Gas negative pressure extraction pipeline leakage evaluation and intelligent disposal method

Abstract

The application provides a gas negative pressure extraction pipeline leakage evaluation and intelligent treatment method, comprising the following steps: constructing a gas negative pressure extraction pipeline leakage positioning evaluation model; wherein the gas negative pressure extraction pipeline leakage positioning evaluation model is built based on a graph neural network and a time series analysis basic model; and constructing an intelligent valve opening degree control model based on a deep reinforcement learning technology. The gas negative pressure extraction pipeline leakage evaluation and intelligent treatment method can accurately predict the leakage section of a gas extraction pipeline and intelligently and quickly treat the leakage section in the process of a leakage accident of a coal mine gas extraction pipeline system, thereby greatly reducing the economic losses caused by a traditional gas leakage treatment process, ensuring the safe operation of the gas extraction system, preventing the accident from further evolving and upgrading, and providing important technical method support for the development of gas extraction pipeline leakage emergency rescue work.

Description

Technical Field

[0001] This invention belongs to the field of gas pipeline technology, and in particular relates to a method for assessing and intelligently handling leaks in gas negative pressure extraction pipelines. Background Technology

[0002] Gas accidents are one of the most serious threats to coal mine safety in my country. By using coal mine gas drainage systems, high-gas mines can be operated under low-gas conditions, avoiding gas accidents. The extracted gas can also create considerable economic value. However, gas drainage pipeline leaks occur frequently due to aging, corrosion, or external impacts. When a leak occurs, the pressure difference between the inside and outside of the pipeline allows outside air to enter, diluting the gas inside, causing a loss of negative pressure, and even leading to the failure of the entire drainage system, severely impacting drainage efficiency. This not only wastes the energy of the drainage pumps and increases the cost of gas extraction, but also, if the leak worsens and the drainage network ruptures, gas may gush out, potentially triggering a gas explosion. Therefore, gas drainage pipeline leaks seriously threaten the normal operation of underground gas transport systems and the safety of mine production. In emergency response to underground gas leaks, the efficient screening and utilization of low-density monitoring data from large-scale, complex gas extraction systems in coal mines to achieve rapid, accurate location, early warning, and intelligent handling of leak sources is crucial for the safe operation of coal mine gas extraction systems. Summary of the Invention

[0003] In view of this, the present invention aims to propose a method for assessing and intelligently handling leaks in gas negative pressure extraction pipelines, so as to achieve rapid and accurate location, early warning and intelligent handling of leak sources.

[0004] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0005] A method for assessing and intelligently handling leaks in a gas negative pressure extraction pipeline, comprising:

[0006] A leakage location assessment model for a gas negative pressure extraction pipeline is constructed; wherein, the leakage location assessment model for a gas negative pressure extraction pipeline is built based on a graph neural network and time series analysis basic model;

[0007] A smart valve opening control model based on deep reinforcement learning technology is constructed and trained to obtain a trained smart valve opening control model. The action of the smart valve opening control model is set as valve opening, and the state is the gas concentration and gas flow rate in the pipeline. The highest excitation state is the highest gas concentration within the specification requirements and the lowest gas flow rate within the safety range.

[0008] A dynamically updated database is constructed based on historical accident and disaster data and real-time monitoring data of basic parameters of gas extraction pipelines;

[0009] The leakage location assessment model of the gas negative pressure extraction pipeline is trained and verified based on the dynamically updated database to obtain the gas negative pressure extraction pipeline leakage location assessment model after training.

[0010] Acquire real-time monitoring data of basic parameters of the gas extraction pipeline, and input the real-time monitoring data into the gas negative pressure extraction pipeline leakage location assessment model after training;

[0011] Obtain the pressure distribution, gas concentration distribution and leak location prediction results of the gas negative pressure extraction pipeline leak location assessment model after training is completed;

[0012] The pressure distribution, gas concentration distribution, and leak location prediction results within the gas extraction pipeline are input into the trained intelligent valve opening control model to achieve intelligent control of the valve opening within the gas extraction pipeline.

[0013] Furthermore, the construction of a gas negative pressure extraction pipeline leakage location assessment model; wherein, the gas negative pressure extraction pipeline leakage location assessment model is built based on a graph neural network and time series analysis basic model, including:

[0014] Construct a topology map of adjacent locations of the monitors within the extraction pipeline to carry the data monitored by the monitors;

[0015] The preset leakage orifice intervals formed between the various monitors are combined to form a leakage interval matrix, and the leakage interval matrix is ​​assigned values ​​according to different leakage states as predicted labels for leakage locations; wherein, the different leakage states include leakage and no leakage.

[0016] The topology graph is processed by graph neural network operations and used as a time series node in the basic time series analysis model. The classification prediction mode in the basic time series analysis model is used to process the prediction task.

[0017] An evaluation index is set for the leakage location assessment model of the gas negative pressure extraction pipeline, and weights are added to different error scenarios using the evaluation index. The most severe error is assigned the largest weight value, while for prediction errors of the leakage range, the weight is increased proportionally according to the distance between the predicted range and the actual range; the greater the distance, the higher the weight. The formula for the evaluation index is as follows:

[0018]

[0019] In the formula, n is the number of intervals with prediction errors, and W iLet be the error weight value for the i-th error interval, and N be the total number of predicted samples.

[0020] Furthermore, a smart valve opening control model based on deep reinforcement learning technology is constructed, and the model is trained to obtain a trained smart valve opening control model. The smart valve opening control model's actions are set as valve opening, its states are the gas concentration and gas flow velocity within the pipeline, and its highest excitation state is the highest gas concentration within the specified range and the lowest gas flow velocity within the safe range, including:

[0021] Flow field simulation of gas negative pressure extraction pipeline based on CFD numerical simulation technology;

[0022] A smart valve opening control model is constructed based on the DQN deep reinforcement learning model;

[0023] The intelligent valve opening control model is trained based on the flow field simulation of the gas negative pressure extraction pipeline.

[0024] The trained intelligent valve opening control model is obtained.

[0025] Furthermore, the dynamically updated database constructed based on historical accident and disaster data and real-time monitoring data of basic parameters of gas extraction pipelines includes:

[0026] A dynamically updated database is obtained by fusing historical accident and disaster data with real-time monitoring data of basic parameters of gas extraction pipelines using multi-source heterogeneous data fusion technology.

[0027] Furthermore, the valve opening degree is divided into 11 opening degree actions in increments of 10%.

[0028] Furthermore, the process of training and validating the gas negative pressure extraction pipeline leakage location assessment model based on the dynamically updated database to obtain the trained gas negative pressure extraction pipeline leakage location assessment model includes:

[0029] Data is obtained from the dynamically updated database, and during the training process, the data is randomly divided into training and test sets according to a certain ratio for training the gas negative pressure extraction pipeline leakage location and evaluation model.

[0030] During training, the initial learning rate was set to 0.01, and cosine annealing was used to automatically adjust the learning rate; the number of iterations was set to 100, and the cross-entropy loss function was used.

[0031] A portion of the data in the training set is used as a validation set, and the validation set is used to validate the gas negative pressure extraction pipeline leakage location assessment model.

[0032] Furthermore, after inputting the pressure distribution, gas concentration distribution, and leak location prediction results within the extraction pipeline into the trained intelligent valve opening control model to achieve intelligent control of the valve opening within the gas extraction pipeline, the method further includes:

[0033] The intelligent valve opening control model is connected to the valve via Internet of Things (IoT) technology to achieve fully automatic intelligent valve opening control in case of an accident.

[0034] Compared with existing technologies, the gas negative pressure extraction pipeline leakage assessment and intelligent handling method described in this invention has the following advantages:

[0035] (1) The gas negative pressure extraction pipeline leakage assessment and intelligent handling method described in this invention can significantly reduce the economic losses caused by traditional gas leakage handling process by accurately predicting and intelligently and quickly handling the leakage section of the gas extraction pipeline during the leakage accident of the coal mine gas extraction pipeline system, and ensure the safe operation of the gas extraction system, prevent the accident from further evolving and escalating, and provide important technical support for the emergency rescue work of gas extraction pipeline leakage.

[0036] (2) The gas negative pressure extraction pipeline leakage assessment and intelligent handling method described in this invention uses the gas negative pressure extraction pipeline valve as the handling terminal, and comprehensively uses multi-source heterogeneous data fusion technology and deep learning technology to achieve rapid and accurate location assessment of gas negative pressure extraction pipeline leakage. On this basis, based on deep reinforcement learning technology, and taking into account economic and safety factors, it achieves accurate matching of handling valves after gas negative pressure extraction pipeline leakage and intelligent control of valve opening, providing the most economical and safe solution for gas negative pressure extraction pipeline leakage accidents, and serving the prediction, early warning and safety prevention and control of gas extraction accidents.

[0037] (3) The gas negative pressure extraction pipeline leakage assessment and intelligent handling method described in this invention can achieve rapid early warning of gas leakage accidents and accurately locate the leakage section. On this basis, by using deep reinforcement learning, the valve opening can be intelligently controlled while ensuring economy and safety. This can significantly reduce the economic losses caused by directly closing valves in the traditional gas leakage handling process and ensure pipeline safety at the accident site. Through the intelligent opening and closing of valves, emergency rescue personnel can avoid directly entering the gas leakage section to operate valves, reducing casualties, and providing technical support for the intelligent and unmanned emergency handling of gas extraction system accidents. Attached Figure Description

[0038] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0039] Figure 1 This is a schematic diagram of a method for assessing and intelligently handling leaks in a gas negative pressure extraction pipeline, as described in an embodiment of the present invention.

[0040] Figure 2 This is a schematic diagram of the handling process in the gas negative pressure extraction pipeline leakage assessment and intelligent handling method described in an embodiment of the present invention;

[0041] Figure 3 This is a schematic diagram of the structure of the gas negative pressure extraction pipeline leakage location assessment model in the gas negative pressure extraction pipeline leakage assessment and intelligent handling method described in an embodiment of the present invention;

[0042] Figure 4 This is a schematic diagram showing the location of the monitor in a gas negative pressure extraction pipeline leakage assessment and intelligent handling method according to an embodiment of the present invention;

[0043] Figure 5 This is a topological diagram of the adjacent positions of the monitor in a gas negative pressure extraction pipeline leakage assessment and intelligent handling method according to an embodiment of the present invention. Detailed Implementation

[0044] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0045] A method for assessing and intelligently handling leaks in gas negative pressure extraction pipelines, such as Figures 1 to 5 As shown, the specific steps include the following:

[0046] Step 101: Construct a leakage location assessment model for gas negative pressure extraction pipeline; wherein, the leakage location assessment model for gas negative pressure extraction pipeline is built based on a graph neural network and time series analysis basic model.

[0047] Specifically, the above-mentioned model for locating and assessing leaks in gas negative pressure extraction pipelines can be constructed by referring to the following steps:

[0048] Step 1011: Construct a topology map of adjacent locations of the monitors in the extraction pipeline to carry the data monitored by the monitors.

[0049] Step 1012: Form a leakage interval matrix by assembling the preset leakage orifice intervals formed between each of the monitors, and assign values ​​to the leakage interval matrix according to different leakage states as prediction labels for leakage locations; wherein, the different leakage states include leakage and no leakage.

[0050] Step 1013: After processing the topology graph through graph neural network operations, it is used as a time-series node in the time-series analysis basic model, and the classification prediction mode in the time-series analysis basic model is used for prediction tasks. In this way, the location relationship information of the monitoring data can be given to the time-series analysis basic model to achieve a spatiotemporal correspondence.

[0051] Step 1014: Set evaluation indicators for the gas negative pressure extraction pipeline leakage location assessment model, and use the evaluation indicators to add weights to different error scenarios; wherein, the most serious error is given the largest weight value, and for the prediction error of the leakage range, the weight is increased proportionally according to the distance between the predicted range and the actual range, the greater the distance, the higher the weight. The formula for the evaluation indicators is as follows:

[0052]

[0053] In the formula, n is the number of intervals with prediction errors, and W i Let be the error weight value of the i-th error interval, and N be the total number of prediction samples (i.e., the total number of intervals to be predicted).

[0054] In the actual model building process, a leak location assessment model for gas negative pressure extraction pipelines can be constructed first, based on the fusion of graph neural network (GCN) and time series analysis model (TimesNet). Figure 3 Technical roadmap for the leakage location assessment model of the gas negative pressure extraction pipeline.

[0055] When a gas extraction pipeline leak occurs, the gas negative pressure extraction pipeline leak location assessment model can quickly predict the pressure distribution, gas concentration distribution, and leak location within the extraction pipeline by using real-time monitoring data of basic parameters of the underground gas extraction pipeline, such as relative pressure, mixed flow rate, concentration, pure flow rate, carbon monoxide, and temperature, monitored by monitoring sensors such as pressure monitors, flow monitors, gas concentration monitors, carbon monoxide concentration monitors, and temperature monitors.

[0056] During data processing, a topology graph of adjacent locations of monitors such as pressure, flow, concentration, and temperature is first constructed. Data from each monitor is then placed into nodes of the topology graph, and data from the same time point are placed into nodes in the graph (e.g.,...). Figure 5 As shown in the figure, the above is the data feature processing procedure.

[0057] Each preset leak orifice interval is used to form a leak interval matrix. The matrix is ​​then assigned a value according to the different leak states (for example, if there is no leak in any of the three leak orifice intervals A, B, and C, the value of the leak orifice interval matrix is ​​[0, 0, 0]; if a leak occurs in interval A, the value of the leak orifice interval matrix is ​​[1, 0, 0]). This matrix serves as a predictive label for the leak location. The labeling method for the pipeline pressure distribution, gas concentration distribution, etc., is the same as that for the leak location label. The above describes the processing procedure for the predictive label.

[0058] During the modeling process, each set of topology graphs is processed by GCN and used as a temporal node in the TimesNet model. The classification prediction mode in TimesNet is then used to perform the prediction task.

[0059] First, the computation process of GCN is as follows: The core of GCN is based on the spectral decomposition of the Laplacian matrix. The Laplacian matrix L of the graph is defined as:

[0060] L=D -1 / 2 (DA)D -1 / 2 =I N -D -1 / 2 AD -1 / 2 ;

[0061] Where I N It is an N×N identity matrix. The degree matrix is ​​defined as D. ii =∑ j A ij .

[0062] The eigenvalue decomposition of matrix L can be described as L = UΛU T , where Λ=diag([λ0,λ1,...,λ N-1 ]), λ i These are the eigenvalues ​​of L. U is a matrix, where each column is an L. The convolution operation of the graph is performed in the Fourier domain and is defined as the product of the input signal x and the filter g:

[0063] g×x=U((U T g)⊙(U T x))=U gθ (Λ)U T x;

[0064] In other words, the input x is transformed into U. T x, the U T x falls within the range [u0, u2, ..., u n-1 In the spectral space U based on ], where ⊙ represents the Huffman product, g θ (Λ)=U T g = diag(θ), where θ ∈ R NTypically, the convolution kernel g θ The computational cost of (Λ) is very high (the computational complexity is O(N). 3 Therefore, some approximation methods have been proposed, such as using only first-order polynomials to simplify the calculation of g*x:

[0065] g×x≈θ(I N +D -1 / 2 AD -1 / 2 )x;

[0066] The above formula is used for approximate calculation. Given: in (Renormalization). Therefore, the l-th layer H (i) The output can be written as:

[0067]

[0068] Where σ is the activation function, such as sigmoid. Let... in and C (l-1) F is the size of the output of the (l-1)th layer. (l -1) This represents the size of the feature vector in each dimension. Then, the above equation can be rewritten as:

[0069]

[0070] Assuming the total number of detectors is R and the number of historical time slots is K, the feature matrix of the graph is X∈RK×R, and the adjacency matrix is ​​A∈RR×R. This module uses two layers of GCN, with a batch normalization layer added between the GCNs. The behavior of the GCN can be represented as follows:

[0071]

[0072] Where X is the characteristic matrix; It is an adjacency matrix with self-loops added. The degree matrix is ​​represented by ; ReLU is the activation function of the network, and W and b are the network parameters.

[0073] After adding a BN layer, the above formula is changed to:

[0074]

[0075] Here, BN is the batch normalization function, and σ is the nonlinear activation function. GCN and BN are grouped together to form functional units, and then two of these units are piled together to extract spatial features. The module output has a size of K×R.

[0076] Secondly, the construction principle of the TimesNet model is as follows:

[0077] The TimesNet model decomposes complex time series changes into different periods through a modular structure (TimesBlock), and achieves unified modeling of intra-period and inter-period changes by transforming the original one-dimensional time series into a two-dimensional space.

[0078] Specifically: First, the frequency components of the time-series data are obtained through Fast Fourier Transform, the Top N low-frequency components are retained, high-frequency noise is removed, and finally the frequency and period of the sub-waveforms after the time-series data decomposition are obtained.

[0079]

[0080] Simplified representation:

[0081] A,{f1,…,f k},{p1,…,p k} = Period(X ID );

[0082] Then, the original one-dimensional time series is folded based on the selected period, a process that can be formalized as follows:

[0083]

[0084] Padding(.) adds zeros to the end of the sequence so that the sequence length is divisible by integers.

[0085] For the two-dimensional vector described above, each column and each row corresponds to an adjacent time and an adjacent period, respectively, and adjacent time and period often imply similar temporal changes. Therefore, the two-dimensional tensor exhibits two-dimensional locality, making it easy to capture information through 2D convolution.

[0086] TimesNet consists of multiple stacked TimesBlocks. The processing steps for each TimesBlock are as follows: (1) transforming the time series data from one dimension to two dimensions; (2) extracting the two-dimensional time series change representation; (3) transforming the two dimensions back to one dimension; and (4) adaptive fusion.

[0087] In the actual process of leak location prediction, since the leak location prediction label used in the experiment is a multi-class label with location attributes and includes both leak and non-leak scenarios, traditional model evaluation methods cannot well describe the prediction effect of the model.

[0088] For example, when a pipeline leaks, if the model fails to detect the leak and predicts that there is no leak (i.e., label 0), or if the pipeline has no leak but the model predicts the leak point, the error in this case is the most serious, much more serious than the model predicting a leak but not the location of the leak.

[0089] For example, suppose the actual leakage area appears in such a location as... Figure 4 If the model predicts a leak but incorrectly predicts the leak interval, then predicting the leak interval as interval B will have a much smaller impact than predicting the leak interval as interval L.

[0090] Therefore, based on the special characteristics of the predicted leak location labels described above, the gas negative pressure extraction pipeline leak location assessment model described in this embodiment also defines a model prediction assessment index for multi-class labels with location attributes during the deep learning calculation process. This index adds weights to different error scenarios, and adds the largest weight value to the most serious error (i.e., the prediction error of whether a leak has occurred, this error weight value is 1). For the prediction error of the leak interval, the weight is increased proportionally according to the distance between the predicted interval and the actual interval. The farther the distance, the higher the weight.

[0091] For example, if the actual leakage interval A is predicted as interval B, and the distance between interval B and interval A is one, the weight of this error is 0.09; then if the actual leakage interval A is predicted as interval L, and the distance between interval L and interval A is 11, the weight of this error is 11 times 0.09, that is, 0.99.

[0092] Step 102: Construct an intelligent valve opening control model based on deep reinforcement learning technology, and train the intelligent valve opening control model to obtain the trained intelligent valve opening control model; wherein, the action of the intelligent valve opening control model is set as valve opening, the state is the gas concentration and gas flow rate in the pipeline, and the highest excitation state is the highest gas concentration within the standard requirement range and the lowest gas flow rate within the safety range.

[0093] Specifically, the above-mentioned intelligent valve opening control model based on deep reinforcement learning technology is constructed, and the intelligent valve opening control model is trained to obtain the trained intelligent valve opening control model. The construction and training can be carried out by referring to the following steps:

[0094] Step 1021: Simulate the flow field of the gas negative pressure extraction pipeline based on CFD numerical simulation technology.

[0095] Step 1022: Construct an intelligent valve opening control model based on the DQN deep reinforcement learning model.

[0096] Step 1023: Train the intelligent valve opening control model based on the flow field simulation of the gas negative pressure extraction pipeline.

[0097] Step 1024: Obtain the intelligent valve opening control model after training.

[0098] In practical applications, to address the issue of intelligent handling of leaks in gas negative pressure extraction pipelines while ensuring both economic and safety benefits, a training framework for an intelligent handling model under accident scenarios can be built using deep reinforcement learning technology combined with CFD flow field simulation of the gas negative pressure extraction pipeline. The action involves controlling the valve opening. After the action is completed, reward feedback is provided based on the simulation results of the flow field. The maximum reward is set to the highest gas concentration within the specified range and the lowest gas flow velocity within the safe range. It should be noted that the aforementioned specified range can be the range stipulated by national regulations or the range required by specific downhole scenarios. Those skilled in the art can set it according to actual needs, which will not be elaborated here.

[0099] The flow field CFD simulation model is built on a one-to-one scale based on the actual pipeline conditions. After the model is built, the flow field gas concentration and gas velocity can be calculated based on the set flow field leakage data and valve opening. The deep reinforcement learning model used is DQN, and the set valve opening is divided into 11 opening actions with a gradient of 10%.

[0100] In practical applications, the model can be built in two stages, as follows:

[0101] Phase 1: CFD-based flow field simulation of gas negative pressure extraction pipeline:

[0102] CFD numerical simulation technology is used to simulate and calculate the pressure, gas concentration, velocity, and temperature inside a coal mine gas drainage pipeline. The gas flow inside the pipeline must follow the laws of conservation of mass, momentum, energy, and composition. The general form of its governing equations is as follows:

[0103]

[0104] In the above formula, Γ is a general variable that can represent solution variables such as velocity, concentration, and pressure; Γ is the generalized diffusion coefficient; and S is the generalized source term. The terms in the equation are, in order, transient, convection, diffusion, and source terms. For a specific equation... Γ and S have specific forms.

[0105] Since the gas flow inside the gas extraction pipeline is turbulent, the widely used standard k-ε turbulence model will be used to simulate the gas flow inside the gas extraction pipeline. Its governing equations are:

[0106]

[0107] In the above formula, μ t Where is the turbulent viscosity, C0 is an empirical constant, k is the turbulent kinetic energy, and ε is the turbulent dissipation rate.

[0108]

[0109] In the above formula, μ is the dynamic viscosity; u i ′ represents the fluctuating velocity along direction i of the grid during discrete computation; x k The turbulent kinetic energy is in the x-axis direction.

[0110] The standard transport equations for k and ε are as follows:

[0111]

[0112]

[0113] In the above formula, u i x represents the flow velocity in the i-th direction of the grid during discrete computation. i x j For discrete computation, the i and j directions of the mesh; G k G represents the turbulent kinetic energy caused by the average velocity gradient. b Y represents the turbulent kinetic energy caused by buoyancy. M The effect of compressible turbulent pulsating expansion on the total dissipation rate; σ k σ ε S represents the Prandtl number corresponding to turbulent kinetic energy and turbulent dissipation rate; C1, C2, and C3 are empirical constants; S k S ε This is a custom number.

[0114] After this stage, the simulated flow field of the gas negative pressure extraction pipeline can be calculated, providing feedback on the gas concentration and gas velocity in the flow field for subsequent DQN model training. For example, according to the "Standards and Related Requirements for Coal Mine Gas Extraction Systems", the positive excitation value obtained by the DQN model is defined as 1 when the gas concentration is greater than or equal to 40% and the gas velocity is less than or equal to 10 m / s; 0 when the gas concentration is greater than or equal to 20% and the gas velocity is greater than or equal to 10 m / s and less than or equal to 15 m / s; and -1 when the gas concentration is less than or equal to 20% and the gas velocity is greater than or equal to 15 m / s.

[0115] The second stage involves building an intelligent valve opening control model based on deep reinforcement learning:

[0116] The deep reinforcement learning model used in this stage is DQN. The action is to control the opening degree of the valve. After the action is completed, the reward feedback is given based on the simulation results of the flow field in the first stage. The highest reward is set as the highest gas concentration within the specification requirements and the lowest gas flow rate within the safety range. The valve opening degree is set in 10% increments and is divided into 11 opening degree actions.

[0117] DQN is a value-based approach rather than a policy-based one; it learns not policies, but a critic. This critic does not directly take actions, but evaluates the quality of those actions. DQN combines deep learning and Q-learning, where the Q-learning algorithm is as follows:

[0118] (1) Initialize Q(s,a), For a∈A(s), first assign arbitrary values ​​to s and a, and Q(terminal-state,·)=0, then initialize Q-table=0;

[0119] (2) Select an action for the current state based on the current Q-table and execute it. The execution process continues until the current training round stops.

[0120] (3) Update the Q-table based on the reward obtained after the action is performed;

[0121] (4) Repeat process (2)-(3) until the optimal Q-table is obtained.

[0122] In DQN, the Q-table is replaced by a neural network to form a Q-network, whose Q-value is updated by using the reward and Q to calculate the Q-value: R t+1 +λmax a Q(S t+1 ,a).

[0123] Specifically, during initialization, DQN initializes two networks: Q and At the beginning, Using the same network as Q, after each training step, when the action executor interacts with the environment, it will obtain a state s. t And take an action a based on the current Q-state. t During training, the participant will explore the environment and receive a reward after exploration. t Then jump to state s t+1 After the above training, a set of s will be obtained. t a t r t and s t+1The data is assembled and placed into a buffer. When the buffer is full, the old buffer is discarded. Next, data will be sampled from the buffer, and the target will be calculated based on the sampled data. The target will then be used... The calculation formula is as follows:

[0124]

[0125] Where a is the value of let The maximum value of 'a' is also in state s. t+1 The lieutenant general will take action, update the value of Q, and after n iterations, will... Replace it with the new Q. After m iterations of training, training will stop when the feedback obtained from the action selected by the model reaches the predetermined effect, and the trained Q is the target model.

[0126] After this phase, the trained intelligent valve opening control model will be able to quickly select the optimal valve opening for gas leak response. In practical applications, it can intelligently control valve opening while ensuring both economy and safety, significantly reducing economic losses caused by directly shutting down valves in traditional gas leak response processes and ensuring safety in accident scenarios.

[0127] Step 103: Construct a dynamically updated database based on historical accident and disaster data and real-time monitoring data of basic parameters of gas extraction pipelines.

[0128] Specifically, a dynamically updated database can be obtained by fusing historical accident and disaster data with real-time monitoring data of basic parameters of gas extraction pipelines based on multi-source heterogeneous data fusion technology.

[0129] In practical applications, a gas negative pressure extraction pipeline accident disaster database, built based on multi-source heterogeneous data fusion technology, is a dynamically updated database that integrates historical accident disaster data, real-time monitoring data from various sensors, and other real-time monitoring data from underground sources. These sensors should include pipeline gas concentration sensors, flow sensors, pressure sensors, temperature sensors, carbon monoxide concentration sensors, valve opening sensors, and other sensors used to measure basic pipeline parameters.

[0130] Step 104: Train and verify the gas negative pressure extraction pipeline leakage location assessment model based on the dynamically updated database to obtain the gas negative pressure extraction pipeline leakage location assessment model after training.

[0131] Specifically, the above-mentioned gas negative pressure extraction pipeline leakage location assessment model is trained and verified based on the dynamically updated database to obtain the trained gas negative pressure extraction pipeline leakage location assessment model. The training can be carried out by referring to the following steps:

[0132] Step 1041: Obtain the data from the dynamically updated database, and during the training process, randomly divide the data into training and testing sets according to the proportions to train the gas negative pressure extraction pipeline leakage location assessment model.

[0133] Step 1042: During training, the initial learning rate is set to 0.01, and the learning rate is automatically adjusted using the cosine annealing method; the number of iterations is set to 100, and the cross-entropy loss function is used.

[0134] Step 1043: Divide a portion of the data in the training set as a validation set, and use the validation set to validate the gas negative pressure extraction pipeline leakage location assessment model.

[0135] In practical applications, data from the dynamically updated database can be fed into the established gas negative pressure extraction pipeline leak location and assessment model for training. The training and test sets are randomly divided using a 7:3 ratio. To prevent overfitting during training, a 30% validation set is allocated from the training set. Those skilled in the art can adjust the ratio of the training and test sets, and the proportion of the validation set in the training set, as needed, to train the gas negative pressure extraction pipeline leak location and assessment model; this will not be elaborated upon here.

[0136] For example, during deep learning training, taking leak location prediction as an example, the hyperparameters of the leak location assessment model for gas negative pressure extraction pipelines are set as follows: the initial learning rate is set to 0.01, and the learning rate is automatically adjusted using the cosine annealing method; the number of iterations (epochs) is set to 100. The loss function used is the categorical cross-entropy loss function. Ultimately, the leak location assessment model for gas negative pressure extraction pipelines achieves good accuracy in predicting leak locations on the test set, meeting expectations well.

[0137] Step 105: Obtain real-time monitoring data of the basic parameters of the gas extraction pipeline, and input the real-time monitoring data into the gas negative pressure extraction pipeline leakage location assessment model after training.

[0138] In practical applications, basic pipeline parameters such as gas concentration, flow rate, pressure, temperature, carbon monoxide concentration, and valve opening can be monitored in real time through monitoring and sensing equipment inside the gas negative pressure extraction pipeline. The constructed gas negative pressure extraction pipeline leakage location and assessment model is based on graph neural network (GCN) and time series analysis basic model (TimesNet), which can quickly and accurately predict gas leakage sections and efficiently match the corresponding treatment valves for subsequent treatment.

[0139] Step 106: Obtain the pressure distribution, gas concentration distribution, and leak location prediction results of the gas negative pressure extraction pipeline leakage location assessment model after training.

[0140] In practical applications, by using various monitoring and sensing devices in the gas negative pressure extraction pipeline to collect real dynamic data from underground, and by using deep learning technology to quickly and accurately predict the leakage section of the gas negative pressure extraction pipeline, the pressure distribution, gas concentration distribution and leakage location prediction results in the extraction pipeline can be obtained. This allows for timely handling using an intelligent valve opening control model, reducing the probability of accidents.

[0141] Step 107: Input the pressure distribution, gas concentration distribution and leakage location prediction results in the extraction pipeline into the trained intelligent valve opening control model to achieve intelligent control of the valve opening in the gas extraction pipeline.

[0142] In practical applications, by inputting the obtained pressure distribution, gas concentration distribution, and leak location prediction results within the gas extraction pipeline into the intelligent valve opening control model, the model can match the appropriate emergency valves to execute emergency actions. Furthermore, based on deep reinforcement learning technology, the valve opening can be intelligently controlled to ensure that the gas concentration within the gas extraction pipeline meets regulatory requirements, reducing economic losses caused by gas leaks. Simultaneously, it ensures that the gas flow velocity within the pipeline remains within a safe range, preventing gas explosions caused by electrostatic sparks due to excessive flow velocity.

[0143] Step 108: Connect the intelligent valve opening control model to the valve through Internet of Things (IoT) technology to achieve fully automatic intelligent valve opening control in case of an accident.

[0144] In practical applications, valves can achieve intelligent opening and closing based on Internet of Things (IoT) technology, enabling fully automatic intelligent valve opening control in emergency situations. For example, existing electrically controlled valves capable of connecting to the IoT can be used. The valve can connect to an intelligent valve opening control model via the IoT to enable the model to control the valve opening, thereby achieving automatic valve opening control in emergency situations. This improves the efficiency of handling pipeline leaks and further reduces the probability of accidents.

[0145] This embodiment provides a method for assessing and intelligently handling leaks in gas negative pressure extraction pipelines. By using the valves of the gas negative pressure extraction pipeline as the handling terminal, it comprehensively utilizes multi-source heterogeneous data fusion technology and deep learning technology to achieve rapid and accurate location assessment of leaks in gas negative pressure extraction pipelines. Based on this, and taking into account economic and safety factors, it achieves precise matching of handling valves after a gas negative pressure extraction pipeline leak and intelligent control of valve opening. This provides the most economical and safe solution for gas negative pressure extraction pipeline leak accidents, serving the prediction, early warning, and safety prevention and control of gas extraction accidents.

[0146] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for assessing and intelligently handling leaks in a gas negative pressure extraction pipeline, characterized in that, include: A leakage location assessment model for a gas negative pressure extraction pipeline is constructed; wherein, the leakage location assessment model for a gas negative pressure extraction pipeline is built based on a graph neural network and a TimesNet time series analysis basic model; A smart valve opening control model based on deep reinforcement learning technology is constructed and trained to obtain a trained smart valve opening control model. The action of the smart valve opening control model is set as valve opening, and the state is the gas concentration and gas flow rate in the pipeline. The highest excitation state is the highest gas concentration within the specification requirements and the lowest gas flow rate within the safety range. A dynamically updated database is constructed based on historical accident and disaster data and real-time monitoring data of basic parameters of gas extraction pipelines; The leakage location assessment model of the gas negative pressure extraction pipeline is trained and verified based on the dynamically updated database to obtain the gas negative pressure extraction pipeline leakage location assessment model after training. Acquire real-time monitoring data of basic parameters of the gas extraction pipeline, and input the real-time monitoring data into the gas negative pressure extraction pipeline leakage location assessment model after training; Obtain the pressure distribution, gas concentration distribution and leak location prediction results of the gas negative pressure extraction pipeline leak location assessment model after training is completed; The pressure distribution, gas concentration distribution, and leakage location prediction results in the gas extraction pipeline are input into the trained intelligent valve opening control model to achieve intelligent control of the valve opening in the gas extraction pipeline. The aforementioned construction of a gas negative pressure extraction pipeline leakage location and assessment model; wherein, the gas negative pressure extraction pipeline leakage location and assessment model is built based on a graph neural network and TimesNet time series analysis basic model, including: Construct a topology map of adjacent locations of the monitors within the extraction pipeline to carry the data monitored by the monitors; The preset leakage orifice intervals formed between the various monitors are combined to form a leakage interval matrix, and the leakage interval matrix is ​​assigned values ​​according to different leakage states as predicted labels for leakage locations; wherein, the different leakage states include leakage and no leakage. The topology graph is processed by graph neural network operations and used as a time series node in the TimesNet time series analysis basic model. The classification prediction mode in the TimesNet time series analysis basic model is used to process the prediction task. An evaluation index is set for the leakage location assessment model of the gas negative pressure extraction pipeline, and weights are added to different error scenarios using the evaluation index. The most severe error is assigned the largest weight value, while for prediction errors of the leakage range, the weight is increased proportionally according to the distance between the predicted range and the actual range; the greater the distance, the higher the weight. The formula for the evaluation index is as follows: In the formula, To predict the number of incorrect intervals, For the first Error weight values ​​for each error interval To predict the total sample size; The process involves constructing an intelligent valve opening control model based on deep reinforcement learning technology and training the model to obtain a trained intelligent valve opening control model. The model's actions are defined as valve opening, its states as gas concentration and gas flow velocity within the pipeline, and its highest excitation state is the highest gas concentration within the specified range and the lowest gas flow velocity within the safe range. This includes: Flow field simulation of gas negative pressure extraction pipeline based on CFD numerical simulation technology; A smart valve opening control model is constructed based on the DQN deep reinforcement learning model; The intelligent valve opening control model is trained based on the flow field simulation of the gas negative pressure extraction pipeline. The trained intelligent valve opening control model is obtained.

2. The method according to claim 1, characterized in that, The dynamically updated database, constructed based on historical accident and disaster data and real-time monitoring data of basic parameters of gas drainage pipelines, includes: A dynamically updated database is obtained by fusing historical accident and disaster data with real-time monitoring data of basic parameters of gas extraction pipelines using multi-source heterogeneous data fusion technology.

3. The method according to claim 1, characterized in that: The valve opening degree is divided into 11 opening degree actions with a 10% gradient.

4. The method according to claim 1, characterized in that, The process of training and validating the gas negative pressure drainage pipeline leakage location assessment model based on the dynamically updated database to obtain the trained gas negative pressure drainage pipeline leakage location assessment model includes: Data is obtained from the dynamically updated database, and during the training process, the data is randomly divided into training and test sets according to a certain ratio for training the gas negative pressure extraction pipeline leakage location and evaluation model. During training, the initial learning rate was set to 0.01, and cosine annealing was used to automatically adjust the learning rate; the number of iterations was set to 100, and the cross-entropy loss function was used. A portion of the data in the training set is used as a validation set, and the validation set is used to validate the gas negative pressure extraction pipeline leakage location assessment model.

5. The method according to claim 1, characterized in that, After inputting the pressure distribution, gas concentration distribution, and leak location prediction results within the extraction pipeline into the trained intelligent valve opening control model to achieve intelligent control of the valve opening within the gas extraction pipeline, the method further includes: The intelligent valve opening control model is connected to the valve via Internet of Things (IoT) technology to achieve fully automatic intelligent valve opening control in case of an accident.

Images