High temporal-spatial resolution leakage behavior modeling and simulation method and platform

Abstract

According to the high-temporal-spatial-resolution leakage behavior modeling and simulation method and platform, preliminary positioning and hot spot recognition are carried out on an easy-to-leak area, an unmanned aerial vehicle and a ground inspection vehicle are combined with an automatic sensor, differential multi-source monitoring is carried out on the ground / underground, and environment and concentration data are collected; and fusing the transposed convolutional neural network and the convolutional long-short-term memory network to construct a space-time prediction model, dynamically simulating and predicting the leakage position, intensity and diffusion trend, and outputting visual early warning. According to the method, the problems of response lag, insufficient space coverage and poor complex environment adaptability are solved, and the leakage behavior recognition and emergency response precision is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of energy transmission facility leakage identification, in particular to a high spatio-temporal resolution leakage behavior modeling and simulation method and platform. BACKGROUND

[0002] In the complex urban space, aboveground and underground pipe networks are widely distributed, and the delivery systems of flammable gases such as gas and oil gas may be overlaid. Once a leak occurs, it is easy to cause major safety accidents such as fire and explosion. Such leakage events have the characteristics of strong suddenness, complex diffusion path, and wide influence range, especially in densely populated areas, which poses a serious threat to public safety. Therefore, after a leak occurs, the leak source location must be quickly identified, effective plugging must be implemented, the crowd in the threatened area must be evacuated in time, and the potential risk area must be dynamically predicted and controlled to minimize the impact of the accident and casualties. The existing technology mainly relies on large-scale sensor deployment to monitor gas and oil gas leakage events in urban aboveground and underground spaces, identify the location and intensity of the leakage source, and identify the diffusion trend. However, under the realistic conditions of complex aboveground-underground space structure and limited sensor deployment, the existing monitoring methods have high deployment costs, poor adaptability, and monitoring accuracy affected by the environment, and still have obvious shortcomings in rapid positioning of leaked gas, diffusion trend perception, and risk range assessment, which cannot meet the needs of "fast, accurate, and wide" in actual emergency disposal.

[0003] The existing patent application document "A gas leakage precise positioning method based on a vehicle-mounted platform" with publication number CN120062557A, the existing method includes: obtaining leakage evaluation data of a plurality of regions of a pipe network to be evaluated, and performing prediction analysis to obtain a leakage prior distribution of the pipe network to be evaluated; continuously obtaining observation gas concentration values of a plurality of path position points of the vehicle-mounted platform on a preset inspection path, and determining whether they are higher than a preset gas concentration threshold; if the observation gas concentration values of each path position point of the vehicle-mounted platform on the preset inspection path are higher than the preset gas concentration threshold, the corresponding is marked as a leakage diffusion point, the diffusion two-dimensional position coordinates and diffusion plume data are obtained, the leakage likelihood function of each leakage diffusion point of the vehicle-mounted platform and the corresponding set leakage region of the pipe network to be evaluated is constructed, and recursive updating processing is performed in combination with the leakage prior distribution to obtain the leakage posterior distribution of the pipe network to be evaluated; if the observation gas concentration values of each path position point of the vehicle-mounted platform on the preset inspection path are not higher than the preset gas concentration threshold, it is not marked; the leakage posterior distribution of the pipe network to be evaluated is comprehensively analyzed to obtain the two-dimensional position coordinates of each leakage source of the pipe network to be evaluated and the corresponding leakage intensity value.

[0004] The existing technology has the following defects: (1) Modeling simplification, insufficient applicability: it adopts a two-dimensional Gaussian diffusion model and a statistical updating method, and the diffusion parameters depend on empirical formulas. Underground environmental parameters such as soil permeability are not considered, making it difficult to cope with complex underground environments and non-Gaussian diffusion characteristics, resulting in large deviations in prediction results in complex scenarios.

[0005] (2) Highly dependent on historical prior data: The inference process relies on the risk database and prior parameters. If the database is not updated in time or the data quality is insufficient, the accuracy and adaptability of the positioning results will be affected, especially in the case of sudden leakage, it is difficult to ensure reliability.

[0006] (3) Only able to speculate on the source of leakage: This method mainly outputs the location of the leakage source, and lacks modeling and prediction of leakage intensity, diffusion trend and evolution process, which is not practical in the face of larger leakage scenarios.

[0007] The existing patent application document with publication number CN120194270A, "Pipeline leakage position positioning method based on excitation signal synchronous acquisition", the existing method includes: loading an excitation signal into a target pipeline to be positioned for a leakage position, and obtaining a pipeline fusion noise signal collected by a plurality of collection points arranged on the target pipeline; wherein the pipeline fusion noise signal is a signal obtained by superimposing an original pipeline noise signal and an excitation signal, the original pipeline noise signal is a noise signal generated by the medium in the target pipeline when leaking from the leakage position during the flow process in the target pipeline; each pipeline fusion noise signal collected by each collection point is respectively divided based on the excitation signal in it, and the pipeline fusion noise signal segment corresponding to each collection point is obtained; the signal transmission time delay of each collection point is calculated according to the setting position of the plurality of collection points, and the signal transmission time delay of each collection point is used to calibrate the phase of the pipeline fusion noise signal segment corresponding to each collection point; the excitation signal is removed from each pipeline fusion noise signal segment corresponding to each collection point after phase calibration, and the original pipeline noise signal segment corresponding to each collection point after phase calibration is obtained; the time delay difference between the original pipeline noise signals reaching the plurality of collection points is obtained by cross-correlation analysis of the original pipeline noise signal segments corresponding to the plurality of collection points after phase calibration, and the leakage position is located in combination with the setting position of the plurality of collection points.

[0008] The existing technology has the following defects: (1) Limited scope of application: This scheme mainly targets the acoustic time delay positioning method inside the pipeline, and is not applicable to open ground environments and ground-underground coupled diffusion scenarios.

[0009] (2) High deployment cost: Excitation signals need to be injected into the pipeline, which not only poses safety risks and operational limitations, but also requires high sensor accuracy, making it difficult to promote and apply in large-scale scenarios.

[0010] The aforementioned existing technologies for locating and monitoring gas and oil leaks in urban above-ground and underground spaces mainly rely on fixed sensors to collect gas concentration data and combine them with algorithms to deduce the location of the leak source. However, this method has significant drawbacks such as response lag, limited spatial coverage, and poor adaptability to complex environments. In particular, in actual leak scenarios, it cannot achieve dynamic, high-resolution reconstruction of the evolution of leaked gas in three-dimensional space over time, limiting a comprehensive understanding of leak behavior and the accuracy of emergency response.

[0011] In summary, existing technologies suffer from technical problems such as delayed response, limited spatial coverage, poor adaptability to complex environments, and limitations in the comprehensive understanding of leakage behavior and the accuracy of emergency response in actual leakage scenarios. Summary of the Invention

[0012] The technical problem to be solved by this invention is: how to solve the technical problems of delayed response, limited spatial coverage, poor adaptability to complex environments, and the limitations on the accuracy of comprehensive understanding of leakage behavior and emergency response in actual leakage scenarios in the prior art.

[0013] This invention solves the above-mentioned technical problems by employing the following technical solution: a high spatiotemporal resolution leakage behavior modeling and simulation method, comprising: S1. Using the area perception module, the area prone to leakage is initially located, the initial location area is delineated, and the leakage hotspot area is identified. S2. Using the monitoring module, gas concentration monitoring is carried out in the leak hotspot area. Differentiated strategies are adopted for above-ground and underground spaces. Combined with automated sensing equipment and data processing operations, high spatiotemporal resolution monitoring is carried out. Through the collaborative operation of UAVs and ground inspection vehicles, multi-source data is collected in the initially located area to obtain environmental information and gas concentration data of the leak area, which are used as input data for leak diffusion prediction. S3. Using the prediction module, a spatiotemporal prediction model for leakage behavior is constructed by integrating transposed convolutional neural networks and convolutional long short-term memory networks. The leakage diffusion prediction data is input into the spatiotemporal prediction model for leakage behavior to perform time prediction, spatial prediction, and leakage scene identification. The location, intensity, and diffusion trend of the leaked gas are dynamically simulated and predicted. S4. Using the visualization module, output a visual representation of the location, intensity, and diffusion trend of the leaked gas, and display information on the prediction of leak behavior.

[0014] This invention models and simulates leakage behavior with high spatiotemporal resolution, enabling dynamic perception and rapid prediction of gas and oil leakage behavior in three-dimensional space with high spatiotemporal resolution under conditions of limited number of sensors and complex environment, thereby improving emergency response capabilities and risk prevention and control levels in leakage scenarios.

[0015] In a more specific technical solution, in S1, the regional perception module, based on the GIS platform, performs large-scale spatial grid division and constructs a local coordinate reference system; In the GIS platform, the initial location area is delineated based on the gas station itself and the overlapping gas pipelines.

[0016] This invention employs a modular architecture, combining remote sensing and GIS technologies to achieve rapid perception and delineation of leak areas, reducing sensor deployment density and system deployment costs, and is suitable for complex above-ground and underground environments. Through drone inspections and collaborative deployment of underground sensors, high spatiotemporal resolution multi-source data is acquired, providing a solid foundation for model training and prediction.

[0017] In a more specific technical solution, in S1, the grid subdivision operation utilizes GIS spatial analysis and vector grid generation tools to recursively divide the initially located area into spatial parts. Based on remote sensing and UAV collaborative operation, leakage area perception is carried out, and potential leakage area images are initially identified through remote sensing.

[0018] In a more specific technical solution, a GIS platform is used to perform spatial overlay analysis on the image and data layers of the potential leakage area and the pre-generated secondary grid layer, and to extract grid cells whose signals exceed the preset threshold in the thermal infrared and microwave bands, and mark them as potential leakage candidate areas. On the GIS platform, spatiotemporal fusion analysis is performed on the data transmitted back by drones to locate leakage hotspots at the three-level grid level.

[0019] This invention utilizes Geographic Information System (GIS) and Remote Sensing (RS) technologies as auxiliary means to perceive and monitor regional characteristics and pinpoint potential leak areas, thereby improving modeling efficiency and prediction accuracy. This invention effectively overcomes the shortcomings of traditional methods in terms of spatial coverage, timeliness, and dynamic perception, providing strong technical support for rapid response and risk management of gas and oil leak incidents. This invention has advantages such as rapid response, flexible deployment, and high prediction accuracy, effectively solving problems such as response delay, insufficient spatial coverage, and limited three-dimensional diffusion modeling capabilities in existing technologies.

[0020] In a more specific technical solution, in S3, training data for a transposed convolutional neural network model is acquired. Specifically, a three-dimensional numerical model of the above-ground and underground space of a gas station is constructed using computational fluid dynamics simulation software. Based on multiphysics coupling simulation technology, the leakage and diffusion process of gas and oil under different environmental conditions is simulated to generate a spatiotemporal evolution feature simulation dataset. Data point features at specific concentration levels are extracted to generate time-series concentration input data.

[0021] In a more specific technical solution, spatial geometric modeling is performed; multi-parameter coupled working condition simulation is conducted; and the diffusion evolution process of gas under different leakage scenarios in above-ground and underground spaces is simulated through a solver to form a high-dimensional parameter feature simulation sample, thereby obtaining a spatiotemporal evolution feature simulation dataset. Based on the simulation dataset, the concentration distribution at preset locations is sliced ​​and extracted to construct a spatial database; An end-to-end deep learning model is constructed based on a transposed convolutional neural network. The input is multi-source time-series samples from the time-series concentration input data, and the output is two-dimensional gas concentration field images from different observation perspectives. After the model is trained, the network structure and parameter weights are saved to form a deployable concentration field prediction subsystem.

[0022] In terms of modeling, this invention employs a fusion of transposed convolutional neural networks and convolutional long short-term memory networks to construct a spatiotemporal evolution prediction model for leaked gases. This model can dynamically simulate and accurately predict the location of the leak source, leak intensity, diffusion trend, and concentration changes at key points. Furthermore, it can trace the cause and timing of the leak. Combined with a 3D visualization module, the system can display the concentration distribution, diffusion process, and early warning information in real time, facilitating management personnel to understand the leak situation and formulate response strategies. Overall, the system possesses strong practicality, engineering adaptability, and promotional value.

[0023] In a more specific technical solution, during the time prediction process in S3, training data of a convolutional long short-term memory network model is acquired. Through computational fluid dynamics simulation, a three-dimensional spatial top-down model of the above-ground and underground areas is constructed. Based on multiphysics coupling simulation technology, the leakage and diffusion process of gas and oil under different environmental conditions is simulated, thereby generating a simulation dataset with complete spatiotemporal evolution characteristics. By inputting historical concentration map sequences and multi-source auxiliary information, dynamic prediction of the gas concentration field is performed for a future preset time period, and the leakage duration and spread rate are back-estimated for leakage behavior tracing and risk identification.

[0024] In a more specific technical solution, during the leakage scenario identification process in S3, based on the output results of the deep learning model and the estimation of physical parameters, leakage scenarios are classified into four quadrants according to the leakage response level, so as to realize the automatic identification and response strategy suggestions for different types of leakage events. By integrating key output variables from the ConvLSTM and TCNN convolutional long short-term memory network model modules, an interpretable scene classification logic is formed.

[0025] This invention proposes an intelligent identification method integrating Transposed Convolutional Neural Network (TCNN) and Convolutional Long Short-Term Memory Network (ConvLSTM) for spatial localization, temporal identification, and flow mode extraction of gas and oil / gas leaks in above-ground and underground spaces. It also proposes a leak cause tracing analysis strategy based on leak intensity and flow mode predicted by TCNN and propagation velocity predicted by ConvLSTM. This invention utilizes collaborative data collection from UAVs and patrol vehicles, combined with the construction of a spatiotemporal prediction model using TCNN and ConvLSTM, to achieve rapid identification and dynamic simulation of the location, intensity, and diffusion trend of leaked gas.

[0026] In a more specific technical solution, S4 integrates and dynamically displays the multi-source heterogeneous data generated by the regional perception module, monitoring module, and prediction module; and builds a three-dimensional interactive visualization system based on WebGIS to display leakage behavior prediction information.

[0027] In a more specific technical solution, the high spatiotemporal resolution leakage behavior modeling and simulation platform includes: The area perception module is used to initially locate areas prone to leakage, delineate the initial location area, and identify leakage hotspots. The monitoring module is used to monitor gas concentration in leak hotspot areas. It adopts differentiated strategies for above-ground and underground spaces, and combines automated sensing equipment and data processing operations to perform high spatiotemporal resolution monitoring. Through the collaborative operation of UAVs and ground inspection vehicles, it collects multi-source data in the initially located area to obtain environmental information and gas concentration data of the leak area, which serve as input data for leak diffusion prediction. The monitoring module is connected to the area perception module. The prediction module is used to integrate transposed convolutional neural networks and convolutional long short-term memory networks to build a spatiotemporal prediction model for leakage behavior. The leakage diffusion prediction input data is input into the spatiotemporal prediction model for leakage behavior to perform time prediction, spatial prediction and leakage scene identification, dynamically simulate and predict the location, intensity and diffusion trend of the leaked gas, and the prediction module is connected to the monitoring module. The visualization module is used to output a visual representation of the location, intensity, and diffusion trend of leaked gas, and to display prediction information on leakage behavior. The visualization module is connected to the prediction module.

[0028] The present invention has the following advantages over the prior art: This invention models and simulates leakage behavior with high spatiotemporal resolution, enabling dynamic perception and rapid prediction of gas and oil leakage behavior in three-dimensional space with high spatiotemporal resolution under conditions of limited number of sensors and complex environment, thereby improving emergency response capabilities and risk prevention and control levels in leakage scenarios.

[0029] This invention employs a modular architecture, combining remote sensing and GIS technologies to achieve rapid perception and delineation of leak areas, reducing sensor deployment density and system deployment costs, and is suitable for complex above-ground and underground environments. Through drone inspections and collaborative deployment of underground sensors, high spatiotemporal resolution multi-source data is acquired, providing a solid foundation for model training and prediction.

[0030] This invention utilizes Geographic Information System (GIS) and Remote Sensing (RS) technologies as auxiliary means to perceive and monitor regional characteristics and pinpoint potential leak areas, thereby improving modeling efficiency and prediction accuracy. This invention effectively overcomes the shortcomings of traditional methods in terms of spatial coverage, timeliness, and dynamic perception, providing strong technical support for rapid response and risk management of gas and oil leak incidents. This invention has advantages such as rapid response, flexible deployment, and high prediction accuracy, effectively solving problems such as response delay, insufficient spatial coverage, and limited three-dimensional diffusion modeling capabilities in existing technologies.

[0031] This invention proposes an intelligent identification method integrating Transposed Convolutional Neural Network (TCNN) and Convolutional Long Short-Term Memory Network (ConvLSTM) for spatial localization, temporal identification, and flow mode extraction of gas and oil / gas leaks in above-ground and underground spaces. It also proposes a leak cause tracing analysis strategy based on leak intensity and flow mode predicted by TCNN and propagation velocity predicted by ConvLSTM. This invention utilizes collaborative data collection from UAVs and patrol vehicles, combined with the construction of a spatiotemporal prediction model using TCNN and ConvLSTM, to achieve rapid identification and dynamic simulation of the location, intensity, and diffusion trend of leaked gas.

[0032] This invention solves the technical problems existing in the prior art, such as delayed response, limited spatial coverage, poor adaptability to complex environments, and limitations in the accuracy of comprehensive understanding of leakage behavior and emergency response in actual leakage scenarios. Attached Figure Description

[0033] Figure 1 A schematic diagram of the basic steps of the high spatiotemporal resolution leakage behavior modeling and simulation method in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the monitoring position of the monitoring module in Embodiment 1 of the present invention; Figure 3This is a side view of the central axis plane in the X / Y axis direction of Embodiment 1 of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, 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 within the scope of protection of the present invention.

[0035] Example 1 like Figure 1 As shown, the high spatiotemporal resolution leakage behavior modeling and simulation method provided by this invention includes the following basic steps: S1. Using the area perception module, the area prone to leakage is initially located, the initial location area is delineated, and the leakage hotspot area is identified. In this embodiment, the high spatiotemporal resolution leakage behavior modeling and simulation platform adopts a modular design concept, including: a region perception module, a monitoring module, a prediction module, and a visualization module, constructing a complete high spatiotemporal resolution leakage behavior monitoring and prediction technology system. The modules work together to form a closed-loop processing flow.

[0036] The platform enables a complete process from area identification, data collection, leak modeling to result visualization, effectively improving the accuracy and timeliness of leak event response.

[0037] The area sensing module in this embodiment aims to quickly and initially locate potential gas leaks within a large-scale space covered by gas stations and areas overlapping with urban gas pipelines. By integrating Geographic Information System (GIS) and remote sensing technology, the area sensing module can significantly reduce the physical area of ​​focus for subsequent high-precision monitoring modules, thereby effectively reducing the deployment density of the sensor network, the number of devices, and long-term operation and maintenance costs. The core implementation operations of the module also include: Large-scale spatial gridding based on GIS involves establishing a local coordinate system. Specifically, with the geographical location of the target gas station as the center, a local coordinate reference system is constructed using a GIS platform (such as ArcGIS or QGIS) to provide basic support for subsequent spatial analysis and layer overlay.

[0038] Define the initial analysis area; specifically, define an initial square area with a side length of 100 m around the center point of the gas station in the GIS. This area should cover the gas station itself and the space of the gas pipelines that overlap with it.

[0039] Grid subdivision strategy: Using the spatial analysis and vector grid generation tools of GIS, recursively divide the initial area. The specific operations include: dividing the initial 100 m × 100 m area into four square sub-areas (i.e., secondary grid cells) with a side length of 50 m; then, further divide each 50 m × 50 m sub-area into four tertiary grid cells with a side length of 25 m; finally, form a multi-level grid data layer with regular structure and distinct levels, providing a spatial reference framework for subsequent remote sensing analysis and UAV inspection.

[0040] In this embodiment, leakage area perception based on the collaboration of remote sensing and UAVs; among which, remote sensing is used for preliminary screening to identify potential leakage areas. For surface leakage, gas or oil and gas has certain infrared radiation characteristics. The leakage area is usually accompanied by a local increase in surface temperature, and large-scale preliminary scanning can be carried out through thermal infrared remote sensing technology.

[0041] For underground leakage, the escaped hydrocarbon gas will diffuse and accumulate in the soil, significantly changing the dielectric constant of the soil. Such changes can be detected through the abnormal backscattering intensity obtained by active microwave remote sensing (such as SAR) or the brightness temperature difference captured by passive microwave remote sensing.

[0042] Using the GIS platform, perform spatial overlay analysis on the images or data layers collected by remote sensing and the previously generated secondary grid layer, extract the grid cells whose signals exceed the preset threshold in the thermal infrared and microwave bands, and mark them as potential leakage candidate areas.

[0043] In this embodiment, UAVs are used for fine positioning of leakage hotspots; specifically, dispatch multiple UAVs equipped with gas concentration sensors to the above-mentioned candidate areas (i.e., 50 m × 50 m secondary grid cells) to perform refined inspection tasks. The UAVs first fly along the four sides of the square area to obtain edge gas concentration data; then, taking the center of the area as the reference, fly along two mutually perpendicular flight paths connecting the midpoints of the four sides to form a "field" - shaped trajectory for concentration sampling inside the area. The UAVs transmit a multi-dimensional data set including spatial position, timestamp, and corresponding gas concentration to the GIS platform in real time during flight.

[0044] In this embodiment, leakage hotspot positioning at the tertiary grid level is carried out; specifically, perform spatio-temporal fusion analysis on the data transmitted back by the UAVs in GIS, and focus on identifying which 25 m × 25 m tertiary grid cell the concentration peak is distributed in to further lock the specific leakage location. This result will be used as the input basis for the subsequent layout of the high-precision monitoring module to guide the precise deployment of sensors.

[0045] S2. Using the monitoring module, gas concentration monitoring is carried out in the leak hotspot area. Differentiated strategies are adopted for above-ground and underground spaces. Combined with automated sensing equipment and data processing operations, high spatiotemporal resolution monitoring is carried out. Through the collaborative operation of UAVs and ground inspection vehicles, multi-source data is collected in the initially located area to obtain environmental information and gas concentration data of the leak area, which are used as input data for leak diffusion prediction. The core objective of the monitoring module in this embodiment is to accurately and continuously monitor gas concentrations in leak hotspots identified by the area sensing module, providing high-quality, standardized input data for the subsequent leak diffusion prediction module. This module employs differentiated strategies for above-ground and underground spaces, combining automated sensing equipment and data processing workflows to construct a high spatiotemporal resolution monitoring system. Specifically, the detection implementation operations include: Ground space monitoring strategy For surface leak scenarios, multiple drones equipped with high-precision gas concentration sensors were deployed to hover and monitor key locations within a three-level grid area (i.e., 25 m × 25 m). For example... Figure 2 As shown, the UAV hovers approximately 1 meter above five key points—the center point and the midpoints of the four sides—within the three-level grid area of ​​the leak zone determined by the regional perception module; it collects gas concentration data within this altitude range, enhancing its ability to capture surface diffusion patterns; the collected data has strong spatial coverage and can construct a grid-level gas concentration field.

[0046] Underground Space Monitoring Strategy For underground leakage scenarios, such as Figure 2 As shown, monitoring holes with a depth of 1m are pre-installed at key locations (center point and four-sided center points) in the three-level grid area; during the system deployment phase, drilling is completed in advance according to the grid structure to form standardized sensor placement positions; when the remote sensing and UAV joint sensing module identifies the leak area, an inspection robot (or mobile work vehicle) is immediately dispatched to place the sensor into the corresponding hole; this enables the monitoring of the concentration of gas and oil in the underground area, adapting to the characteristics of slow diffusion of underground leaks.

[0047] S3. Using the prediction module, a spatiotemporal prediction model for leakage behavior is constructed by integrating transposed convolutional neural networks and convolutional long short-term memory networks. The leakage diffusion prediction data is input into the spatiotemporal prediction model for leakage behavior to perform time prediction, spatial prediction, and leakage scene identification. The location, intensity, and diffusion trend of the leaked gas are dynamically simulated and predicted. In this embodiment, the prediction module includes a spatial prediction module, a temporal prediction module, and a leakage scene identification module. The spatial prediction module is built based on a transposed convolutional neural network model, the temporal prediction module is built based on a convolutional long short-term memory network model, and the leakage scene identification module is built based on the prediction results of the spatial prediction module and the temporal prediction module.

[0048] In this embodiment, the database construction module within the spatial prediction module is the core source of training data for the transposed convolutional neural network model in this method. A three-dimensional numerical model of the above-ground and underground spaces of a gas station is constructed using computational fluid dynamics simulation software. Based on multiphysics coupling simulation technology, the leakage and diffusion processes of natural gas and oil under different environmental conditions are systematically simulated, thereby generating a simulation dataset with complete spatiotemporal evolution characteristics. Subsequently, by extracting data point features at specific concentration levels, time-series concentration input data suitable for model training is formed. Specifically, the database construction operation includes: Spatial modeling: Construct a three-dimensional cuboid spatial model with dimensions of 25 m × 25 m × 2 m (length × width × height). The spatial mesh division accuracy is controlled within 0.1 m to ensure the resolution and accuracy of the simulation results. Multi-parameter coupled operating condition simulation: Based on the completed geometric modeling, a parametric setting method is adopted to systematically consider environmental variables such as wind speed, wind direction, temperature, and soil porosity, as well as typical operating condition parameters such as leaked gas, leak location, and leak rate, comprehensively constructing various combined operating conditions. The solver simulates the diffusion evolution process of gas under different leak scenarios in above-ground and underground spaces. For each operating condition, the spatial distribution and gradient changes of the concentration field are recorded over a complete time series, forming simulation samples with high-dimensional parametric characteristics. Database construction: Based on the simulation dataset, the concentration distribution at specific locations is extracted by slicing. For example... Figure 3 As shown, concentration field slices are extracted along the X and Y axis centerlines of the above-ground / underground space, forming corresponding side-view concentration gradient change cloud maps within each slice. In each slice, concentration feature values ​​are uniformly extracted along the central axis to construct a concentration feature sample database. The extraction process follows the time series of the diffusion process, sampling every second to form continuous temporal concentration samples. Spatial extraction points are distributed at the center point and four sides of the three-level grid region, with the vertical height fixed at the center of the modeling space. After all sample data is extracted, it is shuffled and reorganized, then divided into training and validation sets for model fitting training and generalization performance evaluation, respectively.

[0049] In this embodiment, the model building and training module constructs an end-to-end deep learning model based on a transposed convolutional neural network (TCNN). The input is multi-source time-series samples generated by the database construction module, and the output is two-dimensional gas concentration field images from different observation perspectives (including side-view mid-axis planes from different directions). After model training is complete, the network structure and parameter weights are saved, forming a deployable concentration field prediction subsystem. The model building and training operations include: Model Construction: Specifically, the proposed model adopts a typical encoder-decoder architecture, supporting the mapping from multi-source input information to high-resolution gas concentration images. The model input fusion includes leakage features (such as location and rate), environmental variables (such as wind speed and temperature), and underground working condition parameters (such as soil permeability and burial depth). The target output is a two-dimensional concentration distribution image under different projection views, which is used to reconstruct the gas diffusion morphology from multiple perspectives.

[0050] The encoding phase is primarily responsible for extracting concentration field features from sensor observation data and environmental condition information. The specific process is as follows: a. During the input phase, sensor monitoring data and operating parameters are concatenated into a unified one-dimensional feature vector; b. Extract high-order features through a multi-layer fully connected network and gradually increase the feature dimension; c. When the final layer is encoded, the feature vector is reshaped into a three-dimensional tensor structure to introduce an initial distribution of spatial features, providing a spatial basis for subsequent image generation.

[0051] The decoder employs a multi-layer two-dimensional transposed convolutional network to progressively upsample the encoded features into high-resolution density images. The core structure is designed as follows: a. Each layer of the decoder uses transpose convolution to restore the image; b. To improve the quality and stability of the output image, non-linear activation functions (such as ReLU) and regularization mechanisms (such as Dropout and Batch Normalization) are introduced. c. A skip connection strategy is added between key layers to preserve local boundary features and multi-scale information, thereby enhancing the ability to express regions with abrupt changes in concentration (such as the boundary of a leakage source). d. The output layer uses the Tanh activation function to achieve normalized output, and combines it with the denormalization operation to restore the true physical concentration range, thus realizing concentration map prediction with physical interpretability.

[0052] The model training operation in this embodiment aims to optimize the network weights so that they can accurately reproduce the gas concentration maps from different perspectives generated by simulation. During training, the model input is the full sample set generated by the database construction module, and the output is the target concentration image at the corresponding time point. The main operations include: a. The loss function uses mean squared error (MSE) to measure the pixel-by-pixel error between the predicted image and the real simulated image; b. To enhance the model's ability to fit the overall concentration distribution, the coefficient of determination (R²) is introduced as an evaluation index to measure the fit of image prediction and the global restoration effect. c. The optimizer uses the Adam algorithm, along with appropriate learning rate settings and training epoch control; d. To prevent overfitting and improve generalization ability, a validation set monitoring mechanism and an EarlyStopping strategy are introduced during training to automatically terminate training when performance stagnates.

[0053] After training, the model can automatically predict gas concentration fields at different profile angles, supporting leak identification and concentration trend judgment at key locations. In practical applications, combined with concentration data obtained from the monitoring module, the model inversion predicts the leakage intensity and diffusion trend on the X and Y central axes. By further combining the intersection area of ​​the two lateral profiles and architectural drawing information, the specific leak point can be quickly located, achieving accurate inversion and scene recognition through multi-dimensional information fusion.

[0054] In this embodiment, the database construction module within the time prediction module is the core source of training data for the convolutional long short-term memory network model. Using computational fluid dynamics simulation software, a three-dimensional overhead model of the above-ground and underground areas is constructed. Based on multiphysics coupling simulation technology, the leakage and diffusion processes of natural gas and oil under different environmental conditions are systematically simulated, thereby generating a simulation dataset with complete spatiotemporal evolution characteristics. Database construction operations within the time prediction module: A two-dimensional spatial model with dimensions of 100 m × 100 m was constructed, and a mesh division accuracy of 2 m was used to ensure that the simulated concentration field has sufficient resolution and accuracy to meet the requirements of spatiotemporal continuity for subsequent model training. Operating condition setup and simulation: A parametric method is used to systematically set environmental variables (such as wind speed, wind direction, temperature, and soil porosity) and leakage conditions (such as leaked gas, leak location, leak rate, and leak start and end times) to construct various typical operating condition combinations. The solver simulates the diffusion and evolution process of gas in the above-ground and underground spaces under different scenarios, and records the three-dimensional concentration field distribution and changes over continuous time series in each operating condition set. Auxiliary Feature Extraction: In each simulation scenario, in addition to recording the concentration distribution at each time step, auxiliary information closely related to diffusion behavior is extracted and recorded, including: leak location (planar coordinates); leak intensity (leakage rate per unit time); leak duration (obtained from the time difference between leak start and end); current environmental variables (such as wind speed, wind direction, and temperature); and propagation velocity estimate: calculated based on the propagation path distance between the leak source and the initial sensor monitoring point and the leak duration, reflecting the average propagation velocity of gas from the source point to the target point. These features not only enhance the model's ability to perceive differences in different operating conditions but also provide support for dynamic scenario adaptation in subsequent deployments. Database Construction: The simulated concentration map data is combined with corresponding auxiliary features to construct a training sample set. Specifically, the concentration images of each grid node at each time step are extracted from the simulation dataset to construct a concentration map sequence sample, which is then bound to auxiliary information in a unified data format. All sample data are shuffled and divided into a training set and a validation set. Each training sample consists of two parts: Input: Concentration map sequence of the first T seconds + auxiliary information (leakage location, intensity, duration, environmental parameters, spread rate); Label section: The concentration map sequence after N seconds, used as the model's prediction target.

[0055] In this embodiment, the model building and training module constructs an integrated concentration field modeling model for leakage diffusion prediction and state identification based on a Convolutional Long Short-Term Memory (ConvLSTM) network. Its goal is to: by inputting historical concentration map sequences and multi-source auxiliary information, achieve dynamic prediction of the gas concentration field within the next few seconds; and to inversely estimate the leakage duration and spread rate, providing support for leak behavior tracing and risk identification. After model training, the network structure and parameters are output, constructing an independent prediction unit that can be directly deployed.

[0056] In the model building operation of this embodiment, the proposed model adopts an encoder-decoder architecture. The input includes: the gas concentration map sequence T seconds before the current time; external auxiliary information (leak location, leak intensity, wind speed, wind direction, temperature, permeability, etc.); the real-time estimated leak duration; and the spread rate model obtained by back-calculation based on the leak location and sensor distance. The output includes: a sequence of predicted concentration maps of length N; an estimate of the current leak duration; and an estimate of the current spread rate.

[0057] In this embodiment, the overall model process is divided into an encoding stage and a decoding stage: The encoding stage extracts the spatiotemporal evolution features of the diffusion process from historical concentration map sequences and integrates multi-source auxiliary information to construct hidden states. The main steps are as follows: a. Input preparation: The input is a sequence of concentration maps with shapes (T, H, W, C); b. Auxiliary information fusion: including leakage location, intensity, environmental parameters, leakage time and propagation speed, etc., which are uniformly encoded and introduced into each layer of the ConvLSTM network; c. Spatiotemporal modeling: Multilayer ConvLSTM extracts the spatiotemporal coupling features of concentration changes; d. State output: Finally, the hidden state h and the memory state c are generated, which serve as the input to the decoder and as the feature basis for the leakage time and velocity estimation branches.

[0058] In the decoding phase, without a future concentration map, a sequence of future concentration maps is gradually generated, and the current leakage behavior parameters are regressed. Main steps: a. Initialization state: Set the decoder's initial state to the encoder's output h and c; b. Autoregressive concentration generation: Predicted images are generated frame by frame according to time steps; c. Leakage Status Regression: Introducing a parallel regression channel to output estimates of the current leakage duration and propagation rate; d. Output combination: The final output consists of a concentration map sequence, a leakage duration estimate, and a propagation rate estimate, which are then used by downstream modules.

[0059] The model training in this embodiment employs a multi-task joint learning strategy to collaboratively optimize the concentration map prediction and leakage state estimation tasks, specifically including: Loss function design: The concentration prediction adopts the mean squared error (MSE) and the structural similarity index (SSIM); The regression of leakage duration and propagation rate uses the mean absolute error (MAE); for critical locations (such as protection points), the prediction accuracy can be weighted to emphasize the accuracy. Training process: Supervised learning is performed using a full sample set generated by the database building module; each sample contains the concentration input sequence, auxiliary information, actual leakage time, and actual spread rate; the Adam optimizer is used for parameter updates, and an early stopping mechanism is introduced to prevent overfitting. After training, the model has multiple capabilities, including predicting future concentration trends based on current monitoring data, inferring leakage duration, and estimating spread rate, demonstrating good generalization performance and engineering practical value.

[0060] In this embodiment, the leak scene identification module proposes a "four-quadrant classification method for leak scene classification" based on the output results of a deep learning model and the estimation of physical parameters, enabling automatic identification and response strategy suggestions for different types of leak events. This module integrates key output variables from the Convolutional Long Short-Term Memory (ConvLSTM) and Transposed Convolutional Neural Network (TCNN) models, including leak intensity, propagation speed, and concentration field image features, to form an interpretable scene classification logic. Specifically, the leak scene identification operation includes: Indicator Acquisition and Calculation: The leakage intensity (leakage rate per unit time, in L / min) is obtained from the TCNN module. The concentration front spread rate (in m / s) is obtained from the ConvLSTM module. This rate is derived by dividing the distance between the leak source location and the contaminated sensor by the predicted leak duration. Obtain the two-dimensional concentration image sequence output by TCNN and extract field structure features (such as plume morphology, gradient steepness, isosurface distribution features, etc.).

[0061] In this embodiment, a four-quadrant classification method is defined. Using "leakage intensity" and "spread rate" as two-dimensional coordinate axes, four typical leakage scenarios are divided: (1) Class I: Catastrophic leak (high intensity + rapid spread); (2) Class II: Jet-type leak (low intensity + rapid spread); (3) Class III: Continuous leakage (high intensity + slow spread); (4) Class IV: Micro-leakage (low intensity + slow spread).

[0062] In this embodiment, concentration field image feature matching verification is performed: (1) Establish typical feature templates of concentration images for four types of scenarios, such as Type I presenting sharp fronts and shock wave-like structures, Type II being high-value channels of plumes, Type III being diffused gentle slopes, and Type IV being weak point source signals; (2) If there is a contradiction between the current predicted concentration map and the classification result (e.g., the speed is fast but the image is diffuse), an anomaly flag will be triggered or a manual review will be prompted.

[0063] In this embodiment, a response suggestion is output: Based on the classification results, the system automatically recommends response strategies. For example, it suggests "minute-level emergency shutdown + large-scale evacuation" for Category I and "weekly monitoring and maintenance" for Category IV, thus enabling automatic judgment of leakage levels and decision support for handling.

[0064] This module not only improves the interpretability and practicality of the model output, but also enables dynamic scene recognition during the scene evolution process, which helps to build a more intelligent leakage response system.

[0065] S4. Using the visualization module, output a visual representation of the location, intensity, and diffusion trend of the leaked gas, and display information on the prediction of leak behavior.

[0066] In this embodiment, the visualization module aims to integrate, display, and dynamically represent the multi-source heterogeneous data generated by various functional modules (regional perception, monitoring, and prediction) of the platform, realizing a visualized presentation of the entire leakage behavior process and providing decision support. By constructing a WebGIS-based 3D interactive visualization system, a real-time feedback loop from data perception and prediction to response is achieved, improving users' ability to perceive the leakage situation and their response efficiency.

[0067] In this embodiment, a data flow and transmission mechanism is implemented; specifically, the visualization module accesses upstream module data through a unified data bus, forming a streaming transmission mechanism of "spatiotemporal joint index + module tag + data hierarchy": In this embodiment, the input data sources include: the hierarchical grid layer and remote sensing annotation results output by the area perception module; the sensor spatial positioning data and gas concentration time series returned by the monitoring module; and the concentration field slice images, top-view concentration map sequences, leakage time estimates, and leakage source inference results generated by the prediction module.

[0068] Data transmission paths are implemented through message queues (such as MQTT / Kafka) or database middleware (such as PostGIS+TimeScaleDB) to ensure real-time performance and consistency.

[0069] The data encapsulation format uniformly adopts standard structures such as GeoJSON, TIFF, NetCDF, and PNG (concentration map) to facilitate rendering and interaction on the client side.

[0070] In the visualization carrier and engine architecture of this embodiment, the visualization module is deployed on a WebGL-based browser-side visualization platform, using CesiumJS as the core engine for the 3D map, and combining Three.js + D3.js to construct multi-level dynamic layers and time-series interactive controls. The overall system architecture includes: Data access layer: responsible for reading output data from various modules from the database or streaming channel, and performing format conversion and cache management; Layer rendering engine: responsible for loading the base map, overlaying image layers, rendering 3D objects, and simulating dynamic density fields; User interaction controller: Provides functions such as timeline control, perspective switching, layer showing / hiding, and risk area query; Event linkage mechanism: Supports users clicking on layers to trigger related data queries, animated highlighting, and early warning information push.

[0071] In the visualization content and expression methods of this embodiment, the visualization module supports the joint expression of the following multiple types and multiple levels of data: Regional perception visualization: Displays the first- to third-level grid structure divided by GIS; highlights suspicious grids identified by thermal infrared / microwave remote sensing; and shows the distribution of UAV flight paths and concentration measurement points.

[0072] Visualization of prediction results: (1) Two-dimensional concentration slice diagram: display the concentration distribution on two cross sections X and Y, and superimpose the time axis to realize the diffusion evolution simulation; (2) Top-down concentration evolution diagram: The animation shows the predicted gas diffusion range and intensity changes within the next N seconds; (3) Leakage point location markers: Highlight suspected leak sources based on the inversion results; (4) Protection target early warning display: Once the predicted concentration exceeds the set threshold, the affected area will be automatically marked and an early warning message will pop up.

[0073] (5) Leakage cause display: The leakage cause inferred based on the leakage source intensity and propagation speed will be displayed in the monitoring area. In the interaction and feedback mechanism of this embodiment, spatiotemporal retrospection and prediction are linked: users can view the concentration distribution at any time through the time slider and compare the deviation between the predicted value and the actual monitored value. Multi-view switching: Supports switching to profile view, top view, and 3D perspective view to achieve all-round observation of leakage and diffusion behavior; Key Area Hotspot Tracking: The platform supports user-defined areas of interest (AOI), and the system automatically tracks the concentration dynamics within them and generates concentration-time curves; Information sharing interface: Supports exporting visualization results as images, videos or data packets, facilitating event reporting, emergency response, or access to government platforms.

[0074] In this embodiment, typical visualization effects include, but are not limited to: 3D overlay layers: Overlay building structures, underground pipelines, monitoring points, and concentration layers onto a 3D map; Dynamic diffusion cloud map: Simulates the diffusion and flow process of gas through semi-transparent particle clouds or thermal animation; Multi-point early warning linkage: When the predicted concentration exceeds the set threshold, the platform automatically marks the area exceeding the standard and issues an alarm through pop-up windows and audio-visual prompts; Spatiotemporal evolution animation: Visualizing the leakage and diffusion process in an integrated manner, encompassing the past, present, and future.

[0075] In summary, compared with the prior art document with publication number CN120062557A, this application solves the following technical problems: (1) To address the problem of insufficient applicability of modeling: This application uses transposed convolutional neural network (TCNN) and convolutional long short-term memory network (ConvLSTM) to construct a spatiotemporal prediction model, and uses commercial simulation software to generate a high-quality simulation dataset (considering underground environmental parameters such as soil permeability) for training, so as to dynamically simulate and predict the location, intensity and diffusion trend of leakage, and output visualized early warning results.

[0076] (2) To address the problem of strong reliance on historical priors: This application uses commercial simulation software to conduct parametric simulations and generate a leakage dataset covering multiple factors and scenarios, thereby reducing the reliance on historical databases and improving the adaptability of the model in sudden scenarios.

[0077] (3) To address the problem of only being able to infer the source of leakage: This application achieves spatial localization, temporal identification and flow mode extraction of above-ground and underground space leakage events through deep fusion of TCNN and ConvLSTM, and further predicts the diffusion trend of future concentration fields, which can provide support for emergency response and disposal of large-scale leakage accidents.

[0078] Compared with the prior art document with publication number [number], this application solves the following technical problems: (1) Regarding the problem of limited applicability: This application realizes integrated modeling of above-ground and underground space. Combining CFD simulation and deep learning methods, it can uniformly simulate and predict the continuous diffusion process of underground soil, surface and air media, and its applicability is significantly better than that of acoustic positioning methods.

[0079] (2) Regarding the high deployment cost: This application only requires inputting real-time concentration data collected by sparse sensors to complete the prediction during the deployment and application stage, and does not require large-scale deployment of sensors in advance. It relies on the automatic completion capability of simulation training and deep learning models to achieve concentration field reconstruction, thereby significantly reducing the cost and implementation threshold of on-site deployment.

[0080] This invention solves the technical problems existing in the prior art, such as delayed response, limited spatial coverage, poor adaptability to complex environments, and limitations in the accuracy of comprehensive understanding of leakage behavior and emergency response in actual leakage scenarios.

[0081] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method of high spatio-temporal resolution leakage behavior modeling and simulation, characterized in that, The method comprises: S1, using a region perception module, preliminarily positioning a leakage-prone region, delimiting a preliminary positioning region, and identifying a leakage hotspot region; S2, using a monitoring module, monitoring the gas concentration of the leakage hotspot region, adopting a differentiated strategy for above-ground space and underground space, combining automatic sensing equipment and data processing operations, and performing high-temporal and high-spatial resolution monitoring; through the cooperative operation of a UAV and a ground patrol vehicle, multi-source data of the preliminary positioning region is collected to obtain environmental information and gas concentration data of the leakage region as leakage diffusion prediction input data; S3, using a prediction module, fusing a transposed convolutional neural network and a convolutional long short-term memory network to construct a leakage behavior spatio-temporal prediction model, inputting the leakage diffusion prediction input data into the leakage behavior spatio-temporal prediction model to perform time prediction, space prediction, and leakage scenario identification, and dynamically simulating and predicting the position, intensity, and diffusion trend of the leakage gas; S4, using a visualization module, outputting the visualization expression of the position, intensity, and diffusion trend of the leakage gas, and displaying the leakage behavior prediction information.

2. The high spatio-temporal resolution leakage behavior modeling and simulation method of claim 1, wherein, In S1, the region perception module is based on a GIS platform to perform large-scale spatial grid division and construct a local coordinate reference system; In the GIS platform, the preliminary positioning region is delimited according to the gas station ontology and the overlaid gas pipeline.

3. The high spatio-temporal resolution leakage behavior modeling and simulation method of claim 1, wherein, In S1, the grid subdivision operation uses the spatial analysis and vector grid generation tools of GIS to implement recursive spatial division on the preliminary positioning region; Based on remote sensing operation and UAV cooperative operation, leakage region perception is performed, and the remote sensing operation is used to preliminarily screen and identify potential leakage region images.

4. The high spatio-temporal resolution leakage behavior modeling and simulation method of claim 3, wherein, Using the GIS platform, the potential leakage region images and data layers are spatially overlaid and analyzed with a pre-generated secondary grid layer to extract grid cells with signals exceeding a preset threshold in the thermal infrared and microwave bands, which are marked as potential leakage candidate regions; In the GIS platform, the UAV return data is analyzed for spatio-temporal fusion to locate a leakage hotspot at the third grid level.

5. The high spatio-temporal resolution leakage behavior modeling and simulation method of claim 1, wherein, In S3, training data of a transposed convolutional neural network model is obtained, wherein a three-dimensional numerical model of the above-ground-underground space of a gas station is constructed by using computational fluid dynamics simulation software, a multi-physical field coupling simulation technology is used to simulate the leakage and diffusion process of gas and oil gas under different environmental conditions, and a time-space evolution feature simulation dataset is generated; data point features at a specific concentration level are extracted to generate time-series concentration input data.

6. The high spatio-temporal resolution leakage behavior modeling and simulation method of claim 5, wherein, Spatial geometric modeling is performed; multi-parameter coupling working condition simulation is performed; the diffusion evolution process of gas in different leakage scenarios in the above-ground space and the underground space is simulated by a solver to constitute a high-dimensional parameter feature simulation sample, and the time-space evolution feature simulation dataset is obtained; Based on the simulation dataset, the concentration distribution of a preset position is sliced and extracted to construct a spatial database; Based on the transpose convolutional neural network, an end-to-end deep learning model is constructed, the input is the multi-source time series sample in the time series concentration input data, and the output is the two-dimensional gas concentration field image under different observation angles. After the model training is completed, the network structure and parameter weight are saved to form a deployable concentration field prediction subsystem.

7. The high spatio-temporal resolution leakage behavior modeling and simulation method of claim 1, wherein, In the time prediction process in S3, the training data of the convolutional long short-term memory network model is obtained, a three-dimensional space overhead-underground space overhead- underground space overhead model is constructed through computational fluid dynamics simulation operation, based on multi-physical field coupling simulation technology, the leakage and diffusion process of gas and oil and gas under different environmental conditions is simulated, and a simulation data set with complete space-time evolution characteristics is generated; input the historical concentration map sequence and multi-source auxiliary information, predict the gas concentration field in the future preset period, and inversely estimate the leakage duration and spread speed for leakage behavior tracing and risk identification.

8. The high spatio-temporal resolution leakage behavior modeling and simulation method of claim 1, wherein, In the leakage scene identification process in S3, based on the output results of the deep learning model and the physical parameter estimation, the leakage scene is classified into four quadrants for leakage response classification, and the automatic identification and response strategy suggestion of different types of leakage events are realized. The key output variables from the convolutional long short-term memory network model module ConvLSTM and the transpose convolutional neural network model module TCNN are fused to form an interpretable scene classification logic.

9. The high spatio-temporal resolution leakage behavior modeling and simulation method of claim 1, wherein, In S4, the multi-source heterogeneous data generated by the regional perception module, the monitoring module and the prediction module are integrated and displayed dynamically; a three-dimensional interactive visualization system is built based on WebGIS to display the leakage behavior prediction information.

10. A high spatio-temporal resolution leakage behavior modeling and simulation platform characterized in that, The platform comprises: a regional perception module for preliminarily positioning the easy-leakage region, delimiting the preliminary positioning region, and identifying the leakage hot spot region; a monitoring module for monitoring the gas concentration of the leakage hot spot region, using differentiated strategies for the above-ground space and the underground space, combining automatic sensing equipment and data processing operations for high spatio-temporal resolution monitoring, and acquiring the environmental information and gas concentration data of the leakage region through the cooperative work of unmanned aerial vehicles and ground patrol vehicles as leakage diffusion prediction input data, the monitoring module being connected with the regional perception module; a prediction module for fusing the transpose convolutional neural network and the convolutional long short-term memory network to construct a leakage behavior spatio-temporal prediction model, inputting the leakage diffusion prediction input data into the leakage behavior spatio-temporal prediction model for time prediction, space prediction and leakage scene identification, dynamically simulating and predicting the position, intensity and diffusion trend of the leakage gas, the prediction module being connected with the monitoring module; a visualization module for outputting the visualization expression of the position, intensity and diffusion trend of the leakage gas, displaying the leakage behavior prediction information, the visualization module being connected with the prediction module.

Images