A natural gas pipeline leak detection method based on edge computing

Abstract

The present invention discloses a natural gas pipeline leakage detection method based on edge computing, which relates to the technical field of natural gas pipeline leakage detection. The present invention significantly improves the stability of edge-side leakage detection through the dual innovations of dynamic multi-scale perception and spatiotemporal collaborative verification: the dynamic window mechanism breaks through the fixed analysis window limitation, and through the dual triggering of energy accumulation rate and pulse density, it realizes the adaptive extraction of features of leakage events of different intensities, and reduces the missed reporting rate of small leaks; spatiotemporal fingerprint verification combines the pipeline stress wave propagation characteristics with sensor type matching to construct a physical conduction chain credibility assessment system, effectively suppressing the spread of false alarms caused by regional interference; the dynamic correction of environmental entropy and the sensitivity strategy pool form a closed-loop optimization, so that the system can maintain stable detection performance under complex working conditions, such as construction vibration and sudden weather changes.

Description

Technical Field

[0001] The present invention relates to the technical field of natural gas pipeline leakage detection, and in particular to a natural gas pipeline leakage detection method based on edge computing. Background Art

[0002] Currently, natural gas pipeline leakage monitoring systems are accelerating their evolution towards edge intelligence, relying on computing nodes deployed along the pipeline to achieve localized real-time detection; these edge devices integrate multi-dimensional sensors such as pressure, gas concentration, and vibration, and capture leakage characteristics by continuously collecting physical signals; compared with traditional centralized monitoring models, edge computing significantly reduces data transmission latency and demonstrates significant advantages in long-distance pipelines and harsh outdoor environments; mainstream industry solutions generally use lightweight AI models for signal feature extraction, and combine them with threshold trigger mechanisms to achieve rapid response, forming a closed-loop link of "perception, judgment, and execution."

[0003] Existing solutions face two core bottlenecks: First, the lack of scale adaptability in signal recognition; leakage events include slow changes, such as concentration drift caused by small leaks, and sudden changes, such as vibration impact caused by pipe bursts. Fixed analysis windows cannot take both into account - too large a window smooths high-frequency details and leads to missed reports, while too small a window amplifies low-frequency noise and leads to false alarms; second, there is a lack of trustworthy coordination between edge nodes; when making independent decisions, each node is susceptible to regional interference, such as construction vibration and sudden weather changes, which can mislead and trigger chain reactions; although some studies have attempted to introduce a weighted voting mechanism, they have not considered the spatiotemporal propagation characteristics of environmental interference, resulting in the spread of erroneous signals within the cluster.

[0004] Recent technological improvements have focused on multi-scale signal processing and cluster linkage: some solutions use a dual-window convolution algorithm to capture high-frequency vibrations and low-frequency concentration changes respectively, and then improve the recognition rate through feature fusion; other studies have constructed a lightweight neural network to output a local event credibility score (LECS) and correct the judgment based on the votes of adjacent nodes; however, the former relies on a preset window size and has difficulty in dealing with sudden fluctuations in signal strength; the latter uses static weight distribution and cannot dynamically filter out false signals generated by continuous interference such as construction. Summary of the Invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] The present invention provides a natural gas pipeline leakage detection method based on edge computing to solve the problem that the existing edge detection method cannot adapt to the multi-scale characteristics of the leakage signal due to the fixed analysis window, and the independent decision-making of the node is susceptible to regional interference, resulting in frequent alternation of false alarms and missed alarms.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] An embodiment of the present invention provides a natural gas pipeline leakage detection method based on edge computing, which includes:

[0009] Step S1: synchronously collect pipeline pressure, gas concentration and vibration signals at each edge node, and perform multi-scale feature extraction on the signals through a dynamic window generation module;

[0010] Step S2: Input the extracted features into a lightweight neural network model to generate a local event credibility score, which includes a dual evaluation result of the signal abnormality strength and the probability of environmental interference;

[0011] Step S3: Adjacent edge nodes periodically exchange credibility data packets containing spatiotemporal fingerprints to verify the physical consistency of the event propagation path based on the pipeline stress wave transmission characteristics;

[0012] Step S4: Dynamically adjust the local detection sensitivity strategy based on the neighborhood node credibility data and the environmental entropy value to trigger a leak warning or enter a delayed confirmation mode.

[0013] As a preferred solution of the natural gas pipeline leakage detection method based on edge computing described in the present invention, the dynamic window generation module performs the following operations:

[0014] The window scale switching is triggered according to the real-time signal energy accumulation rate. When the low-frequency signal energy accumulation reaches the first threshold, the large-scale window is enabled. When the high-frequency pulse signal density exceeds the second threshold, the microsecond sampling window is enabled.

[0015] In mixed event scenarios, the dual-scale windows are run synchronously and the fused feature vector is output;

[0016] The triggering condition of the first threshold is: the low-frequency signal energy change rate is continuously lower than the set slope within N consecutive sampling periods, and the baseline drift exceeds the threshold of the pipeline pressure rating.

[0017] As a preferred solution of the natural gas pipeline leakage detection method based on edge computing described in the present invention, the multi-scale feature extraction includes:

[0018] Performing baseline drift trend fitting on the low-frequency signal to generate a first feature group including drift slope and periodic fluctuation amplitude;

[0019] Extracting the peak interval, duration and energy decay rate of the pulse cluster from the high-frequency signal to generate a second feature group;

[0020] The first feature group and the second feature group are aligned in time and then concatenated into a joint feature vector.

[0021] As a preferred solution of the edge computing-based natural gas pipeline leakage detection method described in the present invention, in step S1, during the multi-scale feature extraction of the signal, dynamic window triggering and parallel output are performed, and the dynamic window generation module calculates the energy and pulse density of the low-frequency and high-frequency signals in two windows in real time, and switches or runs in parallel according to the trigger conditions:

[0022] Perform low-frequency energy rate calculations:

[0023] ,

[0024] When continuous cycles satisfy

[0025] ,

[0026] Enable large-scale window ,

[0027] Perform high-frequency pulse density calculations:

[0028] ,

[0029] when Enable microscale window ,

[0030] in, Indicates the The low-frequency signal of the sampling points, represents the accumulated energy in the large-scale window, represents the energy accumulation rate, represents the sampling period, represents the large-scale window length, Indicates the number of sampling cycles for continuous judgment, represents the rate threshold, Indicates the The pressure baseline drift of the cycle, Indicates the pressure offset threshold, represents the number of pulse clusters detected within the microscale window, represents the microscale window length, represents the pulse density threshold;

[0031] If both of the above conditions are met, the dual-scale windows run in parallel and output two original signal segments for subsequent feature extraction.

[0032] As a preferred solution of the natural gas pipeline leakage detection method based on edge computing described in the present invention, in step S1, during the multi-scale feature extraction of the signal, the low-frequency signal segment in the large-scale window is extracted. Perform least squares straight line fitting to obtain the drift slope and fluctuation amplitude:

[0033] ,

[0034] ,

[0035] in,

[0036] ,

[0037] ,

[0038] ,

[0039] represents the baseline drift slope, Indicates the amplitude of periodic fluctuation;

[0040] For the high-frequency signal segment within the microscale window Detect pulse clusters, record the The peak time of the cluster is , the energy within the cluster is:

[0041] ,

[0042] Cluster duration , peak interval , perform an exponential fit to the energy decay:

[0043] ,

[0044] in, represents high-frequency signal samples, Respectively Cluster start and end sample indices, Indicates the Cluster energy, Indicates the Cluster duration, represents the interval between adjacent cluster peaks, represents the number of clusters taken, represents the energy decay rate;

[0045] The low-frequency features With high frequency features

[0046] ,

[0047] Align and concatenate in chronological order to obtain the joint feature vector:

[0048] .

[0049] As a preferred solution of the natural gas pipeline leakage detection method based on edge computing described in the present invention, the generation of the local event credibility score includes:

[0050] Output the original abnormal probability value through the neural network;

[0051] The final credibility score is generated by superimposing the historical false alarm rate correction coefficient of the current node and the real-time intensity value of the environmental interference source;

[0052] The environmental interference sources include wind speed sensor readings and ground vibration sensor readings;

[0053] The historical false alarm rate correction coefficient is obtained by sliding window calculation: the false alarm ratio that has been manually confirmed in the warning events triggered by the current node in the past period of time is counted, and the correction coefficient is generated according to 1-false alarm ratio.

[0054] As a preferred solution of the natural gas pipeline leakage detection method based on edge computing described in the present invention, wherein: the spatiotemporal fingerprint includes the first peak timestamp of the vibration signal, the sensor type identifier and the signal propagation direction code;

[0055] The spatiotemporal fingerprint verification includes:

[0056] When a node detects an abnormal vibration signal, it records the first peak arrival timestamp;

[0057] Receive the wave peak timestamp data of adjacent nodes, calculate the time difference and perform matching verification with the stress wave propagation velocity corresponding to the pipeline material;

[0058] If the time difference matching degree exceeds the set ratio, it is marked as a valid physical conduction event.

[0059] As a preferred solution of the natural gas pipeline leakage detection method based on edge computing described in the present invention, in step S3, during the matching degree verification, the adjacent nodes are received. and After obtaining the spatiotemporal fingerprint, combined with the pipeline stress wave propagation characteristics, the matching degree is calculated and verified according to the following process:

[0060] Calculate the time difference and theoretical conduction time:

[0061] ,

[0062] in, Representation node The first peak arrival timestamp, represents the distance between nodes measured along the centerline of the pipe. Indicates the stress wave propagation velocity corresponding to the pipe material,

[0063] Calculate time matching :

[0064] ,

[0065] in, Indicates the time error, Indicates the maximum allowable time deviation;

[0066] Determine the type Matching degree with direction :

[0067] like ,but ,otherwise ;

[0068] ,

[0069] in, are the sensor type identifiers of the two nodes respectively, are the encoding angles for the signal propagation direction, respectively;

[0070] The overall matching degree and verification criteria are defined as:

[0071] ,

[0072] Calibration standards:

[0073] ,

[0074] in, is the weight coefficient, is the high matching threshold, is the low matching threshold.

[0075] As a preferred solution of the natural gas pipeline leakage detection method based on edge computing described in the present invention, the calculation of the environmental entropy value includes:

[0076] Collect wind speed, temperature gradient and mechanical vibration signals around the current node;

[0077] The standard deviation ratio of each interference source is calculated using a sliding window to generate a 0-1 normalized environmental disorder index;

[0078] The environmental disorder index is weighted averaged with similar data of neighboring nodes to obtain the final environmental entropy value.

[0079] As a preferred solution of the natural gas pipeline leakage detection method based on edge computing described in the present invention, the dynamic adjustment of the sensitivity strategy includes:

[0080] Pre-set multiple sensitivity strategy templates, including night static mode, strong wind interference mode, and construction protection mode;

[0081] Automatically select or mix the current optimal strategy based on the detection accuracy and environmental entropy change trends in the last 24 hours;

[0082] When the credibility scores of more than half of the neighboring nodes are continuously lower than the warning value, the system is forced to switch to high-sensitivity inspection mode;

[0083] The detection accuracy is calculated by performing normal distribution modeling on the time difference between the first reporting time of the node and the actual occurrence time of the event in the confirmed leakage event, and taking the distribution function value as the accuracy index.

[0084] The beneficial effects of the present invention are as follows: the present invention significantly improves the stability of edge-side leakage detection through the dual innovations of dynamic multi-scale perception and spatiotemporal collaborative verification: the dynamic window mechanism breaks through the fixed analysis window limitation, and through the dual triggering of energy accumulation rate and pulse density, it realizes the adaptive extraction of features of leakage events of different intensities, and reduces the missed reporting rate of small leaks; spatiotemporal fingerprint verification combines the propagation characteristics of pipeline stress waves with sensor type matching to construct a physical conduction chain credibility assessment system, which effectively suppresses the spread of false alarms caused by regional interference; the dynamic correction of environmental entropy and the sensitivity strategy pool form a closed-loop optimization, so that the system can maintain stable detection performance under complex working conditions, such as construction vibration and sudden weather changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0086] Figure 1 This is a flow chart of the natural gas pipeline leakage detection method based on edge computing in Example 1. DETAILED DESCRIPTION

[0087] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0088] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0089] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0090] Example 1, with reference to Figure 1 This embodiment provides a natural gas pipeline leakage detection method based on edge computing, including the following steps:

[0091] Step S1: synchronously collect pipeline pressure, gas concentration and vibration signals at each edge node, and perform multi-scale feature extraction on the signals through a dynamic window generation module;

[0092] The dynamic window generation module performs the following operations:

[0093] The window scale switching is triggered according to the real-time signal energy accumulation rate. When the low-frequency signal energy accumulation reaches the first threshold, the large-scale window is enabled. When the high-frequency pulse signal density exceeds the second threshold, the microsecond sampling window is enabled.

[0094] In mixed event scenarios, the dual-scale windows are run synchronously and the fused feature vector is output;

[0095] The triggering condition for the first threshold is: the low-frequency signal energy change rate continues to be lower than the set slope for N consecutive sampling periods, and the baseline drift exceeds the threshold of the pipeline pressure rating;

[0096] Multi-scale feature extraction includes:

[0097] Performing baseline drift trend fitting on the low-frequency signal to generate a first feature group including drift slope and periodic fluctuation amplitude;

[0098] Extracting the peak interval, duration and energy decay rate of the pulse cluster from the high-frequency signal to generate a second feature group;

[0099] Aligning the first feature group and the second feature group in time and concatenating them into a joint feature vector;

[0100] In step S1, during the multi-scale feature extraction of the signal, dynamic window triggering and parallel output are performed. The dynamic window generation module calculates the energy and pulse density of the low-frequency and high-frequency signals under the two windows in real time, and switches or runs in parallel according to the trigger conditions:

[0101] Perform low-frequency energy rate calculations:

[0102] ,

[0103] When continuous cycles satisfy

[0104] ,

[0105] Enable large-scale window ,

[0106] Perform high-frequency pulse density calculations:

[0107] ,

[0108] when Enable microscale window ,

[0109] in, Indicates the The low-frequency signal of the sampling points, represents the accumulated energy in the large-scale window, represents the energy accumulation rate, represents the sampling period, represents the large-scale window length, Indicates the number of sampling cycles for continuous judgment, represents the rate threshold, Indicates the The pressure baseline drift of the cycle, Indicates the pressure offset threshold, represents the number of pulse clusters detected within the microscale window, represents the microscale window length, represents the pulse density threshold;

[0110] If both of the above conditions are met, the dual-scale windows run in parallel and output two original signal segments for subsequent feature extraction;

[0111] In step S1, during the multi-scale feature extraction of the signal, the low-frequency signal segment in the large-scale window is Perform least squares straight line fitting to obtain the drift slope and fluctuation amplitude:

[0112] ,

[0113] ,

[0114] in,

[0115] ,

[0116] ,

[0117] ,

[0118] represents the baseline drift slope, Indicates the amplitude of periodic fluctuation;

[0119] For the high-frequency signal segment within the microscale window Detect pulse clusters, record the The peak time of the cluster is , the energy within the cluster is:

[0120] ,

[0121] Cluster duration , peak interval , perform an exponential fit to the energy decay:

[0122] ,

[0123] in, represents high-frequency signal samples, Respectively Cluster start and end sample indices, Indicates the Cluster energy, Indicates the Cluster duration, represents the interval between adjacent cluster peaks, represents the number of clusters taken, represents the energy decay rate;

[0124] The low-frequency features With high frequency features

[0125] ,

[0126] Align and concatenate in chronological order to obtain the joint feature vector:

[0127] ;

[0128] Specifically, through adaptive window switching and parallel operation, multi-scale and complex scenario perception of signals in natural gas pipelines is achieved. The large-scale window uses least squares fitting to extract the stable trend slope and periodic fluctuation amplitude for low-frequency drift, which can effectively filter out short-term noise. The micro-scale window calculates the peak interval, duration and exponential decay rate for high-frequency pulses to capture transient leakage impact characteristics. The fusion of the two features not only retains long-term trend information but also takes into account the details of transient events, improving the downstream classifier's ability to distinguish leakage events. The innovation lies in the continuous multi-cycle rate judgment trigger and the quadratic fitting of the exponential decay rate, which not only ensures the stability of the characteristics but also enhances the recognition diversity of different leakage intensities and patterns.

[0129] Step S2: Input the extracted features into a lightweight neural network model to generate a local event credibility score. The credibility score includes dual evaluation results of signal anomaly strength and environmental interference probability.

[0130] The generation of local event credibility scores includes:

[0131] Output the original abnormal probability value through the neural network;

[0132] The final credibility score is generated by superimposing the historical false alarm rate correction coefficient of the current node and the real-time intensity value of the environmental interference source;

[0133] Environmental interference sources include wind speed sensor readings and ground vibration sensor readings;

[0134] The historical false alarm rate correction coefficient is obtained through sliding window calculation: the false alarm ratio that has been manually confirmed in the warning events triggered by the current node in the past period of time is counted, and the correction coefficient is generated by dividing the false alarm ratio by 1.

[0135] Step S3: Adjacent edge nodes periodically exchange credibility data packets containing spatiotemporal fingerprints to verify the physical consistency of the event propagation path based on the pipeline stress wave transmission characteristics;

[0136] The spatiotemporal fingerprint contains the timestamp of the first peak of the vibration signal, the sensor type identifier, and the signal propagation direction code;

[0137] Spatiotemporal fingerprint verification includes:

[0138] When a node detects an abnormal vibration signal, it records the first peak arrival timestamp;

[0139] Receive the wave peak timestamp data of adjacent nodes, calculate the time difference and perform matching verification with the stress wave propagation velocity corresponding to the pipeline material;

[0140] If the time difference matching degree exceeds the set ratio, it is marked as a valid physical conduction event.

[0141] In step S3, during the matching degree check, the adjacent nodes are received and After obtaining the spatiotemporal fingerprint, combined with the pipeline stress wave propagation characteristics, the matching degree is calculated and verified according to the following process:

[0142] Calculate the time difference and theoretical conduction time:

[0143] ,

[0144] in, Representation node The first peak arrival timestamp, represents the distance between nodes measured along the centerline of the pipe. Indicates the stress wave propagation velocity corresponding to the pipe material,

[0145] Calculate time matching :

[0146] ,

[0147] in, Indicates the time error, Indicates the maximum allowable time deviation;

[0148] Determine the type Matching degree with direction :

[0149] like ,but ,otherwise ;

[0150] ,

[0151] in, are the sensor type identifiers of the two nodes respectively, are the encoding angles for the signal propagation direction, respectively;

[0152] The overall matching degree and verification criteria are defined as:

[0153] ,

[0154] Calibration standards:

[0155] ,

[0156] in, is the weight coefficient, is the high matching threshold, is the low matching threshold;

[0157] Specifically, through multi-dimensional matching design, strict verification of the physical propagation path of the natural gas leak event was achieved, and the time matching It can accurately reflect the deviation between the arrival time difference of the first wave peak and the theoretical propagation time, and the type matching degree Ensure that only sensor data of the same type is involved in the verification, and the direction matching The consistency of propagation direction is evaluated from the coding perspective, and the overall matching degree is calculated in a weighted manner. High and low thresholds are set to implement a near-real-time physical consistency screening and delayed confirmation mechanism. This method strikes a balance between rigor and flexibility, achieving a balance between the risk of false positives and the risk of missed negatives.

[0158] Step S4: Dynamically adjust the local detection sensitivity strategy based on the neighborhood node credibility data and the environmental entropy value to trigger a leak warning or enter a delayed confirmation mode;

[0159] The calculation of environmental entropy includes:

[0160] Collect wind speed, temperature gradient and mechanical vibration signals around the current node;

[0161] The standard deviation ratio of each interference source is calculated using a sliding window to generate a 0-1 normalized environmental disorder index;

[0162] The environmental disorder index is weighted averaged with similar data of neighboring nodes to obtain the final environmental entropy value;

[0163] Dynamic adjustment of sensitivity strategy includes:

[0164] Pre-set multiple sensitivity strategy templates, including night static mode, strong wind interference mode, and construction protection mode;

[0165] Automatically select or mix the current optimal strategy based on the detection accuracy and environmental entropy change trends in the last 24 hours;

[0166] When the credibility scores of more than half of the neighboring nodes are continuously lower than the warning value, the system is forced to switch to high-sensitivity inspection mode;

[0167] The detection accuracy is calculated as follows: a normal distribution model is built for the time difference between the first reporting time of the node and the actual occurrence time of the incident in a confirmed leakage incident, and the distribution function value is taken as the accuracy indicator.

[0168] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A natural gas pipeline leakage detection method based on edge computing, characterized in that: include, Step S1: synchronously collect pipeline pressure, gas concentration and vibration signals at each edge node, and perform multi-scale feature extraction on the signals through a dynamic window generation module; Step S2: Input the extracted features into a lightweight neural network model to generate a local event credibility score, which includes a dual evaluation result of the signal abnormality strength and the probability of environmental interference; Step S3: Adjacent edge nodes periodically exchange credibility data packets containing spatiotemporal fingerprints to verify the physical consistency of the event propagation path based on the pipeline stress wave transmission characteristics; Step S4: Dynamically adjust the local detection sensitivity strategy based on the neighborhood node credibility data and the environmental entropy value to trigger a leak warning or enter a delayed confirmation mode; The spatiotemporal fingerprint includes the timestamp of the first peak of the vibration signal, the sensor type identifier and the signal propagation direction code; The spatiotemporal fingerprint verification includes: When a node detects an abnormal vibration signal, it records the first peak arrival timestamp; Receive the wave peak timestamp data of adjacent nodes, calculate the time difference and perform matching verification with the stress wave propagation velocity corresponding to the pipeline material; If the time difference matching degree exceeds the set ratio, it is marked as a valid physical conduction event; In step S3, during the matching degree check, the adjacent nodes are received and After obtaining the spatiotemporal fingerprint, combined with the pipeline stress wave propagation characteristics, the matching degree is calculated and verified according to the following process: Calculate the time difference and theoretical conduction time: ， in, Representation node The first peak arrival timestamp, represents the distance between nodes measured along the centerline of the pipe. Indicates the stress wave propagation velocity corresponding to the pipe material, Calculate time matching : ， in, Indicates the time error, Indicates the maximum allowable time deviation; Determine the type Matching degree with direction : like ,but ,otherwise ; ， in, are the sensor type identifiers of the two nodes respectively, are the encoding angles for the signal propagation direction, respectively; The overall matching degree and verification criteria are defined as: ， Calibration standards: ， in, is the weight coefficient, is the high matching threshold, is the low matching threshold.

2. A natural gas pipeline leakage detection method based on edge computing according to claim 1, characterized in that: The dynamic window generation module performs the following operations: The window scale switching is triggered according to the real-time signal energy accumulation rate. When the low-frequency signal energy accumulation reaches the first threshold, the large-scale window is enabled. When the high-frequency pulse signal density exceeds the second threshold, the microsecond sampling window is enabled. In mixed event scenarios, the dual-scale windows are run synchronously and the fused feature vector is output; The triggering condition of the first threshold is: the low-frequency signal energy change rate is continuously lower than the set slope within N consecutive sampling periods, and the baseline drift exceeds the threshold of the pipeline pressure rating.

3. A natural gas pipeline leakage detection method based on edge computing according to claim 2, characterized in that: The multi-scale feature extraction includes: Performing baseline drift trend fitting on the low-frequency signal to generate a first feature group including drift slope and periodic fluctuation amplitude; Extracting the peak interval, duration and energy decay rate of the pulse cluster from the high-frequency signal to generate a second feature group; The first feature group and the second feature group are aligned in time and then concatenated into a joint feature vector.

4. A natural gas pipeline leakage detection method based on edge computing according to claim 3, characterized in that: In step S1, during the multi-scale feature extraction of the signal, dynamic window triggering and parallel output are performed. The dynamic window generation module calculates the energy and pulse density of the low-frequency and high-frequency signals under the two windows in real time, and switches or runs in parallel according to the trigger conditions: Perform low-frequency energy rate calculations: ， When continuous cycles satisfy ， Enable large-scale window , Perform high-frequency pulse density calculations: ， when Enable microscale window , in, Indicates the The low-frequency signal of the sampling points, represents the accumulated energy in the large-scale window, represents the energy accumulation rate, represents the sampling period, represents the large-scale window length, Indicates the number of sampling cycles for continuous judgment, represents the rate threshold, Indicates the The pressure baseline drift of the cycle, Indicates the pressure offset threshold, represents the number of pulse clusters detected within the microscale window, represents the microscale window length, represents the pulse density threshold; If both of the above conditions are met, the dual-scale windows run in parallel and output two original signal segments for subsequent feature extraction.

5. A natural gas pipeline leakage detection method based on edge computing according to claim 4, characterized in that: In step S1, during the multi-scale feature extraction of the signal, the low-frequency signal segment in the large-scale window is extracted. Perform least squares straight line fitting to obtain the drift slope and fluctuation amplitude: ， ， in, ， ， ， represents the baseline drift slope, Indicates the amplitude of periodic fluctuation; For the high-frequency signal segment within the microscale window Detect pulse clusters, record the The peak time of the cluster is , the energy within the cluster is: ， Cluster duration , peak interval , perform an exponential fit to the energy decay: ， in, represents high-frequency signal samples, Respectively Cluster start and end sample indices, Indicates the Cluster energy, Indicates the Cluster duration, represents the interval between adjacent cluster peaks, represents the number of clusters taken, represents the energy decay rate; The low-frequency features With high frequency features ， Align and concatenate in chronological order to obtain the joint feature vector: 。 6. A natural gas pipeline leakage detection method based on edge computing according to claim 1, characterized in that: The generation of the local event credibility score includes: Output the original abnormal probability value through the neural network; The final credibility score is generated by superimposing the historical false alarm rate correction coefficient of the current node and the real-time intensity value of the environmental interference source; The environmental interference sources include wind speed sensor readings and ground vibration sensor readings; The historical false alarm rate correction coefficient is obtained by sliding window calculation: the false alarm ratio that has been manually confirmed in the warning events triggered by the current node in the past period of time is counted, and the correction coefficient is generated according to 1-false alarm ratio.

7. A natural gas pipeline leakage detection method based on edge computing according to claim 1, characterized in that: The calculation of the environmental entropy value includes: Collect wind speed, temperature gradient and mechanical vibration signals around the current node; The standard deviation ratio of each interference source is calculated using a sliding window to generate a 0-1 normalized environmental disorder index; The environmental disorder index is weighted averaged with similar data of neighboring nodes to obtain the final environmental entropy value.

8. The natural gas pipeline leakage detection method based on edge computing according to claim 1, characterized in that: Dynamic adjustment of sensitivity strategy includes: Pre-set multiple sensitivity strategy templates, including night static mode, strong wind interference mode, and construction protection mode; Automatically select or mix the current optimal strategy based on the detection accuracy and environmental entropy change trends in the last 24 hours; When the credibility scores of more than half of the neighboring nodes are continuously lower than the warning value, the system is forced to switch to high-sensitivity inspection mode; The detection accuracy is calculated by performing normal distribution modeling on the time difference between the first reporting time of the node and the actual occurrence time of the event in the confirmed leakage event, and taking the distribution function value as the accuracy index.

Images