Intelligent HSE management method for underground gas storage

Abstract

The invention provides an intelligent HSE management method for an underground gas storage. According to the method, the operation condition of the gas storage is monitored in real time through a multi-source data acquisition network, and risk assessment and personalized HSE guidance generation are performed by using an improved Transform model. The system adopts a multi-agent hierarchical reinforcement learning framework to carry out emergency decision support, and introduces a multi-source heterogeneous federated learning framework to realize continuous optimization. According to the method, advanced technologies such as deep learning, knowledge graph and causal reasoning are integrated, and knowledge sharing among multiple gas storages is realized while data privacy is protected. According to the system, the safe operation level of the gas storage is remarkably improved, the accident rate is reduced, and an innovative solution is provided for the progress of industry safety management practice.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial safety management and relates to an intelligent HSE management method for underground gas storage. Background Art

[0002] As natural gas becomes increasingly important in the global energy mix, underground gas storage (UGS), as critical energy storage facilities, is attracting increasing attention for its safe operation and management. Traditional HSE management of UGS relies primarily on regular inspections, fixed safety procedures, and manual monitoring. While this approach provides a basic guarantee for safe operation, its efficiency and accuracy often fall short of the stringent requirements of modern safety management in the face of complex and changing geological conditions and operating environments.

[0003] In recent years, with the development of the Internet of Things and sensor technology, some gas storage facilities have begun to implement automated monitoring systems. These systems typically include surface equipment status monitoring and wellhead pressure monitoring, enabling real-time data collection of key parameters. However, these systems are mostly limited to data collection and simple threshold alarms, lacking the ability to conduct in-depth analysis and intelligent decision-making support for complex scenarios. Furthermore, existing monitoring systems often treat the surface and underground as relatively independent components, making comprehensive risk assessment difficult.

[0004] Regarding HSE guidance, the current approach primarily relies on establishing unified safety operating procedures and emergency response plans. While this approach provides basic safety guidance for employees, it struggles to adapt to the individual needs of employees with varying positions and experience levels, nor can it adapt safety recommendations to real-time environmental changes and operational conditions. Some advanced gas storage facilities have begun experimenting with digital tools to assist with HSE management, such as electronic checklists and training systems, but these tools still lack intelligence and personalization.

[0005] In the area of ​​emergency response, existing systems primarily rely on pre-set emergency plans and manual decision-making. While some gas storage facilities have established emergency command centers equipped with advanced communications equipment, they still face challenges in rapidly analyzing complex situations and developing optimal response strategies. Especially when dealing with complex scenarios such as multi-source risks and cascading failures, existing systems often struggle with satisfactory response speed and decision-making quality.

[0006] In general, although the HSE management level of underground gas storage is constantly improving, existing technologies still have problems such as insufficient data utilization, insufficient in-depth analysis, lack of personalized guidance, and insufficient support for emergency decision-making. Summary of the Invention

[0007] In order to solve the problems existing in the background technology, the present invention proposes an intelligent HSE management method for underground gas storage.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is as follows: An intelligent HSE management method for underground gas storage, including: First data is collected through a surface personnel positioning system, equipment status monitoring sensors, and an underground reservoir parameter monitoring system, and the first data is preliminarily processed and integrated to obtain second data through a combination of edge computing and cloud computing; Real-time evaluation and prediction of potential risks in the second data through an intelligent risk assessment model based on deep learning; Establishing a knowledge graph that can be dynamically updated and includes static information and dynamic knowledge about the gas storage facility. By using graph embedding technology and reasoning algorithms, the knowledge graph provides knowledge support for potential risk analysis and decision-making in the second data. Based on the risk assessment results of the second data and the knowledge graph, a personalized HSE guidance generation system is established using an improved Transformer model to generate targeted safety guidance by combining factors including the location, nature of the task, and personal characteristics of the worker. Set up an intelligent decision support system based on reinforcement learning. By simulating various emergency situations, it learns the optimal response strategy, provides rapid decision-making suggestions, and assists managers in making correct judgments. Establish an adaptive learning mechanism to continuously optimize its risk assessment model, knowledge graph and decision-making strategy by continuously collecting operational data, security incident information and user feedback.

[0009] Furthermore, the specific method of collecting the first data through the surface personnel positioning system, the equipment status monitoring sensor and the underground reservoir parameter monitoring system is: The personnel positioning system locates personnel and mobile devices. Equipment status monitoring sensors are IoT sensors deployed on fixed equipment and some facilities. IoT sensors transmit data to the central processing system through a low-power wide-area network. Each sensor node is equipped with edge computing capabilities to perform preliminary data processing and anomaly detection. The underground reservoir parameter monitoring system uses distributed fiber optic sensing technology. Fiber optics are installed on the outer wall of the casing of the injection and production wells. Raman scattering is used to achieve distributed measurement of temperature, while Brillouin scattering is used to achieve distributed measurement of strain. The temperature change is calculated based on the observation results. The specific formula is: ; in is the temperature change, is the anti-Stokes scattered light intensity, is the intensity of Stokes scattered light, is the temperature correction factor.

[0010] Furthermore, the specific method for preliminarily processing and integrating the first data by combining edge computing and cloud computing to obtain the second data is as follows: The first data is integrated and classified based on the spatiotemporal labels through the Kalman filter algorithm. Its state equation and observation equation are: , ; in is the first data input, is the system state vector, is the state transition matrix, is the process noise, is the observation vector, is the observation matrix, is the observation noise.

[0011] Furthermore, the specific construction method of the intelligent risk assessment model based on deep learning is: The intelligent risk assessment model based on deep learning includes a spatiotemporal graph neural network and a risk assessment model based on causal discovery; The spatiotemporal graph neural network introduces an adaptive attention mechanism and dynamic graph structure learning. The specific formula is: ; in For the The node feature matrix of the layer, is the adaptive adjacency matrix, is the corresponding degree matrix, is the learnable weight matrix, is the activation function; Adaptive adjacency matrix Calculated as follows: ; in and are the spatial and temporal attention matrices learned from node features, and They are and The transposed matrix of and is a learnable balance parameter; The risk assessment model based on causal discovery uses the PC algorithm to construct an initial causal graph, and then uses interventional learning to verify and refine the causal relationship. The formula is as follows: ; in The function represents a probability distribution, is the output variable, Represents a variable Human intervention, that is, artificial Set to value, is the set of confounding variables, is the value of the latent variable; In the risk assessment model of causal discovery, causal strength is introduced as a regulating factor of sample weight. The specific formula is: ; in is the sample weight, represents the time iteration round, is the base learner weight, is the prediction result of the base learner, For samples The true value of For samples The causal strength index, is an adjustable causal influencing factor.

[0012] Furthermore, the specific method by which the knowledge graph can be dynamically updated is: The knowledge graph dynamically updates the relationship strength and adds temporary nodes based on real-time data. The graph update mechanism is based on the following formula: ; in is the updated relationship strength, is the original relationship strength, Based on current data Calculate the relationship strength, is the balance parameter.

[0013] Furthermore, the specific method of establishing a personalized HSE guidance generation system based on the risk assessment results and the knowledge graph using the improved Transformer model is as follows: Based on the improved Transformer model, a multimodal embedding layer, a risk-aware attention mechanism, a knowledge graph enhancement layer, and a personalized adaptation module are introduced; The multimodal embedding layer converts inputs including text descriptions, numerical features, and categorical features into a unified embedding space. For the location information of the staff, spatial encoding is used. The specific formula is: ; in It is the result of spatial position encoding, combined with position encoding and area code , is the position encoding function, Embedding for the learned region; The risk-aware attention mechanism introduces risk assessment results as a regulating factor in the self-attention calculation: ; in 、 and are query, key, and value matrices respectively, is the transpose of the key matrix, is the risk scoring matrix, is a learnable risk modulation matrix, is the dimension of the key vector; The knowledge graph enhancement layer uses the graph attention network to encode the knowledge graph and fuse it with the intermediate representation of the Transformer. The specific formula is: ; in is the updated node feature matrix, is the hidden state of Transformer, is the adjacency matrix of the knowledge graph, For graph attention network; The personalized adaptation module introduces a meta-learning framework to quickly adapt to the characteristics and preferences of different individuals: ; in are the updated model parameters, are model parameters, Yes The operator for finding the gradient, is the loss function, is user-specific support set data, is the learning rate.

[0014] Furthermore, the personalized HSE guidance generation system also includes a continuous learning module, specifically: By collecting user feedback and actual security incident data, we continuously optimize its generation strategy, adopt security reinforcement learning, and set a reward function for the continuous learning module. , the specific formula is: ; in and Measuring the safety, relevance, personalization, and complexity of guidance, is an adjustable weight coefficient.

[0015] Furthermore, the specific method of setting up an intelligent decision support system based on reinforcement learning is: A hierarchical reinforcement learning framework is used as the core of an intelligent decision support system based on reinforcement learning; A layered reinforcement learning framework, including strategic, tactical, and execution layers; Based on the reinforcement learning framework, a geological risk perception module, a gas diffusion predictor, a multimodal decision interpreter, an automatic emergency plan generator, and a collaborative response optimizer are introduced; In this way, the training of the intelligent decision support system based on reinforcement learning is completed. Conservative strategy constraints, safe exploration mechanisms and human-computer collaborative interfaces are added during the training process to complete the construction of the intelligent decision support system based on reinforcement learning.

[0016] Furthermore, the adaptive learning mechanism includes a local learning module, a federated aggregation server, a knowledge distillation module, a differential privacy protection layer, a heterogeneous data adapter, and a continuous evaluation and adjustment module; Each gas storage facility continuously collects local data, and the local learning module regularly updates the model using new local data while adapting to local special conditions through DANN; The updated model parameters and / or gradients are sent to the federated aggregation server after differential privacy processing; The aggregation server uses the AWFA algorithm to integrate updates from various gas storage facilities to generate a global model; The global model is fed back to each gas storage facility through knowledge distillation and integrated with the local risk assessment model. The continuous evaluation and adjustment module monitors the entire process and dynamically adjusts the learning strategy and privacy protection strength.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This invention utilizes innovative technologies such as multi-source data fusion, intelligent risk assessment, personalized HSE guidance, and federated learning to achieve comprehensive, real-time safety monitoring and intelligent management of underground gas storage facilities. The system accurately identifies potential risks, provides timely and personalized safety recommendations, optimizes emergency decision-making processes, and enables knowledge sharing across multiple gas storage facilities while protecting data privacy. This significantly improves the safe operation of gas storage facilities, reduces accident rates, protects personnel safety, ensures stable operation and environmental protection, and promotes advancements in safety management practices across the industry. DETAILED DESCRIPTION

[0018] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] The technical solution adopted by the present invention is as follows: an intelligent HSE management method for underground gas storage, comprising: The first data is collected through the surface personnel positioning system, equipment status monitoring sensors and underground reservoir parameter monitoring system, and the first data is preliminarily processed and integrated through a combination of edge computing and cloud computing to obtain the second data.

[0020] The potential risks in the second data are evaluated and predicted in real time through an intelligent risk assessment model based on deep learning.

[0021] A knowledge graph is established, which can be dynamically updated and contains static information and dynamic knowledge of the gas storage. Through graph embedding technology and reasoning algorithms, knowledge support is provided for potential risk analysis and decision-making in the second data.

[0022] Based on the risk assessment results of the second data and the knowledge graph, a personalized HSE guidance generation system is established using an improved Transformer model to generate targeted safety guidance by combining factors including the location, nature of the task, and personal characteristics of the worker. Set up an intelligent decision support system based on reinforcement learning to simulate various emergency situations, learn the optimal response strategies, quickly provide decision-making suggestions, and assist managers in making correct judgments.

[0023] Establish an adaptive learning mechanism to continuously optimize its risk assessment model, knowledge graph and decision-making strategy by continuously collecting operational data, security incident information and user feedback.

[0024] The data acquisition network of this invention is designed to comprehensively cover the surface and subsurface environments of underground gas storage facilities, enabling multi-dimensional, high-precision, real-time monitoring. At the surface, the system primarily uses high-precision GNSS technology to locate personnel and mobile devices. Each worker wears a smart device equipped with a GNSS receiver capable of receiving signals from the Beidou satellite system.

[0025] To further improve positioning accuracy, the system uses real-time kinematic positioning (RTK) technology and has set up several base stations in the gas storage area. Mobile devices receive differential correction data from these base stations, improving positioning accuracy to centimeter-level accuracy. The device's position is calculated based on the following equation: ; in is the final position solution, is a single point positioning solution, is the differential correction, is the residual error.

[0026] For monitoring of fixed equipment and critical facilities, the system deploys a large number of IoT sensors. These sensors transmit data to the central processing system through low-power wide area network LPWAN technologies such as LoRaWAN or NB-IoT. Each sensor node is equipped with edge computing capabilities to perform preliminary data processing and anomaly detection. For example, for gas concentration monitoring, the sensor uses electrochemical principles, and its output signal and gas concentration The relationship between can be expressed as: ; in is the sensitivity coefficient, is the nonlinear index, The system can maintain high-precision gas concentration measurement through regular calibration.

[0027] For underground reservoir monitoring, this system innovatively utilizes distributed fiber optic sensing technology. Optical fibers are installed on the outer wall of the injection and production well casing, utilizing Raman scattering to achieve distributed temperature measurement, while Brillouin scattering is also used to achieve distributed strain measurement. For temperature measurement, the system calculates based on the following relationship: ; in is the temperature change, is the anti-Stokes scattered light intensity, is the intensity of Stokes scattered light, This method can realize continuous temperature profile monitoring along the well depth with an accuracy of up to 0.1°C.

[0028] To achieve effective fusion of multi-source data, the system uses data alignment technology based on spatiotemporal tags. Each piece of data is assigned a precise timestamp and spatial coordinates, and spatiotemporal indexing enables rapid retrieval and association analysis. Data fusion uses the Kalman filter algorithm, whose state equation and observation equation are: , ; in is the first data input, is the system state vector, is the state transition matrix, is the process noise, is the observation vector, is the observation matrix, In this way, the system can effectively integrate data from different sensors and provide more accurate and reliable monitoring results.

[0029] The entire data collection network is designed with full consideration of the unique environmental and safety requirements of underground gas storage facilities. Through multi-level and multi-dimensional monitoring methods, combined with advanced data processing algorithms, this system provides a comprehensive, accurate, and real-time data foundation for subsequent risk assessment and decision support, significantly enhancing the HSE management of gas storage facilities.

[0030] The data analysis of this invention adopts an innovative multi-level fusion analysis architecture, combining deep learning, knowledge graph and causal reasoning technology to achieve a comprehensive understanding and accurate prediction of the complex system of underground gas storage.

[0031] First, the system uses an improved spatiotemporal graph neural network (ST-GNN) to process multi-source heterogeneous data. The innovation of this network lies in the introduction of an adaptive attention mechanism and dynamic graph structure learning. The core formula of the network is as follows: ; in For the The node feature matrix of the layer, is the adaptive adjacency matrix, is the corresponding degree matrix, is the learnable weight matrix, is the activation function. The adaptive adjacency matrix Calculated as follows: ; in and are the spatial and temporal attention matrices learned from node features, and They are and The transposed matrix of and is a learnable balance parameter, and This design enables the model to automatically capture the dynamic relationships between different elements in the gas storage system, greatly improving the modeling ability of complex situations.

[0032] Secondly, to integrate domain expert knowledge and data-driven insights, we developed a dynamically evolving knowledge graph. This graph not only contains static domain knowledge but also dynamically updates relationship strengths and adds temporary nodes based on real-time data. The graph update mechanism is based on the following formula: ; in is the updated relationship strength, is the original relationship strength, Based on current data Calculate the relationship strength, is the balance parameter. Function An innovative neural symbolic reasoning method is adopted to improve the flexibility and accuracy of reasoning while maintaining interpretability.

[0033] To further enhance the system's predictive capabilities and interpretability, this solution introduces a risk assessment model based on causal discovery. This model first constructs an initial causal graph using the PC algorithm and then verifies and refines causal relationships through interventional learning. The core interventional learning formula is as follows: ; in The function represents a probability distribution, is the output variable, Represents a variable Human intervention, that is, artificial Set to value, is the set of confounding variables, is the value of the latent variable. This approach allows the system to identify true causal relationships, not just correlations, thus providing a more reliable basis for decision-making in risk management.

[0034] Finally, to integrate the outputs of the aforementioned components, this system uses an innovative ensemble learning framework called Causality-Enhanced Adaptive Boosting (CE-AdaBoost). This method builds on traditional AdaBoost by introducing causal strength as a modulating factor for sample weights. Its update formula is as follows: ; in is the sample weight, represents the time iteration round, is the base learner weight, is the prediction result of the base learner, For samples The true value of For samples The causal strength index, is an adjustable causal influence factor. This approach can improve the sensitivity to key causal paths while maintaining the overall predictive performance of the model.

[0035] Through this multi-level, multi-angle analysis approach, the system extracts valuable insights from massive amounts of monitoring data, enabling not only accurate assessments of current risk status but also early warnings of potential safety hazards. For example, the system can identify high-risk situations arising from seemingly unrelated combinations of parameter changes, or predict long-term safety issues potentially caused by geological changes. This in-depth analytical capability provides unprecedented precision and foresight in HSE management of underground gas storage facilities, significantly improving the efficiency and effectiveness of safety management.

[0036] The personalized HSE guidance generation method uses an innovative multimodal context-aware architecture, centered around an improved Transformer model called the Context-Aware Safety Transformer (CAST). This model comprehensively considers multiple factors to generate highly personalized safety guidance.

[0037] The core innovation of the CAST model lies in its unique encoder structure. In addition to the traditional self-attention mechanism, it also introduces the following key components: 1. Multimodal embedding layer: This layer converts different types of input, such as text descriptions, numerical features, and categorical features, into a unified embedding space. For the location information of staff, spatial encoding technology is used: ; in It is the result of spatial position encoding, combined with position encoding and interval coding , is the position encoding function, is the learned region embedding. This method considers both precise coordinates and regional semantic information.

[0038] 2. Risk-aware attention mechanism: Introducing risk assessment results as a regulatory factor in self-attention calculations: ; in 、 and are matrices of queries, keys, and values, respectively, is the transpose of the key matrix, is the risk scoring matrix, is a learnable risk modulation matrix, is the dimension of the key vector. This enables the model to dynamically adjust attention allocation based on the riskiness of different regions and tasks.

[0039] 3. Knowledge graph enhancement layer: Use the graph attention network (GAT) to encode the knowledge graph and fuse it with the intermediate representation of the Transformer: ; in is the updated node feature matrix, for The hidden state of is the adjacency matrix of the knowledge graph, This design enables the model to effectively utilize structured domain knowledge.

[0040] 4. Personalized Adaptation Module: Introduces a meta-learning framework to quickly adapt to the characteristics and preferences of different individuals: ; in are the updated model parameters, are model parameters, Yes The operator for finding the gradient, is the loss function, For user-specific support datasets, is the learning rate. This allows the model to fine-tune its output based on the historical interaction records of each worker.

[0041] During the decoding phase, the model employs a hierarchical generation strategy. It first generates the overall structure of the safety guidance and then gradually refines the content of each section. This approach improves the coherence and structure of the generated content. A safety constraint checker is also incorporated into the generation process to ensure that the generated guidance complies with predetermined safety rules and standards.

[0042] To improve the system's real-time performance and reliability, an edge computing architecture was adopted. A lightweight CAST model was deployed on staff members' mobile devices or nearby edge servers to handle routine safety guidance generation tasks. For complex situations, the full model in the cloud was used for processing.

[0043] In addition, the system includes a continuous learning module. By collecting user feedback and actual security incident data, the system can continuously optimize its generation strategy. It uses a security reinforcement learning framework, and the reward function is designed as follows: ; and Measuring the safety, relevance, personalization, and complexity of guidance, is an adjustable weight coefficient.

[0044] Through this comprehensive design, the system provides highly personalized, real-time, and safety-compliant HSE guidance to every UGS worker. Whether in routine operations or emergencies, the system generates the most appropriate safety recommendations based on current environmental conditions, individual characteristics, and task requirements, significantly improving the accuracy and effectiveness of safety management.

[0045] This intelligent decision support system utilizes a Multi-Agent Hierarchical Reinforcement Learning (MAHRL) architecture to address the complex emergency management scenarios of underground gas storage facilities. The system not only supports decision-making for individual incidents but also coordinates resources across multiple parties to achieve an optimal overall emergency response.

[0046] The core of the system is a layered reinforcement learning framework, consisting of a strategic layer, a tactical layer, and an execution layer: 1. Strategy Layer: Responsible for formulating the overall emergency response strategy, this layer utilizes a transformer-based Strategy Policy Network (SPN). Its state space includes macro-information such as global risk assessment, resource distribution, and personnel location. Its action space encompasses high-level decisions, such as evacuation initiation and resource allocation. The reward function is designed as follows: ; in Measuring overall security posture, Evaluate response speed, Consider resource consumption, 、 、 corresponding to their weights respectively.

[0047] 2. Tactical layer: Responsible for decision-making on specific areas or tasks, it uses a graph convolutional network (GCN) to model local states and actions. Each node represents an area or key equipment, and edges represent the relationship between them. Its value function is defined as: ; in is the state value function, which represents the evaluation value of the entire system state, is the node value function, is the edge value function, is the equilibrium parameter, Indicates the system The status of a component.

[0048] 3. Execution Layer: Responsible for executing specific actions, this layer utilizes a Deep Q-Network (DQN). The state space includes microscopic information such as local sensor data and device status. The action space contains specific operational instructions, such as valve control and pump station adjustment.

[0049] These three layers interact through an innovative "soft layer communication" mechanism. The upper layer strategy affects the reward function of the lower layer through the attention mechanism: ; in It is the basic reward. and are the hidden states of the upper and lower layers, It is the impact factor.

[0050] To address the particularities of underground gas storage, the system introduces the following innovations: 1. Geological risk perception module: Utilizes seismic wave propagation models and reservoir pressure models to assess geological risks in real time and integrate them into the state space.

[0051] 2. Gas Diffusion Predictor: Based on a hybrid model of computational fluid dynamics (CFD) and deep learning, it quickly predicts potential gas leak diffusion paths.

[0052] 3. Multimodal Decision Explainer: Leveraging attention visualization and natural language generation techniques, it provides intuitive explanations for each decision, enhancing the credibility of the system.

[0053] 4. Automatic Emergency Plan Generator: A variational autoencoder based on a generative adversarial network (GAN) can automatically generate and optimize emergency plans based on the current situation.

[0054] 5. Collaborative Response Optimizer: This uses the multi-agent reinforcement learning (MARL) framework to optimize the collaboration between multiple emergency response teams. Its objective function is: ; in is the reward of each agent, is the mutual information term that measures the coordination of actions, is a trade-off parameter, Indicates the Agents at time action, Indicates the expected value, is the discount factor Power.

[0055] The system is trained using an augmented reality simulation approach. It first conducts large-scale pre-training in a high-fidelity virtual environment, then uses transfer learning techniques to adapt the model to the real environment. Training data includes historical accident records, expert knowledge bases, and large-scale simulation data.

[0056] To ensure the robustness and security of the system, the following mechanisms are introduced: 1. Conservative strategy constraints: Add in strategy optimization Divergence constraints to avoid overly aggressive policy updates.

[0057] 2. Safe exploration mechanism: Adopt an exploration strategy based on uncertainty and encourage exploration while ensuring safety.

[0058] 3. Human-computer collaborative interface: An intuitive visual interface and natural language interaction system are designed to enable human experts to intervene and adjust system decisions in real time.

[0059] The adaptive learning mechanism proposed in this paper utilizes the Multi-Source Heterogeneous Federated Learning (MSHFL) framework, designed to continuously optimize the HSE management system of underground gas storage facilities while protecting the data privacy of each facility. This mechanism not only integrates the experience of multiple facilities but also adapts to the unique circumstances and data structure differences of each facility.

[0060] The core architecture of the system includes the following key components: 1. Local learning module: Each gas storage facility deploys a local learning module responsible for processing local data and updating local models. This module employs a hybrid approach of incremental learning and transfer learning. The incremental learning component uses the Elastic Weight Consolidation (EWC) algorithm to balance the learning of new knowledge with the retention of old knowledge. ; in is the total loss function, are model parameters, is the loss function for new data, is the regularization coefficient, is the Fisher information matrix, are the old model parameters. The transfer learning part uses the Domain Adversarial Neural Network to adapt to local special circumstances.

[0061] 2. Federal aggregation server: The central server is responsible for aggregating model updates from various gas storage facilities. Taking into account the differences in data distribution across different gas storage facilities, the aggregation algorithm used in this application is Adaptive Weighted Federated Averaging, with the specific formula being: ; in are global model parameters, is the data volume of each gas storage, It is The aggregation weight coefficient of the local model, are the model parameters that are not updated.

[0062] 3. Knowledge Distillation Module: To address the issue of inconsistent model structures across different gas storage facilities, a model fusion mechanism based on knowledge distillation is introduced. In addition to updating its own model, each gas storage facility also trains a distillation model with a unified structure: ; in is the knowledge distillation loss function, is the cross entropy loss, yes Divergence loss, and are the outputs of the local proprietary model and the learned unified structure model, is the temperature parameter, is the balancing factor, is the softmax activation function.

[0063] 4. Differential Privacy Protection Layer: To further protect sensitive data, differential privacy noise is added before model updates are transmitted: ; in and are the parameters before and after the update, is the clipping threshold, is the noise scale, according to the preset privacy budget Dynamic adjustment, Represents a normal distribution.

[0064] 5. Heterogeneous data adapter: To address the potential differences in data structures across different gas storage facilities, a data structure alignment module based on a graph neural network (GNN) was designed. This module maps data in different formats into a unified graph structure representation: ; in is the updated feature matrix, is the adjacency matrix, is the node feature, is the degree matrix, is a learnable weight matrix, is the activation function.

[0065] 6. Continuous evaluation and adjustment module: The system includes a Meta-Learning framework for continuously evaluating model performance and automatically adjusting learning strategies. Model-independent reinforcement learning algorithms, such as Proximal Policy Optimization, are used to optimize hyperparameters and learning rates: ; in is the loss function of the CLIP algorithm, Represents expectations in the time dimension, is the ratio of the new and old strategies, Function is used to limit the value to a specified range. is the advantage function, is the cropping parameter.

[0066] The system's operating process is as follows: 1. Each gas storage facility continuously collects local data, including daily operation data, safety incident records and user feedback.

[0067] 2. The local learning module regularly updates the model with new data while adapting to local special circumstances through DANN.

[0068] 3. The updated model parameters and / or gradients are sent to the federated aggregation server after differential privacy processing.

[0069] 4. The aggregation server uses the AWFA algorithm to integrate updates from various gas storage facilities to generate a global model.

[0070] 5. The global model is fed back to each gas storage facility through knowledge distillation and integrated with the local risk assessment model.

[0071] 6. The continuous evaluation and adjustment module monitors the entire process and dynamically adjusts the learning strategy and privacy protection strength.

[0072] This design allows the system to leverage the experience of multiple gas storage facilities for continuous learning and optimization while protecting data privacy. For example, the system can quickly migrate and adapt gas leak warning patterns learned from one gas storage facility to other gas storage facilities, while taking into account the unique geological conditions of each gas storage facility.

[0073] The system also includes an "experience library" module for storing and sharing de-identified, high-value safety incident cases and best practices. This information, after undergoing strict privacy protection, can be shared across different gas storage facilities, further promoting the widespread dissemination of knowledge.

[0074] Through this comprehensive adaptive learning mechanism, the HSE management system continuously evolves, continuously improving the accuracy of risk assessments, the comprehensiveness of knowledge graphs, and the effectiveness of decision-making strategies. This not only enhances the safety management of individual gas storage facilities but also promotes the progress of the entire industry, providing strong technical support for the safe operation of underground gas storage facilities.

[0075] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent HSE management method for underground gas storage, characterized by: Includes: First data is collected through a surface personnel positioning system, equipment status monitoring sensors, and an underground reservoir parameter monitoring system, and the first data is preliminarily processed and integrated to obtain second data through a combination of edge computing and cloud computing; Real-time evaluation and prediction of potential risks in the second data through a deep learning-based risk assessment model; Establishing a knowledge graph that can be dynamically updated and includes static information and dynamic knowledge about the gas storage facility. By using graph embedding technology and reasoning algorithms, the knowledge graph provides knowledge support for potential risk analysis and decision-making in the second data. Based on the risk assessment results of the second data and the knowledge graph, a personalized HSE guidance generation system is established using an improved Transformer model to generate targeted safety guidance by combining factors including the location, nature of the task, and personal characteristics of the worker. Set up an intelligent decision support system based on reinforcement learning to simulate various emergency situations, learn the optimal response strategies, provide decision-making suggestions, and assist managers in making correct judgments; Establish an adaptive learning mechanism to continuously optimize its risk assessment model, knowledge graph and decision-making strategy by continuously collecting operational data, security incident information and user feedback.

2. The intelligent HSE management method for underground gas storage according to claim 1 is characterized in that: The specific method of collecting the first data through the surface personnel positioning system, the equipment status monitoring sensor and the underground reservoir parameter monitoring system is: The personnel positioning system locates personnel and mobile devices. Equipment status monitoring sensors are IoT sensors deployed on fixed equipment and some facilities. IoT sensors transmit data to the central processing system through a low-power wide-area network. Each sensor node is equipped with edge computing capabilities to perform preliminary data processing and anomaly detection. The underground reservoir parameter monitoring system uses distributed fiber optic sensing technology. Fiber optics are installed on the outer wall of the casing of the injection and production wells. Raman scattering is used to achieve distributed measurement of temperature, while Brillouin scattering is used to achieve distributed measurement of strain. The temperature change is calculated based on the observation results. The specific formula is: ； in is the temperature change, is the anti-Stokes scattered light intensity, is the intensity of Stokes scattered light, is the temperature correction factor.

3. The intelligent HSE management method for underground gas storage according to claim 1, characterized in that: The specific method for preliminarily processing and integrating the first data by combining edge computing and cloud computing to obtain the second data is as follows: The first data is integrated and classified based on the spatiotemporal labels through the Kalman filter algorithm. Its state equation and observation equation are: ， ； in is the first data input, is the system state vector, is the state transition matrix, is the process noise, is the observation vector, is the observation matrix, is the observation noise.

4. The intelligent HSE management method for underground gas storage according to claim 1, characterized in that: The specific construction method of the intelligent risk assessment model based on deep learning is: The intelligent risk assessment model based on deep learning includes a spatiotemporal graph neural network and a risk assessment model based on causal discovery; The spatiotemporal graph neural network introduces an adaptive attention mechanism and dynamic graph structure learning. The specific formula is: ； in For the The node feature matrix of the layer, is the adaptive adjacency matrix, is the corresponding degree matrix, is the learnable weight matrix, is the activation function; Adaptive adjacency matrix Calculated as follows: ； in and are the spatial and temporal attention matrices learned from node features, and They are and The transposed matrix of and is a learnable balance parameter; The risk assessment model based on causal discovery uses the PC algorithm to construct an initial causal graph, and then uses interventional learning to verify and refine the causal relationship. The formula is as follows: ； in The function represents a probability distribution, is the output variable, Represents a variable Human intervention, that is, artificial Set to value, is the set of confounding variables, is the value of the latent variable; In the risk assessment model of causal discovery, causal strength is introduced as a regulating factor of sample weight. The specific formula is: ； in is the sample weight, represents the time iteration round, is the base learner weight, is the prediction result of the base learner, For samples The true value of For samples The causal strength index, is an adjustable causal influencing factor.

5. The intelligent HSE management method for underground gas storage according to claim 1, characterized in that: The specific method by which the knowledge graph can be dynamically updated is: The knowledge graph dynamically updates the relationship strength and adds temporary nodes based on real-time data. The graph update mechanism is based on the following formula: ； in is the updated relationship strength, is the original relationship strength, Based on current data Calculate the relationship strength, is the balance parameter.

6. The intelligent HSE management method for underground gas storage according to claim 1, characterized in that: The specific method of establishing a personalized HSE guidance generation system based on the risk assessment results and the knowledge graph using the improved Transformer model is as follows: Based on the improved Transformer model, a multimodal embedding layer, a risk-aware attention mechanism, a knowledge graph enhancement layer, and a personalized adaptation module are introduced; The multimodal embedding layer converts inputs including text descriptions, numerical features, and categorical features into a unified embedding space. For the location information of the staff, spatial encoding is used. The specific formula is: ； in It is the result of spatial position encoding, combined with position encoding and area code , is the position encoding function, Embedding for the learned region; The risk-aware attention mechanism introduces risk assessment results as a regulating factor in the self-attention calculation: ； in 、 and are query, key, and value matrices respectively, is the transpose of the key matrix, is the risk scoring matrix, is a learnable risk modulation matrix, is the dimension of the key vector; The knowledge graph enhancement layer uses the graph attention network to encode the knowledge graph and fuse it with the intermediate representation of the Transformer. The specific formula is: ； in is the updated node feature matrix, is the hidden state of Transformer, is the adjacency matrix of the knowledge graph, For graph attention network; The personalized adaptation module introduces a meta-learning framework to quickly adapt to the characteristics and preferences of different individuals: ； in are the updated model parameters, are model parameters, Yes The operator for finding the gradient, is the loss function, is user-specific support set data, is the learning rate.

7. The intelligent HSE management method for underground gas storage according to claim 6, characterized in that: The personalized HSE guidance generation system also includes a continuous learning module, specifically: By collecting user feedback and actual security incident data, we continuously optimize its generation strategy, adopt security reinforcement learning, and set a reward function for the continuous learning module. , the specific formula is: ； in and Measuring the safety, relevance, personalization, and complexity of guidance, is an adjustable weight coefficient.

8. The intelligent HSE management method for underground gas storage according to claim 1, characterized in that: The specific method of setting up an intelligent decision support system based on reinforcement learning is: A hierarchical reinforcement learning framework is used as the core of an intelligent decision support system based on reinforcement learning; A layered reinforcement learning framework, including strategic, tactical, and execution layers; Based on the reinforcement learning framework, a geological risk perception module, a gas diffusion predictor, a multimodal decision interpreter, an automatic emergency plan generator, and a collaborative response optimizer are introduced; In this way, the training of the intelligent decision support system based on reinforcement learning is completed. Conservative strategy constraints, safe exploration mechanisms and human-computer collaborative interfaces are added during the training process to complete the construction of the intelligent decision support system based on reinforcement learning.

9. The intelligent HSE management method for underground gas storage according to claim 1, characterized in that: The adaptive learning mechanism includes a local learning module, a federated aggregation server, a knowledge distillation module, a differential privacy protection layer, a heterogeneous data adapter, and a continuous evaluation and adjustment module; Each gas storage facility continuously collects local data, and the local learning module regularly updates the model using new local data while adapting to local special conditions through DANN; The updated model parameters and / or gradients are sent to the federated aggregation server after differential privacy processing; The aggregation server uses the AWFA algorithm to integrate updates from various gas storage facilities to generate a global model; The global model is fed back to each gas storage facility through knowledge distillation and integrated with the local risk assessment model. The continuous evaluation and adjustment module monitors the entire process and dynamically adjusts the learning strategy and privacy protection strength.

Images