earlier, reducing leak detection lag.

[0160] Regarding false alarms, missed alarms, and interference resistance, the false alarm rate for leakage status in this invention decreased from 12.8% to 4.6%, the missed alarm rate decreased from 9.5% to 3.2%, and the identification accuracy under sensor noise interference increased from 71.6% to 88.9%. These changes indicate that this invention, through processing sensor health status, historical noise levels, and confidence discounting, corrects anomalous evidence, thereby reducing the impact of sensor drift, communication anomalies, and noise interference on the judgment results. After fusing evidence from different sources using the DS evidence theory fusion model, the proportion of conflict windows among multi-source evidence decreased from 23.7% to 9.8%, indicating that conflicts between residual evidence are effectively mitigated.

[0161] Regarding location capabilities, traditional methods typically only output leak alarms, failing to provide accurate leak locations. This invention, by combining calculations based on abnormal sensor nodes, disturbance arrival time difference, pipe segment propagation velocity, and candidate pipe segment errors, reduces the average leak location error from 42.5m to 13.8m, and can output leak status, abnormal evidence sources, candidate pipe segments, and leak location estimation results. Although the average processing time per monitoring window increases from 1.8s to 2.6s, the increased computational load mainly comes from state evolution prediction, residual evidence generation, and evidence fusion processes, still meeting online monitoring requirements. In summary, this invention achieves higher identification accuracy, lower false alarm / false negative rates, and more refined location results with a slight increase in processing time, making it suitable for continuous monitoring of pressure pipeline leaks under complex operating conditions.

[0162] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring leakage in pressure pipelines based on multi-sensor fusion, characterized in that, Includes the following steps: S1. Obtain multi-sensor raw data of the target pressure pipeline during the monitoring period, and perform time calibration, anomaly removal, normalization and pipe segment mapping on the multi-sensor raw data to obtain multi-sensor aligned data. S2. Based on the pressure sequence, flow sequence, vibration sequence, acoustic emission sequence, temperature sequence and valve state sequence in the multi-sensor aligned data, extract the fluctuation feature, offset feature, impact feature and consistency feature to obtain the operation fusion feature; S3. Input the fusion features into the improved Koopman pipeline state evolution model, select the state evolution matrix according to the operating condition code, valve state code and pipe segment pressure level, map and predict the current operating state, and obtain the predicted state features. S4. Based on the operational fusion characteristics and predicted state characteristics, perform residual calculation, standardization, and evidence splitting to obtain pressure residual evidence, flow residual evidence, vibration residual evidence, and acoustic emission residual evidence. S5. Input the residual evidence into the DS evidence theory fusion model, and perform probability allocation based on sensor health status, distance weight and historical noise level to obtain leakage candidate fusion evidence. S6. Based on the leakage probability allocation, non-leakage probability allocation, and uncertain probability allocation in the leakage candidate fusion evidence, the state is determined and the leakage monitoring results are obtained. S7. Based on the abnormal sensor nodes, disturbance arrival time difference and propagation speed corresponding to the leakage monitoring results, perform inversion to obtain the leakage location estimation result.

2. The pressure pipeline leakage monitoring method based on multi-sensor fusion according to claim 1, characterized in that, S1 includes the following steps: S11. Acquire pressure, flow, vibration, acoustic emission, temperature sampling data and valve status data collected by sensor nodes along the target pressure pipeline, and obtain pipeline basic parameters, sensor configuration parameters and historical operating parameters to obtain multi-sensor raw dataset; S12. Based on a unified clock reference and sensor sampling frequency, perform timestamp correction, resampling, and monitoring window segmentation on the multi-sensor raw dataset to obtain synchronized sensing data. S13. Based on the sensor range, historical drift range and abrupt change amplitude limitations, perform anomaly removal, missing data filling, drift correction and normalization on the synchronous sensing data to obtain clean sensing data. S14. Based on the sensor node installation location, pipe segment connection relationship, medium flow direction and valve position, perform spatial mapping and pipe segment attribution marking on the cleaning sensor data to obtain multi-sensor alignment data.

3. The pressure pipeline leakage monitoring method based on multi-sensor fusion according to claim 1, characterized in that, S2 includes the following steps: S21. Calculate the sampling point difference, window mean offset, peak-valley variation amplitude and pressure-flow matching difference based on the pressure sequence and flow sequence in the multi-sensor aligned data to obtain the pressure-flow fluctuation offset characteristics. S22. Based on the vibration sequence and acoustic emission sequence in the multi-sensor aligned data, perform frequency band decomposition, energy statistics, impact peak extraction and duration statistics to obtain the mechanical disturbance impact characteristics. S23. Based on the valve state sequence, the temperature change sequence calculated from the temperature sequence, the pressure and flow fluctuation offset characteristics, and the mechanical disturbance and impact characteristics, the working condition segments are divided to obtain steady-state delivery segments, valve regulation segments, and abnormal candidate segments. S24. Based on the change direction, change amplitude, occurrence order and propagation interval of each sensor node within the same working condition segment, perform consistency calculation to obtain the operation fusion characteristics.

4. The pressure pipeline leakage monitoring method based on multi-sensor fusion according to claim 1, characterized in that, S3 includes the following steps: S31. Based on the pressure-flow combination characteristics, valve status characteristics, temperature compensation characteristics, and cross-sensor consistency characteristics in the operation fusion characteristics, and combined with the pipe section pressure level, medium flow velocity range, and pipe section connection relationship, the current monitoring window is coded to obtain the operating condition vector. S32. Input the running fusion features and operating condition vectors into the improved Koopman pipeline state evolution model's up-dimensional mapping network, and perform nonlinear embedding, operating condition splicing, and state normalization on the low-dimensional sensing features to obtain the high-dimensional state vector. The up-dimensional mapping network includes a multi-branch up-dimensional mapping structure, a working condition coding splicing layer, and a state normalization layer; The multi-branch up-dimensional mapping structure includes a pressure-flow mapping branch and a vibration-acoustic emission mapping branch; S33. Based on the operating condition vector, valve opening change, medium flow rate and pipe pressure level, select the target state evolution matrix from the candidate state evolution matrix to obtain the operating condition matching state evolution parameters. S34. Perform state recursion based on the high-dimensional state vector and the working condition matching state evolution parameters, and map the recursive state back to the original feature space through a dimensionality reduction mapping network to obtain the predicted state features.

5. A pressure pipeline leakage monitoring method based on multi-sensor fusion according to claim 4, characterized in that, S32 includes the following steps: S321. Establish a pressure and flow input vector based on the pressure and flow combination characteristics in the operation fusion characteristics, and establish a vibration and acoustic emission input vector based on the vibration energy increment and acoustic emission peak intensity in the mechanical disturbance and impact characteristics to obtain the branch input characteristics; S322. Input the branch input features into the multi-branch up-dimensional mapping structure of the improved Koopman pipeline state evolution model, perform nonlinear embedding processing on the pressure and flow input vector through the pressure and flow mapping branch to obtain the pressure and flow embedding vector, and perform nonlinear embedding processing on the vibration and acoustic emission input vector through the vibration and acoustic emission mapping branch to obtain the vibration and acoustic emission embedding vector. S323. Generate a working condition embedding vector based on the working condition vector, and input the working condition embedding vector, pressure-flow embedding vector, and vibration-acoustic emission embedding vector into the working condition encoding splicing layer of the improved Koopman pipeline state evolution model to obtain a joint state input vector. S324. Input the joint state input vector into the state normalization layer of the improved Koopman pipeline state evolution model to perform scale unification and state distribution correction on the feature amplitudes formed by different sensing branches, and obtain a high-dimensional state vector.

6. The pressure pipeline leakage monitoring method based on multi-sensor fusion according to claim 4, characterized in that, S33 includes the following steps: S331. Based on the conveying load level, valve action range, temperature change range and pipe section pressure level in the working condition vector, the working condition index is encoded to obtain the working condition index value. The operating condition index value is the matrix selection identifier for the operating condition type corresponding to the current monitoring window; S332. Based on the operating condition index value, call the candidate library of operating condition grouping matrices in the improved Koopman pipeline state evolution model to obtain the initial evolution matrix; The candidate library of working condition grouping matrices is a set of state evolution matrices stored according to steady-state delivery, valve regulation, load fluctuation and pressure transition working conditions respectively; S333. The initial evolution matrix is ​​scaled according to the pipe section length, medium flow velocity and pipe section pressure level to obtain the pressure-velocity correction matrix and scale correction parameters. The scale correction parameter is a matrix scaling parameter calculated from the physical properties of the pipe section and the operating state of the medium. S334. Calculate the feature difference data based on the operation fusion characteristics of the current monitoring window and the previous monitoring window, and calculate the valve state change based on the valve state data of the two monitoring windows. Input the feature difference data and the valve state change into the gating parameter generation unit in the improved Koopman pipeline state evolution model to obtain the gating coefficient. S335. Adjust the local evolution parameters in the pressure-velocity correction matrix according to the gating coefficient to obtain the target state evolution matrix; S336. Based on the target state evolution matrix, scale correction parameters and gating coefficients, the parameters are sorted to obtain the working condition matching state evolution parameters.

7. The pressure pipeline leakage monitoring method based on multi-sensor fusion according to claim 1, characterized in that, S4 includes the following steps: S41. Based on the difference between the operational fusion characteristics and the predicted state characteristics in the same dimension, residual calculation is performed to obtain the pressure residual sequence, flow residual sequence, vibration residual sequence and acoustic emission residual sequence; S42. Based on the pipe section pressure level, medium flow rate, distance between adjacent sensing nodes and historical fluctuation range, each residual sequence is segmented and standardized to obtain a standardized residual sequence. S43. Based on the residual amplitude, duration, mutation direction, cross-node propagation order and sensor node location in the standardized residual sequence, the evidence is split to obtain the initial pressure residual evidence, initial flow residual evidence, initial vibration residual evidence and initial acoustic emission residual evidence. S44. Based on the matching relationship between each initial residual evidence and the valve state change, the operating condition disturbance component caused by valve regulation is eliminated to obtain pressure residual evidence, flow residual evidence, vibration residual evidence and acoustic emission residual evidence.

8. The pressure pipeline leakage monitoring method based on multi-sensor fusion according to claim 1, characterized in that, S5 includes the following steps: S51. Calculate the leakage support, non-leakage support, and uncertainty support based on the pressure residual evidence, flow residual evidence, vibration residual evidence, and acoustic emission residual evidence, respectively, to obtain single-source evidence support data; S52. Based on the sensor health status, drift range, noise level, number of intermittent samplings, and communication delay, perform credibility discounting on the single-source evidence support data to obtain discounted evidence data. S53. Based on the abnormal sensing nodes and pipe segment connection relationships corresponding to the residual evidence, candidate pipe segments are determined, and the discounted evidence data is spatially weighted according to the candidate pipe segments, the direction of medium propagation, and the degree of signal attenuation to obtain weighted evidence data. S54. Input the weighted evidence data into the DS evidence theory fusion model, perform combination calculations on the basic probability allocation of multiple evidence sources, and obtain leaked candidate fusion evidence.

9. The pressure pipeline leakage monitoring method based on multi-sensor fusion according to claim 1, characterized in that, S6 includes the following steps: S61. Calculate the state difference based on the leakage probability allocation and non-leakage probability allocation in the leakage candidate fusion evidence, and make an initial judgment based on the changing trend of the leakage probability allocation in adjacent monitoring windows to obtain the leakage state judgment value. S62. Based on the uncertainty probability distribution in the leakage candidate fusion evidence, the sensor health status, and the statistical results of the leakage candidate fusion evidence within the continuous monitoring window, the leakage status judgment value is corrected for the confidence interval to obtain the corrected judgment value. The statistical results include the number of abnormal persistence windows and the number of consecutive increases in leakage probability; S63. Based on the comparison between the correction judgment value and the suspected leakage threshold and leakage confirmation threshold in the preset state threshold library, and combined with the number of abnormal evidence sources and the consistency of cross-sensor evidence, the target pressure pipeline is classified into states to obtain state judgment data. S64. Based on the state determination data, the fusion probability allocation in the leakage candidate fusion evidence, the source of abnormal evidence corresponding to the residual evidence, the number of abnormal persistence windows and the corresponding monitoring window number, the results are encapsulated to obtain the leakage monitoring results.

10. A pressure pipeline leakage monitoring method based on multi-sensor fusion according to claim 1, characterized in that, S7 includes the following steps: S71. Extract the residual peak time based on the abnormal sensor nodes and residual evidence corresponding to the leakage monitoring results, and process the disturbance arrival time of the residual peak time to obtain the multi-node disturbance arrival time series. S72. Based on the arrival time sequence of multi-node disturbances, pipe segment connection relationship and medium flow direction, and combined with the pipe segment propagation velocity determined by medium type, pipe material and pipe segment pressure level, calculate the propagation time difference from candidate leak point to adjacent sensing node to obtain candidate location error data. S73. Select the candidate pipe segment position with the smallest error based on the candidate position error data, and correct the position by combining the source nodes of pressure residual evidence, acoustic emission residual evidence and vibration residual evidence to obtain corrected candidate position data. S74. Based on the corrected candidate location data, candidate pipe segment number and the length ratio from the upstream sensor node, the location results are sorted to obtain the leakage location estimation result.

Images