Pipeline leakage risk early warning system based on multi-source data fusion

Abstract

This invention relates to the field of pipeline monitoring technology, specifically disclosing a pipeline leakage risk early warning system based on multi-source data fusion. The system includes: a data acquisition and monitoring switching module, a feature extraction module, a probability analysis module, and a risk classification module. It constructs a heterogeneous sensor network for the target pipeline, collects multi-source feature parameters of the pipeline and its surroundings in real time, generates data quality labels, adaptively switches monitoring modes, extracts corresponding dominant feature groups from the multi-source feature parameters based on the monitoring mode, and performs dynamic verification of the physical model and multi-source data fusion based on the monitoring mode, combined with the data quality labels and dominant feature groups. It calculates the physical model fit and comprehensive leakage probability, and performs risk classification and early warning based on the physical model fit and comprehensive leakage probability. This achieves accurate identification and early warning of leakage risks in complex environments, effectively reducing false alarms and missed alarms.

Description

Technical Field

[0001] This invention relates to the field of pipeline monitoring technology, specifically to a pipeline leakage risk early warning system based on multi-source data fusion. Background Technology

[0002] With the acceleration of urbanization, municipal water supply pipelines, as a crucial component of urban lifeline engineering, are directly related to residents' livelihoods and public safety. During long-term service, underground cavities can easily form beneath pipelines due to geological subsidence, construction disturbance, or water seepage and erosion. In severe cases, this can lead to pipeline rupture or even ground collapse. Simultaneously, with the large-scale deployment of fiber optic broadband services and the rapid advancement of next-generation mobile communication core and access networks, communication infrastructure and water supply networks coexist densely in urban underground spaces. Pipeline leakage accidents not only threaten public safety but may also have secondary impacts on nearby communication lines and telecommunications services such as mobile voice and mobile data communication. Therefore, achieving early warning of underground cavities and pipeline leakage risks has become an urgent need for urban infrastructure maintenance and management.

[0003] Currently, pipeline leakage monitoring mainly relies on manual inspections or traditional point sensors, which suffers from limited monitoring range, poor real-time performance, and difficulty in detecting hidden cavities. Distributed fiber optic vibration sensing technology, due to its advantages such as long distance, resistance to electromagnetic interference, and high sensitivity, is gradually being applied to the field of pipeline safety monitoring. However, existing fiber optic monitoring systems still have the following technical problems in practical engineering applications: the underground environment is complex and variable, and the propagation of vibration signals is significantly affected by factors such as soil type and water saturation. Studies have shown that vibration attenuation is relatively small in clay, while it increases sharply in sandy soil or saturated soil. This leads to a significant decrease in the signal-to-noise ratio of single vibration monitoring under soil saturation conditions, making it difficult to effectively identify abnormal information from cavities. It is also difficult to use geographic information mapping, underground pipeline mapping, and other technologies to perform detailed spatial modeling of pipelines and their surrounding environment, resulting in limited accuracy in locating leakage points.

[0004] Existing systems generally lack effective integration of big data resource services, making it difficult to leverage internet data services such as cloud computing software, databases, and cloud database services to achieve in-depth mining and fusion analysis of monitoring data. Furthermore, existing monitoring methods mostly rely on single physical field judgments, lack multi-source data fusion mechanisms, and have not formed a systematic fault prediction and health management capability. They are unable to continuously assess and predict the health status of pipelines before leakage occurs, making it difficult to achieve true pre-diagnosis and health management.

[0005] To address the above problems, this invention proposes a pipeline leakage risk early warning system based on multi-source data fusion. Summary of the Invention

[0006] The purpose of this invention is to provide a pipeline leakage risk early warning system based on multi-source data fusion to solve the aforementioned background problems.

[0007] The objective of this invention can be achieved through the following technical solution: a pipeline leakage risk early warning system based on multi-source data fusion, comprising: Acquisition and monitoring switching module: Constructs a heterogeneous sensor network for the target pipeline, collects multi-source characteristic parameters of the pipeline and its surroundings in real time and generates data quality labels, and adaptively switches monitoring modes; Feature extraction module: Based on the monitoring mode, extracts the corresponding dominant feature groups from multi-source feature parameters; Probability Analysis Module: Based on the monitoring mode, it combines data quality labels and dominant feature groups to perform dynamic verification of the physical model and multi-source data fusion, and calculates the physical model fit and overall leakage probability. Risk classification module: Provides risk classification and early warning based on the fit of the physical model and the overall leakage probability; Furthermore, the adaptive switching of monitoring modes is achieved as follows: Based on real-time soil saturation data obtained from multi-source feature parameters, the saturation state of the soil is dynamically assessed. Soil saturation S is defined as the percentage of the average volumetric water content of each soil volume in the soil saturation data to the preset saturated volumetric water content. When S is below 60%, the monitoring mode is activated as vibration-dominated mode. When S is between 60% and 85%, the vibration-electrical method co-dominated mode is activated. When S is above 85%, the mode is switched to electrical method-temperature co-dominated mode. Furthermore, the method for extracting the dominant feature set corresponding to the dominant vibration mode is as follows: The vibration data, including the multi-source feature parameters, is subjected to DC removal and bandpass filtering to obtain a continuous vibration signal. A sliding window framing technique is used to divide the signal into fixed-length short frames. Vibration features are extracted from the current short frame as the dominant feature group. These features include the time-domain features, frequency-domain features, and vibration energy distribution curve of the continuous vibration signal. The time-domain features include the root mean square value, peak factor, and kurtosis index. The spectrum of the continuous vibration signal is obtained through Fast Fourier Transform, and the frequency band energy and centroid frequency of the preset leakage feature frequency band are extracted based on the spectrum. These are then integrated to obtain the frequency domain features. The frequency band energy of the preset leakage feature frequency band is used as the vibration energy measure of the corresponding pipe location, and the vibration energy distribution curve is plotted along the longitudinal direction of the target pipe. Furthermore, the extraction method for the dominant feature set corresponding to the vibration electrical method cooperative dominant mode is as follows: For the current short frame, vibration features and resistivity features are extracted as the dominant feature group; For vibration characteristics, the continuous vibration signal compensation process is initiated. The compensated continuous vibration signal is used as input and the signal-to-noise ratio in the leakage characteristic frequency band is calculated. If it is lower than the preset available working condition threshold, the data confidence of the vibration data is set to zero. Otherwise, the vibration characteristics are extracted from the compensated continuous vibration signal in the same way as the vibration-dominant mode. For resistivity features, a resistivity distribution profile is generated using a high-density resistivity inversion algorithm for resistivity data included in multi-source feature data. The resistivity temporal gradient and resistivity spatial gradient are calculated and integrated to obtain an electrical resistivity feature vector. Resistivity anomaly regions that meet a preset area threshold are identified through image segmentation and connected component analysis. The geometric center, area, and average anomaly amplitude of the resistivity anomaly regions are extracted to obtain resistivity anomaly region information. If there are no resistivity anomaly regions, the anomaly confidence of the resistivity features is set to 0, and the resistivity features are integrated to obtain the resistivity features. Furthermore, the specific method for continuous vibration signal compensation is as follows: A Biot theoretical model is constructed to describe the mapping relationship between the attenuation coefficients of different frequency components with propagation distance and soil saturation, offline calibrated soil intrinsic parameters, and excitation frequency. Inverse Q filtering technology is used to decompose the continuous vibration signal into the time-frequency domain through short-time Fourier transform to obtain a time-frequency representation. The attenuation coefficients corresponding to each frequency component are calculated based on the propagation distance and the Biot theoretical model. The corresponding gain function is determined to perform amplitude recovery and phase correction for each time-frequency point in the time-frequency representation to obtain the compensated time-frequency representation. The compensated continuous vibration signal is then reconstructed through inverse short-time Fourier transform. Furthermore, the extraction method for the dominant feature set corresponding to the electrochemical temperature-coordinated dominant mode is as follows: For the current short frame, resistivity and temperature features are extracted as the dominant feature group, and resistivity features are calculated and processed in the same way as the vibration electrical method co-dominant mode. For temperature features, background temperature field subtraction and drift correction are performed on the temperature data included in the multi-source feature data. The temporal gradient and spatial gradient of the temperature data are extracted and integrated to obtain the temperature feature vector. An adaptive threshold method is used to remove false positive temperature anomalies caused by normal pipeline operating condition fluctuations. Temperature anomaly regions that meet the preset area threshold and duration are identified. The geometric center, area and average anomaly amplitude of the temperature anomaly regions are extracted to obtain the temperature anomaly region information. If there is no temperature anomaly region, the anomaly confidence of the temperature feature is set to 0. The temperature feature vector and the temperature anomaly region information are the extracted temperature features. Furthermore, the method for obtaining the overall leakage probability under the vibration-dominated mode is as follows: A single evidence source decision strategy is adopted. For the vibration features included in the dominant feature group, the ratio R of the frequency band energy of the current short-time intra-frame leakage feature frequency band included in the vibration features to the preset historical background energy is calculated. The ratio R is converted into a probability value in the interval [0,1] through a nonlinear mapping function. The obtained probability value is corrected by calculating the data confidence level in combination with the data quality label, and the vibration leakage probability value is obtained as the comprehensive leakage probability under the vibration dominant mode. Furthermore, the method for obtaining the overall leakage probability under the vibration-electrical resistivity tomography (PET) synergistic dominant mode is as follows: The physical model fit of resistivity data is obtained, and the physical model fit is linearly mapped to the fit modulation factor. A weighted fusion strategy is adopted, and the initial weight of vibration features is dynamically adjusted with soil saturation. The data confidence of vibration data and resistivity data is obtained or calculated through data quality labels. The product of the fit modulation factor, data confidence and initial weight is used as the actual weight. The vibration leakage probability value is calculated in the same way as the vibration-dominant mode. If the anomaly confidence of the resistivity feature is not 0, the resistivity leakage probability value is calculated based on the resistivity anomaly area information. The vibration leakage probability value and the resistivity leakage probability value are weighted and fused using actual weights to obtain the comprehensive leakage probability. Furthermore, the method for obtaining the physical model fit of the resistivity data is as follows: A saturated soil seepage-electrical method coupled model is constructed. Multi-source characteristic parameters are used as inputs. The abnormal resistivity characteristic vector is predicted by finite element numerical solution. The fit is calculated by comparing the resistivity characteristic vector and the physical model fit of resistivity data is obtained. Furthermore, the method for obtaining the overall leakage probability under the electrochemical temperature-coordinated dominant mode is as follows: A heat conduction model is constructed, using multi-source characteristic parameters as input. The abnormal temperature characteristic vector is predicted by finite element numerical solution. The fit is calculated by comparing the temperature characteristic vectors to obtain the physical model fit of the temperature data. The physical model fit of the resistivity data is calculated and the physical model fit is linearly mapped to the corresponding fit modulation factor. An evidence theory fusion framework is initiated, and an identification framework is constructed that includes two mutually exclusive propositions: the existence of leakage and the non-existence of leakage. Resistivity and temperature features are used as evidence sources. The anomaly confidence levels of resistivity and temperature features are obtained or calculated based on the evidence sources and multiplied using a matching modulation factor to obtain the basic probability allocation supporting the existence of leakage. The corresponding quality confidence level is calculated based on the corresponding data quality label. The Dempster-Shafer evidence synthesis rule is used to perform orthogonal summation on the basic probability allocations of each evidence source to calculate the comprehensive leakage probability.

[0008] The beneficial effects of this invention are as follows: 1. This invention collects multi-source feature parameters and generates data quality labels by deploying a heterogeneous sensor network. Based on the soil saturation adaptive switching monitoring mode, it solves the problem that traditional single monitoring methods cannot adapt to changes in the soil environment. It extracts the dominant feature groups for different modes, and uses the Biot theory model and inverse Q filtering to compensate for the vibration signal in the vibration electrical method co-dominant mode, thereby restoring the effective information of attenuation in saturated soil and ensuring the reliability and effectiveness of data acquisition under complex working conditions.

[0009] 2. This invention calculates the physical model fit by constructing a seepage-electrical coupling model and a heat conduction model, verifies the physical laws of measured characteristics, effectively suppresses false alarms caused by environmental interference, and uses a weighted fusion strategy and Dempster-Shafer evidence theory to handle conflicts of multiple evidence sources, obtaining a reliable comprehensive leakage probability. It also combines risk classification and a sliding window mechanism for early warning decision-making. When the model fit does not meet the standard, the system self-check is triggered first, avoiding false alarms caused by sensor failure. Attached Figure Description

[0010] The invention will now be further described with reference to the accompanying drawings.

[0011] Figure 1 This is a module architecture diagram of a pipeline leakage risk early warning system based on multi-source data fusion, as described in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the specific steps of a pipeline leakage risk early warning system based on multi-source data fusion, as described in an embodiment of the present invention. Detailed Implementation

[0012] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0013] Example 1 Please see Figure 1 and Figure 2 As shown in the embodiment of the present invention, a pipeline leakage risk early warning system based on multi-source data fusion aims to solve the problem of existing pipeline leakage monitoring methods being singular and unable to adapt to changes in soil environment, resulting in a high false alarm rate. This system collects multi-source feature parameters and generates data quality labels by constructing a heterogeneous sensor network. Based on soil saturation, it adaptively switches monitoring modes, extracts corresponding dominant feature groups, and combines dynamic verification of physical models with multi-source data fusion to calculate the physical model's fit and comprehensive leakage probability, and performs risk classification and early warning. This achieves accurate identification and early warning of leakage risks in complex environments, effectively reducing false alarm and missed alarm rates. Specifically, it includes the following modules: Acquisition and monitoring switching module: Constructs a heterogeneous sensor network for the target pipeline, collects multi-source characteristic parameters of the pipeline and its surroundings in real time and generates data quality labels, and adaptively switches monitoring modes; Specifically, a heterogeneous sensor network is deployed at key sections and upstream and downstream areas of the target pipeline, including distributed fiber optic vibration sensors, distributed fiber optic temperature sensors, resistivity detection electrode arrays, pressure sensors, flow sensors, and soil moisture sensors. All sensors are connected to a central processing platform via a unified data acquisition unit to collect multi-source characteristic parameters of the pipeline and its surroundings in real time, including vibration data, temperature data, resistivity data, pressure data, flow data, and soil saturation data. Among them, distributed fiber optic vibration sensors and distributed fiber optic temperature sensors collect continuous vibration and temperature data along the longitudinal direction of the pipeline; resistivity detection electrode arrays are arranged on the ground surface on both sides of the pipeline according to a preset electrode spacing to collect resistivity data; pressure sensors are arranged at key hydraulic characteristic points of the target pipeline to collect pressure data; flow sensors are distributed along the target pipeline to collect flow data; and soil moisture sensors are buried at preset typical locations around the pipeline to monitor soil saturation data in real time. The soil saturation data includes the volumetric water content of the soil at each preset typical location. Preprocessing and spatiotemporal alignment of the collected multi-source feature parameters: For discrete data collected by point sensors such as resistivity detection electrode arrays and soil moisture sensors, the Kriging interpolation algorithm is used to extend the spatial domain into a continuous profile that matches the multi-source feature parameters collected by distributed optical fiber sensors. A sliding time window mechanism is used to perform time registration of multi-source feature parameters in units of fixed length time windows to ensure that multi-source feature parameters of the same time and the same monitoring profile correspond accurately. The quality of each type of multi-source feature parameter within the time window is evaluated to generate data quality labels. The data quality labels include signal-to-noise ratio, data integrity rate, and abnormal glitch identifier. The data integrity rate is obtained by statistically analyzing the percentage of the actual number of data sampling points received to the theoretical number of sampling points that should be received. The abnormal glitch identifier is identified and marked based on the signal amplitude mutation detection algorithm. If no abnormal glitch is detected, the abnormal glitch identifier is 0; otherwise, it is 1. An example data quality label table is shown below; Table 1: Example of data quality labels (excerpt from the table); Based on real-time soil saturation data obtained from multi-source feature parameters, the saturation state of the soil is dynamically assessed, and the monitoring mode is adaptively switched according to the preset threshold range. Specifically, soil saturation S is defined as the percentage of the average volumetric water content of each soil volume in the soil saturation data to the preset saturated volumetric water content. When S is below 60%, the soil is determined to be in an unsaturated state, and the monitoring mode is activated as vibration-dominated mode. When S is between 60% and 85%, the soil is determined to be in a transitional state, and the monitoring mode is activated as vibration-electrical method-coordinated dominant mode. When S is above 85%, the soil is determined to be in a saturated state. At this time, the vibration signal is rapidly attenuated due to the viscosity of the pore water in the soil, and the monitoring mode is switched to electrical method-temperature-coordinated dominant mode. It should be noted that the 60% and 85% thresholds are pre-set based on the critical points of capillary water saturation and pore connectivity of typical soils. In actual engineering applications, they need to be calibrated and adjusted according to the results of on-site geotechnical tests. It should be noted that the adaptive switching monitoring mode ensures that the system can collect effective data with the optimal combination of physical fields under different soil saturation levels, laying the foundation for subsequent leakage feature extraction and fusion early warning. Feature extraction module: Based on the monitoring mode, extracts the corresponding dominant feature groups from multi-source feature parameters; Specifically, the vibration data, including the multi-source characteristic parameters, is subjected to DC removal and bandpass filtering. A typical passband is set to eliminate power frequency interference and high-frequency environmental noise to obtain a continuous vibration signal. The continuous vibration signal is divided into short time frames of fixed length as data units using sliding window framing technology. If the monitoring mode is in vibration-dominant mode, vibration features are extracted as the dominant feature group for the current short frame. The vibration features include the time-domain and frequency-domain features of the continuous vibration signal and the vibration energy distribution curve. The time-domain features include the root mean square value, peak factor and kurtosis index. The spectrum of the continuous vibration signal is obtained by fast Fourier transform, and the frequency band energy and centroid frequency of the leakage feature frequency band preset according to the target pipe material and size are extracted based on the spectrum. The frequency domain features are integrated and the frequency band energy of the leakage feature frequency band is used as the vibration energy measure of the corresponding pipe position. The vibration energy distribution curve is drawn along the longitudinal direction of the target pipe. It should be noted that the center of gravity frequency refers to the frequency value obtained by calculating the energy-weighted average of the spectrum of the vibration signal within the leakage characteristic frequency band, which is used to characterize the concentration trend of vibration energy within this frequency band. If the monitoring mode is in the vibration-electrical resistivity combined dominant mode, the vibration features and resistivity features are extracted as the dominant feature group for the current short frame. Specifically, for vibration characteristics, a continuous vibration signal compensation process is initiated; A Biot theory model is constructed, which analyzes the vibration propagation attenuation in saturated porous media based on Biot theory. The model describes the mapping relationship between the attenuation coefficient of different frequency components with propagation distance and soil saturation, soil porosity, soil permeability and excitation frequency. The propagation distance is preset based on the mileage coordinates of the target pipeline. The soil porosity and soil permeability are determined by on-site geotechnical tests and are inherent parameters of the soil that are calibrated offline. The excitation frequency is provided by the spectrum of the continuous vibration signal, which is the frequency value of each frequency component. The inverse Q-filtering technique is used to decompose the continuous vibration signal into the time-frequency domain through short-time Fourier transform to obtain a time-frequency representation. The attenuation coefficients corresponding to each frequency component are calculated based on the propagation distance and the Biot theoretical model. Based on the attenuation coefficients, the corresponding gain functions are determined to perform amplitude recovery and phase correction for each time-frequency point in the time-frequency representation to obtain the compensated time-frequency representation. The compensated continuous vibration signal is then reconstructed through inverse short-time Fourier transform. It should be noted that, as an inverse filtering method, the compensation accuracy of inverse Q-filtering is highly dependent on the accuracy of the attenuation model Q-value. The Biot theoretical model provides a theoretical basis for the dynamic estimation of the Q-value. The Q-value represents the quality factor, which is an indicator describing the rate at which a medium such as soil consumes energy during fluctuations. Using the compensated continuous vibration signal as input, the signal-to-noise ratio of the continuous vibration signal in the leakage characteristic frequency band is calculated. If the signal-to-noise ratio is lower than the preset usable working condition threshold, the continuous vibration signal is determined to be in failure, and the data confidence of the vibration data is set to zero. Otherwise, the vibration characteristics are extracted from the compensated continuous vibration signal in the same way as the vibration-dominant mode. For resistivity features, a resistivity distribution profile of the target pipeline and surrounding soil is generated using a high-density resistivity inversion algorithm, targeting the resistivity data included in the multi-source feature data. Iterative inversion is performed using a least squares method based on smooth constraints to eliminate interference from terrain and electrode arrangement on the detection results, yielding the inverted resistivity distribution profile. Based on this profile, the difference between the current resistivity distribution profile and the historical baseline profile is calculated to obtain the resistivity temporal gradient. The spatial gradient of resistivity along the longitudinal direction of the target pipeline is then calculated based on the resistivity distribution profile to obtain the resistivity spatial gradient. The resistivity temporal gradient and resistivity spatial gradient are integrated to obtain the electrical resistivity feature vector. Image segmentation and connected component analysis are used to identify resistivity anomaly regions that meet a preset area threshold. The geometric center, area, and average anomaly amplitude of these anomaly regions are extracted to obtain resistivity anomaly region information. The electrical resistivity feature vector and the resistivity anomaly region information constitute the extracted resistivity features. It should be noted that if image segmentation and connected component analysis fail to identify regions with abnormal resistivity, the information on regions with abnormal resistivity will be marked as empty, and the anomaly confidence of the resistivity feature will be set to 0. The anomaly confidence represents the degree of certainty that the resistivity feature supports the proposition that leakage exists. It should be noted that the historical reference profile represents the standard profile of the spatial distribution of background resistivity of the soil, obtained by inversion using the high-density resistivity method within a pre-selected historical period. It is used to compare with the current measured profile to identify the temporal changes in resistivity. If the monitoring mode is in the electrical-temperature co-dominant mode, the resistivity and temperature features are extracted as the dominant feature group for the current short frame. Specifically, for resistivity characteristics, the resistivity characteristics are calculated and organized in the same way as the vibration electrical method co-dominant mode; For temperature features, the temperature data included in the multi-source feature data is subjected to background temperature field subtraction and drift correction. The temporal gradient of the temperature data is extracted, that is, the rate of change of temperature data at the same location per unit time. The spatial gradient of the temperature data is extracted, that is, the rate of change of temperature data along the axial direction of the target pipeline. The temporal and spatial gradients of the temperature data are integrated to obtain the temperature feature vector. Combined with the pipeline operating conditions, an adaptive threshold method is used to remove false positive temperature anomalies caused by normal pipeline operating condition fluctuations. The adaptive threshold is taken as ±3 times the standard deviation of the mean of temperature data within a preset time period with the current time as the endpoint. Temperature anomaly regions that meet the preset area threshold and duration are identified, and the geometric center, area and average anomaly amplitude of the temperature anomaly regions are extracted to obtain the temperature anomaly region information. The temperature feature vector and the temperature anomaly region information are the extracted temperature features. Exemplary pipeline operating conditions include changes in the temperature of the transported medium and changes in the ambient temperature; It should be noted that if the adaptive threshold method fails to identify the temperature anomaly area, the temperature anomaly area information will be marked as empty, and the anomaly confidence of the temperature feature will be set to 0. Probability Analysis Module: Based on the monitoring mode, it combines data quality labels and dominant feature groups to perform dynamic verification of the physical model and multi-source data fusion, and calculates the physical model fit and overall leakage probability. If the monitoring mode is in the vibration-dominant mode, a single evidence source decision strategy is adopted, with the vibration characteristics included in the dominant feature group as the core basis, supplemented by data quality labels for credibility correction. Specifically, the vibration characteristics include the ratio R of the frequency band energy of the current short-time intra-frame leakage characteristic frequency band to the historical background energy. The historical background energy is taken from the statistical mean of the frequency band energy of the leakage characteristics during a preset historical period without leakage. The ratio R is converted into a probability value in the interval [0,1] through a nonlinear mapping function. The nonlinear mapping function is designed as an S-shaped curve, so that when the ratio R is lower than the preset normal fluctuation range, the probability value approaches 0, and when the ratio exceeds the preset normal fluctuation range, the probability value rises rapidly to close to 1. The probability value is most sensitive to changes within the preset normal fluctuation range, thereby achieving an effective response to small frequency band energy changes. The probability values ​​obtained are corrected by combining data quality labels. The signal-to-noise ratio, data integrity rate, and abnormal spur identifiers of the vibration data are normalized and weighted to obtain the data confidence of the vibration data. The data confidence is used as a modulation factor and multiplied with the obtained probability values ​​to obtain the vibration leakage probability value, which is directly used as the comprehensive leakage probability under the vibration-dominated mode. Under the vibration-dominated mode, the soil is in an unsaturated state, and no physical model fit calculation is performed. The physical model fit of the vibration data is set to 1 by default. If the monitoring mode is in the vibration-electrical method co-dominant mode, a saturated soil seepage-electrical method coupled model is constructed. The seepage-electrical method coupled model is established based on the saturated soil seepage equation and the Archie conductivity relationship. It describes the spatiotemporal evolution law of resistivity caused by the change of pore water saturation after seepage water intrudes into the soil. The pressure data, flow data, soil saturation and soil porosity contained in the soil's inherent parameters are used as inputs to the saturated soil seepage-electrical method coupled model. The abnormal resistivity feature vector that should appear at the current time is predicted by finite element numerical solution under the assumption of seepage. The abnormal resistivity feature vector is compared with the resistivity feature vector. The structural similarity index is used to calculate the degree of fit, with a value range of [0,1], to obtain the physical model fit of the resistivity data. The physical model fit is linearly mapped to the fit modulation factor between [0.5,1]. A weighted fusion strategy is adopted, and the initial weight of vibration features is dynamically adjusted with soil saturation. When soil saturation increases linearly from 60% to 85%, the initial weight of vibration features decreases linearly from 1 to 0. The sum of the initial weights of vibration features and electrical resistivity features is 1. If the data confidence of vibration data is not set to 0, the corresponding data confidence is obtained after normalization and weighted fusion based on the data quality labels of vibration data and resistivity data, including signal-to-noise ratio, data integrity rate, and abnormal spurs. Otherwise, the data confidence of resistivity data is set to 1, and the product of the matching modulation factor, data confidence, and initial weight is used as the actual weight. It should be noted that the role of the matching modulation factor is to enable the fusion system to autonomously question abnormal signals that deviate significantly from the physical model, effectively suppressing false alarms caused by sensor drift and environmental interference. The vibration leakage probability value is calculated in the same way as the vibration-dominant mode. If the anomaly confidence of the resistivity feature is not set to 0, the average anomaly amplitude and area included in the resistivity anomaly area information are mapped to the [0,1] interval through a preset membership function to obtain the resistivity leakage probability value. Otherwise, the resistivity leakage probability value is 0. The vibration leakage probability value and the resistivity leakage probability value are weighted and fused using actual weights to obtain the comprehensive leakage probability. If the monitoring mode is in the electrical temperature-coordinated dominant mode, a heat conduction model is constructed. The heat conduction model is based on the heat diffusion equation and describes the heat exchange process between the seepage medium and the surrounding soil. The temperature data, flow data, soil saturation, and soil porosity contained in the soil's inherent parameters are used as inputs to the heat conduction model. The abnormal temperature feature vector that should appear at the current moment is predicted by finite element numerical solution under the assumption of seepage. The abnormal temperature feature vector is compared with the temperature feature vector, and the degree of fit is calculated by normalized cross-correlation. The physical model fit of the temperature data is obtained. The physical model fit of the resistivity data is calculated. The physical model fit is linearly mapped to a fit modulation factor between [0.5, 1]. An evidence theory fusion framework is initiated, and an identification framework is constructed. The identification framework contains two mutually exclusive propositions: the existence of leakage and the non-existence of leakage, which serve as the target set for fusion decision. Resistivity features and temperature features are used as evidence sources. The anomaly confidence of resistivity features and temperature features is obtained. If they are not 0, they are mapped to the [0,1] interval through a preset membership function based on the average anomaly amplitude and area of ​​the resistivity anomaly region and the temperature change amplitude and area of ​​the temperature anomaly region, respectively, and multiplied using a matching modulation factor to obtain the basic probability allocation that supports the existence of leakage proposition. The corresponding quality confidence is calculated based on the corresponding data quality label and is also expressed in the form of basic probability allocation. The Dempster-Shafer evidence synthesis rule is adopted to perform orthogonal summation on the basic probability allocation of each evidence source. By comprehensively considering the basic probability allocation of each evidence source to support the proposition with leakage and the degree of conflict between evidence sources, the comprehensive leakage probability is calculated. The degree of conflict between evidence sources is obtained by calculating the sum of the products of the basic probability allocations of each evidence source to the mutually exclusive proposition, which reflects the degree of contradiction between each evidence source under the same identification framework. It should be noted that this synthesis process can reasonably handle inconsistencies between evidence sources. When the evidence is consistent in its judgment of leakage, the fusion result strengthens the common support direction. When there are conflicts between the evidence, the fusion result will appropriately reduce the credibility of the fusion probability according to the degree of conflict, so as to avoid being misled by a single contradictory piece of evidence. Risk classification module: Provides risk classification and early warning based on the fit of the physical model and the overall leakage probability; Specifically, the warning level of the overall leakage probability is divided into three levels according to the preset risk range: low risk, medium risk and high risk. If the overall leakage probability is lower than the preset risk range, it is judged as low risk, indicating that the possibility of leakage in the target pipeline is extremely low. If the overall leakage probability is within the preset risk range, it is judged as medium risk, indicating that there is a potential leakage possibility. If the overall leakage probability exceeds the preset risk range, it is judged as high risk, indicating that a high probability leakage event will occur. If the fit of all physical models corresponding to the overall leakage probability fails to meet the fit standard, it indicates that the current monitoring data deviates significantly from the physical model. The system will output a self-check prompt, reminding the user to prioritize checking the sensor status, model parameters, or sudden environmental changes, and will not issue a leakage warning to avoid false alarms. Otherwise, a sliding window mechanism is introduced, using N consecutive short frames as a statistical window. If more than half of the short frames in the statistical window are determined to be of medium risk, on-site inspection is arranged. If more than half of the short frames in the statistical window are determined to be of high risk, an alarm is immediately triggered and sent to the administrator terminal. Otherwise, monitoring continues. The technical solution of this invention is as follows: a heterogeneous sensor network is constructed for the target pipeline, multi-source feature parameters of the pipeline and its surroundings are collected in real time and data quality labels are generated, monitoring modes are adaptively switched, and based on the monitoring modes, the corresponding dominant feature groups are extracted from the multi-source feature parameters. The physical model is dynamically verified and multi-source data is fused by combining the data quality labels and the dominant feature groups, the physical model fit and comprehensive leakage probability are calculated, and risk classification and early warning are performed based on the physical model fit and comprehensive leakage probability.

[0014] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A pipeline leakage risk early warning system based on multi-source data fusion, characterized in that: Includes the following modules: Acquisition and monitoring switching module: Constructs a heterogeneous sensor network for the target pipeline, collects multi-source characteristic parameters of the pipeline and its surroundings in real time and generates data quality labels, and adaptively switches monitoring modes; Feature extraction module: Based on the monitoring mode, extracts the corresponding dominant feature groups from multi-source feature parameters; Probability Analysis Module: Based on the monitoring mode, it combines data quality labels and dominant feature groups to perform dynamic verification of the physical model and multi-source data fusion, and calculates the physical model fit and overall leakage probability. Risk classification module: Risk classification and early warning are conducted based on the fit of the physical model and the overall leakage probability.

2. The pipeline leakage risk early warning system based on multi-source data fusion according to claim 1, characterized in that: The adaptive switching monitoring mode is as follows: Based on real-time soil saturation data obtained from multi-source feature parameters, the saturation state of the soil is dynamically assessed. Soil saturation S is defined as the percentage of the average volumetric water content of each soil volume in the soil saturation data to the preset saturated volumetric water content. When S is below 60%, the monitoring mode is activated as vibration-dominated mode. When S is between 60% and 85%, the vibration-electrical method co-dominated mode is activated. When S is above 85%, the mode is switched to electrical method-temperature co-dominated mode.

3. The pipeline leakage risk early warning system based on multi-source data fusion according to claim 2, characterized in that: The extraction method for the dominant feature set corresponding to the vibration-dominant mode is as follows: The vibration data, including multi-source feature parameters, is subjected to DC removal and bandpass filtering to obtain a continuous vibration signal. A sliding window framing technique is used to divide the signal into fixed-length short frames. Vibration features are extracted from the current short frame as the dominant feature group. These features include the time-domain features, frequency-domain features, and vibration energy distribution curve of the continuous vibration signal. The time-domain features include the root mean square value, peak factor, and kurtosis index. The spectrum of the continuous vibration signal is obtained through Fast Fourier Transform, and the frequency band energy and centroid frequency of the preset leakage feature frequency band are extracted based on the spectrum. These are then integrated to obtain the frequency domain features. The frequency band energy of the preset leakage feature frequency band is used as the vibration energy measure of the corresponding pipe location, and the vibration energy distribution curve is plotted along the longitudinal direction of the target pipe.

4. The pipeline leakage risk early warning system based on multi-source data fusion according to claim 3, characterized in that: The extraction method for the dominant feature set corresponding to the vibration electrical method cooperative dominant mode is as follows: For the current short frame, vibration features and resistivity features are extracted as the dominant feature group; For vibration characteristics, the continuous vibration signal compensation process is initiated. The compensated continuous vibration signal is used as input and the signal-to-noise ratio in the leakage characteristic frequency band is calculated. If it is lower than the preset available working condition threshold, the data confidence of the vibration data is set to zero. Otherwise, the vibration characteristics are extracted from the compensated continuous vibration signal in the same way as the vibration-dominant mode. For resistivity features, a resistivity distribution profile is generated using a high-density resistivity inversion algorithm for resistivity data included in the multi-source feature data. The resistivity temporal gradient and resistivity spatial gradient are calculated and integrated to obtain the resistivity feature vector. Resistivity anomaly regions that meet the preset area threshold are identified through image segmentation and connected component analysis. The geometric center, area, and average anomaly amplitude of the resistivity anomaly regions are extracted to obtain resistivity anomaly region information. If there are no resistivity anomaly regions, the anomaly confidence level of the resistivity feature is set to 0, and the resistivity features are integrated to obtain the resistivity features.

5. A pipeline leakage risk early warning system based on multi-source data fusion according to claim 4, characterized in that: The specific method for continuous vibration signal compensation is as follows: A Biot theoretical model was constructed to describe the mapping relationship between the attenuation coefficients of different frequency components with propagation distance and soil saturation, offline calibrated soil intrinsic parameters, and excitation frequency. Inverse Q filtering technology was used to decompose the continuous vibration signal into the time-frequency domain through short-time Fourier transform to obtain a time-frequency representation. The attenuation coefficients corresponding to each frequency component were calculated based on the propagation distance and the Biot theoretical model. The corresponding gain function was determined to perform amplitude recovery and phase correction for each time-frequency point in the time-frequency representation to obtain the compensated time-frequency representation. The compensated continuous vibration signal was then reconstructed through inverse short-time Fourier transform.

6. A pipeline leakage risk early warning system based on multi-source data fusion according to claim 4, characterized in that: The extraction method for the dominant feature set corresponding to the electro-thermal co-dominant mode is as follows: For the current short frame, resistivity and temperature features are extracted as the dominant feature group, and resistivity features are calculated and sorted in the same way as the vibration electrical method co-dominant mode. For temperature features, background temperature field subtraction and drift correction are performed on the temperature data included in the multi-source feature data. The temporal gradient and spatial gradient of the temperature data are extracted and integrated to obtain the temperature feature vector. An adaptive threshold method is used to remove false positive temperature anomalies caused by normal pipeline operating condition fluctuations. Temperature anomaly regions that meet the preset area threshold and duration are identified. The geometric center, area and average anomaly amplitude of the temperature anomaly regions are extracted to obtain the temperature anomaly region information. If there is no temperature anomaly region, the anomaly confidence of the temperature feature is set to 0. The temperature feature vector and the temperature anomaly region information are the extracted temperature features.

7. A pipeline leakage risk early warning system based on multi-source data fusion according to claim 3, characterized in that: The method for obtaining the overall leakage probability under vibration-dominant mode is as follows: A single-evidence-source decision strategy is adopted. For the vibration features included in the dominant feature group, the ratio R of the frequency band energy of the current short-time intra-frame leakage feature frequency band included in the vibration features to the preset historical background energy is calculated. The ratio R is converted into a probability value in the interval [0,1] through a nonlinear mapping function. The obtained probability value is corrected by calculating the data confidence level in combination with the data quality label, and the vibration leakage probability value is obtained as the comprehensive leakage probability under the vibration-dominant mode.

8. A pipeline leakage risk early warning system based on multi-source data fusion according to claim 7, characterized in that: The method for obtaining the overall leakage probability under the vibration-electrical resistivity-based co-dominant mode is as follows: The physical model fit of resistivity data is obtained, and the physical model fit is linearly mapped to the fit modulation factor. A weighted fusion strategy is adopted, and the initial weight of vibration features is dynamically adjusted with soil saturation. The data confidence of vibration data and resistivity data is obtained or calculated through data quality labels. The product of the fit modulation factor, data confidence and initial weight is used as the actual weight. The vibration leakage probability value is calculated in the same way as the vibration-dominant mode. If the anomaly confidence of the resistivity feature is not 0, the resistivity leakage probability value is calculated based on the resistivity anomaly area information. The vibration leakage probability value and the resistivity leakage probability value are weighted and fused using actual weights to obtain the comprehensive leakage probability.

9. A pipeline leakage risk early warning system based on multi-source data fusion according to claim 8, characterized in that: The method for obtaining the physical model fit of resistivity data is as follows: A saturated soil seepage-electrical method coupled model was constructed. Using multi-source characteristic parameters as input, the abnormal resistivity characteristic vector was predicted by finite element numerical solution. The fit was calculated by comparing the resistivity characteristic vector, and the physical model fit of the resistivity data was obtained.

10. A pipeline leakage risk early warning system based on multi-source data fusion according to claim 8, characterized in that: The method for obtaining the overall leakage probability under the electrochemical temperature-dependent synergistic mode is as follows: A heat conduction model is constructed, using multi-source characteristic parameters as input. The abnormal temperature characteristic vector is predicted by finite element numerical solution. The fit is calculated by comparing the temperature characteristic vectors to obtain the physical model fit of the temperature data. The physical model fit of the resistivity data is calculated and the physical model fit is linearly mapped to the corresponding fit modulation factor. An evidence theory fusion framework is initiated, and an identification framework is constructed that includes two mutually exclusive propositions: the existence of leakage and the non-existence of leakage. Resistivity and temperature features are used as evidence sources. The anomaly confidence levels of resistivity and temperature features are obtained or calculated based on the evidence sources and multiplied using a matching modulation factor to obtain the basic probability allocation supporting the existence of leakage. The corresponding quality confidence level is calculated based on the corresponding data quality label. The Dempster-Shafer evidence synthesis rule is used to perform orthogonal summation on the basic probability allocations of each evidence source to calculate the comprehensive leakage probability.

Images