Buried gas pipeline multi-leakage parameter identification and concentration distribution rapid prediction method

Abstract

The invention discloses a buried gas pipeline multi-leakage parameter identification and concentration distribution rapid prediction method, and belongs to the technical field of buried gas pipeline safety detection. The method comprises the following steps: firstly, based on a porous medium flow control equation and a component transport equation, establishing a leakage diffusion physical model by utilizing computational fluid mechanics and generating a multi-working-condition sample data set; training a neural network model by using the sample data, and constructing a rapid concentration prediction agent model; in combination with measured data of a monitoring area, an ensemble Kalman filtering algorithm is adopted, the agent model serves as a forward prediction operator, and collaborative inversion of parameters such as the soil type, the leakage direction and the flow is achieved through iterative updating; and finally, outputting complete underground gas concentration distribution in the monitoring area based on the leakage parameters obtained through identification. According to the method, rapid identification of multiple leakage parameters and accurate prediction of concentration distribution are realized under the condition of limited measuring points, and the method has remarkable disaster early warning and engineering application values.

Description

Technical Field

[0001] This invention belongs to the field of buried gas pipeline safety detection technology, specifically involving a method for reverse identification of leakage parameters and rapid prediction of concentration distribution in buried gas pipelines that combines computational fluid dynamics (CFD), deep learning neural network surrogate models, and ensemble Kalman filtering (EnKF). Background Technology

[0002] There are many methods for transmitting and distributing natural gas, among which laying underground pipelines is a common method used in most areas. Because underground pipelines often traverse long geological zones or pass through areas with frequent human activity, buried gas pipelines can age over time or develop undetected ruptures due to municipal construction, creating potential hazards for various gas leaks. Therefore, accurately and effectively identifying the gas concentration distribution and leakage parameters at the location of a leak in a buried gas pipeline is crucial for routine maintenance and emergency response.

[0003] Traditional gas leak detection methods mainly include fixed-location sensors and manual inspection. Fixed-location sensors have limited detection range and do not provide sufficient information on the gas concentration distribution at the leak location. Manual inspection is limited by the detection cycle, cannot reflect the gas concentration distribution at the leak location in a timely manner, and its detection accuracy is greatly affected by the inspector. CFD technology can obtain detailed information on the gas concentration distribution at the leak location, but the calculation results are greatly affected by the user's skill level and the calculation time is long, making it difficult to use for real-time prediction.

[0004] For the inversion of leakage parameters of buried gas pipelines, traditional data assimilation methods are based on Kalman filtering and specific CFD / neural network models. These methods are sensitive to the initial parameters of the model, which can easily lead to large estimation biases in practical applications. They lack good transferability and are greatly limited by computational resources. Secondly, conventional Kalman filtering methods are not good at handling high-dimensional nonlinear problems, which limits their ability to invert leakage parameters and makes it impossible to invert multiple leakage parameters simultaneously.

[0005] In summary, existing methods for detecting leaks in buried gas pipelines suffer from the following problems: (1) incomplete acquisition of concentration distribution at the leak location, and long acquisition time for high-precision concentration field information, which cannot meet the needs of emergency situations; (2) rapid concentration field prediction relies too much on trained samples and cannot cope with complex actual environments; (3) leakage parameter identification cannot achieve simultaneous inversion of multiple parameters, resulting in an inability to accurately assess the scope of the accident's impact. To address these problems, this invention provides a method that can simultaneously achieve high-precision diffusion prediction, real-time performance, and joint inversion of multiple leakage parameters, enabling the identification of parameters such as soil type, leakage direction, and leakage amount under limited monitoring point conditions, and reconstructing the concentration distribution across the entire space. Summary of the Invention

[0006] This invention provides a method for identifying leakage parameters and rapidly predicting the concentration field of buried gas pipelines. It achieves rapid reconstruction of the concentration field distribution while accurately inverting unknown leakage parameters (soil type, leakage direction, and leakage flow rate). The method constructs a multi-condition training set using a CFD model, rapidly calculates the underground gas concentration distribution using a neural network, and combines ensemble Kalman filtering (EnKF) to achieve collaborative inversion of multiple leakage parameters under limited monitoring point conditions. This enables real-time prediction of buried gas leakage source parameters and reconstruction of the concentration field.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] 1. A method for identifying multiple leakage parameters and rapidly predicting concentration distribution in buried gas pipelines, characterized by comprising the following steps:

[0009] S1: Construct a physical model of buried gas pipeline leakage. Based on the flow control equation and component transport equation of porous media, use computational fluid dynamics (CFD) to simulate the diffusion process of gas in porous soil media under various working conditions, and generate sample datasets of spatial location and concentration distribution under different soil types, leakage directions, leakage flow rates and burial depths.

[0010] S2: Preprocess the sample data, construct a neural network model, train the neural network using the sample data, establish a nonlinear mapping relationship between input parameters (soil type, leakage direction, leakage flow rate, burial depth and spatial location) and gas concentration at the monitoring point, and form a rapid calculation model for gas leakage concentration;

[0011] S3: Deploy gas concentration sensors in the target monitoring area to obtain real-time concentration data and set the initial state set of the leakage parameters to be identified (soil type, leakage direction, leakage flow rate);

[0012] S4: Using the Ensemble Kalman Filter (EnKF) algorithm, the fast calculation model is used as the prediction operator, and the real-time observed concentration data is used as the observation value. The state set is iteratively updated through prediction and analysis steps to reconstruct the leakage parameters of the buried gas pipeline.

[0013] S5: Input the leakage parameters identified in step S4 into the fast calculation model, and output the complete concentration field distribution in the soil after the gas leak within the monitoring area.

[0014] 2. The method for identifying multiple leakage parameters and rapidly predicting concentration distribution of buried gas pipelines according to claim 1, characterized in that, in step S1, the physical model of gas pipeline leakage is a porous media model, which simulates the diffusion process of gas in the soil by setting parameters such as soil porosity, viscous resistance coefficient, and inertial resistance coefficient; the multiple working conditions are working conditions formed by an orthogonal combination of four factors: soil type, leakage direction, leakage flow rate, and burial depth; the porous media flow control equation includes the mass conservation equation, momentum conservation equation, and component transport equation, wherein the momentum conservation equation introduces a source term to characterize the obstruction effect of the soil medium on the gas, and includes an inertial resistance term and a viscous resistance term, and the inertial resistance coefficient... The calculation formula is:

[0015]

[0016] coefficient of viscous resistance term The calculation formula is:

[0017]

[0018] in, The diameter of soil particles, Soil porosity.

[0019] 3. The method for identifying multiple leakage parameters and rapidly predicting concentration distribution of buried gas pipelines according to claim 1, characterized in that, in step S2, the neural network model is preferably a BP neural network; the training process includes: setting the number of hidden layers to [15, 20], using TanSigmoid as the activation function, selecting the Levenberg-Marquardt algorithm as the training algorithm, and adjusting the network weights and biases through gradient descent until the prediction error is lower than a preset threshold.

[0020] 4. The method for identifying multiple leakage parameters and rapidly predicting concentration distribution of buried gas pipelines according to claim 1, characterized in that, in step S4, the specific process of the ensemble Kalman filter (EnKF) algorithm is as follows:

[0021] Prediction Steps: Based on the analysis value from the previous time step, propagate forward using the fast calculation model to obtain the predicted state matrix for the current time step:

[0022]

[0023] Analysis steps: Calculate the Kalman gain By combining real-time observed concentration data with the predicted state matrix, the analytical value at the current moment is obtained:

[0024]

[0025] in, The state matrix typically consists of n rows and N columns, where n and N correspond to the number of parameters and the size of the set, respectively. It is a theoretical value corrected by observation data. M represents a nonlinear dynamic model used to propagate the state matrix over time. K is the Kalman gain, and H is a nonlinear observation operator that converts the state matrix into data for the corresponding observation points.

[0026] 5. The method according to claim 4, characterized in that a memory factor is introduced when updating the parameter set. To suppress filter divergence and improve dynamic tracking speed, the memory factor The preferred setting is 0.9.

[0027] 6. The method for identifying multiple leakage parameters and rapidly predicting concentration distribution of buried gas pipelines according to claim 1, characterized in that, in step S3, the number of gas concentration sensors is 7, and the number of state sets is set to 30, in order to balance computational cost and inversion accuracy.

[0028] The core of this invention lies in constructing an efficient prediction operator through a neural network and introducing an improved EnKF algorithm with a memory factor α to quickly predict the gas concentration distribution in the leak area across the entire field, thereby improving the stability of the multi-parameter joint inversion process. Under limited measurement point conditions, it enables the synchronous identification of multiple leakage parameters and the reconstruction of the concentration distribution across the entire space, making it suitable for early warning of leak risks in urban buried pipelines and providing a scientific basis for practical natural gas leak monitoring and early warning systems. Attached Figure Description

[0029] Figure 1 Physical model diagram of gas pipeline leakage

[0030] Figure 2 Flowcharts of different neural networks: (a) LSTM neural network; (b) BP neural network; (c) RBF neural network

[0031] Figure 3Comparison of neural network computations in four cases: (a) Case 1; (b) Case 2; (c) Case 3; (d) Case 4

[0032] Figure 4 Data assimilation method framework diagram

[0033] Figure 5 The impact of different parameters on the results in EnKF: (a) number of sets; (b) number of measurement points; (c) Re value; (d) value

[0034] Figure 6 Graph showing the variation of RMSE (Resolution Rate) with iteration number for soil type, leakage direction, flow rate, and concentration data.

[0035] Figure 7 Case 1 Concentration Distribution Comparison: (a) True Distribution; (b) Assimilation Distribution

[0036] Figure 8 Flowchart of a method for identifying multiple leakage parameters and rapidly predicting concentration distribution in buried gas pipelines Detailed Implementation

[0037] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention. After reading the teachings of this invention, those skilled in the art can make various modifications or alterations to the present invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0038] Step 1: Establish a CFD computational model

[0039] Taking a buried methane pipeline leak as an example, a cylindrical physical model of the leaking gas diffusion in the soil is first established, centered on the leak point with a radius of 10 meters and a burial depth of 0.9 meters (see [reference]). Figure 1 This model is used as the computational domain. Specifically, when a gas leak occurs, its diffusion behavior is mainly concentrated within a 10-meter radius. Changes in the concentration of leaked gas within this range have a direct impact on the environment and safety.

[0040] Based on the established physical model of leaked gas diffusion in soil, the boundary conditions of the model are set. The optimal boundary condition settings are shown in Table 1:

[0041] Table 1 Boundary Condition Settings

[0042]

[0043] Preferably, the ground surface, lateral surrounding boundaries, and other soil bottom boundaries are all uniformly set as pressure outlets to simulate the natural diffusion boundary in a semi-infinite space. The physical model of leaked gas diffusion in the soil simplifies the leak outlet to a circle with a diameter of 15 mm, satisfying the conditions of the orifice leak model. The methane velocity at the leak outlet is calculated according to equations (1), (2), and (3):

[0044] The methane velocity at the leak point is calculated using different methods depending on the critical pressure ratio (CPR). The CPR calculation formula is as follows:

[0045]

[0046] Where k is the adiabatic index of the gas; ρ is the critical pressure (Pa); p is the absolute pressure (Pa).

[0047] When the pressure ratio satisfies the condition of formula (2), the methane velocity is calculated using subsonic speed:

[0048]

[0049] When the pressure ratio satisfies the condition of formula (3), the methane velocity is calculated using the speed of sound:

[0050]

[0051] The leak velocity calculation process comprehensively considers parameters such as the gas flow coefficient at the leak point (preferably 1), the leak area (A), the amount of gaseous substance (M), and the adiabatic index (k), ensuring the physical consistency of the source input.

[0052] Based on the established physical model of leaked gas diffusion in soil, this invention introduces a porous media model for setting up the fluid domain, specifically for common urban soils. The viscous drag coefficient (D) and inertial drag coefficient (C) are defined to characterize their blocking effect on the gas. Preferably, the soil types selected are sandy soil, loam, and clay. The inertial drag coefficient formula and viscous drag coefficient formula involved in claim 2 are used to calculate the parameters for each of the three soil types. Specific soil parameters are shown in Table 2.

[0053] Table 2 Soil Parameters

[0054]

[0055] The established physical model of leaked gas diffusion in the soil, boundary conditions, and fluid domain parameters were imported into Fluent software. Steady-state calculations were performed using the standard k-ε turbulence model. The turbulent kinetic energy k and turbulent dissipation rate ε were calculated according to equations (4) and (5), respectively.

[0056]

[0057]

[0058] Where v is the fluid velocity in m / s, I is the turbulence intensity calculated from the Reynolds number of the leaking gas, and l is the turbulence intensity scale.

[0059] The pressure-velocity coupling method selected is the semi-implicit method (SIMPLE) of the pressure coupling equations. The numerical scheme mainly adopts the first-order upwind scheme, and the pressure interpolation method is the pressure staggered scheme (PRESTO!). The model uses standard wall functions, and the optimal value of the average yplus of the wall surface in the leak soil diffusion model is 14. The mesh type is hexahedral mesh, and local refinement is required near the leak point during mesh generation. After mesh independence verification, the optimal mesh size determined in this embodiment is 2.42 million elements. At this mesh density, the deviation between the simulated methane concentration field distribution and the calculation results of a higher density mesh (4.3 million elements) is within the preset tolerance range, thus optimizing computational resource consumption while ensuring simulation accuracy.

[0060] Step 2: Training the Neural Network

[0061] Based on the CFD calculation model established in Step 1, this invention sets different soil types, leakage size, leakage outlet direction, pipeline operating pressure, and pipeline burial depth as five key factors affecting the diffusion and distribution of methane in porous media. In order to cover a variety of actual working conditions, this embodiment establishes a multi-factor level table (see Table 3) and determines the parameter boundaries of the technical solution.

[0062] Table 3 Multifactor Level Table

[0063]

[0064] This embodiment designed a total of 27 working conditions using the orthogonal experimental method. The specific experimental training working conditions are shown in the attached table.

[0065] Steady-state simulations were performed on the 27 experimental conditions using Fluent software to obtain the methane concentration distribution at each measuring point in space under each condition. In extracting the characteristic data of the methane concentration distribution, this embodiment employs a non-equidistant spatial sampling strategy: vertically, a monitoring point is preferably set every 0.1 m from the leak point upwards to the soil surface; horizontally, a monitoring point is preferably set every 0.5 m extending from the leak point towards the edge of the model. This method yields the methane concentration at each point in space under each condition, forming a basic sample database.

[0066] Boundary conditions (including soil type, leak diameter, leak direction, burial depth, and flow rate) from the basic sample database were selected as input parameters for neural network training. The methane concentration at the monitoring point was preferred as the output value of the corresponding neural network. Twenty-seven operating conditions from the basic sample database were used as the training set, and new operating conditions were used to form a test set to test the trained neural network, ensuring the model's generalization ability.

[0067] This invention evaluates the performance of three neural networks suitable for concentration prediction: Long Short-Term Memory (LSTM), Back Propagation (BP), and Radial Basis Function (RBF) Network. To eliminate the impact of concentration value range on accuracy, all three neural networks undergo logarithmic preprocessing of the concentration data during training. Optimal hyperparameter settings were implemented for each of the three different neural networks (see Table 4).

[0068] Table 4 Hyperparameter Settings

[0069]

[0070] This invention utilizes root mean square error (RMSE), relative root mean square error (RRMSE), mean absolute error (MAE), mean bias error (MBE), and coefficient of determination (R²). 2 Five performance metrics were used to quantitatively evaluate three neural networks trained on a basic sample database, specifically by comparing four different test set cases. See details... Figure 3 The calculation comparison chart and the quantitative comparison results shown in Table 5 demonstrate that, in the buried gas leak prediction involved in this invention, the BP neural network exhibits significantly better goodness of fit between the true and predicted values ​​on the test set than the RBF neural network and the LSTM neural network. The BP neural network maintains a fast computation speed while achieving good prediction accuracy, meeting the requirements for real-time prediction. Therefore, this invention preferably employs the BP neural network as a rapid calculation method for predicting the concentration distribution of buried gas leaks.

[0071] Table 5 Comparison of Neural Network Training Results

[0072]

[0073] Step 3: Parameter assimilation based on buried sensor concentration data

[0074] The BP neural network trained by deep learning in step two is used as a fast gas diffusion prediction model and embedded in the ensemble Kalman filter (EnKF) framework. This replaces the high computational cost of traditional numerical diffusion models during the iterative process, enabling real-time correction of unknown leakage parameters based on sensor measurement data. The specific implementation process is as follows:

[0075] First establish The state matrix contains n rows and N columns, where n is the number of parameters and N is the set size. It includes: physical parameters to be identified (soil type, leakage direction, mass flow rate) and the predicted methane concentration for each monitoring point.

[0076] Prediction phase: [This refers to the already established...] The state matrix is ​​predicted using a BP neural network model based on formula (6), according to the parameter states of the previous time step. Propagation to the current moment, prior estimate of the generated set

[0077]

[0078] Where M(∙) represents the BP neural network prediction model, used to propagate the state matrix over time.

[0079] Analysis phase: Prior estimates Substitute into formula (7) and combine with sensor observation data The prior estimate is calculated. Updated correction values (Corrected parameters) Through an iterative cycle of prediction and analysis, the leakage parameters gradually converge to the true value during multiple assimilation processes, thereby achieving dynamic inversion of unknown leakage parameters and prediction of methane concentration.

[0080]

[0081] Among them, H( ) is a nonlinear observation operator used to map the state matrix to the observation point, and K is the Kalman gain.

[0082] Through an iterative cycle of prediction and analysis, the leakage parameters gradually converge to the true value during multiple assimilation processes, thereby achieving dynamic inversion of unknown leakage parameters and prediction of methane concentration.

[0083] Error covariance and gain calculation: This step avoids explicitly constructing a high-dimensional error covariance matrix by using ensemble covariance, thereby improving computational efficiency and enhancing the model's adaptability to nonlinear systems. Prior estimates are then used. Substitute into formula (8) to calculate the prediction error covariance matrix :

[0084]

[0085] It is the mean of each row. The Kalman gain K is then calculated using formula (9).

[0086]

[0087] In the formula, R e This represents the covariance matrix of the observation set. It is derived from the observation error vector. (10) According to formula (11):

[0088]

[0089]

[0090] in, It is the amount of disturbance in the observed data.

[0091] Parameter Update: After each assimilation step, the system updates the parameters in the state vector. First, the latest ensemble average parameter is calculated according to equation (12).

[0092]

[0093] in It is the correction parameter for the latest iteration step of the corresponding set i.

[0094] Then, a memory factor is introduced when updating the parameter set. Add noise according to equations (13) and (14) and update the parameter set according to equation (15):

[0095]

[0096]

[0097]

[0098] in, This represents the noise added to the new set parameter. The parameter (between 0 and 1) represents the degree to which the effects of previous states (i.e., "memories") are retained, and s is the latest set of parameters. standard deviation These are random numbers that follow a normal distribution N(0,1).

[0099] EnKF assimilation of sensor data is affected by parameter uncertainties, among which the number of ensembles, the number of measurement points, the measurement error value, and the model memory parameter value have a significant impact on the assimilation results. The optimal parameter values ​​for the suitable model are determined by designing and comparing different sets of parameters.

[0100] Based on the empirical method that the estimated value of the unknown parameters should not be less than ten times the number of unknown parameters, the initial set size was chosen to be 20, 30, and 40 to analyze its impact on assimilation. Figure 5 (a) shows that the assimilation effect is relatively optimal with a set size of 30, and the relative error range is within 20%.

[0101] The number of measurement points was selected as 5, 6, 7, and 8 to analyze their impact on assimilation. Figure 5 (b) shows that when the number of measuring points is 7 and 8, the assimilation results of soil parameters are better. At this time, the relative error range is within 20%. The assimilation effect of 7 and 8 measuring points is not much different. It can be concluded that the assimilation effect has reached the optimal effect when the number of measuring points reaches 7. After that, the effect remains unchanged when the number of measuring points is increased.

[0102] The measurement error Re values ​​of 0.05, 0.1, 0.2, and 0.3 were selected to analyze their impact on assimilation. Figure 5 (c) When the diagonal elements are 0.05, 0.1, 0.2, and 0.3, the soil types correspond to 2.83, 3.01, 2.96, and 2.86, respectively; the leakage directions correspond to 1.21, 1.17, 1.15, and 1.25, respectively; and the leakage flows correspond to 0.00172, 0.00169, 0.00171, and 0.00164, respectively. This is compared with the actual parameters of soil type 3, leakage direction 1, and leakage flow of 0.0017. It is determined that when the diagonal element is 0.1, the assimilation results of the soil parameters are better, with a relative error range within 20%.

[0103] choose The values ​​of 0.7, 0.8, 0.88, 0.92, 0.94, and 0.96 were used to analyze their impact on assimilation. Figure 5 (d) shows When the values ​​are as described above, the corresponding soil types are 2.97, 2.5, 3.86, 3.01, 2.93, 2.9, and 2.87, respectively; the corresponding leakage directions are 0, 0, 0, 1.16, 1.18, 1.18, and 1.188, respectively; and the corresponding leakage flow rates are 0, 0, 0.00161, 0.00168, 0.00168, 0.00157, and 0.0016. Compare these values ​​with the actual parameters of soil type 3, leakage direction 1, and leakage flow rate of 0.0017. Judgment. When the value is 0.9, the assimilation results of soil parameters are good, with a relative error range within 20%.

[0104] by Figure 6Using the assimilation parameter iteration count diagram of the test set as shown as an example, the process of identifying multiple leakage parameters is illustrated. A state matrix is ​​constructed containing the multiple parameters to be identified (soil type, leakage direction, leakage flow rate) and the corresponding concentrations at the measuring points. The set size is determined to be 30, the number of measuring points to be 7, and the Re value to be 0.1. After setting the optimal EnKF parameter to 0.9, the optimal gas diffusion prediction model trained in step 2 and the observed concentration data are iteratively updated multiple times within the ensemble Kalman filter framework. When the leakage parameters tend to stabilize, the assimilation process is considered to have converged, and the final leakage parameters are output. As can be seen from the figure, with the increase of the number of iterations, the soil type parameter, leakage direction parameter, and leakage flow rate gradually approach the true values ​​and stabilize near them, and the RMSE decreases rapidly with the increase of the number of iterations.

[0105] by Figure 7 The concentration distribution comparison chart shown is used as an example to illustrate the concentration distribution prediction process of the present invention. Figure 7 The actual soil parameters were set as follows: soil type 3, leakage direction 2, and flow rate 0.0034. When an unknown leakage of methane gas occurred, the neural network quickly predicted the gas using the selected optimal assimilation parameters. Based on the observed partial concentration data, it performed rapid iterative updates within an ensemble Kalman filter framework. When the leakage parameters to be identified stabilized after iteration, the rapid concentration prediction model (a BP neural network trained with deep learning) output the complete underground gas concentration distribution. Figure 7 In the case shown, the soil type assimilated by the ensemble Kalman filter framework converged to 3.09 (true value 3), the leakage direction converged to 2.2 (true value 2), and the flow rate converged to 0.0035 kg / s (true value 0.0034 kg / s). The relative errors of the parameters obtained after assimilation were all kept within 20%, and the concentration prediction showed good consistency.

[0106] The results show that the method of the present invention can achieve rapid identification of underground gas leakage parameters and prediction of concentration distribution under limited measurement points, and has high accuracy and engineering application value.

[0107] Appendix:

[0108] Experimental Training Condition Table

[0109]

Claims

1. A method for identifying multiple leakage parameters and rapidly predicting concentration distribution in buried gas pipelines, characterized in that, Includes the following steps: S1: Construct a physical model of buried gas pipeline leakage. Based on the flow control equation and component transport equation of porous media, use computational fluid dynamics (CFD) to simulate the diffusion process of gas in porous soil media under various working conditions, and generate sample datasets of spatial location and concentration distribution under different soil types, leakage directions, leakage flow rates and burial depths. S2: Preprocess the sample data, construct a neural network model, train the neural network using the sample data, establish a nonlinear mapping relationship between input parameters (soil type, leakage direction, leakage flow rate, burial depth and spatial location) and gas concentration at the monitoring point, and form a rapid calculation model for gas leakage concentration; S3: Deploy gas concentration sensors in the target monitoring area to obtain real-time concentration data and set the initial state set of the leakage parameters to be identified (soil type, leakage direction, leakage flow rate); S4: The Ensemble Kalman Filter (EnKF) algorithm is adopted, and the fast calculation model is used as the prediction operator to replace the numerical diffusion model. The real-time observed concentration data is used as the observation value. The state set is iteratively updated through prediction and analysis steps to reconstruct the multiple leakage parameters of the buried gas pipeline. S5: Input the leakage parameters identified in step S4 into the fast calculation model, and output the complete concentration field distribution in the soil after the gas leak within the monitoring area.

2. The method for identifying multiple leakage parameters and rapidly predicting concentration distribution of buried gas pipelines according to claim 1, characterized in that, In step S1, the physical model for gas pipeline leakage is a porous media model. By setting parameters such as soil porosity, viscous drag coefficient, and inertial drag coefficient, the diffusion process of gas in the soil is simulated. The various operating conditions are formed by an orthogonal combination of four factors: soil type, leakage direction, leakage flow rate, and burial depth. The porous media flow control equations include the mass conservation equation, momentum conservation equation, and component transport equation. The momentum conservation equation introduces a source term to characterize the obstruction effect of the soil medium on the gas, and includes inertial drag term and viscous drag term. The inertial drag coefficient... The calculation formula is: coefficient of viscous resistance term The calculation formula is: in, The diameter of soil particles, Soil porosity.

3. The method for identifying multiple leakage parameters and rapidly predicting concentration distribution of buried gas pipelines according to claim 1, characterized in that, In step S2, the neural network model is preferably a BP neural network; the training process includes: setting the number of hidden layers to [15, 20], using TanSigmoid as the activation function, selecting the Levenberg-Marquardt algorithm as the training algorithm, and adjusting the network weights and biases through gradient descent until the prediction error is lower than a preset threshold.

4. The method for identifying multiple leakage parameters and rapidly predicting concentration distribution of buried gas pipelines according to claim 1, characterized in that, In step S4, the specific process of the ensemble Kalman filter (EnKF) algorithm is as follows: Prediction Steps: Based on the analysis value from the previous time step, propagate forward using the fast calculation model to obtain the predicted state matrix for the current time step: Analysis steps: Calculate the Kalman gain By combining real-time observed concentration data with the predicted state matrix, the analytical value at the current moment is obtained: in, The state matrix typically consists of n rows and N columns, where n and N correspond to the number of parameters and the size of the set, respectively. It is a theoretical value corrected by observation data. M represents a nonlinear dynamic model used to propagate the state matrix over time. K is the Kalman gain, and H is a nonlinear observation operator that converts the state matrix into data for the corresponding observation points.

5. The method according to claim 4, characterized in that, Introducing a memory factor when updating the parameter set To suppress filter divergence and improve dynamic tracking speed, the memory factor The preferred setting is 0.

9.

6. The method for identifying multiple leakage parameters and rapidly predicting concentration distribution of buried gas pipelines according to claim 1, characterized in that, In step S3, the number of gas concentration sensors is 7, and the number of state sets is set to 30, in order to balance computational cost and inversion accuracy.

Images