Intelligent gas monitoring system based on artificial intelligence

Abstract

The application relates to the field of intelligent gas monitoring, in particular to an intelligent gas monitoring system based on artificial intelligence, which comprises a gas sensing module, a voiceprint positioning module, a decision module and a linkage control module; the gas sensing module obtains boundary parameters and environmental parameters through a sensor, generates virtual data in combination with digital twinning and PINN, and realizes early fault prediction; the voiceprint positioning module loads corresponding voiceprint models according to different fault types, and then corrects the spatial conflict positioning of multiple fault types through a wave direction fusion algorithm; the decision module inputs the environmental parameters into XGBoost to quantize the danger level, introduces a simulated annealing algorithm to optimize tree splitting, and realizes spatial risk analysis in combination with GIS; and the linkage control module obtains the optimal control strategy of an actuator based on a genetic algorithm, processes multiple target conflicts and is optimized through a Pareto frontier; the application realizes accurate prediction, positioning and intelligent disposal of gas pipeline faults, and provides a full-process intelligent solution for gas safety.

Description

Technical Field

[0001] This invention relates to the field of intelligent gas monitoring, specifically to an intelligent gas monitoring system based on artificial intelligence. Background Technology

[0002] In the field of intelligent gas monitoring, traditional monitoring systems are limited by technology and cannot meet the safety management needs under complex operating conditions. Specifically:

[0003] When locating multiple fault types, single acoustic fingerprint localization technology cannot handle spatial conflicts of complex faults such as leaks, cracks, and thermal anomalies. It is prone to misjudgment when the distance between the location point is less than twice the pipe diameter. Hazard level assessment only learns global features, ignores spatial clustering, and lacks coupling with GIS. It cannot quantify spatial distribution patterns through Moran's index, resulting in a disconnect between regional assessment and operation and maintenance needs. The physical equation modeling and calculation are complex and difficult to solve in real time. Pure data-driven algorithms lack physical constraints and are prone to deviating from engineering mechanisms. The separation of the two leads to insufficient model generalization ability under extreme conditions. These technical bottlenecks result in existing systems having problems such as lack of physical field perception, insufficient multi-source data fusion, weak spatial analysis capabilities, and low degree of intelligent decision optimization in the entire process of "fault prediction-location-assessment-disposal". Therefore, an artificial intelligence-based intelligent gas monitoring system is needed to solve the above problems. Summary of the Invention

[0004] To address the technical problems mentioned in the background section, this invention provides an artificial intelligence-based intelligent gas monitoring system. The objective of this invention can be achieved through the following technical solution: This invention provides an artificial intelligence-based intelligent gas monitoring system, including a gas sensing module, an acoustic signature localization module, a decision-making module, a linkage control module, and a database.

[0005] The gas sensing module acquires the set of boundary parameters for the operation of the gas equipment and the set of environmental parameters in the scenario based on various types of sensors. It then combines the boundary parameter set with a digital twin to perform early fault prediction. Specifically:

[0006] The environmental parameters near the gas pipeline are obtained by various environmental sensors. The environmental parameters include, but are not limited to, soil moisture, pH value, resistivity and stray current.

[0007] By deploying sensor arrays at key boundary nodes of gas equipment, the set of boundary parameters of equipment operation is collected in real time. The set of boundary parameters includes, but is not limited to, pipe wall pressure, temperature, fluid velocity, and flow data at equipment interfaces.

[0008] The boundary parameter set is input into a physical information neural network (PINN) containing constraints from the physical equations of the equipment to generate virtual sensor data. Specifically, PINN adopts a multi-layer fully connected neural network architecture with the network structure: u(x,t) = NN(x; t; θ), where u(x,t) is the output vector, which includes the pressure field P(x,t), stress field σ(x,t), and temperature field T(x,t) of the gas pipeline; θ is the trainable parameters of the network, including weights and biases; NN is a multi-layer perceptron connected by activation functions; x and t are spatiotemporal coordinates, including the axial position of the pipeline and the time point; a loss function is set in the model, which is composed of data fitting terms. With constraints of the physical equations of the equipment The weighted summation is constructed, and its calculation logic is as follows: Where ω date and ω physics The fixed weighting coefficients are 0.46 and 0.54, respectively. The data fitting term between the boundary sensor data and the PINN predicted values ​​is calculated using the L2 norm, and its calculation logic is as follows: in Let i be the spatiotemporal coordinates of the i-th boundary sensor. N represents the actual measured value from the sensor. b is the total number of boundary sensors, and i is the boundary sensor number; The L2 norm of the residuals of the equipment physical equations is given. The equipment physical equations include the transient flow model of the gas pipeline network, the fluid-structure coupling equations, and the heat conduction equations. The calculation logic for the residuals RF(t) of the transient flow model of the gas pipeline network is as follows: in The time derivative of pressure, Let be the dot product of velocity and pressure gradient, and k be the fluid diffusion coefficient. The product of the pressure and the Laplace, f(t) is the time perturbation term; therefore, the physical constraint term is set as follows: in N represents the constraint points randomly sampled within the computational domain, used to enforce the satisfaction of physical equations. c The total number of constraint points is given by , j is the constraint point number, and so on. Physical constraint terms for the fluid-structure coupling equation and heat conduction equation are obtained. Virtual data is generated using the predicted pressure field P(x,t), stress field σ(x,t), and temperature field T(x,t) output by PINN. This virtual data includes the gas pipeline corrosion rate, cover sealing index, and thermal stress risk index, with the gas pipeline corrosion rate C... r Combining the stress and temperature fields, and based on the electrochemical corrosion kinetics formula, the following calculation is performed: C r =β·σx,t)·e -ε / T(x,t)Where β is the material corrosion sensitivity coefficient and ε is the temperature activation energy coefficient; the sealing index δ(x,t) of the gas pipeline cover is obtained through the pressure field P(x,t) predicted by PINN, and its calculation logic is as follows: Where η P (x,t) represents the standard deviation of the contact pressure at the sealing surface, κ P (x,t0 is the mean value of the contact pressure of the sealing surface, δ(x,t)∈[0,1]; its thermal stress risk index η) P (x,t) is obtained by combining the temperature field and the material's thermal expansion coefficient FY. Its calculation logic is as follows: η P (x,t)=E·FY·(TT ref ), where T ref For reference temperature;

[0009] The average distribution of residual change rate under normal operating conditions is statistically analyzed using historical data as the first-level alarm threshold ψ. The residual change rate RQ is set as the time first derivative of the residual norm L2. If the residual change rate RQ exceeds the first-level alarm threshold ψ, the physical equation type corresponding to the residual abnormality is obtained and the corresponding fault type is determined. The corresponding fault types include leakage, structural cracks, and local overheating.

[0010] The acoustic signature localization module loads the corresponding acoustic signature model according to different fault types to locate the fault location point in the gas pipeline. The specific process is as follows:

[0011] Upon receiving detection commands and fault types, the system loads corresponding acoustic signature models, including leak acoustic signature localization networks, crack acoustic signature localization networks, and thermal anomaly acoustic signature detection networks. The system spatiotemporally aligns the collected raw acoustic signature data with virtual data. A Mel filter bank is used to convert the raw acoustic signature data of the gas pipeline into a Mel spectrogram. Two-dimensional convolution is employed to extract spatial features from the Mel spectrogram, which are then input into the leak acoustic signature classification network. A fully connected layer outputs the gas leak location DT. Residual connections are used to mitigate gradient vanishing. The calculation logic is: DT = x avg +Tanh(W r c+b r )·LR, where x avg The reference point for the acoustic signature sensor array is LR, the total length of the pipe is LR, and the Tanh activation function limits the output to LR.

[0012] Continuous wavelet transform is used to extract the time-frequency features of non-steady-state acoustic signatures. The signal yc from the m-th acoustic signature sensor is analyzed. m Its wavelet coefficients C m The calculation logic for (ay,r) is as follows: Where ∞ and -∞ represent the entire time axis, ay is the scale parameter, r is the translation parameter, t1 is the continuous variable of the time axis, and the square of the wavelet coefficients represents the energy distribution of the signal in the time-frequency domain. The actual time-frequency energy EK of the acoustic sensor at scale ay is set, and its calculation logic is as follows: Where dr is the integration variable; a region scanning algorithm is set within the network to divide the pipe into grid points along the axial and circumferential directions, and the crack initiation location is set as B. h If h is the grid point number, then the time it takes for the signal to propagate to the m-th sensor is t. h,m Its calculation logic is as follows: Where vf is the speed of sound, s m Let be the coordinates of the m-th sensor; the acoustic emission source signal is Fe, and the theoretical signal propagating to sensor m is: Where z is the energy decay coefficient; for Perform wavelet transform to obtain the theoretical time-frequency energy. Based on actual time-frequency energy EK and theoretical time-frequency energy Cosine similarity calculation of time-frequency energy matching degree M wavelet,h Its calculation logic is as follows:

[0013] Where g represents the total number of sensors, and g2 represents the total number of grid points; select M wavelet,h The smallest grid point is used as the location of the crack in the gas pipeline;

[0014] A clustering detection algorithm is set up in the thermal anomaly soundprint detection network to obtain the time-frequency features of the soundprint. K-means clustering is used to divide the time-frequency features into K classes, one of which is the normal mode and the rest are the abnormal modes. Then, the K-means is trained with the soundprint under normal operating conditions to obtain the center of the normal mode. The Euclidean distance ds between the unknown soundprint feature and the center of the normal mode is calculated. If ds > the first preset threshold, it is judged as abnormal. Combined with the sensor position corresponding to the abnormal soundprint, the location of the thermal anomaly in the gas pipeline is obtained through the triangulation algorithm.

[0015] If two or three fault types exist simultaneously, the direction-of-arrival fusion algorithm is used to correct spatial conflicts in the location results. This correction is made when the distance difference between the gas leak location point and the crack point is less than the preset minimum distance threshold D. min When this occurs, the fusion correction process is triggered, and its calculation logic is as follows: in To estimate the direction of arrival angle, argminτ is a parameter that minimizes τ. The theoretical time delay of the m-th sensor, The measured time delay of the m-th sensor is used to complete the positioning correction, and the positioning point is sent to the decision module.

[0016] The decision-making module inputs the set of environmental parameters corresponding to the location point into the gradient boosting tree (XGBoost) to quantify the hazard level, introduces the simulated annealing algorithm to optimize the tree splitting process, and combines GIS geographic information to realize spatial risk analysis. The specific process is as follows: The hazard level generation logic is: Where Ω is the characteristic engineering function. For fusion analysis, SA is a simulated annealing algorithm; the original features are the set of environmental parameters and virtual data of the location points, and the empirical features are the historical burial depth, historical coating damage rate, and historical cathodic protection potential of the gas pipeline in the database; XGBoost constructs multiple decision trees through an additive model. During the generation of each tree in XGBoost, for each candidate split point s, the objective function gain RT(s) after splitting is calculated, and SA is used to determine whether to accept the split. There are three judgment principles: the first principle is that if RT(s) ≥ 0, it is considered a good solution and the split is accepted; the second principle is that if RT(s) < 0, it is considered a bad solution and the split is not accepted; the third principle is that the split is not accepted in other cases. For any node I, the candidate split point s divides I into left child nodes I. L and right child node I R The calculation logic for its objective function gain RT(s) is as follows: Where γ is the leaf node number penalty factor, λ1 is the regularization coefficient, and G and H are the gradient and the diagonal elements of the Hessian matrix, respectively; the formula for the overall prediction function is: Where f k ∈Ω, K is the total number of trees, f k (qn1) is the hazard prediction value of the k-th tree for the n1-th sample; measured by the Moran index. The spatial clustering of hazard prediction values ​​is quantified, and the hazard level of location points is identified by combining GIS geographic information. Specifically: if This indicates that the location is a high-risk pipeline concentration area, and a processing command is generated and sent to the linkage control module. This indicates that the risk level is randomly distributed. This indicates a low-risk pipe section.

[0017] The linkage control module obtains the optimal control strategy for each actuator in the gas pipeline based on a genetic algorithm. The specific process is as follows: Upon receiving a processing command, the actuators, including valves and ventilation equipment, are defined as decision variables and encoded as chromosomes. These control parameters include the action sequence, time, and intensity. The weighted sum of the action sequence, time, and intensity is set as the fitness function. For multi-objective conflicts, non-dominated solutions are retained through Pareto optimality. A non-dominated solution is defined as follows: if solution A is no worse than solution B on all objectives, and is better on at least one objective, then A dominates B. The undominated solutions constitute the Pareto front, which is incorporated into the genetic algorithm. The genetic algorithm generates 100 random action sequences, times, and intensities. The leakage amount and time for each sequence are calculated using a pipeline simulation model and substituted into the fitness function. The top 10% of fitness values ​​are selected, and sequential crossover is performed on the selected individuals to generate new individuals. Mutation is performed on the new individuals in reverse order to correct any sequences that violate constraints, until the optimal control strategy is obtained through iteration. The optimal strategy includes the optimal action sequence, optimal time, and optimal intensity.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: The decision module inputs environmental parameters into the XGBoost model optimized by the simulated annealing algorithm to quantify the hazard level. The simulated annealing algorithm optimizes the tree splitting process to avoid getting trapped in local optima. At the same time, combined with GIS geographic information, the Moran index is used to quantify the spatial clustering of hazard prediction values. If the Moran index is greater than 0, it is identified as a high-risk pipeline concentration area, and a processing instruction is generated and sent to the linkage control module. This realizes accurate assessment of hazard level and spatial risk analysis, and provides a scientific basis for subsequent disposal.

[0019] The acoustic signature localization module employs corresponding acoustic signature models for different fault types. For example, the leak acoustic signature localization network extracts features through Mel spectrum and two-dimensional convolution to locate the leak location, while the crack acoustic signature localization network uses continuous wavelet transform and region scanning algorithms to determine the crack location. The thermal anomaly acoustic signature detection network identifies anomalies using K-means clustering. Furthermore, when multiple fault types exist simultaneously, the direction-of-arrival fusion algorithm is used to correct spatial conflicts in the localization results. When the distance difference between the gas leak location point and the crack point is less than a preset minimum distance threshold, the fusion correction process is triggered, thereby improving the localization accuracy under multiple fault types. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. The following drawings are not drawn to scale according to the actual size, but are intended to illustrate the main idea of ​​the present invention. Figure 1 This is a schematic diagram showing the connections of the various modules of the present invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are also within the scope of protection of the present invention.

[0022] Please refer to Figure 1 As shown, the present invention provides an intelligent gas monitoring system based on artificial intelligence, including a gas sensing module, an acoustic positioning module, a decision-making module, a linkage control module, and a database.

[0023] The gas sensing module acquires the set of boundary parameters for the operation of the gas equipment and the set of environmental parameters in the scenario based on various types of sensors. It then combines the boundary parameter set with a digital twin to perform early fault prediction. Specifically:

[0024] The environmental parameters near the gas pipeline are obtained by various environmental sensors. The environmental parameter set includes soil moisture, pH value, resistivity and stray current.

[0025] By deploying sensor arrays at key boundary nodes of gas equipment, the set of boundary parameters of equipment operation is collected in real time. The set of boundary parameters includes, but is not limited to, pipe wall pressure, temperature, fluid velocity, and flow data at equipment interfaces.

[0026] The boundary parameter set is input into a physical information neural network (PINN) containing constraints from the physical equations of the equipment to generate virtual sensor data. Specifically, PINN adopts a multi-layer fully connected neural network architecture with the network structure: u(x,t) = NN(x; t; θ), where u(x,t) is the output vector, which includes the pressure field P(x,t), stress field σ(x,t), and temperature field T(x,t) of the gas pipeline; θ is the trainable parameters of the network, including weights and biases; NN is a multi-layer perceptron connected by activation functions; x and t are spatiotemporal coordinates, including the axial position of the pipeline and the time point; a loss function is set in the model, which is composed of data fitting terms. With constraints of the physical equations of the equipment The weighted summation is constructed, and its calculation logic is as follows: Where ω date and ω physics The fixed weighting coefficients are 0.46 and 0.54, respectively. The data fitting term between the boundary sensor data and the PINN predicted values ​​is calculated using the L2 norm, and its calculation logic is as follows: in Let i be the spatiotemporal coordinates of the i-th boundary sensor. N represents the actual measured value from the sensor. bis the total number of boundary sensors, and i is the boundary sensor number; The L2 norm of the residuals of the equipment physical equations is given. The equipment physical equations include the transient flow model of the gas pipeline network, the fluid-structure coupling equations, and the heat conduction equations. The calculation logic for the residuals RF(t) of the transient flow model of the gas pipeline network is as follows: in The time derivative of pressure, Let be the dot product of velocity and pressure gradient, and k be the fluid diffusion coefficient. The product of pressure and Laplace, f(t0) is the time perturbation term; therefore, the physical constraint term is set as follows: in N represents the constraint points randomly sampled within the computational domain, used to enforce the satisfaction of physical equations. c Let j be the total number of constraint points, j be the constraint point number, and so on, to obtain the physical constraint terms of the fluid-structure coupling equation and the heat conduction equation. Virtual data is generated using the predicted pressure field P(x,t), stress field σ(x,t), and temperature field T(x,t) output by PINN. The virtual data includes the gas pipeline corrosion rate, cover sealing index, and thermal stress risk index, with the gas pipeline corrosion rate C... r Combining the stress and temperature fields, and based on the electrochemical corrosion kinetics formula, the following calculation is performed: C r =β·σ(x,t)·e -ε / T(x,t) Where β is the material corrosion sensitivity coefficient and ε is the temperature activation energy coefficient; the sealing index δ(x,t) of the gas pipeline cover is obtained through the pressure field P(x,t) predicted by PINN, and its calculation logic is as follows: Where η P (x,t) represents the standard deviation of the contact pressure at the sealing surface, κ P (x,t) represents the average contact pressure of the sealing surface, δ(x,t)∈[0,1]. The closer the value is to 1, the more uniform the contact pressure of the sealing surface and the lower the risk of leakage; its thermal stress risk index η P (x,t) is obtained by combining the temperature field and the material's thermal expansion coefficient FY. Its calculation logic is as follows: η P (x,t)=E·FY·(TT ref ), where T ref The reference temperature is used as a reference. It should be noted that the role of virtual data is to fit physical equations using the PINN model and, based on measurable boundary data such as inlet and outlet pressure, flow rate, and temperature, to calculate key parameters of unmeasurable locations inside the gas pipeline, thus filling monitoring blind spots.

[0027] The average distribution of residual change rate under normal operating conditions is statistically analyzed based on historical data and used as the first-level alarm threshold ψ. The residual change rate RQ is set as the time first derivative of the residual norm L2, reflecting the rate of deviation of the physical field from the law. If the residual change rate RQ exceeds the first-level alarm threshold ψ, the physical equation type corresponding to the residual anomaly is obtained and the corresponding fault type is determined. The corresponding fault types include leakage, structural cracks, and local overheating. It should be noted that the residual anomaly of the transient flow model points to fluid disturbance, the residual anomaly of the stress equation points to structural damage, and the residual anomaly of the heat conduction equation points to local overheating of the pipeline. The detection command and fault type are generated and sent to the acoustic fingerprint localization module in sequence.

[0028] The acoustic signature localization module loads the corresponding acoustic signature model according to different fault types to locate the fault location point in the gas pipeline. The specific process is as follows:

[0029] Upon receiving detection commands and fault types, the system loads corresponding acoustic signature models, including leak acoustic signature localization networks, crack acoustic signature localization networks, and thermal anomaly acoustic signature detection networks. The system spatiotemporally aligns the collected raw acoustic signature data with virtual data. A Mel filter bank is used to convert the raw acoustic signature data of the gas pipeline into a Mel spectrogram. Two-dimensional convolution is employed to extract spatial features from the Mel spectrogram, which are then input into the leak acoustic signature classification network. A fully connected layer outputs the gas leak location DT. Residual connections are used to mitigate gradient vanishing. The calculation logic is: DT = x avg +Tanh(W r c+b r )·LR, where x avg The reference point for the acoustic signature sensor array is LR, the total length of the pipe is LR, and the Tanh activation function limits the output to LR.

[0030] Continuous wavelet transform is used to extract the time-frequency features of non-steady-state acoustic signatures. The signal yc from the m-th acoustic signature sensor is analyzed. m Its wavelet coefficients C m The calculation logic for (ay,r) is as follows: Where ∞ and -∞ represent the entire time axis, ay is the scale parameter, r is the translation parameter, t1 is the continuous variable of the time axis, and the square of the wavelet coefficients represents the energy distribution of the signal in the time-frequency domain. The actual time-frequency energy EK of the acoustic sensor at scale ay is set, and its calculation logic is as follows: Where dr is the integration variable; a region scanning algorithm is set within the network to divide the pipe into grid points along the axial and circumferential directions, and the crack initiation location is set as B. h If h is the grid point number, then the time it takes for the signal to propagate to the m-th sensor is t. h,m Its calculation logic is as follows: Where vf is the speed of sound, s mLet be the coordinates of the m-th sensor; the acoustic emission source signal is Fe, and the theoretical signal propagating to sensor m is: Where z is the energy decay coefficient; for Perform wavelet transform to obtain the theoretical time-frequency energy. Based on actual time-frequency energy EK and theoretical time-frequency energy Cosine similarity calculation of time-frequency energy matching degree M wavelet,h Its calculation logic is as follows:

[0031] Where g represents the total number of sensors, and g2 represents the total number of grid points; select M wavelet,h The smallest grid point is used as the location of the crack in the gas pipeline.

[0032] A clustering detection algorithm is incorporated into the thermal anomaly voiceprint detection network to acquire the time-frequency features of the voiceprint. K-means clustering is then used to divide these features into K classes, one representing normal patterns and the rest representing abnormal patterns. The K-means algorithm is then trained using voiceprints from normal operating conditions to obtain the center y of the normal pattern. normal Calculate unknown voiceprint features and y normal The Euclidean distance ds is used to determine the location of the abnormality in the gas pipeline. If ds > the first preset threshold, it is judged as abnormal. Combined with the sensor location corresponding to the abnormal sound pattern, the location of the thermal abnormality in the gas pipeline is obtained through the triangulation algorithm.

[0033] If two or three fault types exist simultaneously, the direction-of-arrival fusion algorithm is used to correct spatial conflicts in the location results. This correction is made when the distance difference between the gas leak location point and the crack point is less than the preset minimum distance threshold D. min When this occurs, the fusion correction process is triggered. The preset minimum distance threshold is usually twice the pipe diameter, such as a steel pipe with a diameter of 0.5m. min =1m, and its calculation logic is as follows: in To estimate the direction of arrival angle, argminτ is a parameter that minimizes τ. The theoretical time delay of the m-th sensor, The measured time delay of the m-th sensor is used to complete the positioning correction, and the positioning point is sent to the decision module.

[0034] The decision-making module inputs the set of environmental parameters corresponding to the location point into the gradient boosting tree (XGBoost) to quantify the hazard level, introduces the simulated annealing algorithm to optimize the tree splitting process, and combines GIS geographic information to realize spatial risk analysis. The specific process is as follows: The hazard level generation logic is: Where Ω is the characteristic engineering function. For fusion analysis, SA is a simulated annealing algorithm; the original features are the set of environmental parameters and virtual data of the location points, and the empirical features are the historical burial depth, historical coating damage rate, and historical cathodic protection potential of the gas pipeline in the database; XGBoost constructs multiple decision trees through an additive model. During the generation of each tree in XGBoost, for each candidate split point s, the objective function gain RT(s) after splitting is calculated, and SA is used to determine whether to accept the split. There are three judgment principles: the first principle is that if RT(s) ≥ 0, it is considered a good solution and the split is accepted; the second principle is that if RT(s) < 0, it is considered a bad solution and the split is not accepted; the third principle is that the split is not accepted in other cases. For any node I, the candidate split point s divides I into left child nodes I. L and right child node I R The calculation logic for its objective function gain RT(s) is as follows: Where γ is the leaf node number penalty factor, λ1 is the regularization coefficient, and G and H are the gradient and the diagonal elements of the Hessian matrix, respectively; the formula for the overall prediction function is: Where f k ∈Ω, K is the total number of trees, f k (qn1) is the hazard prediction value of the k-th tree for the n1-th sample; measured by the Moran index. The spatial clustering of hazard prediction values ​​is quantified, and the hazard level of location points is identified by combining GIS geographic information. Specifically: if This indicates that the location is a high-risk pipeline concentration area, and a processing command is generated and sent to the linkage control module. This indicates that the risk level is randomly distributed. This indicates a low-risk pipe section.

[0035] The linkage control module obtains the optimal control strategy for each actuator in the gas pipeline based on a genetic algorithm. The specific process is as follows: Upon receiving processing instructions, the actuators, including valves and ventilation equipment, are defined as decision variables and encoded as chromosomes. These control parameters include the sequence, time, and intensity of actions. Specifically: multiple leak points require the closure of multiple valves; optimizing the closure sequence minimizes leakage. Multiple actuators need to coordinate actions, such as valve closure and sprinkler activation; optimizing the action time difference avoids conflicts. Limited resources need to be allocated to multiple actuators; optimizing the allocation scheme maximizes coverage of high-risk points. The weighted sum of the action sequence, time, and intensity is then set as the fitness function. For multi-objective conflicts, Pareto optimality is used. Non-dominated solutions are retained and ultimately selected by the decision-maker. A non-dominated solution is defined as follows: solution A is no worse than solution B on all objectives, and is better on at least one objective; therefore, A dominates B. The undominated solutions constitute the Pareto front, which is incorporated into the genetic algorithm. The genetic algorithm generates 100 random action sequences, times, and intensities. The leakage rate and time for each sequence are calculated using a pipeline simulation model and substituted into the fitness function. The top 10% of fitness values ​​are selected, and sequential crossover is performed on the selected individuals to generate new individuals. Mutation is performed on the new individuals in reverse order to correct any sequences that violate constraints, until the optimal control strategy is obtained through iteration. The optimal control strategy includes the optimal action sequence, optimal time, and optimal intensity. The optimal control strategy is then sent to the actuator.

[0036] The foregoing description is illustrative of the invention and should not be construed as limiting it. Although several exemplary embodiments of the invention have been described, those skilled in the art will readily understand that many modifications can be made to the exemplary embodiments without departing from the novel teachings and advantages of the invention. Therefore, all such modifications are intended to be included within the scope of the invention as defined in the claims. It should be understood that the foregoing description is illustrative of the invention and should not be construed as limiting it to the specific embodiments disclosed, and modifications to the disclosed embodiments and other embodiments are intended to be included within the scope of the appended claims. The invention is defined by the claims and their equivalents.

Claims

1. An artificial intelligence-based intelligent gas monitoring system, comprising a gas sensing module, an acoustic signature positioning module, a decision-making module, a linkage control module, and a database, characterized in that: The gas sensing module acquires the set of boundary parameters and the set of environmental parameters in the scene when the gas equipment is running based on various types of sensors. The set of environmental parameters includes soil moisture, pH value, resistivity and stray current. The set of boundary parameters includes pipe wall pressure, temperature, fluid velocity and flow data at the equipment interface. The boundary parameter set is combined with digital twin to predict faults. The acoustic signature localization module loads the corresponding acoustic signature model according to different fault types to locate the fault location of the gas pipeline. If two or three fault types exist at the same time, the direction-of-arrival fusion algorithm is used to correct the spatial conflict of the localization results. The decision module inputs the set of environmental parameters corresponding to the gas pipeline location point into the gradient boosting tree XGBoost to quantify the hazard level, introduces the simulated annealing algorithm to optimize the tree splitting process, and combines GIS geographic information to analyze spatial risks. The linkage control module obtains the optimal control strategy for each actuator in the gas pipeline based on a genetic algorithm; If the voiceprint localization module has two or three fault types simultaneously, a direction-of-arrival fusion algorithm is used to correct spatial conflicts in the localization results, specifically: When the distance difference between the gas leak location point and the crack point is less than the preset minimum distance threshold, the fusion correction process is triggered, the estimated direction of arrival angles are calculated and fed back to each model to correct the location, and the corrected location point is sent to the decision module. The gas sensing module combines boundary parameter sets with digital twins to predict faults in their early stages, specifically: The boundary parameter set is input into a physical information neural network (PINN) containing constraints from the equipment's physical equations to generate virtual sensor data. Specifically, PINN employs a multi-layer fully connected neural network architecture, which includes an output vector, trainable parameters, and activation functions connected to a multi-layer perceptron. The output vector includes the pressure, stress, and temperature fields of the gas pipeline. A loss function is set in the model, which is a weighted sum of the data fitting term and the constraints from the equipment's physical equations. The data fitting term is obtained through... The norm is calculated; the constraint terms of the equipment physical equation are the residuals of the equipment physical equation. Norms, equipment physical equations include transient flow models of gas pipeline networks, fluid-structure coupling equations, and heat conduction equations; Virtual data is generated by using the predicted pressure, stress, and temperature fields output by PINN. The virtual data includes the corrosion rate of the gas pipeline, the sealing index of the gas cover, and the thermal stress risk index. The corrosion rate of the gas pipeline is calculated based on the electrochemical corrosion kinetics formula by combining the stress and temperature fields. The sealing index of the gas pipeline is calculated by using the pressure field predicted by PINN, and the thermal stress risk index is obtained by combining the temperature field and the thermal expansion coefficient of the material.

2. The intelligent gas monitoring system based on artificial intelligence according to claim 1, characterized in that, The specific process of the decision-making module is as follows: The hazard level generation logic is: wherein is a feature engineering function, is a fusion analysis, is a simulated annealing algorithm; The original features are the set of environmental parameters and virtual data of the location points, while the empirical features are the historical burial depth, historical coating damage rate, and historical cathodic protection potential of the gas pipelines in the database. XGBoost constructs multiple decision trees through an additive model. During the generation of each tree in XGBoost, for each candidate split point s, the objective function gain after splitting is calculated. The decision to accept the split is made through the SA (Self-Assessment) mechanism, which uses three principles: the first principle is... If the value is ≥0, then the solution is considered optimal and the split is accepted; the second principle is... If the value is less than 0, the solution is considered inferior and not accepted. The third principle is that splitting is not accepted under other circumstances; a formula for the overall prediction function is set to obtain the risk prediction value of each sample in the decision tree; and the Moran index is used. The spatial clustering of hazard prediction values ​​is quantified, and the hazard level of location points is identified by combining GIS geographic information. Specifically: if If the value is greater than 0, it indicates that the location is a high-risk pipeline concentration area, and a processing command is generated and sent to the linkage control module; if... =0 indicates that the risk level is randomly distributed; if If the value is less than 0, it indicates that the location point is a low-risk pipe section. 3.The AI-based intelligent gas monitoring system according to claim 1, wherein, The acoustic signature localization module loads corresponding acoustic signature models according to different fault types to locate the fault location of the gas pipeline. The specific process is as follows: The system receives detection commands and fault types, loads corresponding acoustic fingerprint models, including leak acoustic fingerprint localization networks, crack acoustic fingerprint localization networks, and thermal anomaly acoustic fingerprint detection networks. It aligns the collected raw acoustic fingerprint data with virtual data in a spatiotemporal manner, converts the raw acoustic fingerprint data of the gas pipeline into a Mel spectrogram through a Mel filter bank, extracts the spatial features of the Mel spectrogram using two-dimensional convolution, inputs the spatial features into a leak acoustic fingerprint classification network, outputs the gas leak location through a fully connected layer, mitigates gradient vanishing using residual connections, and sends the leak location to the decision module. The time-frequency features of the non-steady-state acoustic signature are extracted using continuous wavelet transform. Wavelet coefficients are obtained through continuous variables of the acoustic signature sensor's signal, scale parameters, translation parameters, and time axis. The square of the wavelet coefficients represents the energy distribution of the signal in the time-frequency domain. The actual time-frequency energy of the acoustic signature sensor under the scale parameters is set. A region scanning algorithm is set in the network to divide the pipe into grid points in the axial and circumferential directions, set the crack source location, and mark the time when the signal propagates to the sensor. Acquire the acoustic emission source signal and calculate the theoretical signal propagating to the sensor; The theoretical time-frequency energy is obtained by performing wavelet transform on the theoretical signal; The time-frequency energy matching degree is calculated based on the cosine similarity between the actual time-frequency energy and the theoretical time-frequency energy. The grid point with the smallest time-frequency energy matching degree is selected as the location of the gas pipeline crack, and the crack location is sent to the decision module. A clustering detection algorithm is set up in the thermal anomaly soundprint detection network to obtain the time-frequency features of the soundprint. K-means clustering is used to divide the time-frequency features into K classes, one of which is the normal mode and the rest are the abnormal modes. Then, the K-means is trained with the soundprint under normal operating conditions to obtain the center of the normal mode. The Euclidean distance between the unknown soundprint feature and the center of the normal mode is calculated. If the Euclidean distance is greater than a first preset threshold, it is judged as an anomaly. Combined with the sensor position corresponding to the abnormal soundprint, the location of the thermal anomaly in the gas pipeline is obtained through the triangulation algorithm and the thermal anomaly location is sent to the decision module.

4. The artificial intelligence based smart gas monitoring system as claimed in claim 1, wherein, The gas sensing module acquires the set of boundary parameters and the set of environmental parameters in the scene based on various types of sensors during the operation of the gas equipment. Specifically: The environmental parameters near the gas pipeline are acquired by various environmental sensors. These environmental parameters include, but are not limited to, soil moisture, pH value, resistivity, and stray current. By deploying sensor arrays at key boundary nodes of gas equipment, a set of boundary parameters for equipment operation is collected in real time. These boundary parameters include, but are not limited to, pipe wall pressure, temperature, fluid velocity, and flow data at equipment interfaces.

5. The artificial intelligence based smart gas monitoring system as claimed in claim 1 wherein, The gas sensing module sets the residual change rate as a residual norm by taking the residual change rate distribution mean value under normal working conditions as a first-level alarm threshold If the residual change rate exceeds the first-level alarm threshold, the corresponding physical equation type of the current residual anomaly is obtained, and the corresponding fault type is determined, including leakage, structural crack, and local overheating. 6.The AI-based intelligent gas monitoring system according to claim 1, wherein, The linkage control module obtains the optimal control strategy for each actuator in the gas pipeline based on a genetic algorithm. The specific process is as follows: Upon receiving the processing instruction, the actuators, including valves and ventilation equipment, define the actuator's control parameters as decision variables and encode them as chromosomes. These control parameters include action sequence, time, and intensity. The weighted sum of action sequence, time, and intensity is set as the fitness function. For multi-objective conflicts, non-dominated solutions are retained through Pareto optimality. A non-dominated solution is defined as follows: solution A is not worse than solution B on all objectives, and is better on at least one objective; therefore, A dominates B. The undominated solutions constitute the Pareto front, which is incorporated into the genetic algorithm. The genetic algorithm generates 100 random action sequences, times, and intensities. The leakage amount and time for each sequence are calculated using a pipeline simulation model and substituted into the fitness function. The top 10% of fitness values ​​are selected, and sequential crossover is performed on the selected individuals to generate new individuals. Mutation is performed on the new individuals in reverse order to correct any constraint violations, until the optimal control strategy is obtained through iteration. The optimal control strategy includes the optimal action sequence, optimal time, and optimal intensity. The optimal control strategy is then sent to the actuators.

Images