Comprehensive gas management system based on large model

Abstract

The present invention discloses a comprehensive gas management system based on a large model, which relates to the field of gas management technology. The comprehensive gas management system based on a large model includes: a gas monitoring module and a gas safety management module. The present invention automatically calculates and adjusts the monitoring frequency according to the gas usage pattern of the target household and the real-time status of the gas pipeline network, thereby reducing unnecessary data collection and reducing potential risks to user privacy; ensuring the real-time and accuracy of gas data; by integrating necessary data such as gas concentration, pipeline pressure, valve status, and optional data such as ambient sound spectrum, combined with graph attention networks and digital twin models, it realizes deep fusion and deep analysis of multi-source data, thereby improving the comprehensiveness and accuracy of gas safety judgments; and adopts a multi-level safety warning mechanism to effectively improve the sensitivity and reliability of gas safety warnings.

Description

Technical Field

[0001] The present invention relates to the technical field of gas management, and in particular to a large model-based integrated gas management system. Background Art

[0002] Gas, a vital energy source in modern society, is widely used in residential and industrial production. However, its use also carries a number of safety hazards, such as explosions and fires caused by gas leaks, which pose a serious threat to life and property. Therefore, gas safety management is crucial. To effectively prevent and manage these safety hazards, real-time gas data collection and analysis is required, including user gas usage behavior, the operating status of gas equipment, and real-time pipeline network data. However, the frequent collection and use of gas data also raises concerns about user privacy and wastes computing resources. Summary of the Invention

[0003] The present invention aims to provide a comprehensive gas management system based on a large model, which can dynamically regulate gas safety management and adopt a multi-level early warning mechanism to prevent false alarms and missed alarms.

[0004] The integrated gas management system based on the large model includes:

[0005] The gas monitoring module includes a data acquisition unit and a monitoring frequency update unit. The data acquisition unit is used to obtain necessary gas monitoring data and optional gas monitoring data for the target household using an intelligent analysis frequency. At the same time, it monitors the sudden increase in current data of the target household within the current continuous time.

[0006] The monitoring frequency update unit is used to update the intelligent analysis frequency in real time based on abnormal gas fluctuations of the target household and the gas network where the target household is located;

[0007] The gas safety management module includes a basic safety management unit and a modified safety management unit. The basic safety management unit is used to perform the first gas safety management based on the necessary gas monitoring data, the optional gas monitoring data, and the sudden current data to obtain the gas safety judgment result. If the gas safety judgment result exceeds the gas safety judgment range, the second gas safety management is performed; otherwise, the gas safety monitoring continues.

[0008] The specific steps of the revised safety management unit for conducting the second gas safety management include: obtaining historical gas maintenance data and gas safety external data of the gas pipeline network where the target residents are located; revising the gas safety judgment results based on the historical gas maintenance data and gas safety external data to obtain revised gas safety judgment results; and conducting further gas safety management based on the revised gas safety judgment results.

[0009] As a preferred technical solution of the present invention, the specific steps of updating the intelligent analysis frequency in real time include:

[0010] Set the initial analysis frequency;

[0011] Obtain the target household gas concentration at the initial analysis frequency, and calculate the temporal and spatial gradient of gas concentration based on the target household gas concentration;

[0012] Obtain gas data from the target household's gas pipeline network at the initial analysis frequency; construct a gas fluctuation propagation graph based on the gas pipeline network data; and calculate the gas node risk entropy based on the gas fluctuation propagation graph and the graph attention network.

[0013] The initial analysis frequency is updated based on the spatiotemporal gradient of gas concentration, the risk entropy of gas nodes and the dynamic frequency feedback calibration model to obtain the intelligent analysis frequency for the target households.

[0014] As a preferred technical solution of the present invention, the dynamic frequency feedback calibration model includes a data encoding sublayer, a multimodal fusion layer and a feedback control layer;

[0015] The data coding sublayer includes a spatiotemporal gradient coding sublayer and a graph risk coding sublayer. The spatiotemporal gradient coding sublayer is used to extract features of the target household gas concentration and the spatiotemporal gradient of gas concentration to obtain a spatiotemporal gas feature vector. The graph risk coding sublayer is used to extract features of the pipeline network gas fluctuation propagation graph and the gas node risk entropy to obtain a graph risk feature vector.

[0016] The multimodal fusion layer is used to fuse the spatiotemporal gas feature vector and the graph risk feature vector to obtain the gas fluctuation fusion feature;

[0017] The feedback control layer is used to perform feedback control based on the gas fluctuation fusion characteristics to obtain the intelligent analysis frequency.

[0018] As a preferred technical solution of the present invention, the specific steps of performing the first gas safety management include:

[0019] Necessary data for gas monitoring include gas concentration, gas pipeline pressure and gas valve status;

[0020] Optional data for gas monitoring include ambient sound spectrum;

[0021] Determine the safe concentration range corresponding to the gas concentration; determine the safe pressure range corresponding to the gas pipeline pressure;

[0022] A safety judgment is made based on the concentration safety range, pressure safety range, and gas valve status to obtain the gas monitoring judgment result. If the gas monitoring judgment result meets the first-level gas warning, the ambient sound spectrum is matched with the voiceprint to obtain the voiceprint recognition result. Otherwise, no operation is performed.

[0023] If the voiceprint recognition result meets the requirements of the second-level gas warning, the sudden increase current data is combined with the trained abnormal current capture model to analyze the abnormal current judgment result; otherwise, the first-level gas warning is maintained;

[0024] The gas monitoring judgment results, voiceprint recognition results, and abnormal current judgment results output after the first-level gas warning and the second-level gas warning are combined to obtain the gas safety judgment result.

[0025] As a preferred technical solution of the present invention, in the second gas safety management, the specific steps of correcting the gas safety judgment result include:

[0026] Output gas anomaly prior probability based on historical gas maintenance data;

[0027] Based on external gas safety data, feature extraction is performed to obtain the risk density of the pipeline network area and the gas anomaly impact factor;

[0028] Based on the prior probability of gas anomalies, the risk density of the pipeline network area, and the gas anomaly impact factor, a digital twin model is combined to simulate and obtain the impact of gas anomaly diffusion;

[0029] The gas safety judgment result is corrected based on the influence degree of abnormal gas diffusion to obtain a corrected gas safety judgment result.

[0030] As a preferred technical solution of the present invention, the specific steps of training the abnormal current capture model include:

[0031] Collecting several groups of abnormal current identification training samples, each group of abnormal current identification training samples contains stage current data and corresponding current anomaly identification results; combining the several groups of abnormal current identification training samples to obtain an abnormal current identification training set;

[0032] Model training is performed based on the abnormal current recognition training set to obtain the abnormal current capture model to be verified; a model evaluation is performed on the abnormal current capture model to be verified. If the abnormal current capture model to be verified passes the model evaluation, the abnormal current capture model to be verified is used as the trained abnormal current capture model; otherwise, model training is continued using the abnormal current recognition training set.

[0033] As a preferred technical solution of the present invention, the first gas safety management and the second gas safety management are implemented based on the trained large model.

[0034] The present invention has the following advantages:

[0035] 1. This invention automatically calculates and adjusts the monitoring frequency based on the target household's gas usage patterns and the real-time status of the gas pipeline network, thereby reducing unnecessary data collection and potential risks to user privacy. It also ensures the real-time and accuracy of gas data. Through real-time updates of intelligent analysis frequencies, it optimizes resource allocation and improves monitoring efficiency. By integrating necessary data such as gas concentration, pipeline pressure, and valve status, as well as optional data such as ambient sound spectrum, combined with graph attention networks and digital twin models, it achieves deep fusion and analysis of multi-source data, improving the comprehensiveness and accuracy of gas safety assessments. It also adopts a multi-level safety warning mechanism, performing two gas safety assessments and management through a basic safety management unit and a modified safety management unit. On the basis of the first-level warning, it further achieves accurate identification of the second-level warning through voiceprint matching and abnormal current capture models, effectively improving the sensitivity and reliability of gas safety warnings.

[0036] 2. Through the comprehensive application of the dynamic frequency feedback calibration model and the large model, the present invention can automatically adjust the monitoring strategy according to the abnormal fluctuations of the gas pipeline network and historical maintenance data, optimize the gas safety management process, and adapt to the gas safety needs in different scenarios; by adopting advanced data coding, multimodal fusion and feedback control technologies, it improves data processing efficiency and model training effects; through the training and application of the abnormal current capture model, it enhances the system's ability to identify gas anomalies, further improving the intelligent level of gas safety management. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a schematic structural diagram of the large model-based integrated gas management system adopted in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0039] Gas integrated management system based on large model, see Figure 1 Shown, including:

[0040] The gas monitoring module includes a data acquisition unit and a monitoring frequency update unit. The data acquisition unit is used to acquire necessary and optional gas monitoring data for the target household at an intelligent analysis frequency. At the same time, it monitors the target household's current surge data within the current continuous time. The monitoring frequency update unit is used to update the intelligent analysis frequency in real time based on abnormal gas fluctuations in the target household and the gas pipeline network where the target household is located.

[0041] Necessary data for gas monitoring include gas concentration, gas pipeline pressure and gas valve status;

[0042] Optional data for gas monitoring include ambient sound spectrum;

[0043] Essential gas monitoring data is directly collected by gas concentration sensors, high-precision pressure transmitters, and intelligent valve status monitoring units deployed at key gas pipeline nodes. These sensors continuously monitor methane volume concentration percentage, dynamic changes in pipeline pressure, and digital signals of valve opening and closing status. Optional gas monitoring data is obtained by capturing ambient sound waveforms using broadband acoustic sensors installed in residents' kitchens.

[0044] The specific steps for updating the intelligent analysis frequency in real time include:

[0045] Set the initial analysis frequency; the initial analysis frequency is uniform for all households within the gas network where the target household is located. The initial analysis frequency is set uniformly by professional technicians in the field of gas safety management based on industry standards, equipment performance indicators, and historical accident statistics. The setting of the initial frequency must comprehensively consider hardware limitations such as sensor accuracy, communication bandwidth, and edge computing capabilities, while also meeting the minimum monitoring density required by regulations.

[0046] Obtain the target household gas concentration at the initial analysis frequency, and calculate the temporal and spatial gradient of gas concentration based on the target household gas concentration;

[0047] The target household gas concentration is time series data within a certain time range. The concentration change rate at adjacent time points is analyzed based on the target household gas concentration to obtain gas concentration time gradient data, reflecting the rapidity of gas concentration growth over time. The concentration change rate at adjacent time points can be approximately calculated using the difference method, and the gradient of the difference is smoothed using the sliding window method to obtain gas concentration time gradient data. The target household gas concentration is analyzed based on the difference in gas concentration with the adjacent pipeline network nodes to obtain gas concentration spatial gradient data for calculating spatial diffusion trends. The relative distance between the target household and the adjacent pipeline network nodes is calculated. The difference calculation is performed based on the relative distance and gas concentration difference to obtain gas concentration spatial gradient data.

[0048] The gas concentration time gradient data and gas concentration space gradient data are dynamically updated through a sliding window algorithm to generate a two-dimensional matrix to obtain the gas concentration time-space gradient, where each element of the gas concentration time-space gradient represents the concentration change intensity at a specific time and space position.

[0049] Obtain gas data from the target household's gas pipeline network at the initial analysis frequency; construct a gas fluctuation propagation graph based on the gas pipeline network data; and calculate the gas node risk entropy based on the gas fluctuation propagation graph and the graph attention network.

[0050] Gas data in a gas pipeline network contains information such as gas pressure, gas flow rate, and gas valve opening and closing status for each user node in the network. The network gas fluctuation propagation graph is represented by a graph structure. Nodes in the graph represent key locations in the gas pipeline network, such as household access points, pressure regulating stations, and valves. Each node contains attributes such as real-time pressure and flow. Edges in the graph represent pipelines connecting nodes. Edge weights are determined by physical parameters such as pipeline length, diameter, and material, reflecting the propagation resistance of gas fluctuations.

[0051] The pipeline gas fluctuation propagation graph is fed into a graph attention network, where an adaptive attention mechanism analyzes interactions between nodes. Specifically, the graph attention network automatically learns attention weights between adjacent nodes to determine the intensity of gas propagation impacts, such as the specific impact of an upstream valve closure on downstream pressure changes. Furthermore, when aggregating neighbor information, each node in the graph combines its own attributes with the weighted features of its neighbors to achieve feature enhancement and extract high-order representations. During this process, the adaptive attention mechanism dynamically assigns the degree of influence of different neighbors on the target node, thereby improving risk identification capabilities. Furthermore, the risk entropy of each node is calculated using the information entropy formula, reflecting the complexity and uncertainty of the node's gas fluctuation state. Higher entropy values ​​indicate greater potential risk. Finally, an entropy vector is output for each node to form a node risk distribution map. The entropy vectors of all nodes are then fused to obtain a comprehensive risk entropy indicator for the entire gas pipeline network, providing data support for gas safety supervision and dynamic frequency adjustment.

[0052] The initial analysis frequency is recalibrated based on the spatiotemporal gradient of gas concentration, the risk entropy of gas nodes, and the dynamic frequency feedback calibration model to obtain an intelligent analysis frequency for the target household. The dynamic frequency feedback calibration model uses a hierarchical coding architecture to extract and fuse features from multi-source gas data.

[0053] The dynamic frequency feedback calibration model includes a data coding sublayer, a multimodal fusion layer, and a feedback control layer. The data coding sublayer includes a spatiotemporal gradient coding sublayer and a graph risk coding sublayer.

[0054] The spatiotemporal gradient encoding sublayer extracts features from the target household's gas concentration and its spatiotemporal gradient, generating a spatiotemporal gas feature vector. The spatiotemporal gradient encoding sublayer uses a one-dimensional CNN to extract the gas concentration time series features of the target household's gas concentration, generating a burst fluctuation feature map. The spatiotemporal gradient of gas concentration is input into the LSTM network, which iterates through time steps to output a spatial correlation feature vector. The burst fluctuation feature map and the spatial correlation feature vector are fused through a fully connected layer to ultimately output a spatiotemporal gas feature vector of the same dimension.

[0055] A one-dimensional convolutional neural network is used to scan the target household's gas concentration, extracting local temporal patterns such as sudden fluctuations and gradual trends to capture temporal features. A long short-term memory network is used to process the spatiotemporal gradient of gas concentration, identify the direction and rate of concentration diffusion, and capture spatial correlation features. The temporal and spatial correlation features are integrated to output a spatiotemporal gas feature vector. This vector condenses the spatiotemporal evolution characteristics of the household's gas anomaly, reflecting the suddenness or gradualness of the leak. It can also indicate the possible location and diffusion range of the leak point, helping to distinguish between local leaks and pipeline network cascade fluctuations. The graph risk coding sublayer is used to extract features from the pipeline network gas fluctuation propagation graph and the gas node risk entropy to obtain a graph risk feature vector.

[0056] The graph risk encoding sublayer uses a graph attention network to process the pipeline gas fluctuation propagation graph. The input is a graph structured data consisting of node attributes and edge weights. First, each node in the pipeline gas fluctuation propagation graph is traversed. The attention weight of each node and its neighboring nodes is calculated using a multi-head attention mechanism. This captures the gas propagation influence relationship between nodes and outputs a high-order node representation.

[0057] At the same time, a weighted aggregation is performed based on the calculated gas node risk entropy and the high-order node representation of all nodes to obtain a graph risk feature vector; in the weighted aggregation, the physical connection strength and real-time pressure gradient information between nodes are weighted and fused to comprehensively reflect the mutual influence and risk transmission between nodes, which is used to encode the risk distribution and propagation path of the entire pipeline network; the graph risk feature vector can quantify the mutual influence relationship between nodes and identify high-risk areas, providing a pipeline-level basis for frequency adjustment; to ensure the dimensional consistency of the node high-order representation and risk entropy in the weighted aggregation process, the scalar entropy value can be expanded into a vector of the same dimension as the node feature through linear mapping, and then feature fusion is performed; or in the aggregation stage, the entropy value is used as the node weighting coefficient to participate in the construction of the graph-level representation, so as to maintain a reasonable structure and calculation specifications, and ensure that the final graph risk feature vector has a unified dimension and risk interpretability.

[0058] The multimodal fusion layer is used to fuse the spatiotemporal gas feature vector and the graph risk feature vector to obtain the gas fluctuation fusion feature;

[0059] Through the cross-attention mechanism, the weights of the spatiotemporal gas feature vector and the graph risk feature vector are dynamically assigned to perform feature fusion; the spatiotemporal gas feature vector and the graph risk feature vector are used as query and key respectively, and the attention weight matrix is ​​obtained through the dot product attention calculation method, and the weighted gas fluctuation fusion feature is output; wherein, the multimodal fusion layer also includes residual connection and layer normalization operations. For example, when the overall risk of the pipeline network is low, the focus is on the local characteristics of the residents; when there is a cascade risk in the pipeline network, the decision weight of the global feature is increased; after splicing the two feature vectors, a fully connected network is used to compress them into a joint feature of a unified dimension to obtain the gas fluctuation fusion feature; the gas fluctuation fusion feature contains a comprehensive representation of both local anomalies and global risks;

[0060] The feedback control layer is used to perform feedback control based on the gas fluctuation fusion characteristics to obtain the intelligent analysis frequency. The PID algorithm is used in the feedback control layer to process the gas fluctuation fusion characteristics. The gas fluctuation fusion characteristics are first mapped to risk scores through an MLP (multi-layer perceptron), and then the frequency is adjusted through a discrete PID controller: , e(t)=s(t)-threshold; s(t) is the risk score output by the current MLP, and threshold is the preset risk threshold, which is set manually; after continuously receiving the risk score sequence s(t), s(t-1), s(t-2)... output by the dynamic frequency feedback calibration model after analysis at the initial analysis frequency; using the sliding time window integration strategy, the deviation values ​​of the data obtained in multiple time steps of the most recent data acquisition at the initial analysis frequency are stored to obtain Σe(t) to achieve numerical accumulation; K p is the proportional gain coefficient, K i is the integral gain coefficient, K d is the differential gain coefficient, and the specific values ​​of the three coefficients are set manually. Δf is the frequency adjustment amount; after superimposing Δf and the initial analysis frequency, the intelligent analysis frequency is obtained. In this embodiment, the risk threshold threshold=0.65 can be set, and the initial analysis frequency is preferably set to 2Hz (that is, data is collected every 0.5 seconds), and the risk score sequence is continuously obtained; based on the Z risk score values ​​(s(t) (0.72), s(t-1) (0.68), s(t-2) (0.63), ...) obtained when the initial analysis frequency is 2Hz, the calculation is performed (Z data are obtained every 0.5 seconds during the control period, when the initial analysis frequency is 2Hz and the control period is 30 seconds, Z=60), it can be obtained that e(t)=X t (take 0.72-0.65=0.07), the historical cumulative amount Σe(t) is weighted summed with an attenuation coefficient of 0.95, and Σe(t)=Y (take 0.382); based on the set proportional gain coefficient, integral gain coefficient, and differential gain coefficient, calculation is performed (take Kp =1,K i =0.15, K d =0.25), then Δf = 1*0.07+0.15*0.382+0.25*(0.07-0.03) = 0.1373 Hz; the intelligent analysis frequency is 2.14 Hz (reduced after superposition). The above only illustrates the frequency variation of a specific gas fluctuation; actual data will be determined by measurement data during use.

[0061] When the intelligent analysis frequency is 2.14Hz, the sampling interval will be changed until the next control cycle and the intelligent analysis frequency will be recalculated to form a dynamic balance. The control cycle is set manually and can be set to 30 seconds as a cycle.

[0062] The gas fluctuation fusion features are input into the multi-layer perceptron for risk identification and output of risk scores. The PID mechanism is combined to perform dynamic frequency calculation, and the initial analysis frequency is adjusted based on the risk score, ensuring that the frequency is within the hardware allowable range to obtain an intelligent analysis frequency.

[0063] Based on the spatiotemporal gradient of gas concentration and the risk entropy of gas nodes, a dynamic frequency feedback calibration model is combined to adjust the monitoring frequency in real time. This dynamic adjustment ensures real-time and accurate monitoring while avoiding over-monitoring, reducing the frequency and volume of data collection, and protecting user privacy. The system integrates multiple information, including target household gas concentration and gas network data, and extracts spatiotemporal gas and graph risk characteristics through spatiotemporal gradient encoding and graph risk encoding. These characteristics are then fused in a multimodal fusion layer. This fusion of multi-source data improves the comprehensiveness and accuracy of monitoring and provides a reliable basis for subsequent frequency adjustments. The dynamic frequency feedback calibration model, through the feedback control layer, intelligently adjusts the monitoring frequency based on the fused gas fluctuation characteristics, enabling it to adaptively adapt to changes in the gas network and target household gas usage. This intelligent adjustment mechanism not only improves monitoring efficiency but also reduces unnecessary data collection. Dynamic frequency adjustment allows for increased monitoring during peak gas usage periods and decreased monitoring during off-peak periods. This optimized resource utilization not only improves monitoring efficiency but also reduces operating costs, enhancing the overall intelligence of gas management.

[0064] The gas safety management module includes a basic safety management unit and a modified safety management unit. The basic safety management unit is used to perform the first gas safety management based on the necessary gas monitoring data, the optional gas monitoring data, and the sudden current data to obtain the gas safety judgment result. If the gas safety judgment result exceeds the gas safety judgment range, the second gas safety management is performed; otherwise, the gas safety monitoring continues.

[0065] The specific steps of the revised safety management unit for the second gas safety management include: obtaining historical gas maintenance data and gas safety external data of the gas pipeline network where the target residents are located; revising the gas safety judgment results based on the historical gas maintenance data and gas safety external data to obtain revised gas safety judgment results; conducting further gas safety management based on the revised gas safety judgment results; the first gas safety management and the second gas safety management are implemented based on the trained large model.

[0066] Specific steps for conducting the first gas safety management include:

[0067] Determine the safe concentration range corresponding to the gas concentration; determine the safe pressure range corresponding to the gas pipeline pressure;

[0068] The safe concentration range corresponding to the gas concentration and the safe pressure range corresponding to the gas pipeline pressure are determined by manually preset grade rules, which are set according to the pipeline network design standards;

[0069] A safety judgment is made based on the concentration safety range, pressure safety range, and gas valve status to obtain the gas monitoring judgment result. If the gas monitoring judgment result meets the requirements of the first-level gas warning, a voiceprint matching is performed on the ambient sound spectrum to obtain the voiceprint recognition result. Otherwise, no operation is performed. The triggering condition for the first-level gas warning is that any one of the gas concentration, pressure, or valve exceeds the safety range, which triggers the first-level gas warning.

[0070] When a level 1 gas warning is triggered, the system starts collecting ambient sound spectra and performing voiceprint matching. It calculates the Mel-frequency cepstral coefficients of the ambient sound spectrum, extracts gas leakage features, and obtains voiceprint recognition results.

[0071] The step of identifying gas leak characteristics is to train a voiceprint recognition classifier to identify the sound of gas leaks. During the training process, the model learns the differences between gas leak characteristics and other sound characteristics, so that it can accurately identify gas leaks. The training samples of the classifier are several groups of voiceprint recognition samples with labeled results. The training goal is to achieve a set accuracy rate for the model recognition.

[0072] For example, if the concentration and pressure are abnormal and the valve is open, there may be a risk of gas leakage or pipe burst. If the concentration and pressure are abnormal but the valve is closed, the gas may be shut off normally and requires manual review. Among them, pressure drop is the core characteristic of gas leakage and is mutually verified with concentration data.

[0073] If the voiceprint recognition result meets the requirements of the second-level gas warning, the sudden increase current data is combined with the trained abnormal current capture model to analyze the abnormal current judgment result; otherwise, the first-level gas warning is maintained;

[0074] The abnormal current capture model is constructed using an architecture that combines the CNN and BiLSTM models. The model inputs sudden current data and outputs abnormal current judgment results. Within the abnormal current capture model, multi-layer one-dimensional convolution is used to extract local features from the sudden current data, generating a local feature map of the sudden current data. This local feature map is then input into a bidirectional LSTM unit to capture long-term dependencies and output the temporal features of the sudden current data. Finally, the temporal features of the sudden current data are analyzed through weighted aggregation using an attention mechanism and a fully connected layer to obtain the abnormal current probability distribution and the abnormal current judgment result.

[0075] The triggering condition for the secondary gas warning is that the gas leakage matching probability obtained by voiceprint recognition reaches the preset judgment value; after the secondary gas warning is triggered, the sudden increase current data is collected for judgment; the preset judgment value is set manually;

[0076] The specific steps for training the abnormal current capture model include:

[0077] Collecting several groups of abnormal current identification training samples, each group of abnormal current identification training samples contains stage current data and corresponding current anomaly identification results; combining the several groups of abnormal current identification training samples to obtain an abnormal current identification training set;

[0078] Model training is performed based on the abnormal current identification training set to obtain an abnormal current capture model to be verified; a model evaluation is performed on the abnormal current capture model to be verified. If the abnormal current capture model to be verified passes the model evaluation, the abnormal current capture model to be verified is used as the trained abnormal current capture model; otherwise, model training is continued using the abnormal current identification training set; the condition for passing the model evaluation is that the abnormal current capture model's abnormality recognition accuracy for the abnormal current identification training set reaches a preset accuracy value, which is manually set;

[0079] The gas monitoring judgment results, voiceprint recognition results, and abnormal current judgment results output after the first-level gas warning and the second-level gas warning are combined to obtain the gas safety judgment result;

[0080] Among them, if only the first-level gas warning is triggered but the second-level gas warning is not triggered, the gas monitoring judgment result is output as the gas safety judgment result;

[0081] If a level 2 gas warning is triggered after a level 1 gas warning is triggered, a judgment is made based on the voiceprint recognition result, including two situations: if the voiceprint recognition result meets the level 2 gas warning, the result at this time is a level 1 gas warning and a level 2 gas warning, and the gas monitoring judgment result, voiceprint recognition result, and abnormal current judgment result are output as the gas safety judgment result; if the voiceprint recognition result does not meet the level 2 gas warning, the result at this time is still a level 1 gas warning, and the gas monitoring judgment result and voiceprint recognition result are output as the gas safety judgment result;

[0082] By collecting a large number of training samples containing stage current data and corresponding current anomaly identification results, a highly accurate abnormal current capture model can be trained. This model can effectively identify current anomalies and accurately determine whether there is a gas leak or other abnormal situation after the second-level gas warning is triggered. Based on the first-level warning, voiceprint matching and abnormal current capture model are further used to achieve accurate identification of the second-level warning, effectively improving the sensitivity and reliability of gas safety warnings.

[0083] In the second gas safety management, the specific steps for revising the gas safety judgment results include:

[0084] Outputs the prior probability of gas anomalies based on historical gas maintenance data. This data contains key information such as historical maintenance records, leak reports, and equipment replacement logs for the target household and the network area where it is located. By analyzing this historical data, patterns and trends of abnormal events can be identified. For example, certain areas or seasons are more prone to gas leaks. Using statistical analysis and machine learning algorithms, the prior probability of a gas anomaly is calculated—that is, the likelihood of a gas anomaly occurring in the absence of current real-time data. The prior probability provides a historical reference for subsequent anomaly judgments, helping to understand which areas or conditions are more prone to gas anomalies, thereby giving greater weight to these areas when revising gas safety judgments.

[0085] Based on external gas safety data, feature extraction is performed to obtain the risk density of the pipeline network area and the gas anomaly impact factor;

[0086] A risk assessment matrix is ​​constructed based on external gas safety data, and the contribution of each factor to risk is weighted to determine the regional risk density of the pipeline network. The diffusion rate of leaked gas is simulated based on wind speed and building layout, and the impact level is defined by combining population density and the distribution of key facilities. The impact factor of gas anomalies is the combined calculation result of the diffusion coefficient and the severity coefficient.

[0087] Gas safety external data includes gas-related data such as building density, underground pipeline distribution, traffic vibration intensity, temperature and humidity, wind speed, population density, key facility distribution, and emergency resource coverage radius;

[0088] A digital twin model is used to simulate the impact of gas anomaly diffusion based on the prior probability of gas anomalies, the regional risk density of the pipeline network, and the influencing factors of gas anomalies. The digital twin model includes a high-precision three-dimensional simulation system for the gas pipeline network, which contains physical parameters such as pipeline material, diameter, pressure level, and connection topology. The prior probability of gas anomalies, the regional risk density of the pipeline network, and the influencing factors of gas anomalies are input into the digital twin model to simulate the diffusion path and impact range of the anomaly in the pipeline network. For example, if an abnormal increase in gas concentration is detected in a certain area, the model can predict the diffusion direction and speed of the abnormal gas based on the pipeline network structure and meteorological conditions in that area. The simulation can assess the specific impact of the anomaly on the gas pipeline network, including the potentially affected areas, the number of users, and potential safety hazards. This provides dynamic and real-time support for revising gas safety judgment results.

[0089] The digital twin model is built based on digital twin technology. Based on publicly available gas design drawings and geographic information data, a parametric 3D model is generated using BIM modeling software. The 3D model of the gas pipeline network is then generated, including physical parameters such as pipe material, diameter, pressure level, and connection topology. The output is a gas pipeline network grid model that includes topological connection relationships.

[0090] In the digital twin model, the diffusion path and impact range of abnormal conditions in the pipeline network are simulated. The specific steps are as follows: first, the prior probability of gas anomaly is mapped to the pipeline network spatial grid, and the initial leakage probability field is generated through the probability weighted diffusion algorithm; the initial leakage probability field represents the mapping of the prior probability distribution of gas anomaly to the pipeline network spatial grid, reflecting the possibility of leakage at different locations; this field is generated through the probability weighted diffusion algorithm, taking into account the node topology relationship and physical connection path, to form a spatial leakage probability distribution map.

[0091] The pipeline network regional risk density, gas anomaly impact factors, and initial leak probability fields are fed into an adapted CFD solver for simulation output, which produces a dynamically updated gas diffusion thermogram. Based on these inputs, the CFD solver constructs a gas propagation and diffusion model within the pipeline network, simulating the dynamic evolution of gas leaks in complex spatial structures. The resulting thermogram displays a time-varying gas concentration distribution for subsequent risk assessment and impact calculation.

[0092] All generated gas diffusion heat maps (including generation timestamps) are weightedly fused using a time decay function to obtain a time-series cumulative gas diffusion heat map. The time-series cumulative gas diffusion heat map is divided into multiple grids using a quadtree decomposition algorithm, and the gas concentration extreme value ratio, gradient direction consistency, and danger zone coverage are extracted for each grid cell. A spatial convolution kernel (3*3*3 cube) is used to extract the gas concentration extreme value ratio, gradient direction consistency, and danger zone coverage to calculate the instantaneous risk value and obtain the local abnormal diffusion impact. The gas concentration extreme value ratio is the ratio of the maximum concentration in the current grid to the average concentration in the adjacent area. The gradient direction consistency is the degree of consistency in the direction of the local concentration gradient, which is used to measure whether the diffusion is concentrated. The danger zone coverage is the volume or area ratio of the grid that exceeds the safety concentration threshold. The convolution operation extracts the stability and change trend of this value in the local spatial neighborhood. This operation is performed on each grid to form the instantaneous risk value of the local space.

[0093] All calculated local abnormal diffusion influences are normalized according to the maximum possible diffusion influence of gas diffusion to obtain the final gas abnormal diffusion influence, and the final result is standardized to the interval [0, 1].

[0094] Correcting the gas safety judgment result based on the influence of abnormal gas diffusion to obtain a corrected gas safety judgment result;

[0095] Based on the diffusion impact output by the digital twin model, the gas safety judgment results are recalibrated, the risk level of the gas safety judgment can be corrected, and the graded emergency response measures can be re-matched; for example: when the gas safety judgment result is high, medium, or low risk, a re-judgment is made based on the abnormal diffusion impact of the gas, and the original risk level is increased or decreased.

[0096] In this example, assume that an abnormality occurs in the gas pipeline network of a residential area. The following is the specific implementation process of the gas integrated management system based on the large model in this event:

[0097] The monitoring frequency is dynamically adjusted based on the real-time status of the gas pipeline network and the gas usage of target households. In this residential area, the system discovered abnormal fluctuations in gas concentration in a certain area, so the monitoring frequency in that area was increased. Ambient sounds were collected and analyzed, and gas leak characteristics were extracted by calculating the Mel-frequency cepstral coefficients. The system detected high-frequency, harsh sound characteristics in the environment, which highly matched the voiceprint characteristics of gas leaks, triggering a secondary gas warning. After the secondary gas warning was triggered, the system collected sudden current data. Using the trained abnormal current capture model, the system analyzed the current data. Based on historical current data and the abnormal current identification training set, the model identified the correlation between current fluctuations and gas leaks, further confirming the existence of the abnormal situation. Combining the voiceprint recognition results and the abnormal current judgment results, it was preliminarily determined that there was a high possibility of gas leaks in the area, triggering a second round of gas safety management.

[0098] The system retrieved historical gas maintenance data and analyzed the probability of gas anomalies occurring under similar conditions in the past. The results showed that the area had experienced d gas leaks over the past m years, with most of these leaks occurring during the rainy or hot seasons. Combined with current meteorological data, the system calculated a priori probability of n% for the gas anomaly. The system also integrated external gas safety data, including: the current temperature is a°C, the humidity is b%, and the wind speed is c (m / s); the area is low-lying and prone to gas accumulation; the surrounding area is densely populated with residential areas and underground parking lots, resulting in a high risk of gas leaks; the gas pipeline network in the area is old, and some pipelines are corroded. The system extracted features from this data and calculated a risk density of 0.8 for the pipeline network area, indicating a high risk; and a gas anomaly impact factor of 0.9, indicating a high impact.

[0099] The prior probability of gas anomalies, the risk density of the pipeline area, and the gas anomaly impact factor were input into the digital twin model to simulate the diffusion path and impact range of gas anomalies. The simulation results showed that gas leaks could quickly spread to low-lying areas, affecting residents within a radius of e meters. Due to the influence of high temperature and high humidity, the diffusion rate of gas anomalies was relatively fast, potentially affecting more areas within f hours. Due to the dense population in the surrounding area, gas leaks could pose a serious safety hazard. Based on the simulation results, the output gas anomaly diffusion impact was 0.95, and the preliminary gas safety judgment results were revised, the severity of the gas safety judgment results was upgraded, and the warning content was further refined to obtain a revised gas safety judgment result. Among them, m, n, a, b, c, d, e, and f represent specific numbers.

[0100] Among them, large models generally refer to deep learning models with a large number of parameters, such as the GPT series, BERT, etc. These models can understand and generate human language through training on massive data, and are applied to natural language processing tasks. In this embodiment, they are used for comprehensive gas management. In the above embodiment, the first gas safety management and the second gas safety management are carried out based on the trained large model, realizing dynamic adjustment, multi-source data fusion, intelligent anomaly detection, historical and external data-assisted judgment, multi-level safety warning, optimized resource utilization, enhanced emergency response capabilities, support for decision optimization, and enhanced user experience. It significantly improves the intelligence level of gas safety monitoring and management, and provides strong support for the safety management and sustainable development of the gas industry.

[0101] It should be understood that those skilled in the art may make improvements or modifications based on the above description, and all such improvements and modifications shall fall within the scope of protection of the appended claims. Any portion of this specification not described in detail is prior art known to those skilled in the art.

Claims

1. A comprehensive gas management system based on a large model, characterized by: include: The gas monitoring module includes a data acquisition unit and a monitoring frequency update unit. The data acquisition unit is used to obtain necessary gas monitoring data and optional gas monitoring data for the target household using an intelligent analysis frequency. At the same time, it monitors the sudden increase in current data of the target household within the current continuous time. Necessary data for gas monitoring include gas concentration, gas pipeline pressure and gas valve status; Optional data for gas monitoring include ambient sound spectrum; The monitoring frequency update unit is used to update the intelligent analysis frequency in real time based on abnormal gas fluctuations of the target household and the gas network where the target household is located; Gas safety management module, including basic safety management unit and modified safety management unit; The basic safety management unit is used to perform the first gas safety management based on the necessary gas monitoring data, optional gas monitoring data and sudden current data, and obtain the gas safety judgment result; If the gas safety judgment result exceeds the gas safety judgment range, a second gas safety management will be carried out; otherwise, gas safety monitoring will continue; The specific steps for the revised safety management unit to perform the second gas safety management include: obtaining historical gas maintenance data and external gas safety data of the gas pipeline network where the target household is located; revising the gas safety judgment result based on the historical gas maintenance data and the external gas safety data to obtain a revised gas safety judgment result; Conduct further gas safety management based on the revised gas safety judgment results; The specific steps for updating the intelligent analysis frequency in real time include: Set the initial analysis frequency; Obtain the target household gas concentration at the initial analysis frequency, and calculate the temporal and spatial gradient of gas concentration based on the target household gas concentration; Obtain gas data from the target household's gas pipeline network at the initial analysis frequency; construct a gas fluctuation propagation graph based on the gas pipeline network data; and calculate the gas node risk entropy based on the gas fluctuation propagation graph and the graph attention network. The initial analysis frequency is updated based on the spatiotemporal gradient of gas concentration, the risk entropy of gas nodes and the dynamic frequency feedback calibration model to obtain the intelligent analysis frequency for the target households.

2. The large model-based integrated gas management system according to claim 1 is characterized in that: The dynamic frequency feedback calibration model includes a data encoding sublayer, a multimodal fusion layer, and a feedback control layer; The data coding sublayer includes a spatiotemporal gradient coding sublayer and a graph risk coding sublayer; The spatiotemporal gradient coding sublayer is used to extract features of the target household gas concentration and the spatiotemporal gradient of gas concentration to obtain the spatiotemporal gas feature vector. The graph risk coding sublayer is used to extract features of the pipeline network gas fluctuation propagation graph and the gas node risk entropy to obtain the graph risk feature vector. The multimodal fusion layer is used to fuse the spatiotemporal gas feature vector and the graph risk feature vector to obtain the gas fluctuation fusion feature; The feedback control layer is used to perform feedback control based on the gas fluctuation fusion characteristics to obtain the intelligent analysis frequency.

3. The large model-based integrated gas management system according to claim 2 is characterized in that: Specific steps for conducting the first gas safety management include: Determine the safe concentration range corresponding to the gas concentration; determine the safe pressure range corresponding to the gas pipeline pressure; A safety judgment is made based on the concentration safety range, pressure safety range, and gas valve status to obtain the gas monitoring judgment result. If the gas monitoring judgment result meets the first-level gas warning, the ambient sound spectrum is matched with the voiceprint to obtain the voiceprint recognition result. Otherwise, no operation is performed. If the voiceprint recognition result meets the requirements of the second-level gas warning, the sudden increase current data is combined with the trained abnormal current capture model to analyze the abnormal current judgment result; otherwise, the first-level gas warning is maintained; The gas monitoring judgment results, voiceprint recognition results, and abnormal current judgment results output after the first-level gas warning and the second-level gas warning are combined to obtain the gas safety judgment result.

4. The large model-based integrated gas management system according to claim 3 is characterized in that: In the second gas safety management, the specific steps for revising the gas safety judgment results include: Output gas anomaly prior probability based on historical gas maintenance data; Based on external gas safety data, feature extraction is performed to obtain the risk density of the pipeline network area and the gas anomaly impact factor; Based on the prior probability of gas anomalies, the risk density of the pipeline network area, and the gas anomaly impact factor, a digital twin model is combined to simulate and obtain the impact of gas anomaly diffusion; The gas safety judgment result is corrected based on the influence degree of abnormal gas diffusion to obtain a corrected gas safety judgment result.

5. The large model-based integrated gas management system according to claim 4 is characterized in that: The specific steps for training the abnormal current capture model include: Collecting several groups of abnormal current identification training samples, each group of abnormal current identification training samples contains stage current data and corresponding current anomaly identification results; combining the several groups of abnormal current identification training samples to obtain an abnormal current identification training set; Model training is performed based on the abnormal current recognition training set to obtain the abnormal current capture model to be verified; a model evaluation is performed on the abnormal current capture model to be verified. If the abnormal current capture model to be verified passes the model evaluation, the abnormal current capture model to be verified is used as the trained abnormal current capture model; otherwise, model training is continued using the abnormal current recognition training set.

6. The large model-based integrated gas management system according to claim 5, characterized in that: The first gas safety management and the second gas safety management are implemented based on the trained large model.

Images