Urban gas pipeline micro-leakage early warning method and system

Abstract

The invention provides an urban gas pipeline micro-leakage early warning method and system, and relates to the technical field of data processing, and the method comprises the steps: employing a fuzzy analytic hierarchy process to calculate a dynamic risk index based on an evaluation matrix, and triggering the three-stage linkage alarm of an audible and visual alarm, a mobile terminal and an emergency platform when the index exceeds a threshold value; for three-level linkage alarm, a mobile robot is dispatched, a target area is navigated based on thermodynamic diagram positioning and reference coordinates, gas concentration is remeasured through a methane laser detector, a pipeline surface temperature field is scanned in combination with infrared thermal imaging, and a verification result and correction data are bound through a block chain evidence storage technology. And a tamper-resistant full-link evidence chain is formed. According to the invention, through cooperation of early warning and calculation, full-process automation of gas pipeline monitoring, early warning and emergency verification is realized, and the intelligent level of safety monitoring is improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method and system for early warning of micro-leakage in urban gas pipelines. Background Art

[0002] With the acceleration of urbanization, the scale of urban gas pipeline networks continues to expand. However, due to long-term underground service, gas pipelines are affected by factors such as soil corrosion, geological subsidence, and external force damage, making them prone to micro-leaks. If they are not discovered and handled in time, micro-leaks may gradually expand and cause serious safety accidents such as fires, explosions, and poisoning.

[0003] At present, urban gas pipeline leakage detection technologies mainly include manual inspections, pressure monitoring, and fiber optic sensing. Manual inspections rely on manual experience, are inefficient, and have difficulty detecting early micro-leaks. Pressure monitoring is easily affected by environmental factors and has a high false alarm rate. Although fiber optic sensing technology has high sensitivity, it is expensive and difficult to deploy on a large scale. In addition, existing technologies mostly use a single parameter or a single detection method, which cannot fully and accurately reflect the actual situation of gas leakage, and the assessment accuracy of leakage location, diffusion trend, and risk level is insufficient. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and system for early warning of micro-leakage in urban gas pipelines. Through the coordination of early warning and calculation, the whole process of gas pipeline monitoring, early warning and emergency verification can be automated, thereby improving the intelligent level of safety monitoring.

[0005] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0006] In a first aspect, a method for early warning of micro-leakage in a city gas pipeline is provided, the method comprising:

[0007] Step S1: Deploy a multi-parameter array of integrated pressure fluctuation, trace gas spectrum, acoustic wave or vibration, and pipeline deformation sensors along the gas pipeline, with node spacing less than or equal to 30 meters. Wide-area coverage is achieved through a LoRa or NB-IoT hybrid network. Select two fixed target points with a spacing greater than or equal to 50 meters and located on the same straight line segment, and calibrate their three-dimensional coordinates through precision positioning to form a reference line.

[0008] Step S2: Based on the reference line, the pressure, gas concentration, and vibration data between the two points are collected in real time, and the environmental noise baseline, signal attenuation coefficient, and spatial consistency parameter are calculated to generate a dynamic correction value. The original signal is adaptively filtered to eliminate environmental noise through wavelet packet decomposition. The dynamic baseline is calibrated in combination with Kalman filtering. The correction value is used to perform spatial compensation for the pressure gradient, gas concentration gradient, and vibration spectrum characteristics, and the corrected spatiotemporal feature vector is output.

[0009] Step S3: The corrected spatiotemporal feature vector is input into the deep spatiotemporal convolutional network. The temporal evolution law is extracted through the bidirectional LSTM, and the spatial correlation features are mined by combining the 3D convolution kernel. The GIS pipeline topology and the reference line parameters are simultaneously integrated to generate a 3D heat map of the leakage probability and diffusion trend.

[0010] Step S4: Based on the evaluation matrix, the dynamic risk index is calculated using the fuzzy analytic hierarchy process. When the index exceeds the threshold, a three-level linkage alarm is triggered, including the sound and light alarm, mobile terminal, and emergency platform.

[0011] Step S5: In response to the three-level linkage alarm, the mobile robot is dispatched to navigate to the target area based on the thermal map positioning and reference coordinates. The gas concentration is remeasured using a methane laser detector, and the pipeline surface temperature field is scanned using infrared thermal imaging. The verification results and correction data are then bound using blockchain evidence storage technology to form an unalterable full-link evidence chain.

[0012] Furthermore, step S1: a multi-parameter array integrating pressure fluctuation, trace gas spectrum, sound wave or vibration, and pipeline deformation sensors is distributed along the gas pipeline, with the node spacing being less than or equal to 30 meters, and wide-area coverage is achieved through a LoRa or NB-IoT hybrid network; two fixed target points with a spacing of greater than or equal to 50 meters and located on the same straight line are selected, and the three-dimensional coordinates are calibrated through precision positioning to form a reference line, including:

[0013] Sensor nodes are deployed along the gas pipeline at intervals of 30 meters or less. Each node integrates a pressure fluctuation sensor, a trace gas spectrum sensor, an acoustic or vibration sensor, and a pipeline deformation monitoring unit. LoRa wireless and NB-IoT cellular communication functions are pre-installed in each sensor node to achieve low-power wide-area communication coverage.

[0014] Based on low-power wide-area communication coverage, a dual-channel communication link is established through LoRa modulation and NB-IoT cellular network. LoRa is used for short-distance networking between nodes, transmitting pressure fluctuations, gas concentration and vibration data in real time. NB-IoT is used for wide-area data backhaul, prioritizing the transmission of high-frequency deformation data from pipeline deformation monitoring units and synchronously uploading the location coordinates of each sensor node. This dual-channel link ensures that the upload delay of all sensor data is less than or equal to 5 seconds, forming a dynamically updated pipeline network spatial topology map.

[0015] Based on the spatial topology map of the pipe network, two target points with a distance of 50 meters or more and located on the same straight line segment are selected from the deployed sensor nodes. The three-dimensional coordinates of the target points are calibrated using a GNSS positioning device, and the coordinate data is transmitted back to the control center in real time via the NB-IoT channel. Based on the returned coordinate data, the actual distance between the two points is measured using a laser rangefinder, and the deviation from the theoretical straight-line distance is calculated. If the deviation is less than or equal to 0.1%, the line connecting the two points is confirmed as a valid reference line. If the deviation is greater than 0.1%, the target points are reselected from the topology map and the process is iterated until the deviation is less than or equal to 0.1% and a reference line is formed.

[0016] Furthermore, step S2: based on the reference straight line, the pressure, gas concentration and vibration data between the two points are collected in real time, the environmental noise baseline, the signal attenuation coefficient and the spatial consistency parameter are calculated, and a dynamic correction value is generated; the original signal is adaptively filtered, the environmental noise is eliminated by wavelet packet decomposition, the dynamic baseline is calibrated in combination with Kalman filtering, and the pressure gradient, gas concentration gradient and vibration spectrum characteristics are spatially compensated using the correction value, and the corrected spatiotemporal feature vector is output, including:

[0017] Based on the pressure fluctuation data between two points on the reference line, the sliding window average method is used to calculate the ambient noise baseline; combined with the gas concentration gradient, the signal attenuation coefficient is calculated based on the Fresnel reflection model; and the spatial consistency parameter is determined through covariance analysis.

[0018] Based on the calculated environmental noise baseline, the original pressure signal is decomposed by wavelet packets to extract the effective signal in the 1-100 Hz frequency band; based on the signal attenuation coefficient, the dynamic baseline of the gas concentration data is calibrated using Kalman filtering;

[0019] Based on the spatial consistency parameters, the dynamic correction values ​​are distributed to the pressure gradient, gas concentration gradient and vibration spectrum characteristics according to the weights; the weighted eigenvectors are normalized and the corrected spatiotemporal eigenvectors are output.

[0020] Furthermore, step S3: the corrected spatiotemporal feature vector is input into the deep spatiotemporal convolutional network, the temporal evolution law is extracted through the bidirectional LSTM, the spatial correlation features are mined by combining the 3D convolution kernel, and the GIS pipeline topology and the reference line parameters are synchronously integrated to generate a 3D heat map of the leakage probability and diffusion trend, including:

[0021] A deep spatiotemporal convolutional network model was constructed, consisting of three bidirectional LSTM layers and two three-dimensional convolutional layers. The model was supervised and trained using a leaky sample dataset. The loss function was a weighted combination of cross entropy and mean squared error. The network weights were optimized using a backpropagation algorithm to obtain the optimized DSTCN model weights.

[0022] Based on the optimized DSTCN model weights, the modified spatiotemporal feature vectors are sliced ​​into time windows and input into the model, outputting a leakage probability sequence in the time dimension. The elbow and valve coordinates and the reference line spatial parameters in the GIS pipeline topology are simultaneously integrated, and a three-dimensional heat map is generated through inverse distance weighted interpolation, where the leakage probability values ​​are mapped to the range of 0-1.

[0023] Based on the three-dimensional heat map, areas with a leakage probability greater than or equal to 0.7 are extracted as core leakage points. Combined with the gas diffusion model, the core leakage points are used as initial conditions to predict the diffusion path in the next 30 minutes.

[0024] Furthermore, step S4: based on the evaluation matrix, a dynamic risk index is calculated using the fuzzy analytic hierarchy process. When the index exceeds a threshold, a three-level linkage alarm is triggered, including:

[0025] The original data of risk factors are mapped to the range of 0-1 using the range normalization method to obtain the normalized factor value;

[0026] For the normalized factor values, a machine learning model trained based on historical data is used to generate triangular fuzzy number judgment results, and then a 12×12 fuzzy judgment matrix is ​​constructed;

[0027] Based on the fuzzy judgment matrix, the logarithmic least squares method is used to solve the fuzzy weight vector of each risk factor. The triangular fuzzy numbers in the fuzzy weight vector are converted into precise values ​​through the defuzzification method to form an accurate weight vector.

[0028] The precise weight vector is checked for consistency and the fuzzy consistency ratio is calculated. If the ratio is less than 0.1, it indicates that the consistency of the fuzzy judgment matrix is ​​good and the weight vector is valid.

[0029] The leakage probability value of each grid in the three-dimensional heat map is weighted and aggregated with the normalized value and precise weight of the corresponding risk factor to calculate the dynamic risk index of each grid; when the calculated dynamic risk index exceeds the preset threshold, a three-level linkage alarm is triggered, including the sound and light alarm, mobile terminal APP and emergency platform.

[0030] Furthermore, in step S5, in response to the three-level linkage alarm, a mobile robot is dispatched to navigate to the target area based on thermal map positioning and reference coordinates. The methane laser detector is used to re-measure the gas concentration, and infrared thermal imaging is used to scan the pipeline surface temperature field. The verification results and correction data are then bound together using blockchain evidence storage technology to form an unalterable full-link evidence chain, including:

[0031] Based on the abnormal concentration coordinates of the leakage heat map and the coordinates of the endpoints of the benchmark pipe network topology, a dynamic path planning algorithm is used to generate a robot collaborative navigation strategy, and a three-dimensional spatial benchmark coordinate system is constructed through multi-source positioning fusion technology.

[0032] Based on a three-dimensional reference coordinate system, a high-precision methane laser detector continuously collects gas concentration gradient data every 0.5m along the pipeline axis. Simultaneously, it drives an infrared thermal imager to scan for abnormal temperature fluctuations within a 2m radius around the detection point, obtaining temperature field fluctuation characteristics.

[0033] The temperature field fluctuation characteristics and concentration gradient data are aligned and fused in time and space, combined with the leakage diffusion model correction parameters and risk level judgment results, and a distributed hash tree structure is used to generate data fingerprints. The data is then written into the cross-domain evidence storage platform through the blockchain consensus node to form a full-link evidence chain.

[0034] In a second aspect, a city gas pipeline micro-leakage early warning system includes:

[0035] The reference line module is used to deploy a multi-parameter array of integrated pressure fluctuation, trace gas spectrum, acoustic wave or vibration, and pipeline deformation sensors along the gas pipeline. The node spacing is less than or equal to 30 meters, and wide-area coverage is achieved through LoRa or NB-IoT hybrid networking. Two fixed target points are selected with a spacing of more than or equal to 50 meters and located on the same straight line. The three-dimensional coordinates are calibrated through precision positioning to form a reference line.

[0036] The output module is used to collect pressure, gas concentration, and vibration data between two points in real time based on a reference line, calculate the environmental noise baseline, signal attenuation coefficient, and spatial consistency parameter, and generate dynamic correction values. It also performs adaptive filtering on the original signal, eliminates environmental noise through wavelet packet decomposition, calibrates the dynamic baseline in conjunction with Kalman filtering, and uses the correction values ​​to perform spatial compensation for pressure gradients, gas concentration gradients, and vibration spectrum characteristics, outputting the corrected spatiotemporal feature vectors.

[0037] The generation module is used to input the corrected spatiotemporal feature vector into the deep spatiotemporal convolutional network, extract the temporal evolution law through the bidirectional LSTM, combine it with the three-dimensional convolution kernel to mine the spatial correlation features, and synchronously integrate the GIS pipeline topology and benchmark line parameters to generate a three-dimensional heat map of leakage probability and diffusion trend;

[0038] The trigger module is used to calculate the dynamic risk index based on the evaluation matrix using the fuzzy analytic hierarchy process. When the index exceeds the threshold, it triggers a three-level linkage alarm of the sound and light alarm, mobile terminal, and emergency platform;

[0039] The verification module is used to dispatch mobile robots in response to three-level linkage alarms, navigate to the target area based on thermal map positioning and reference coordinates, remeasure gas concentration using a methane laser detector, and scan the pipeline surface temperature field using infrared thermal imaging. The verification results and correction data are then bound together using blockchain evidence storage technology to form an unalterable full-link evidence chain.

[0040] According to a third aspect, a computing device includes:

[0041] one or more processors;

[0042] The storage device is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the method.

[0043] In a fourth aspect, a computer-readable storage medium stores a program, which implements the method when executed by a processor.

[0044] The above solution of the present invention includes at least the following beneficial effects:

[0045] By deploying a multi-parameter sensor array and combining it with LoRa or NB-IoT networking, wide-area real-time monitoring of gas pipelines is achieved, ensuring that even small leaks within a 30-meter node spacing are not missed. High-precision positioning and data processing algorithms improve data collection accuracy, accurately capture early leak characteristics, and reduce false alarm rates. Technologies such as deep spatiotemporal convolutional networks and bidirectional LSTMs are used to mine data features and integrate GIS information to generate three-dimensional heat maps. Fuzzy analytic hierarchy process (AHP) assesses risk by integrating 12 factors to accurately determine the leak level and trigger a three-level linkage alarm. After the alarm is triggered, robots are dispatched to conduct re-surveys based on the heat map and reference coordinates. High-precision instrument scanning rapidly verifies and locates the leak point, making it more efficient and safer than manual inspections. Blockchain technology is used to bind verification results to data, forming an unalterable chain of evidence, ensuring data authenticity and integrity, and providing a reliable basis for accident analysis. Early warning and computing equipment collaborate to automate the entire process from monitoring and early warning to emergency verification, improving the intelligence and management efficiency of gas pipeline safety monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 The present invention provides a flow chart of a method for early warning of micro-leakage in urban gas pipelines.

[0047] Figure 2 The present invention provides a schematic diagram of a city gas pipeline micro-leakage early warning system. DETAILED DESCRIPTION

[0048] The following describes exemplary embodiments of the present disclosure in more detail with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0049] like Figure 1 As shown, an embodiment of the present invention provides a city gas pipeline micro-leakage early warning method, the method comprising the following steps:

[0050] Step S1: Deploy a multi-parameter array of integrated pressure fluctuation, trace gas spectrum, acoustic wave or vibration, and pipeline deformation sensors along the gas pipeline, with node spacing less than or equal to 30 meters. Wide-area coverage is achieved through a LoRa or NB-IoT hybrid network. Select two fixed target points with a spacing greater than or equal to 50 meters and located on the same straight line segment, and calibrate their three-dimensional coordinates through precision positioning to form a reference line.

[0051] Step S2: Based on the reference line, the pressure, gas concentration, and vibration data between the two points are collected in real time, and the environmental noise baseline, signal attenuation coefficient, and spatial consistency parameter are calculated to generate a dynamic correction value. The original signal is adaptively filtered to eliminate environmental noise through wavelet packet decomposition. The dynamic baseline is calibrated in combination with Kalman filtering. The correction value is used to perform spatial compensation for the pressure gradient, gas concentration gradient, and vibration spectrum characteristics, and the corrected spatiotemporal feature vector is output.

[0052] Step S3: The corrected spatiotemporal feature vector is input into the deep spatiotemporal convolutional network. The temporal evolution law is extracted through the bidirectional LSTM, and the spatial correlation features are mined by combining the 3D convolution kernel. The GIS pipeline topology and the reference line parameters are simultaneously integrated to generate a 3D heat map of the leakage probability and diffusion trend.

[0053] Step S4: Based on the evaluation matrix, the dynamic risk index is calculated using the fuzzy analytic hierarchy process. When the index exceeds the threshold, a three-level linkage alarm is triggered, including the sound and light alarm, mobile terminal, and emergency platform.

[0054] Step S5: In response to the three-level linkage alarm, the mobile robot is dispatched to navigate to the target area based on the thermal map positioning and reference coordinates. The gas concentration is remeasured using a methane laser detector, and the pipeline surface temperature field is scanned using infrared thermal imaging. The verification results and correction data are then bound using blockchain evidence storage technology to form an unalterable full-link evidence chain.

[0055] In an embodiment of the present invention, multi-parameter fusion, high-density deployment of sensor nodes, and the use of adaptive filtering and spatial compensation technology are used to achieve multi-dimensional and precise monitoring of micro-leaks in gas pipelines, avoid missed detections, and reduce false alarms; with the help of bidirectional LSTM and three-dimensional convolution kernels, spatiotemporal intelligent diagnosis is performed, and the risk index is quantified in combination with the fuzzy hierarchical analysis method to trigger a three-level linkage alarm to achieve intelligent analysis and rapid early warning; robots carrying high-precision equipment are used to accurately review the leakage area, and full-link evidence is stored through blockchain technology to ensure the reliability of disposal; relying on LoRa or NB-IoT hybrid networking to achieve low-power wide-area communication, build end-edge-cloud integrated closed-loop management, promote the intelligent transformation of inspections and support urban emergency coordination, and improve the safety and intelligence level of the gas system through multi-technology integration and full-process closed-loop.

[0056] In a preferred embodiment of the present invention, step S1: deploying a multi-parameter array integrating pressure fluctuation, trace gas spectrum, acoustic wave or vibration, and pipeline deformation sensors along the gas pipeline, with node spacing of less than or equal to 30 meters, and achieving wide-area coverage through a LoRa or NB-IoT hybrid network; selecting two fixed target points with a spacing of greater than or equal to 50 meters and located on the same straight line segment, and calibrating their three-dimensional coordinates through precise positioning to form a reference line, may include:

[0057] Step S11: Deploy sensor nodes along the gas pipeline at intervals of 30 meters or less. Each node integrates a pressure fluctuation sensor, a trace gas spectrum sensor, an acoustic or vibration sensor, and a pipeline deformation monitoring unit. LoRa wireless and NB-IoT cellular communication functions are pre-installed in each sensor node to achieve low-power wide-area communication coverage.

[0058] Step S12: Based on low-power wide-area communication coverage, a dual-channel communication link is established through the LoRa modulation method and the NB-IoT cellular network. LoRa is used for short-distance networking between nodes to transmit pressure fluctuations, gas concentrations, and vibration data in real time. NB-IoT is used for wide-area data backhaul, prioritizing the transmission of high-frequency deformation data from pipeline deformation monitoring units and synchronously uploading the position coordinates of each sensor node. The dual-channel link ensures that the upload delay of all sensor data is less than or equal to 5 seconds, forming a dynamically updated pipeline network spatial topology map.

[0059] In step S13, based on the spatial topology map of the pipe network, two target points with a distance of 50 meters or more and located on the same straight line segment are selected from the deployed sensor nodes; the GNSS positioning device is used to calibrate the three-dimensional coordinates of the target points, and the coordinate data is transmitted back to the control center in real time through the NB-IoT channel; based on the returned coordinate data, the actual distance between the two points is measured using a laser rangefinder, and the deviation from the theoretical straight line distance is calculated: if the deviation is less than or equal to 0.1%, the line connecting the two points is confirmed as a valid reference line; if the deviation is greater than 0.1%, the target points are reselected from the topology map and iteratively executed until a reference line is formed when the deviation is less than or equal to 0.1%.

[0060] In this embodiment of the present invention, high-density deployment of multi-parameter sensor nodes enables comprehensive collection of key gas pipeline data, eliminating monitoring blind spots. Low-power wide-area communication ensures stable data transmission over long distances, reducing equipment energy consumption and maintenance costs, and strengthening the data foundation. A dual-channel communication link allocates transmission tasks based on data characteristics, leveraging the advantages of LoRa and NB-IoT technologies to improve transmission efficiency and stability. Combined with strict latency control, it enables real-time monitoring of pipeline operations. Dynamically updated pipeline network topology maps provide intuitive geographic information for analysis, positioning, and management. A reference line, formed by the deviation between calculated and theoretical straight-line distances, ensures accurate and consistent data collection, laying a solid foundation for precise early warning of micro-leaks.

[0061] In the embodiment of the present invention, when it is specifically applied, the above steps can be implemented by the following steps:

[0062] In step S11, sensor nodes are deployed along the gas pipeline at intervals no greater than 30 meters. Each node must integrate multiple sensors, including pressure fluctuation sensors, trace gas spectrometer sensors, acoustic or vibration sensors, and pipeline deformation monitoring units, to collect multi-dimensional data from the gas pipeline. Furthermore, each sensor node is pre-installed with LoRa wireless and NB-IoT cellular communication capabilities. During communication implementation, the relevant parameters of LoRa and NB-IoT must be configured and debugged. These include setting LoRa frequency bands and spreading factors to accommodate low-power, long-distance communication requirements, and configuring NB-IoT access modes and data transmission protocols. This ensures stable, low-power, wide-area communication connections between sensor nodes and with the control center, enabling efficient data transmission.

[0063] Step S12: Based on the achieved low-power wide-area communication coverage, a dual-channel communication link is constructed. For the LoRa communication link, its role in short-distance networking between nodes must be clearly defined. Therefore, the data types transmitted must be configured and managed. The transmission of pressure fluctuation, gas concentration, and vibration data is assigned to the LoRa link. Parameters such as data packaging, compression, and transmission frequency are optimized to ensure real-time and efficient transmission. For the NB-IoT communication link, high-frequency deformation data (e.g., sampling data 10 times per second or more) generated by the pipeline deformation monitoring unit and the location coordinate data of each sensor node are prioritized for transmission. Data caching, priority scheduling, and transmission protocols for the NB-IoT link are configured to ensure reliable transmission of high-frequency data and critical coordinate data. During the dual-channel data transmission process, data upload latency is monitored in real time. By adjusting LoRa and NB-IoT transmission parameters and data queue scheduling, the upload latency for all sensor data is kept within 5 seconds. Simultaneously, based on the received sensor node location coordinates and data transmission relationships, a spatial topology map of the pipeline network is gradually constructed and updated, recording the connection relationships, location information, and data flow of each node.

[0064] In step S13, based on the established spatial topology map of the pipe network, two target points located on the same straight line segment with a distance of 50 meters or greater are selected from the numerous deployed sensor nodes. A GNSS positioning device is used to accurately measure the three-dimensional coordinates (including longitude, latitude, and elevation) of the selected target points, obtaining initial coordinate data. This data is then transmitted back to the control center in real time via the NB-IoT communication channel. After receiving the coordinate data, the control center uses a laser rangefinder to measure the actual distance between the two points. The actual measured distance is compared with the theoretical straight-line distance calculated from the coordinate data, and the deviation between the two is calculated. The specific calculation process involves first calculating the theoretical straight-line distance, then subtracting the actual distance from the theoretical distance, and then dividing the difference by the theoretical distance and multiplying it by 100% to obtain the deviation percentage. If the deviation value is less than or equal to 0.1%, the line connecting the two points is deemed to meet the accuracy requirements and is determined as a valid reference line. If the deviation value is greater than 0.1%, other target points that meet the conditions are selected from the topological map, and the above operations such as coordinate measurement, distance calculation, and deviation judgment are repeated until a line connecting the two points with a deviation value that meets the requirements is found to form the final reference line.

[0065] In a preferred embodiment of the present invention, the above step S2: based on the reference straight line, real-time acquisition of pressure, gas concentration and vibration data between two points, calculation of the environmental noise baseline, signal attenuation coefficient and spatial consistency parameter, and generation of a dynamic correction value; adaptive filtering of the original signal, elimination of environmental noise by wavelet packet decomposition, calibration of the dynamic baseline in combination with Kalman filtering, and spatial compensation of the pressure gradient, gas concentration gradient and vibration spectrum characteristics using the correction value, and output of the corrected spatiotemporal feature vector, may include:

[0066] Step S21, based on the pressure fluctuation data between two points on the reference line, the ambient noise baseline is calculated using the sliding window average method; the signal attenuation coefficient is calculated based on the Fresnel reflection model in combination with the gas concentration gradient; and the spatial consistency parameter is determined through covariance analysis;

[0067] Step S22: Based on the calculated ambient noise baseline, the original pressure signal is subjected to wavelet packet decomposition to extract the effective signal in the 1-100 Hz frequency band; based on the signal attenuation coefficient, the dynamic baseline is calibrated on the gas concentration data using Kalman filtering;

[0068] In step S23, based on the spatial consistency parameter, the dynamic correction value is distributed to the pressure gradient, gas concentration gradient and vibration spectrum characteristics according to the weight; the weighted feature vector is normalized, and the corrected spatiotemporal feature vector is output.

[0069] In an embodiment of the present invention, the environmental noise baseline, signal attenuation coefficient and spatial consistency parameters are obtained through the sliding window averaging method, Fresnel reflection model and covariance analysis, accurately characterizing the normal operation characteristics of the pipeline and providing a reliable basis for leakage judgment; wavelet packet decomposition and Kalman filtering are used to denoise and calibrate the pressure and gas concentration data to improve data quality; and then dynamic correction value weighting and normalization are performed to eliminate the influence of spatial factors, making multi-dimensional data more accurate and comparable, and finally outputting high-quality spatiotemporal feature vectors to enhance the accuracy, sensitivity and overall performance of micro-leakage warning.

[0070] In the embodiment of the present invention, when it is specifically applied, the above steps can be implemented by the following steps:

[0071] Step S21, ambient noise baseline calculation (pressure fluctuation data):

[0072] Data range, extract the pressure fluctuation data of all sensor nodes between two points of the reference line, and the time window is set to the last 30 minutes (which can be adjusted according to the pipeline operation characteristics).

[0073] Sliding window mean method:

[0074] The time window is divided into multiple sub-windows (e.g., each sub-window is 5 minutes), and the mean and standard deviation of the pressure data in each sub-window are calculated; outliers outside the mean ± 3 times the standard deviation are eliminated (considered as transient interference), and the mean of the remaining data is used as the environmental noise baseline (representing the pressure fluctuation benchmark during normal operation).

[0075] Signal attenuation coefficient calculation (gas concentration gradient + Fresnel reflection model):

[0076] For the methane concentration data of all nodes on the baseline, the values ​​that exceed the measurement range or have obvious jumps (such as concentration changes greater than 50% between adjacent moments) are eliminated, and interpolation methods (such as linear interpolation or cubic spline interpolation) are used to fill in missing data to ensure the continuity of the concentration series.

[0077] Concentration gradient calculation:

[0078] For the nodes arranged in sequence on the reference line, the concentration gradient between adjacent nodes is calculated, along the one-dimensional coordinates of the reference line (with the starting point as 0 meters and the end point as the total length D meters).

[0079] Fresnel diffraction model adaptation and curve fitting, ideal model assumptions:

[0080] In an ideal scenario with no wind and no obstacles, the gas leaks and diffuses in all directions in the form of a point source, which conforms to the Fresnel diffraction theory. The concentration decays with distance x as follows: C(x) = C0 / x 2 (inverse square law of distance), where C0 is the initial concentration at the leakage source.

[0081] Actual data fitting:

[0082] The coordinates of each node on the reference line and concentration As sample points, the exponential decay curve is fitted using the nonlinear least squares method. (Closer to the actual diffusion scenario), where k represents the fitting parameter and e represents the base of the natural logarithm.

[0083] Fitting steps:

[0084] Take the natural logarithm of the concentration data and transform it into a linear model: .

[0085] Calculate the slope using the least squares method and intercept , and obtain the fitting parameter k (i.e., signal attenuation coefficient).

[0086] Goodness of fit test:

[0087] Calculate the coefficient of determination R 2 , if R 2If R is greater than or equal to 0.9, the fitting is considered valid; 2 If the value is less than 0.9, you need to check whether there is local interference (such as ventilation or terrain influence), and remove abnormal nodes and refit. If there are branch pipes or valves on the reference line, the attenuation coefficient k needs to be corrected according to the node position:

[0088] For nodes close to branch points, the attenuation coefficient increases by 20% (to simulate the effect of changes in gas diffusion paths); for nodes located upstream of valves, the attenuation coefficient decreases by 10% (the valve may hinder gas diffusion).

[0089] Step S22, pressure signal wavelet packet decomposition (environmental noise baseline application):

[0090] Collect pressure data from each sensor on the reference line, with a sampling rate of 200 Hz (to ensure that signals within 100 Hz can be captured). Use the ambient noise baseline calculated in step S21 (i.e., the average and standard deviation of pressure fluctuations during normal operation).

[0091] Three-layer wavelet packet decomposition

[0092] First-level decomposition: The original pressure signal is decomposed into a "low-frequency component" and a "high-frequency component." The low-frequency component contains signals between 0 and 100 Hz (which may contain noise and leakage characteristics), while the high-frequency component contains signals between 100 and 200 Hz (usually irrelevant noise, which is directly discarded).

[0093] Second layer decomposition: decompose the low-frequency part of the first layer again to obtain lower-frequency 0-50Hz and 50-100Hz signals.

[0094] The third layer of decomposition: decompose the 0-50Hz signal into 0-25Hz and 25-50Hz, and decompose the 50-100Hz signal into 50-75Hz and 75-100Hz, resulting in a total of 4 frequency bands.

[0095] Frequency band noise evaluation:

[0096] Calculate the "energy" of each frequency band (i.e., the sum of the squares of the signal strengths) and compare the energy of each frequency band with the ambient noise baseline. If the energy of a frequency band exceeds 80% of the baseline energy, it is determined that the frequency band contains mainly noise (usually the 0-25Hz band).

[0097] Noise band processing:

[0098] The 0-25Hz frequency band is completely discarded or significantly attenuated (equivalent to filtering out steady-state fluid flow noise).

[0099] 25-50Hz frequency band: Partially preserves the signal and only attenuates very small fluctuations (possibly residual noise).

[0100] 50-100 Hz band: Almost completely preserved (this band most likely contains pressure pulses from leaks).

[0101] Signal reconstruction:

[0102] The processed four-band signals are recombined to generate a pressure fluctuation curve after noise filtering. The signal waveforms before and after processing are compared to ensure that low-frequency noise is reduced while high-frequency pulse characteristics are retained.

[0103] Kalman filtering of gas concentration data (signal attenuation coefficient application):

[0104] initialization:

[0105] The actual measured concentration value at the current moment is used as the initial estimate, and the initial "uncertainty" (i.e., estimation error) is set to half of the sensor accuracy.

[0106] Prediction stage:

[0107] Based on historical estimates: Based on the concentration estimate at the previous moment, predict the concentration at the current moment (assuming that the concentration change is continuous, such as the current concentration is approximately equal to 95% of the concentration at the previous moment), multiply the predicted value by the attenuation coefficient (calculated in step S21), and correct the concentration difference caused by the distance from the leak source.

[0108] Update phase:

[0109] Compare the predictions with the actual measurements and calculate the difference between the predicted concentration and the actual measured concentration;

[0110] Dynamically adjust weights:

[0111] If the difference is small, it means the prediction is accurate and the predicted value is given a higher weight;

[0112] If the difference is large, it means that the measurement is more reliable and the measured value is given a higher weight.

[0113] Based on the weights, the predicted values ​​and measured values ​​are weighted averaged to obtain the calibrated concentration estimate (i.e., dynamic baseline). Based on the calibration effect, the estimation error is appropriately reduced, and the signal attenuation coefficient is recalculated every 10 minutes. If the measurement value of a sensor continues to deviate from the predicted value, the location is marked as a suspected leak point, and its attenuation coefficient is temporarily increased to enhance sensitivity.

[0114] Step S23, dynamic correction value weight allocation (spatial consistency parameter application):

[0115] The correction value consists of the dynamic correction value, which includes the pressure gradient correction, the concentration gradient correction, and the vibration spectrum correction, which are calculated by the ambient noise baseline, the signal attenuation coefficient, and the spatial consistency parameter respectively.

[0116] Weight distribution logic:

[0117] The higher the spatial consistency parameter (such as greater than or equal to 0.8), the stronger the correlation between the data between nodes, and the lower the weight of the correction value (to avoid over-correction); the lower the spatial consistency parameter (such as less than 0.5), the more likely the data is subject to local interference, and the higher the weight of the correction value (to enhance the correction strength).

[0118] Specific allocation method: weight w = 1 - spatial consistency parameter (for example, a consistency parameter of 0.6 corresponds to a weight of 0.4, which means that the contribution of the correction value is 40%).

[0119] Eigenvector weighting and normalized weighting processing:

[0120] The pressure gradient, gas concentration gradient, and vibration spectrum characteristics (such as main frequency and energy distribution) are multiplied by the corresponding correction value weights to adjust the spatial consistency of each parameter (for example, the closer the node is to the leakage point, the higher the weight and the greater the correction amount).

[0121] Normalization processing:

[0122] The weighted eigenvalues ​​are mapped to the interval [0, 1] to eliminate the dimension effects (e.g., the pressure gradient unit is kPa / m, the concentration gradient unit is ppm / m, and the vibration spectrum is a dimensionless energy value). Through the above-mentioned layered calculation, noise suppression, spatial deviation calibration and multi-parameter consistency adjustment of the original signal are achieved.

[0123] In a preferred embodiment of the present invention, step S3 above: inputting the corrected spatiotemporal feature vector into a deep spatiotemporal convolutional network, extracting temporal evolution patterns through a bidirectional LSTM, mining spatial correlation features in combination with a three-dimensional convolution kernel, and synchronously integrating GIS pipeline topology and reference line parameters to generate a three-dimensional heat map of leakage probability and diffusion trend, may include:

[0124] Step S31: construct a deep spatiotemporal convolutional network model, which includes three bidirectional LSTM layers and two three-dimensional convolutional layers; use the leakage sample dataset to supervise the model training, and the loss function is a weighted combination of cross entropy and mean square error. The network weights are optimized by the back propagation algorithm to obtain the optimized DSTCN model weights;

[0125] In step S32, based on the optimized DSTCN model weights, the modified spatiotemporal feature vector is sliced ​​into the model according to the time window, and a leakage probability sequence in the time dimension is output; the elbow and valve coordinates and the reference line spatial parameters in the GIS pipeline topology are synchronously integrated, and a three-dimensional heat map is generated through inverse distance weighted interpolation, in which the leakage probability value is mapped to the range of 0-1;

[0126] Step S33: Based on the three-dimensional heat map, extract the area with a leakage probability greater than or equal to 0.7 as the core leakage point; combine the gas diffusion model, use the core leakage point as the initial condition, and predict the diffusion path in the next 30 minutes.

[0127] In an embodiment of the present invention, by scientifically designing a deep spatiotemporal convolutional network architecture and optimizing training, the model effectively learns the spatiotemporal characteristics of gas pipeline leakage data and improves the leakage feature extraction and identification capabilities; the spatiotemporal feature vector slices are input into the model, combined with the GIS pipeline topology and reference line parameters, the leakage probability sequence is output and a three-dimensional heat map is generated to realize leakage risk visualization; the core leakage points are accurately identified, and the trend is predicted in combination with the gas diffusion model, providing a scientific basis for emergency decision-making and rescue planning, enhancing the effectiveness of emergency response and reducing the harm of accidents.

[0128] In the embodiment of the present invention, when it is specifically applied, the above steps can be implemented by the following steps:

[0129] In step S31, the construction of the deep spatiotemporal convolutional network (DSTCN) model architecture is carried out around three parts: network design, data preparation, and model training optimization, as follows:

[0130] Network architecture design:

[0131] DSTCN adopts a hierarchical stacked structure, with its core consisting of three bidirectional LSTM layers and two three-dimensional convolutional layers. The bidirectional LSTM layer serves as a temporal feature extraction module. Each bidirectional LSTM layer contains two LSTM units, one forward and one backward, which can simultaneously capture data change characteristics from the forward order (past to future) and the reverse order (future to past) of the time series, solving the problem that traditional unidirectional LSTM cannot utilize future information. The stacking of multiple layers of bidirectional LSTM can achieve hierarchical abstraction of temporal features. The bottom layer captures short-term anomalies such as second-level pressure fluctuations, and the upper layer mines long-term trends such as hour-level gas concentration changes.

[0132] The three-dimensional convolution layer focuses on mining the spatial correlation characteristics of the data. Its input is the space-time tensor converted from the output of the bidirectional LSTM layer, covering three dimensions: time step, number of spatial nodes, and feature dimension. The three-dimensional convolution kernel slides in the time and space dimensions, and can simultaneously extract the coordinated changes between adjacent sensor nodes within a certain period of time. For example, it can identify the simultaneous increase in gas concentration at multiple nodes in a certain area, or analyze the concentration diffusion pattern within the entire baseline range, so as to accurately capture the spatial characteristics of the leakage.

[0133] Leakage sample dataset preparation:

[0134] A dataset was constructed that included the normal operating status of gas pipelines and different leakage scenarios, covering various situations such as micro-leaks and sudden large-break leaks. The dataset was divided into three subsets according to a certain ratio: the training set was used for model parameter learning, the validation set was used to evaluate model performance and adjust hyperparameters during training, and the test set was used to independently test the generalization ability of the model after training was completed to ensure the effectiveness of the model on unknown data.

[0135] Model training and optimization:

[0136] During the training phase, a loss function that is a weighted combination of cross entropy and mean squared error is used. Cross entropy is used to measure the difference between the model's predicted category probability (e.g., whether a leak has occurred) and the true label, making it suitable for classification tasks. Mean squared error measures the error between the predicted value (e.g., the specific value of the leak probability) and the true value, making it suitable for regression tasks. The combination of the two can simultaneously optimize the model's classification accuracy and numerical prediction accuracy.

[0137] During training, the training data set is first fed into the model. The model then performs forward propagation based on the current weights to calculate the predicted results. A loss function is then used to calculate the error between the predicted and true values. The backpropagation algorithm then propagates the error back through the network layer by layer, calculating the gradient for each weight. The network weights are then adjusted based on the gradient, completing a parameter update.

[0138] To prevent the model from overfitting on the training data, the validation set is regularly used to evaluate the model performance during the training process, and hyperparameters such as the weighted combination coefficient of the loss function and the learning rate are dynamically adjusted. After multiple rounds of iterative training, the optimized DSTCN model weights are finally obtained.

[0139] In step S32, based on the optimized DSTCN model weights, the corrected spatiotemporal feature vector is sliced ​​according to a fixed time window length, and the data in each time window is input into the model in sequence; the bidirectional LSTM layer in the model processes the sliced ​​data, extracts the leakage probability change pattern in the time dimension, and outputs the leakage probability sequence in the time dimension.

[0140] At the same time, GIS pipeline topology data is obtained, including the coordinate information of key components such as elbows and valves, as well as the spatial parameters of the reference line (such as two-point coordinates, length, etc.); the inverse distance weighted interpolation method is used to perform data interpolation calculations in three-dimensional space based on these known coordinate information and leakage probability sequences; the leakage probability values ​​at different locations are mapped to the range of 0-1, and a three-dimensional heat map is generated based on the mapped probability values, where areas with higher probability values ​​appear brighter in color in the heat map, intuitively showing the possibility of leakage at different locations.

[0141] In step S33, the three-dimensional thermal map generated in step S32 is analyzed to screen out areas with a leakage probability greater than or equal to 0.7. These areas are identified as core leakage points. A gas diffusion model is selected (such as a diffusion model based on fluid mechanics or an empirical diffusion model), and information such as the location of the core leakage point and the current gas concentration is input into the model as initial conditions. The model is used to simulate and calculate the gas diffusion process based on pipeline environmental parameters (such as wind speed, temperature, pipeline pressure, etc.), and the gas diffusion path in the next 30 minutes is predicted. The prediction results containing information such as the diffusion direction and diffusion range are output.

[0142] In a preferred embodiment of the present invention, the above step S4: based on the evaluation matrix, the dynamic risk index is calculated using the fuzzy analytic hierarchy process. When the index exceeds the threshold, the three-level linkage alarm of the sound and light alarm, the mobile terminal and the emergency platform is triggered, which may include:

[0143] Step S41: Map all the original data of the risk factors to the range of 0-1 using the range normalization method to obtain the normalized factor values;

[0144] Step S42: for the normalized factor values, a machine learning model trained based on historical data is used to generate triangular fuzzy number judgment results, thereby constructing a 12×12 fuzzy judgment matrix;

[0145] Step S43: Using the logarithmic least squares method to solve the fuzzy judgment matrix, a fuzzy weight vector of each risk factor is obtained. The triangular fuzzy numbers in the fuzzy weight vector are converted into precise values ​​through a defuzzification method to form a precise weight vector.

[0146] Step S44: Perform consistency check on the precise weight vector and calculate the fuzzy consistency ratio. If the ratio is less than 0.1, it indicates that the consistency of the fuzzy judgment matrix is ​​good and the weight vector is valid.

[0147] In step S45, the leakage probability value of each grid in the three-dimensional heat map is weighted and aggregated with the normalized value and precise weight of the corresponding risk factor to calculate the dynamic risk index of each grid; when the calculated dynamic risk index exceeds the preset threshold, a three-level linkage alarm is triggered, including the sound and light alarm, mobile terminal APP and emergency platform.

[0148] In an embodiment of the present invention, the dimensional differences of different risk factors of leakage rate and building vulnerability are eliminated through range normalization, the data range is unified, the foundation is laid for weighted calculation and model processing, and the assessment accuracy is improved; a data-driven machine learning model is used, combined with the uncertainty of triangular fuzzy quantitative factor comparison, to reduce subjective bias; the logarithmic least squares method is used to rigorously solve the weights, and the operability of the results is enhanced through defuzzification; logical conflicts are avoided with the help of consistency testing, and automated processes are implemented to improve efficiency; finally, the leakage probability and multi-factor weights are integrated to dynamically assess the risk, and a hierarchical alarm mechanism is used to respond quickly to reduce the hazards of gas leakage accidents.

[0149] In the embodiment of the present invention, when it is specifically applied, the above steps can be implemented by the following steps:

[0150] In the above steps, the 12 risk factors are divided into four first-level indicators, each of which contains three second-level factors. The raw data of the specific risk factors are described as follows:

[0151] Leakage characteristics (3 factors):

[0152] Leak rate:

[0153] Definition: Gas leakage per unit time (unit: cubic meters / hour, m 3 / h).

[0154] The data source is calculated by reverse calculation based on the pressure gradient change and gas concentration diffusion model (such as the thermal map output in step S3).

[0155] Diffusion rate:

[0156] Definition: Diffusion rate of leaked gas in the air (unit: meters per minute, m / min).

[0157] Data source: Predicted rate of change of diffusion range using a combination of wind speed, terrain data (e.g., GIS maps), and 3D thermal maps.

[0158] Gas concentration peak:

[0159] Definition: The maximum methane concentration in the leakage area (unit: %LEL, percentage of lower explosion limit).

[0160] Data source: sensor measured concentration data or robot retest results (step S5).

[0161] Environmental sensitivity (3 factors):

[0162] Population density:

[0163] Definition: The average population within 500 meters of the leak point (unit: people / square kilometer).

[0164] Data source: Urban population distribution statistics (such as civil affairs departments or GIS databases).

[0165] Building density:

[0166] Definition: The proportion of the building area in the area surrounding the leakage point (unit: %).

[0167] Data source: urban planning data (such as building volume ratio and satellite remote sensing interpretation results).

[0168] Complexity of underground pipe network:

[0169] Definition: The density of other underground pipelines (such as water supply, electricity, and communications) in the area where the leak is located (level: low, medium, high, quantifiable as 1-3 points).

[0170] Data source: underground pipeline network topology map (recording the number of pipelines and intersection density).

[0171] Pipeline vulnerability (3 factors):

[0172] Pipe aging degree:

[0173] Definition: The ratio of the service life of the pipeline to its design life (unit: %).

[0174] Data source: Pipeline archival records (such as laying time, material type, maintenance history).

[0175] Corrosion rate:

[0176] Definition: Average annual corrosion thickness of pipeline wall (unit: mm / year, mm / a).

[0177] Data source: pipeline internal inspection report (such as magnetic flux leakage detection, ultrasonic thickness measurement).

[0178] Deformation rate:

[0179] Definition: Annual change in axial or circumferential deformation of a pipeline (unit: mm / year, mm / a).

[0180] Data source: high-frequency data from the pipeline deformation monitoring unit in step S1 (such as fiber Bragg grating strain measurement).

[0181] Emergency response difficulty (3 factors):

[0182] Arrival Time:

[0183] Definition: The shortest driving time for emergency teams from their base to the leak site (unit: minutes).

[0184] Data source: Emergency resource database (calculated based on real-time traffic network congestion conditions).

[0185] Disposal space:

[0186] Definition: The area around the leak point that can be used for the deployment of emergency vehicles and equipment (unit: square meters, ㎡).

[0187] Data source: Spatial data such as vacant land and road width in GIS maps.

[0188] Communication coverage quality:

[0189] Definition: The wireless communication signal strength in the leakage area (level: strong, medium, weak, quantifiable as 1-3 points).

[0190] Data source: Communication operators’ signal coverage heat maps or communication quality indicators of field sensor nodes (such as the signal-to-noise ratio of LoRa or NB-IoT).

[0191] In step S41, the original data of 12 risk factors (such as leakage rate, population density, etc.) are extracted from multiple source systems such as the heat map in step S3, the GIS database, and the sensor network. An independent data sequence is established for each factor (for example, the "population density" of all monitoring points forms an array).

[0192] Extreme value determination:

[0193] For the data series of each factor, its historical maximum and minimum values ​​are counted (for example, the historical maximum value of population density is 10,000 people / square kilometer, and the minimum value is 1,000 people / square kilometer), and these extreme values ​​are stored as benchmark parameters for normalization of all subsequent data points.

[0194] Linear mapping calculation:

[0195] For each raw data point, perform the following operations: subtract the minimum value of the corresponding factor from the value of the data point; divide the result by the difference between the maximum and minimum values ​​of the factor; the quotient is the normalized value (range 0-1).

[0196] Example: The original value of population density is 5000, and the normalized value = (5000-1000) / (10000-1000) = 0.44.

[0197] Matrix construction:

[0198] The normalized values ​​of all 12 factors are arranged according to monitoring points to form a two-dimensional matrix (rows: monitoring points, columns: factors), which serves as the input of the subsequent fuzzy hierarchical analysis.

[0199] In the above step S42, historical risk assessment reports (including expert scores), pipeline accident records (including the actual values ​​of 12 factors and the final risk level), and industry standards (such as the "Urban Gas Design Code") are collected.

[0200] Construct a data set, where each sample contains the normalized values ​​of 12 factors (output from step S41) and corresponding risk level labels (such as low, medium, and high risk); add ±5% random noise to the normalized value of each factor to generate three times the number of enhanced samples, and for each sample, generate a 12×12 factor importance matrix (if factor i has a greater impact on risk in historical cases, then the matrix element (i, j) = 1, otherwise = 0).

[0201] Dataset division:

[0202] The dataset is divided into training set, validation set, and test set in a ratio of 8:1:1. For example, for 1000 historical cases, 800 are used for training, 100 for validation, and 100 for testing.

[0203] Neural network architecture design:

[0204] Input layer:

[0205] The 12 nodes correspond to the normalized values ​​of the 12 risk factors.

[0206] Hidden layer design:

[0207] The first layer has 32 neurons and ReLU activation function to capture the nonlinear relationship between factors.

[0208] The second layer has 64 neurons and a LeakyReLU activation function (α=0.01) to enhance the perception of small differences in importance.

[0209] The third layer has 32 neurons and Tanh activation function, which maps the features to the interval [-1, 1].

[0210] Output layer:

[0211] 72 nodes (corresponding to the lower triangular part of the 12×12 factor pair matrix, excluding the diagonal).

[0212] Each output node represents the importance score of factor i relative to j, and the Sigmoid activation function maps the score to the [0,1] interval.

[0213] Training process:

[0214] Loss function design:

[0215] The main loss, mean squared error (MSE), measures how important the factor pair predicted by the model differs from the historical case labels.

[0216] Regularization term: L2 regularization (λ=0.001) is added to prevent overfitting.

[0217] Consistency constraint: If model prediction factor i is more important than j, and j is more important than k, then the prediction score of i is forced to be higher than that of k, and a penalty term is added when it is violated.

[0218] Optimizer configuration:

[0219] The Adam optimizer is used, with an initial learning rate of 0.001 and a learning rate decay strategy: the learning rate decays to 0.9 times the original value every 50 rounds.

[0220] Batch size: 64; training epochs: 200, early stopping mechanism: stop training if the validation set loss does not decrease for 10 consecutive epochs.

[0221] Training tips:

[0222] Weight initialization: Use He normal distribution to initialize the hidden layer weights.

[0223] Batch normalization, adding a BatchNorm layer after each hidden layer to accelerate convergence.

[0224] Dropout, set the dropout rate to 0.2 in the second and third layers to enhance the generalization ability.

[0225] For the trained model, the same test sample is input 100 times, and Dropout is enabled each time during forward propagation to simulate the uncertainty of the model. 100 importance scores of each factor pair (i, j) are collected to form a distribution, which is represented by triangular fuzzy numbers:

[0226] Lower bound: the probability that factor i is “slightly more important” than factor j;

[0227] Most likely value: the probability that factor i is "significantly more important" than factor j;

[0228] Upper limit: the probability that factor i is "extremely important" than factor j.

[0229] For example, the model might determine that the importance of "leakage rate" relative to "population density" is (0.6, 0.8, 0.9).

[0230] Matrix symmetry processing:

[0231] If the importance of factor i to j is (a, b, c), then the importance of factor j to i is automatically set to (1 / c, 1 / b, 1 / a), filling in the relative importance values ​​of all factor pairs to construct a complete 12×12 fuzzy judgment matrix.

[0232] In the above step S43, the logarithmic least squares method is used to solve:

[0233] Based on the fuzzy judgment matrix, the fuzzy weight of each factor is calculated through an iterative optimization algorithm (such as gradient descent). Each weight is expressed as a triangular fuzzy number (for example, the weight of the leakage rate may be (0.15, 0.25, 0.35)). For each triangular fuzzy number weight, its exact value is calculated:

[0234] Take the average of the lower limit, most likely value, and upper limit to get the corresponding exact weight (for example, in the above example, it is converted to (0.15 + 0.25 + 0.35) / 3 = 0.25).

[0235] Weight vector generation:

[0236] Arrange the exact weights of the 12 factors in order to form the final exact weight vector (such as [0.25, 0.18, …, 0.05]).

[0237] In the above step S44, it is checked whether the judgment matrix satisfies the transitivity: if factor i is more important than j, and j is more important than k, then i should be more important than k. For example, if "leakage rate" > "population density" and "population density" > "communication quality", then "leakage rate" must be > "communication quality".

[0238] Fuzzy Consistency Ratio (FCR) calculation:

[0239] The FCR value is calculated by comparing the difference between the judgment matrix and the ideal consistency matrix.

[0240] FCR < 0.1 indicates that the matrix logic is reasonable and the weights are valid; otherwise, the model needs to be retrained to adjust the judgment matrix.

[0241] In step S45, for each grid in the three-dimensional heat map, the leakage probability value (from step S3) is obtained. The normalized values ​​of the 12 factors are then obtained. Each normalized value is multiplied by its corresponding precise weight, then added together and multiplied by the leakage probability value to obtain the final risk index. Threshold determination and graded alarms are performed. If the risk index is ≥ 0.6, the audible and visual alarms flash and sound at a 2Hz frequency. Mobile terminals push an emergency notification containing the leak location, risk level, and estimated spread. The emergency platform automatically establishes a 50-meter evacuation radius centered on the leak point, generates a valve closure command, and sends it to the site for execution. This calculation process is repeated every five minutes, adjusting the risk index and alarm status based on the latest data. Once the alarm is triggered, the risk index is continuously monitored until it falls below the threshold, at which point the alarm is lifted.

[0242] In a preferred embodiment of the present invention, step S5 above: in response to the three-level linkage alarm, dispatching a mobile robot to navigate to the target area based on thermal map positioning and reference coordinates, re-measuring gas concentration using a methane laser detector, combining infrared thermal imaging to scan the pipeline surface temperature field, and binding the verification results and correction data using blockchain evidence storage technology to form an unalterable full-link evidence chain, which may include:

[0243] Step S51: Based on the abnormal concentration coordinates of the leakage heat map and the coordinates of the endpoints of the reference pipe network topology, a dynamic path planning algorithm is used to generate a robot collaborative navigation strategy, and a three-dimensional space reference coordinate system is constructed through multi-source positioning fusion technology;

[0244] In step S52, based on the three-dimensional reference coordinate system, a high-precision methane laser detector continuously collects gas concentration gradient data every 0.5 m along the pipeline axis. At the same time, an infrared thermal imager is driven to scan the temperature field for abnormal fluctuations within a range of 2 m around the detection point to obtain temperature field fluctuation characteristics.

[0245] In step S53, the temperature field fluctuation characteristics and concentration gradient data are spatially and temporally aligned and fused, and the leakage diffusion model correction parameters and risk level judgment results are combined to generate data fingerprints using a distributed hash tree structure. The data is then written into the cross-domain evidence storage platform through the blockchain consensus node to form a full-link evidence chain.

[0246] In an embodiment of the present invention, through a dynamic path planning algorithm and multi-source positioning fusion technology, the mobile robot can quickly and accurately reach the target area. Combined with the collaborative detection of a high-precision methane laser detector and an infrared thermal imager, the leakage situation is comprehensively judged from the two dimensions of gas concentration and temperature field. Compared with a single detection method, the accuracy and reliability of the detection are greatly improved. By using blockchain evidence storage technology, the detection data is generated into an unalterable data fingerprint and a full-link evidence chain is formed, which ensures the authenticity and integrity of the data, provides a reliable basis for subsequent accident analysis, responsibility determination, etc., and effectively avoids the risk of data tampering or loss. The automated detection and data processing process can quickly complete retesting and data verification after the alarm is triggered, and provide accurate information for emergency decision-making in a timely manner, so that the emergency platform can more efficiently formulate and implement response measures, which helps to reduce accident losses and hazards.

[0247] In the embodiment of the present invention, when it is specifically applied, the above steps can be implemented by the following steps:

[0248] In step S51, the coordinates of abnormal concentrations are extracted from the leakage heat map. These coordinates reflect the location information of the suspected leakage area. At the same time, the coordinates of the endpoints of the baseline pipe network topology are obtained as the basic reference for the pipe network layout. The coordinate information is input into the dynamic path planning algorithm. The algorithm will comprehensively consider factors such as pipeline distribution and obstacle location, plan the path of the mobile robot from the current position to the target leakage area, generate a robot collaborative navigation strategy, and clarify the robot's route. Multi-source positioning fusion technology is used to integrate data from various positioning methods such as satellite positioning and inertial navigation, calibrate and fuse positioning information from different sources, and construct a three-dimensional spatial reference coordinate system.

[0249] In step S52, based on the constructed three-dimensional space reference coordinate system, the high-precision methane laser detector is controlled to move along the axial direction of the pipeline, and gas concentration detection is performed every 0.5 meters. The gas concentration data at multiple positions are continuously collected to form gas concentration gradient data, thereby reflecting the changes in gas concentration along the pipeline; at the same time, the infrared thermal imager is driven to scan the 2-meter range around each detection point to capture the temperature information in the area. By analyzing the temperature data, it is identified whether there are abnormal fluctuations in the temperature field, and then the temperature field fluctuation characteristics are obtained, which are used to determine whether there are temperature anomalies on the pipeline surface caused by leakage and other problems.

[0250] In step S53, the temperature field fluctuation characteristics and gas concentration gradient data obtained in step S52 are aligned according to the time and space dimensions to ensure accurate correspondence between the two data types. A fusion analysis is then performed. The fused data is parsed based on the leak diffusion model's modified parameters and correlated with the risk level determination results, thereby achieving a comprehensive assessment of the leak situation. A distributed hash tree structure is used to calculate the processed data and generate a unique data fingerprint that represents the characteristics of the original data. Finally, through the blockchain consensus node, the data fingerprint and related data information are written to the cross-domain evidence storage platform, leveraging the characteristics of blockchain technology to form an unalterable, full-link evidence chain.

[0251] like Figure 2 As shown, an embodiment of the present invention further provides a city gas pipeline micro-leakage early warning system, comprising:

[0252] The reference line module is used to deploy a multi-parameter array of integrated pressure fluctuation, trace gas spectrum, acoustic wave or vibration, and pipeline deformation sensors along the gas pipeline. The node spacing is less than or equal to 30 meters, and wide-area coverage is achieved through LoRa or NB-IoT hybrid networking. Two fixed target points are selected with a spacing of more than or equal to 50 meters and located on the same straight line. The three-dimensional coordinates are calibrated through precision positioning to form a reference line.

[0253] The output module is used to collect pressure, gas concentration, and vibration data between two points in real time based on a reference line, calculate the environmental noise baseline, signal attenuation coefficient, and spatial consistency parameter, and generate dynamic correction values. It also performs adaptive filtering on the original signal, eliminates environmental noise through wavelet packet decomposition, calibrates the dynamic baseline in conjunction with Kalman filtering, and uses the correction values ​​to perform spatial compensation for pressure gradients, gas concentration gradients, and vibration spectrum characteristics, outputting the corrected spatiotemporal feature vectors.

[0254] The generation module is used to input the corrected spatiotemporal feature vector into the deep spatiotemporal convolutional network, extract the temporal evolution law through the bidirectional LSTM, combine it with the three-dimensional convolution kernel to mine the spatial correlation features, and synchronously integrate the GIS pipeline topology and benchmark line parameters to generate a three-dimensional heat map of leakage probability and diffusion trend;

[0255] The trigger module is used to calculate the dynamic risk index based on the evaluation matrix using the fuzzy analytic hierarchy process. When the index exceeds the threshold, it triggers a three-level linkage alarm of the sound and light alarm, mobile terminal, and emergency platform;

[0256] The verification module is used to dispatch mobile robots in response to three-level linkage alarms, navigate to the target area based on thermal map positioning and reference coordinates, remeasure gas concentration using a methane laser detector, and scan the pipeline surface temperature field using infrared thermal imaging. The verification results and correction data are then bound together using blockchain evidence storage technology to form an unalterable full-link evidence chain.

[0257] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A city gas pipeline micro-leakage early warning method, characterized in that: The method comprises: Step S1: Deploy a multi-parameter array of integrated pressure fluctuation, trace gas spectrum, acoustic wave or vibration, and pipeline deformation sensors along the gas pipeline, with node spacing less than or equal to 30 meters. Wide-area coverage is achieved through a LoRa or NB-IoT hybrid network. Select two fixed target points with a spacing greater than or equal to 50 meters and located on the same straight line segment, and calibrate their three-dimensional coordinates through precision positioning to form a reference line. Step S2: Based on the reference line, the pressure, gas concentration, and vibration data between the two points are collected in real time, and the environmental noise baseline, signal attenuation coefficient, and spatial consistency parameter are calculated to generate a dynamic correction value. The original signal is adaptively filtered to eliminate environmental noise through wavelet packet decomposition. The dynamic baseline is calibrated in combination with Kalman filtering. The correction value is used to perform spatial compensation for the pressure gradient, gas concentration gradient, and vibration spectrum characteristics, and the corrected spatiotemporal feature vector is output. Step S3: The corrected spatiotemporal feature vector is input into the deep spatiotemporal convolutional network. The temporal evolution law is extracted through the bidirectional LSTM, and the spatial correlation features are mined by combining the 3D convolution kernel. The GIS pipeline topology and the reference line parameters are simultaneously integrated to generate a 3D heat map of the leakage probability and diffusion trend. Step S4: Based on the evaluation matrix, the fuzzy analytic hierarchy process is used to calculate the dynamic risk index. When the index exceeds the threshold, a three-level linkage alarm is triggered, including the sound and light alarm, mobile terminal, and emergency platform.

2. The city gas pipeline micro-leakage early warning method according to claim 1 is characterized in that: Step S1: Deploy a multi-parameter array of integrated pressure fluctuation, trace gas spectrum, acoustic wave or vibration, and pipeline deformation sensors along the gas pipeline, with node spacing less than or equal to 30 meters, and achieve wide-area coverage through LoRa or NB-IoT hybrid networking; select two fixed target points with a spacing greater than or equal to 50 meters and located on the same straight line segment, and calibrate their three-dimensional coordinates through precise positioning to form a reference line, including: Sensor nodes are deployed along the gas pipeline at intervals of 30 meters or less. Each node integrates a pressure fluctuation sensor, a trace gas spectrum sensor, an acoustic or vibration sensor, and a pipeline deformation monitoring unit. LoRa wireless and NB-IoT cellular communication functions are pre-installed in each sensor node to achieve low-power wide-area communication coverage. Based on low-power wide-area communication coverage, a dual-channel communication link is established through LoRa modulation and NB-IoT cellular network. LoRa is used for short-distance networking between nodes, transmitting pressure fluctuations, gas concentration and vibration data in real time. NB-IoT is used for wide-area data backhaul, prioritizing the transmission of high-frequency deformation data from pipeline deformation monitoring units and synchronously uploading the location coordinates of each sensor node. This dual-channel link ensures that the upload delay of all sensor data is less than or equal to 5 seconds, forming a dynamically updated pipeline network spatial topology map. Based on the spatial topology map of the pipe network, two target points with a distance of 50 meters or more and located on the same straight line segment are selected from the deployed sensor nodes. The three-dimensional coordinates of the target points are calibrated using a GNSS positioning device, and the coordinate data is transmitted back to the control center in real time via the NB-IoT channel. Based on the returned coordinate data, the actual distance between the two points is measured using a laser rangefinder, and the deviation from the theoretical straight-line distance is calculated. If the deviation is less than or equal to 0.1%, the line connecting the two points is confirmed as a valid reference line. If the deviation is greater than 0.1%, the target points are reselected from the topology map and the process is iterated until the deviation is less than or equal to 0.1% and a reference line is formed.

3. The city gas pipeline micro-leakage early warning method according to claim 2 is characterized in that: Step S2: Based on the reference line, the pressure, gas concentration and vibration data between the two points are collected in real time, the ambient noise baseline, signal attenuation coefficient and spatial consistency parameter are calculated, and a dynamic correction value is generated; Adaptive filtering is performed on the original signal, and environmental noise is eliminated through wavelet packet decomposition. The dynamic baseline is calibrated in combination with Kalman filtering. The pressure gradient, gas concentration gradient and vibration spectrum characteristics are spatially compensated using correction values, and the corrected spatiotemporal feature vector is output, including: Based on the pressure fluctuation data between two points on the reference line, the sliding window average method is used to calculate the ambient noise baseline; combined with the gas concentration gradient, the signal attenuation coefficient is calculated based on the Fresnel reflection model; and the spatial consistency parameter is determined through covariance analysis. Based on the calculated environmental noise baseline, the original pressure signal is decomposed by wavelet packets to extract the effective signal in the 1-100 Hz frequency band; based on the signal attenuation coefficient, the dynamic baseline of the gas concentration data is calibrated using Kalman filtering; Based on the spatial consistency parameters, the dynamic correction values ​​are distributed to the pressure gradient, gas concentration gradient and vibration spectrum characteristics according to the weights; the weighted eigenvectors are normalized and the corrected spatiotemporal eigenvectors are output.

4. The city gas pipeline micro-leakage early warning method according to claim 3 is characterized in that: Step S3: Input the corrected spatiotemporal feature vector into the deep spatiotemporal convolutional network. The bidirectional LSTM is used to extract the temporal evolution law. The three-dimensional convolution kernel is combined to mine the spatial correlation features. The GIS pipeline topology and the reference line parameters are simultaneously integrated to generate a three-dimensional heat map of the leakage probability and diffusion trend, including: A deep spatiotemporal convolutional network model was constructed, consisting of three bidirectional LSTM layers and two three-dimensional convolutional layers. The model was supervised and trained using a leaky sample dataset. The loss function was a weighted combination of cross entropy and mean squared error. The network weights were optimized using a backpropagation algorithm to obtain the optimized DSTCN model weights. Based on the optimized DSTCN model weights, the modified spatiotemporal feature vectors are sliced ​​into time windows and input into the model, outputting a leakage probability sequence in the time dimension. The elbow and valve coordinates and the reference line spatial parameters in the GIS pipeline topology are simultaneously integrated, and a three-dimensional heat map is generated through inverse distance weighted interpolation, where the leakage probability values ​​are mapped to the range of 0-1. Based on the three-dimensional heat map, areas with a leakage probability greater than or equal to 0.7 are extracted as core leakage points. Combined with the gas diffusion model, the core leakage points are used as initial conditions to predict the diffusion path in the next 30 minutes.

5. The city gas pipeline micro-leakage early warning method according to claim 4 is characterized in that: Step S4: Based on the evaluation matrix, the dynamic risk index is calculated using the fuzzy analytic hierarchy process. When the index exceeds the threshold, a three-level linkage alarm is triggered, including the sound and light alarm, mobile terminal, and emergency platform, including: The original data of risk factors are mapped to the range of 0-1 using the range normalization method to obtain the normalized factor value; For the normalized factor values, a machine learning model trained based on historical data is used to generate triangular fuzzy number judgment results, and then a 12×12 fuzzy judgment matrix is ​​constructed; Based on the fuzzy judgment matrix, the logarithmic least squares method is used to solve the fuzzy weight vector of each risk factor. The triangular fuzzy numbers in the fuzzy weight vector are converted into precise values ​​through the defuzzification method to form an accurate weight vector. The precise weight vector is checked for consistency and the fuzzy consistency ratio is calculated. If the ratio is less than 0.1, it indicates that the consistency of the fuzzy judgment matrix is ​​good and the weight vector is valid. The leakage probability value of each grid in the three-dimensional heat map is weighted and aggregated with the normalized value and precise weight of the corresponding risk factor to calculate the dynamic risk index of each grid; when the calculated dynamic risk index exceeds the preset threshold, a three-level linkage alarm is triggered, including the sound and light alarm, mobile terminal APP and emergency platform.

6. The city gas pipeline micro-leakage early warning method according to claim 5, characterized in that: After step S4, the method further includes: Step S5: In response to the three-level linkage alarm, the mobile robot is dispatched to navigate to the target area based on the thermal map positioning and reference coordinates. The gas concentration is remeasured using a methane laser detector, and the pipeline surface temperature field is scanned using infrared thermal imaging. The verification results and correction data are then bound using blockchain evidence storage technology to form an unalterable full-link evidence chain.

7. A city gas pipeline micro-leakage early warning system, which implements the method according to any one of claims 1 to 6, characterized in that: include: The reference line module is used to deploy a multi-parameter array of integrated pressure fluctuation, trace gas spectrum, acoustic wave or vibration, and pipeline deformation sensors along the gas pipeline. The node spacing is less than or equal to 30 meters, and wide-area coverage is achieved through LoRa or NB-IoT hybrid networking. Two fixed target points are selected with a spacing of more than or equal to 50 meters and located on the same straight line. The three-dimensional coordinates are calibrated through precision positioning to form a reference line. The output module is used to collect pressure, gas concentration and vibration data between two points in real time based on the reference line, calculate the environmental noise baseline, signal attenuation coefficient and spatial consistency parameter, and generate dynamic correction values; Adaptive filtering is performed on the original signal to eliminate environmental noise through wavelet packet decomposition, and the dynamic baseline is calibrated in combination with Kalman filtering. The pressure gradient, gas concentration gradient and vibration spectrum characteristics are spatially compensated using correction values, and the corrected spatiotemporal feature vector is output; The generation module is used to input the corrected spatiotemporal feature vector into the deep spatiotemporal convolutional network, extract the temporal evolution law through the bidirectional LSTM, combine it with the three-dimensional convolution kernel to mine the spatial correlation features, and synchronously integrate the GIS pipeline topology and benchmark line parameters to generate a three-dimensional heat map of leakage probability and diffusion trend; The trigger module is used to calculate the dynamic risk index based on the evaluation matrix using the fuzzy hierarchical analysis method. When the index exceeds the threshold, it triggers the three-level linkage alarm of the sound and light alarm, mobile terminal and emergency platform.

8. A computing device, characterized in that include: one or more processors; A storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a program, which, when executed by a processor, implements the method according to any one of claims 1 to 6.

Images