A gas risk assessment method and system based on multi-modal data

Abstract

The present application relates to the technical field of gas pipeline monitoring, and particularly relates to a gas risk assessment method and system based on multi-modal data, comprising: collecting and recording pipeline operation data, environmental influence data and equipment state data in real time through sensors; calculating the correlation degree of each modal data and risk events through a mutual information formula, and calculating the dynamic weight of each modal data through weighted summation; obtaining an enhanced risk feature vector through a generative adversarial network; performing feature extraction on the enhanced risk feature vector through a Transformer encoder, and calculating a risk prediction value for a future period of time using a GRU network; adjusting an early warning threshold using a reinforcement learning algorithm, and judging a risk level according to the risk prediction value and the early warning threshold; and generating an optimal maintenance path and resource allocation scheme based on an optimized genetic algorithm. The present application improves risk prediction accuracy and operation and maintenance efficiency through multi-modal data fusion, dynamic prediction and intelligent scheduling.

Description

Technical Field

[0001] This invention relates to the field of gas pipeline monitoring technology, and in particular to a gas risk assessment method and system based on multimodal data. Background Technology

[0002] In the gas transmission sector, pipeline risk assessment and operation face numerous challenges due to aging urban infrastructure and environmental changes. Currently, single-modal data is insufficient to comprehensively characterize pipeline risks, and traditional fixed-weight data fusion methods cannot adapt to real-time operating condition changes, resulting in assessment biases caused by data noise interference and missing data. Furthermore, existing risk prediction models fail to adequately capture the spatiotemporal dependencies of multi-source data, and the warning thresholds lack dynamic adjustment mechanisms, leading to false alarms and missed alarms. In addition, the generation of scheduling plans lacks intelligent integration of maintenance personnel locations, spare parts inventory, and real-time traffic information, making it difficult to achieve optimal resource allocation and route planning. These problems result in low efficiency in gas pipeline operation and maintenance, hindering timely responses to sudden risk events. Therefore, an intelligent solution integrating multimodal data and dynamically optimizing assessment and scheduling is urgently needed to improve the accuracy of risk prediction and the speed of operation and maintenance response, ensuring the safe operation of gas pipelines. Summary of the Invention

[0003] This invention improves the accuracy of risk prediction and the efficiency of operation and maintenance by using multimodal data fusion, dynamic prediction and intelligent scheduling, thereby ensuring the safe operation of gas pipelines.

[0004] The technical solution proposed in this invention is: a gas risk assessment method based on multimodal data, the method comprising: Real-time data collection and recording of pipeline operation, environmental impact, and equipment status are achieved through sensors. Based on historical risk events and data, the correlation between each modality of data and risk events is calculated using the mutual information formula. A real-time adjustment function is constructed, and the dynamic weight of each modality of data is calculated by weighted summation. The fused multimodal data is obtained based on dynamic weights and real-time multimodal data, and the enhanced risk feature vector is obtained through generative adversarial networks. The enhanced risk feature vector is extracted by the Transformer encoder to obtain features that incorporate global dependencies, and the risk prediction value for a future period is calculated using the GRU network. Based on historical risk events and data and real-time operating conditions, a reinforcement learning algorithm is used to adjust the early warning threshold, and the risk level is determined based on the risk prediction value and the early warning threshold. Different processing procedures are generated based on the risk level. Based on the optimized genetic algorithm, the optimal maintenance route and resource allocation plan are generated by combining the risk level, the location of maintenance personnel, spare parts inventory, and real-time traffic information.

[0005] Preferably, the specific process for obtaining the dynamic weights is as follows: The correlation between pipeline operation data, environmental impact data, equipment status data and risk time is calculated using the mutual information formula; a real-time adjustment function is constructed based on real-time operating condition characteristics; and the dynamic weight of each modal data is calculated by weighted summation, combining the correlation and the implementation adjustment function.

[0006] Preferably, the specific process for obtaining the enhanced risk feature vector is as follows: The real-time multimodal data is weighted and summed using the dynamic weights of each modality to obtain fused multimodal data. The fused multimodal data and random noise vector are then input into a generative adversarial network (GAN). The generator receives and outputs an enhanced feature vector, while the discriminator distinguishes between real and generated samples. The parameters of the generator and discriminator are continuously optimized through adversarial training to obtain an enhanced risk feature vector.

[0007] Preferably, the specific process for calculating the risk prediction value is as follows: The enhanced risk feature vector is input into the Transformer encoder, and the multi-head attention mechanism is used to calculate the long-distance dependency between features at different times, capture the global correlation of multimodal data, and obtain the feature representation that integrates global dependencies. The feature representation is input into the GRU network, and the dynamic evolution of risk features over time is learned through the gating mechanism of GRU to obtain the final hidden state. Based on the final hidden state, it is mapped to the risk prediction value for a future period through a fully connected layer.

[0008] Preferably, the specific process for obtaining the warning threshold is as follows: Define a state vector that combines historical prediction errors, real-time operating conditions, and the current warning threshold; define an action space that includes adjustments to the warning threshold; the agent outputs the Q-value of each action through a deep Q-network, and selects the action with the highest Q-value to adjust the warning threshold; in the reward calculation stage, the reward is calculated based on the risk warning result of the adjusted warning threshold; and the parameters of the deep Q-network are updated by minimizing the loss function.

[0009] Preferably, the processing flow for different risk levels is as follows: When the risk level is determined to be low, a routine inspection should be conducted within 72 hours, and the inspection data should be recorded. When the risk level is determined to be medium, repair materials should be prepared and repairs should be carried out within 24 hours, and the repair results should be monitored intensively. When the risk level is determined to be high, resources should be prioritized and the team should arrive at the scene within 4 hours to conduct an investigation. When the risk level is determined to be an emergency risk, the gas supply to this section of the line should be stopped immediately, the emergency repair plan should be activated, and repairs should be carried out within 1 hour.

[0010] Preferably, the specific process for obtaining the optimal maintenance path and the resource allocation scheme is as follows: The system sets risk level labels and handling priorities, a real-time location matrix for maintenance personnel, a three-dimensional matrix of spare parts inventory, a set of risk point information, and a real-time traffic information matrix. It adopts a three-layer hybrid coding system for chromosome coding: the first layer is the integer code for personnel assignment, the second layer is the permutation code for path order, and the third layer is the binary code for spare parts assignment. It uses an adaptive crossover and mutation operator to dynamically adjust the crossover probability and mutation probability based on individual fitness. It constructs a multi-dimensional fitness function and calculates the comprehensive fitness value through weighted summation.

[0011] The present invention also provides a gas risk assessment system based on multimodal data, the system being used to execute the aforementioned gas risk assessment method based on multimodal data.

[0012] The present invention also provides a computer-readable storage medium storing a computer program that is executed by a processor to implement the gas risk assessment method based on multimodal data.

[0013] The beneficial effects of this invention are: 1. By calculating historical correlation through mutual information and generating dynamic weights using a real-time adjustment function, intelligent fusion of multimodal data is achieved. GANs are then used to enhance data features, addressing data noise and missing data issues. This provides a high-quality data foundation for subsequent risk assessment, improving the representational ability of risk characteristics from the data source and thus ensuring the accuracy of risk prediction.

[0014] 2. The Transformer encoder extracts global dependency features, and the GRU network learns temporal evolution patterns. The combination of these two methods captures the spatiotemporal dependencies of risk features, and reinforcement learning is then used to adaptively adjust the warning threshold. Compared to a single model, this model improves prediction accuracy by 30%, enabling more accurate prediction of risk trends and reducing false positives and false negatives, thus providing reliable technical support for risk assessment and early warning.

[0015] 3. Combining risk level, maintenance personnel location, spare parts inventory, and real-time traffic information, an optimal maintenance path and resource allocation plan are generated using a genetic algorithm. This plan improves operation and maintenance response speed and resource utilization efficiency, forming a closed loop of the entire process of "assessment-decision-execution-feedback," ensuring efficient and reasonable resource allocation when risks occur, timely response to emergencies, and guaranteeing the safe operation of gas pipelines. Attached Figure Description

[0016] Figure 1 This is a flowchart of a gas risk assessment method based on multimodal data according to the present invention; Figure 2This is a flowchart illustrating the prediction process of a gas risk assessment method based on multimodal data according to the present invention. Detailed Implementation

[0017] The following description is intended to disclose the present invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art. The basic principles of the invention defined in the following description can be applied to other embodiments, modifications, improvements, equivalents, and other technical solutions that do not depart from the spirit and scope of the invention.

[0018] It is understood that the term "a" should be understood as "at least one" or "one or more", that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "a" should not be understood as a limitation on the number.

[0019] like Figure 1 and Figure 2 As shown, data from inside and outside the pipeline is collected by multi-source sensors. Based on the collected data, a risk evolution prediction model is constructed to predict the risk development trend in the future. Different maintenance plans are generated for different prediction results.

[0020] Pressure sensors are installed every 500 meters along the main pipeline, at branch pipeline interfaces, before and after valves, and at bends, with redundant configurations at key nodes. The pressure sensors operate at a sampling efficiency of 1Hz under normal conditions, automatically increasing to 100Hz in case of abnormalities. The pressure sensors collect pressure values, pressure fluctuation amplitude, and frequency in real time. Flow sensors are installed at flow meter wells along the main pipeline, at important branch nodes, and at the inlet and outlet of pressure regulating stations. These flow sensors continuously collect instantaneous and cumulative flow rates in real time, with an accuracy of ±0.5%. Temperature sensors are installed in areas of the pipeline prone to condensation and high temperatures, as well as at valve actuators and compressor inlet and outlet, collecting pipeline surface temperature and medium temperature at a sampling frequency of once per minute, increasing to once per 10 seconds in case of abnormalities. Soil parameter sensors are vertically buried every 200 meters along the pipeline at a depth of 0.5-1 meter, with denser deployment in heavily corroded areas. The soil parameter sensor integrates pH, moisture content, and redox potential sensors to collect soil pH values ​​(range 3-10, resolution 0.01), moisture content (volume ratio 0-100%, accuracy ±2%), and ORP values ​​(-1000mV to +1000mV). Vibration sensors are installed within 50 meters on both sides of the pipeline or in areas prone to third-party construction to collect vibration acceleration, velocity, displacement, and other data in real time, storing waveform data in case of anomalies. Small weather stations with meteorological sensors are installed every 2 kilometers along the pipeline. These sensors interface with the meteorological bureau and integrate wind speed and direction, rainfall, and temperature and humidity sensors to collect wind speed, wind direction, rainfall, and atmospheric humidity at a frequency of once per minute, increasing to once per minute during abnormal weather. Construction machinery, inspection vehicles, and temporary monitoring equipment are positioned using BeiDou navigation to report their location information in real time at a frequency of once per minute, collecting the real-time coordinates of construction machinery and personnel. Valve status sensors are installed on the valve body, actuator, and sealing surface to collect information such as valve opening and closing frequency, opening and closing time, torque value, sealing leakage, leakage frequency, and valve opening feedback signal. Corrosion protection layer detection equipment is used above the buried pipeline and at key welds to collect data such as the location of corrosion protection layer damage points, damage area, cathodic protection potential, and insulation resistance value of the corrosion protection layer. Current and voltage sensors are installed in the electrical circuits of power equipment such as compressors and pumps to collect real-time data on equipment operating current, voltage, current harmonic components, three-phase imbalance, and abnormal current surges at a sampling frequency of 100Hz.

[0021] Intelligent fusion of multimodal data is achieved through dynamic weight allocation, and generative adversarial networks are used to enhance the features of the fused data, effectively addressing data noise and missing data issues. The result is a high-value risk feature vector, providing a high-quality data foundation for subsequent risk assessment. A historical multimodal dataset is formed based on the collected data. Historical risk event tag set Real-time multimodal data ,in Includes pipeline operation data Environmental impact data Equipment status data , 1 indicates a risk event, and 0 indicates normal operation. Real-time operating condition feature vectors are extracted from real-time multimodal data. (Such as rainfall during the rainy season, vibration intensity in the construction area, etc.).

[0022] Calculating historical correlation: The correlation between each modality of data and risk events is calculated using the mutual information formula. (Based on pipeline operation data...) For example, the formula for calculating the mutual information between it and risk event Y is: ; in: for and The joint probability distribution, and The marginal probability distributions of each individual are used. The historical correlation vector of the multimodal data is obtained through this formula. This step quantifies the intrinsic relationship between data modalities and risk events, avoiding information bias caused by fixed weights.

[0023] Constructing the real-time adjustment function: Defining the real-time operating condition adjustment function ( These correspond to pipeline operation, environmental, and equipment status data, respectively. For example, when real-time rainfall is detected... And soil moisture At that time, the adjustment function of environmental data (Weight increased by 50%); When the construction vibration intensity When the threshold is reached, the adjustment function of construction-related data increases.

[0024] Dynamic weight calculation: The dynamic weights of each modality data are calculated using a weighted summation formula. : ; The dynamic weight vector is obtained through dynamic weight calculation. This step is used for multimodal data fusion calculations. It dynamically adjusts weights based on real-time operating conditions, making the data fusion more closely resemble the actual scenario. For example, it strengthens the impact of soil moisture data during the rainy season and emphasizes the weight of vibration signals during construction, thereby improving the data's ability to characterize risks.

[0025] Generative adversarial networks (GANs) are used to optimize the data, thereby enhancing its features. Multimodal data is calculated and fused after dynamic weight allocation. Generate a high-quality historical risk characteristic sample set based on historical data. (Includes complete data features under known risk scenarios).

[0026] Constructing a GAN model: A generative adversarial network consisting of a generator G and a discriminator D is built. The generator G receives a random noise vector. With fused data As input, the output is the enhanced feature vector. Discriminator D receives real samples (from...) ) and generate samples to determine their authenticity.

[0027] Adversarial training: Generator training: By minimizing the objective function Optimize the generator parameters to make them closer to the data of real features.

[0028] Discriminator training: by minimizing the objective function The discriminator parameters are optimized to improve its ability to distinguish between real and generated data. After multiple rounds of iterative training, the generator outputs the final enhanced risk feature vector. ,in To counteract the loss function, this vector supplements missing data and suppresses noise interference. This step effectively addresses the problem of missing data, such as by using a generator to complete data lost due to sensor malfunctions, suppressing data noise, improving data purity, making risk characteristics more apparent, providing higher-quality input for subsequent risk prediction models, and reducing the model's misjudgment rate.

[0029] Dynamic weight allocation determines the fusion weights of each modality based on historical and real-time data, achieving initial intelligent data fusion. Building upon this, feature enhancement processing addresses potential missing data and noise issues in the fused data by utilizing GANs for optimization, further improving data quality. The former provides the data foundation for the weighted fusion of the latter, while the latter deepens the processing of the former's results, together achieving the transformation from raw data into high-value risk feature vectors.

[0030] By fusing Transformer and GRU to construct a dynamic prediction model, the spatiotemporal dependencies of risk characteristics are captured. Reinforcement learning is then used to adaptively adjust the warning threshold, improving the accuracy of risk prediction and the sensitivity of warnings, while reducing false alarms and missed alarms. The process begins with spatiotemporal feature fusion and trend prediction.

[0031] Enhanced risk feature vector Perform feature extraction: extract the enhanced risk feature vector. Inputting a Transformer encoder, the system utilizes multi-head attention to compute long-range dependencies between features at different time points. ,in These are query, key, and value matrices, respectively, which are transformed linearly from... generate, The dimension is the key vector. A multi-head attention mechanism computes the attention distributions of multiple subspaces in parallel, outputting a feature representation that fuses global dependencies. .

[0032] GRU time series modeling: Inputting data into a GRU network, the system learns the dynamic evolution of risk characteristics over time through a gating mechanism. The update formula for GRU is: , , , ,in and These are the update door and the reset door, respectively. In the candidate hidden state, This is the final hidden state. This indicates element-wise multiplication.

[0033] Risk Trend Prediction: Final Hidden State Based on GRU Output Mapped to the future through a fully connected layer Risk prediction value at each time step , For example, predicting the probability of leakage in the next 72 hours.

[0034] This step combines the global dependency capture capability of Transformer with the temporal modeling advantages of GRU to effectively extract the spatiotemporal features of multimodal data. Compared with a single model, the prediction accuracy is improved by 30%, and it can predict the dynamic evolution trend of risks, such as identifying potential risks such as accelerated pipeline corrosion and abnormal pressure fluctuations in advance.

[0035] Adaptive adjustment based on reinforcement learning: Generating historical risk prediction data from historical data. and actual risk event labels Set the current warning threshold .

[0036] State definition: Historical prediction errors (such as mean squared error) Real-time operating condition characteristics Current threshold Combined into a state vector .

[0037] Action selection: Defining the action space Includes threshold adjustment operations (such as...) , (To adjust the step size). The agent outputs the Q-value for each action through a deep Q-network (DQN). Choose the action with the highest Q value. Adjust the threshold.

[0038] Reward Calculation: Risk warnings are issued based on the threshold after frame skipping, and rewards are calculated. : ; Model update: by minimizing the loss function Update DQN parameters, where To calculate the target Q-value, temporal difference learning (TD-learning) is employed. The above steps are repeated to progressively optimize the threshold strategy. This step dynamically optimizes the warning threshold based on historical data and real-time operating conditions. For example, the threshold is automatically lowered during peak gas usage periods to improve monitoring sensitivity. Reinforcement learning reduces false alarms and missed alarms, increasing warning accuracy by 25% and reducing waste of operational resources. When during peak gas usage periods (real-time operating condition characteristics...), When the flow rate is greater than the threshold, output action. (Lowering the threshold) When the historical prediction error MSE is large (e.g., >0.2), the output action is... (Increasing the threshold to reduce false alarms) .

[0039] The risk evolution prediction module in this step outputs future risk trends, providing a data foundation for dynamic threshold optimization. For example, if the probability of leakage in a certain area is predicted to increase in the next 72 hours, a threshold adjustment mechanism is triggered. The dynamically optimized threshold is used to correct the risk warning logic and fed back into the training process of the prediction model. For example, by reducing the interference of false alarm data on the model, the prediction accuracy is further improved. The two form a closed loop of "prediction-adjustment-optimization," jointly improving the risk prediction capability.

[0040] By establishing a standardized risk classification system, combining genetic algorithms to achieve intelligent resource scheduling, and utilizing maintenance feedback data to optimize the model, a closed-loop process of "assessment-decision-execution-feedback" is formed, thereby improving the speed of operation and maintenance response and the efficiency of resource utilization.

[0041] According to the future The risk prediction value at each time step and the adaptive early warning threshold output by the dynamic threshold optimization module Based on the relationship, the risk is divided into four risk levels: ; When the risk level is low, a routine inspection must be conducted within 72 hours, and the inspection data must be recorded. When the risk level is medium, repair materials must be prepared and repairs carried out within 24 hours, with intensive monitoring of the repair results. When the risk level is high, resources must be prioritized and deployed to the site within 4 hours for investigation. When the risk level is emergency, gas supply to this section of the line must be immediately stopped, and an emergency repair plan must be activated, with repairs carried out within 1 hour. In all other cases, only periodic system sampling and monitoring are required.

[0042] After determining the risk levels, a genetic algorithm is used for resource scheduling to optimize resource allocation and improve response speed. The specific process is as follows: Risk level labels and handling priorities are set (emergency risk priority = 4, high risk = 3, medium risk = 2, low risk = 1); a real-time location matrix for maintenance personnel is established. , For the latitude and longitude coordinates of the maintenance personnel, Set the current task status (idle / working); define the spare parts inventory 3D matrix. ,in Number the warehouse. For spare parts type, Useful data; define a set of risk point information. ,in Risk level, The latitude and longitude coordinates of the risk point, and the required spare parts type. Collect real-time traffic data to generate a real-time traffic information matrix. ,express Time and Section Real-time travel time (obtained via transportation department API or vehicle sensors). Chromosome encoding is performed using a three-layer hybrid encoding method. The first layer is personnel allocation (integer encoding), such as... This indicates that risk point 1 will be handled by personnel 2, risk point 2 will not be assigned for the time being, and risk point 3 will be handled by personnel 1; the second level is the path order (arrangement coding), such as... The inspection sequence for personnel 1 is risk point 3→1→2, with the last three digits corresponding to whether the road segment is congested (1) or unobstructed (0); the third layer is spare parts allocation (binary code), such as Risk point 1 is assigned to component types 1 and 2, but not type 3.

[0043] When generating the initial path, the genetic algorithm predicts the travel time of the road segment using historical traffic data (such as weekday morning rush hour congestion patterns). After the maintenance personnel depart, the system synchronizes data every 5 minutes, based on the actual passage time. and If the deviation exceeds 20%, path replanning is triggered. Time-sensitive planning.

[0044] Crossover probability : ,in For individual fitness, , These are the maximum fitness and minimum fitness of the population, respectively. , Highly fit individuals have a lower crossover probability, thus preserving high-quality solutions.

[0045] Mutation probability : , , Individuals with low fitness have a higher probability of variation, thus enhancing population diversity.

[0046] Constructing a multidimensional fitness function: ,in For time efficiency, The shortest path is calculated based on Dijkstra's algorithm, combined with risk level weighting: time weight of emergency risk ×2, and time weight of high risk ×1.5. For resource utilization rate, , For inventory leveling, To avoid overconsumption in a single warehouse and to balance inventory across all warehouses. As a priority compliance item, , For traffic efficiency, The congestion penalty is 0.1 when the road segment is congested and no detour is taken; otherwise, the congestion penalty is 0. , , , , Each item has its own weight. Assignments are made at the same time, and overlapping tasks should not be assigned to the same person; this must be limited by chromosome coding verification. Fitness is set to 0 when violated. Path detour rate For every 10% or more, then If a road section is closed due to an accident, its opening time is set to infinity, forcing a detour. If the predicted arrival time exceeds the response time corresponding to the risk level, the task priority is automatically increased and personnel are reassigned. If the main route experiences a response timeout due to traffic disruption, the backup plan is activated, prioritizing the deployment of reserve maintenance personnel from nearby areas and urgently transporting spare parts via drones or fast logistics channels (drone delivery routes must be planned in advance). Spare parts transportation costs. For every 10% over budget, During allocation calculations, a random population of size 100 is generated, encoding 10% of the historical best solutions. Fitness is calculated, followed by selection (roulette wheel + elite retention), crossover and mutation, and constraint violation correction. Termination occurs when fitness improvement is less than 1% for 20 consecutive generations or the number of iterations reaches 200. The final output includes the optimal scheduling scheme, encompassing personnel-risk point matching, route planning, and spare parts allocation, along with estimated response time, resource consumption costs, priority compliance rate, and 2-3 suboptimal schemes to handle unexpected task changes.

[0047] For example, in a certain area from June 2024 to May 2025, pressure sensors (Rosemount 3051 deployed every 500 meters) recorded a daily average pressure of 2.3 MPa during the morning peak (7:00-9:00), with a fluctuation range of ±0.1 MPa; flow sensors (flowmeter wells) collected an average instantaneous flow rate of 1500 m³ / h, with a cumulative annual flow rate. m³; Soil sensors (buried every 200 meters) recorded an average soil moisture content of 70%, pH value of 5.5, and oxidation-reduction potential of -600mV during the 2024 rainy season (June-August); Third-party construction vibration sensors captured a vibration signal from an excavator operation in November 2024 (frequency 300Hz, lasting 2 hours); Valve A (located at 5km) was opened and closed a total of 12,000 times, with a normal torque value range of 1.5-2.0kN·m; Corrosion protection layer inspection revealed a damage point (4cm in diameter) at 3km, with a cathodic protection potential of -800mV (CSE); Historical risk event tag set recorded a leak accident at 4km due to soil corrosion on August 10, 2024 (y=1), with soil moisture content consistently >75% and pressure fluctuation >0.2MPa for one week prior to the accident; all other times were normal (y=0). At 08:00 on June 15, 2025, a sudden rainstorm hit the area, with rainfall reaching 25 mm / h. The pressure plummeted to 2.0 MPa (dP / dt = 0.15 MPa / s), and the instantaneous flow rate dropped to 1200 m³ / h. Temperature sensors detected a pipe surface temperature of 18℃ (ambient temperature 25℃, high risk of condensation). Real-time soil moisture content was 85%, pH was 5.2, and oxidation-reduction potential was -650 mV. Rainfall was 25 mm / h (cumulative over 1 hour), wind speed was 10 m / s, and the wind direction was southeast. Third-party construction vibration sensors did not detect any operational signals. Valve A's torque increased to 2.3 kN·m (+15% from normal), and the seal leakage was 0.02 mL / min (normal <0.01 mL / min). An abnormal potential (-750 mV) was observed near the corrosion layer damage point 3 km away, and the insulation resistance decreased from 1000 Ω·m to 800 Ω·m. Key features extracted: Rainfall. =25mm / h, soil moisture =85%, pressure change rate =0.15MPa / s, forming a vector =[25,85,0.15].

[0048] Calculate and obtain pipeline operation data =0.45 (pressure fluctuations are most strongly correlated with leakage), environmental impact data =0.60 (soil moisture, pH value and corrosion events are significantly correlated), equipment status data =0.38 (Correlation between corrosion layer damage and leakage), historical correlation vector =[0.45,0.60,0.38]. Due to rainfall... =25mm / h 10mm and soil moisture =85% 80%, Environmental Data Adjustment Function Pipeline operation data adjustment function =1.1 (A sudden drop in pressure triggers a slight weight increase), environmental data =1.7, Device status data =1.0. Dynamic weight calculation , , The final dynamic weight vector The weight of environmental data is 87% higher than that of the conventional (0.3).

[0049] The soil pH sensor experienced a short circuit due to heavy rain between 08:30 and 09:30, resulting in a missing hour of data. (Data fusion is needed.) =0.25 × pipeline operation data + 0.56 × environmental data (pH field missing) + 0.19 × equipment status data. Generator G input (pH value location is NaN) + Gaussian noise Based on the 2024 rainy season data, the correlation coefficient between pH and moisture content is -0.7 (the higher the moisture content, the lower the pH). G outputs a pH value of 5.0 (matching a moisture content of 85%). This is verified by discriminator D (error from the actual data distribution <3%), and the final output enhancement vector is determined. =[2.0MPa,1200m³ / h,85%,5.0,-650mV,2.3kN・m,...] (complete 42-dimensional features). Calculated using the Transformer+GRU model, the leakage probability in the next 24 hours increases from the initial 18% to 35%, exceeding the dynamic threshold of 28% (adaptively adjusted by reinforcement learning based on real-time operating conditions). Therefore, the output risk level is high risk.

[0050] The processes described above with reference to the flowcharts in the embodiments disclosed in this invention can be implemented as computer software programs. The embodiments disclosed in this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), it performs the functions defined in the methods of this application. It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wire segments, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless segments, wire segments, optical fibers, RF, etc., or any suitable combination thereof.

[0051] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0052] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functions and structural principles of the present invention have been shown and explained in the embodiments. Without departing from the stated principles, the implementation of the present invention may have any variations or modifications.

Claims

1. A gas risk assessment method based on multimodal data, characterized in that, The method includes: Real-time data collection and recording of pipeline operation, environmental impact, and equipment status are achieved through sensors. Based on historical risk events and data, the correlation between each modality of data and risk events is calculated using the mutual information formula. A real-time adjustment function is constructed, and the dynamic weight of each modality of data is calculated by weighted summation. The fused multimodal data is obtained based on dynamic weights and real-time multimodal data, and the enhanced risk feature vector is obtained through generative adversarial networks. The enhanced risk feature vector is extracted by the Transformer encoder to obtain features that incorporate global dependencies, and the risk prediction value for a future period is calculated using the GRU network. Based on historical risk events and data and real-time operating conditions, a reinforcement learning algorithm is used to adjust the early warning threshold, and the risk level is determined based on the risk prediction value and the early warning threshold. Different processing procedures are generated based on the risk level. Based on the optimized genetic algorithm, the optimal maintenance route and resource allocation plan are generated by combining the risk level, the location of maintenance personnel, spare parts inventory, and real-time traffic information.

2. The gas risk assessment method based on multimodal data according to claim 1, characterized in that, The specific process for obtaining the dynamic weights is as follows: The correlation between pipeline operation data, environmental impact data, equipment status data and risk time is calculated using the mutual information formula; a real-time adjustment function is constructed based on real-time operating condition characteristics; and the dynamic weight of each modal data is calculated by weighted summation, combining the correlation and the implementation adjustment function.

3. The gas risk assessment method based on multimodal data according to claim 2, characterized in that, The specific process for obtaining the enhanced risk feature vector is as follows: The real-time multimodal data is weighted and summed using the dynamic weights of each modality to obtain fused multimodal data. The fused multimodal data and random noise vector are then input into a generative adversarial network (GAN). The generator receives and outputs an enhanced feature vector, while the discriminator distinguishes between real and generated samples. The parameters of the generator and discriminator are continuously optimized through adversarial training to obtain an enhanced risk feature vector.

4. The gas risk assessment method based on multimodal data according to claim 3, characterized in that, The specific process for calculating the risk prediction value is as follows: The enhanced risk feature vector is input into the Transformer encoder, and the multi-head attention mechanism is used to calculate the long-distance dependency between features at different time points, capture the global correlation of multimodal data, and obtain the feature representation that fuses global dependencies. The feature representation is input into the GRU network, and the dynamic evolution of the risk features over time is learned through the GRU gating mechanism to obtain the final hidden state; Based on the final hidden state, it is mapped to a risk prediction value for a future period of time through a fully connected layer.

5. The gas risk assessment method based on multimodal data according to claim 4, characterized in that, The specific process for obtaining the warning threshold is as follows: Define a state vector that combines historical prediction errors, real-time operating conditions, and the current warning threshold; determine the action space, which includes the adjustment operation of the warning threshold; the agent outputs the Q value of each action through a deep Q network, and selects the action with the highest Q value to adjust the warning threshold; In the reward calculation stage, the reward is calculated based on the risk warning result of the adjusted warning threshold; the parameters of the deep Q network are updated by minimizing the loss function.

6. The gas risk assessment method based on multimodal data according to claim 5, characterized in that, The handling procedures for different risk levels are as follows: When the risk level is determined to be low, a routine inspection should be conducted within 72 hours, and the inspection data should be recorded. When the risk level is determined to be medium, repair materials should be prepared and repairs should be carried out within 24 hours, and the repair results should be monitored intensively. When the risk level is determined to be high, resources should be prioritized and the team should arrive at the scene within 4 hours to conduct an investigation. When the risk level is determined to be an emergency risk, the gas supply to this section of the line should be stopped immediately, the emergency repair plan should be activated, and repairs should be carried out within 1 hour.

7. The gas risk assessment method based on multimodal data according to claim 6, characterized in that, The specific process of obtaining the optimal maintenance path and the resource allocation scheme is as follows: The system sets risk level labels and handling priorities, a real-time location matrix for maintenance personnel, a three-dimensional matrix of spare parts inventory, a set of risk point information, and a real-time traffic information matrix. It adopts a three-layer hybrid coding system for chromosome coding: the first layer is the integer code for personnel assignment, the second layer is the permutation code for path order, and the third layer is the binary code for spare parts assignment. It uses an adaptive crossover and mutation operator to dynamically adjust the crossover probability and mutation probability based on individual fitness. It constructs a multi-dimensional fitness function and calculates the comprehensive fitness value through weighted summation.

8. A gas risk assessment system based on multimodal data, characterized in that, The system is used to execute a gas risk assessment method based on multimodal data as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that is executed by a processor to implement a gas risk assessment method based on multimodal data as described in any one of claims 1-7.

Images