Marine oil and gas equipment data monitoring method and system based on enhanced graph learning

Abstract

The invention relates to the technical field of marine environment monitoring, in particular to a marine oil and gas equipment data monitoring method and system based on enhanced graph learning. The method comprises the following steps: carrying out preprocessing and textualization conversion on original node attributes of a dynamic network structure chart; analyzing the converted dynamic network structure chart by using a large language model (LLM); feature extraction is carried out from the semantic embedding side and the structure embedding side; performing semantic embedding and structure embedding feature fusion by using a multi-head attention mechanism; downstream tasks are executed based on fusion embedding, and data monitoring is achieved. The LLM-GNN collaborative optimization system provided by the invention realizes the intelligent breakthrough of data monitoring of the offshore oil and gas system.

Description

Technical Field

[0001] The present invention relates to the technical field of marine environment monitoring, and in particular to a marine oil and gas equipment data monitoring method and system based on enhanced graph learning. Background Art

[0002] As deep-sea oil and gas development extends to ultra-deep waters above 1,500 meters, the system faces three major technical bottlenecks: First, the challenge of processing ocean-specific data. The digital nodes of deep-sea oil and gas fields form a three-dimensional monitoring network covering thousands of square kilometers, of which unstructured data accounts for more than 65%. Traditional numerical analysis has difficulty in resolving the nonlinear noise interference unique to the marine environment; second, the accuracy of dynamic modeling is insufficient. The existing static model cannot adapt to accelerated aging in high-pressure and low-temperature environments, the synergistic effect of seawater corrosion and microbial attachment, and dynamic load changes during the typhoon season, resulting in an 18-25% increase in the false alarm rate in scenarios such as deep-water pipeline stress monitoring; third, the real-time response requirement in extreme environments. For high-risk scenarios such as blowout warning and submarine landslide monitoring, the decision-making loop from data collection to emergency shutdown must be completed within 300 milliseconds, but the current multimodal data processing delay cannot meet the millisecond-level response requirements in strong ocean current areas.

[0003] While data analysis methods based on graph neural networks (GNNs) have made some progress in device relationship modeling and data monitoring, their application in real-world scenarios still faces several key technical bottlenecks. Specifically, existing GNN solutions face the following challenges: First, in terms of topology modeling, traditional methods generally assume static graph structures, failing to effectively capture the essential characteristics of the dynamic evolution of device nodes. Second, in terms of data utilization efficiency, existing systems suffer from severe modal imbalance. Industry research reports indicate that approximately 78% of industrial monitoring systems still rely on traditional numerical signal processing, simply archiving or discarding text data such as equipment logs and maintenance records. This processing approach directly results in the loss of critical semantic information such as device operating context and maintenance history, making it difficult for models to identify complex failure modes. Furthermore, in terms of multimodal fusion methods, currently mainstream shallow fusion strategies such as feature concatenation or ensemble voting struggle to establish deep correlations between semantic features and physical signals. This superficial fusion approach fails to capture cross-modal causal relationships, such as those described in logs regarding insulation degradation and increased partial discharge signals. Ultimately, the generated maintenance recommendations deviate significantly from actual conditions. These limitations severely constrain the effectiveness of offshore oil and gas equipment systems, necessitating technological innovation to achieve breakthroughs in the following areas: real-time evolution of dynamic topology, deep semantic understanding of multimodal data, collaborative reasoning mechanisms for cross-modal features, and improved industrial-grade computing efficiency. Only by addressing these issues can the potential of GNNs for intelligent equipment operation and maintenance be truly unleashed.

[0004] At the same time, existing methods combining large language models (LLMs) with graph neural networks (GNNs) face the following challenges in practical applications: First, LLMs are good at processing semantic information but lack the ability to model graph structures, while GNNs, while able to capture topological relationships, have difficulty understanding complex semantics, resulting in difficulty in feature alignment when fusing the two. Second, data in industrial network scenarios (such as device status and user behavior) often changes dynamically, making it difficult for existing methods to adaptively adjust graph structures or efficiently fuse multimodal inputs, limiting the model's generalization capabilities. In addition, LLMs have a large number of parameters, and direct coupling with GNNs will lead to a surge in computational overhead, especially in scenarios with large-scale data (such as millions of device nodes), resulting in low training and inference efficiency. At the same time, both LLMs and GNNs suffer from black-box nature, and their joint decision-making processes (such as node classification and anomaly detection) lack interpretability, making it difficult to meet the model credibility requirements of industrial scenarios. Finally, existing methods mostly use fixed architectures, making it difficult to dynamically adjust the interaction between LLMs and GNNs for different tasks (such as fault diagnosis and supply chain optimization), resulting in performance degradation when migrating across tasks. Therefore, there is an urgent need for an efficient, interpretable and dynamic scene-adaptive LLM-GNN fusion method to solve the above problems. Summary of the Invention

[0005] In order to solve the above-mentioned problems, the present invention provides a method and system for monitoring marine oil and gas equipment data based on enhanced graph learning.

[0006] In a first aspect, the present invention provides a method for monitoring marine oil and gas equipment data based on enhanced graph learning, which adopts the following technical solutions: A method for monitoring marine oil and gas equipment data based on enhanced graph learning, comprising: Acquire multi-source sensor data and construct a dynamic network structure graph with spatiotemporal characteristics; Preprocess and convert the original node attributes of the dynamic network structure graph into text; Use the large language model LLM to parse the converted dynamic network structure graph; Feature extraction is performed from the semantic embedding side and the structural embedding side respectively; Use multi-head attention mechanism to fuse semantic embedding and structural embedding features; Execute downstream tasks based on fusion embedding to achieve data monitoring.

[0007] Furthermore, the method of acquiring multi-source sensor data and constructing a dynamic network structure diagram with spatiotemporal characteristics includes deploying a multi-source heterogeneous sensor network to collect device operation status data in real time, constructing a dynamic network structure diagram G=(V,A,X,Y) with spatiotemporal characteristics, and expanding the basic node set V, upgrading each device node vi∈V to an enhanced node entity with spatiotemporal perception capabilities, upgrading the adjacency matrix A from a static definition to a spatiotemporal correlation matrix, and integrating the four types of modal features into the feature matrix X to form a unified representation.

[0008] Furthermore, the original node attributes of the dynamic network structure diagram are preprocessed and converted into text, including mapping continuous attributes to the [0, 1] interval using the min-max normalization method; for discrete attributes, converting them into binary vectors using one-hot encoding; and generating a text segment for each node, which contains all the attribute information of the node, and sorting and describing the attributes according to the weight of the semantic association. The description is expressed as: The attributes of node [node number] include: [attribute 1 name]: [attribute 1 value], [attribute 2 name]: [attribute 2 value], ..., [attribute n name]: [attribute n value]; Among them, attribute names and attribute values are sorted according to the weight of semantic association, and the attributes with higher weights are ranked higher.

[0009] Furthermore, the converted dynamic network structure diagram is parsed using the large language model LLM, including using the LLM to analyze the node attribute text, inferring the potential relationship between the nodes, for each pair of nodes, inputting their attribute text into the LLM, and obtaining the relationship description output by the LLM, and calculating the relationship weight between each pair of nodes based on the similarity of the attribute text vectors of the nodes, which is expressed as: , in, represents the relationship weight between node i and node j, and Represent the attribute text vectors of node i and node j respectively, represents the similarity between the attribute text vectors of node i and node j, and n represents the total number of nodes.

[0010] Furthermore, the use of the large language model (LLM) to parse the converted dynamic network structure graph also includes extracting the topological features of the underwater equipment nodes using a graph calculation method based on the adjacency matrix. For the deep-sea production network, the degree centrality of each node is calculated as the basic structural attribute, and the dynamic weighted PageRank algorithm is used to accurately evaluate the global importance of the underwater equipment nodes. The time-varying communication weight and environmental disturbance factor are introduced to adapt to the complex working conditions of the deep sea. The iterative calculation process is expressed as follows: , Among them, N ( i） Represents the nodes constructed by the marine oil and gas equipment system you Neighbor set, damping coefficient α ∈(0,1) is optimized to 0.8 according to the characteristics of the ocean environment, N The total number of device nodes for all builds, is the time-varying edge weight, reflecting the node you and vj At the moment t Real-time communication load, is the environmental disturbance factor, Set to 0.8.

[0011] Furthermore, the use of the large language model LLM to parse the converted dynamic network structure graph also includes using the LLM to perform edge adjustment and adaptive graph structure optimization, wherein the LLM is first used to parse the node pair ( you , vj ), output the relationship type and confidence and construct the time series relationship tensor , aggregate historical information through LSTM, and then dynamically fuse topological features with semantic features, expressed as: , The balance factor β Learned jointly by topological and semantic features, .

[0012] Furthermore, feature extraction is performed from both the semantic embedding and structural embedding sides. On the semantic embedding side, hierarchical semantic encoding is performed on the nodes building the marine oil and gas system. For equipment working at a specific time, local time window feature extraction is performed. Equipment logs are divided into blocks according to fixed time windows, and a lightweight LLM is input into the blocks to obtain local semantic representations. A sliding window strategy is used to cover all time series data. For continuously working equipment, global time series feature extraction is performed. The entire equipment history log is input into the LLM to extract a global context representation, and dynamic gating is designed to fuse local and global features, expressed as follows: , in, is the Sigmoid function, are learnable parameters.

[0013] Furthermore, the feature extraction from the semantic embedding side and the structural embedding side respectively also includes, on the structural embedding side, using a spatiotemporal graph convolutional network to obtain information at the temporal and spatial levels. First, spatial dependency modeling is performed, and a multi-scale graph attention network is used to aggregate the topological associations of the underwater production system. According to the ocean characteristic parameters, an environmental disturbance factor is introduced into the attention coefficient, and the probability of connection between nodes is calculated by node similarity. For the long-period characteristics of deep-water equipment signals, a void causal convolution is used to capture long-term temporal patterns, and residual connections and hierarchical normalization are performed. The node similarity is expressed as: , in is the edge weight.

[0014] Furthermore, the multi-head attention mechanism is used to fuse semantic embedding and structural embedding features, including distinguishing the embedding sources based on adversarial cross-modal alignment identification, and performing linear transformations on the semantic embedding and structural embedding respectively to generate query vectors, key vectors and value vectors. The attention weights between the semantic embedding and the structural embedding are then calculated to measure the importance of the two in the fusion process, and residual connections are used to fuse the semantic information back to the original features to avoid information loss caused by direct coverage. Finally, downstream tasks are performed based on the fused embedding, and LLM and GNN parameters are jointly optimized through end-to-end training to achieve real-time monitoring of marine oil and gas equipment node data.

[0015] The second aspect is a marine oil and gas equipment data monitoring system based on enhanced graph learning, including: The data acquisition module is configured to acquire multi-source sensor data and construct a dynamic network structure graph with spatiotemporal characteristics; The preprocessing module is configured to preprocess and convert the original node attributes of the dynamic network structure graph into text; The parsing module is configured to parse the converted dynamic network structure graph using a large language model (LLM); The feature extraction module is configured to extract features from the semantic embedding side and the structural embedding side respectively; The feature fusion module is configured to use the multi-head attention mechanism to fuse semantic embedding and structural embedding features; The monitoring module is configured to perform downstream tasks based on the fused embedding to achieve data monitoring.

[0016] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, a method for monitoring marine oil and gas equipment data based on enhanced graph learning.

[0017] In a fourth aspect, the present invention provides a terminal device comprising a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor to implement the method for monitoring marine oil and gas equipment data based on enhanced graph learning.

[0018] In summary, the present invention has the following beneficial technical effects: (1) This invention innovatively integrates the Large Language Model (LLM) with the Graph Neural Network (GNN). Through the LLM's powerful semantic understanding and reasoning capabilities, it breaks through the limitations of traditional marine oil and gas equipment system data monitoring, which relies solely on structured data. The LLM can deeply analyze unstructured text such as equipment logs and maintenance work orders, extract key semantic relationships such as "abnormal motor vibration leads to increased bearing temperature", and convert them into dynamic weights to be injected into the GNN's adjacency matrix. This enables the graph structure, which originally only modeled explicit physical connections, to capture implicit fault propagation paths. This fusion architecture can effectively improve the accuracy of fault diagnosis in actual measurements of multiple industrial scenarios.

[0019] (2) The LLM-GNN collaborative optimization system proposed in this paper achieves an intelligent breakthrough in marine oil and gas system data monitoring. The system uses LLM to parse the equipment status description text in real time. Combined with the Granger causality test and dynamic time warping algorithm, it automatically identifies potential equipment associations and dynamically optimizes the graph structure: when a motor current waveform and bearing temperature sequence are detected to have a significant causal relationship, the edge connection between the corresponding nodes is automatically established and strengthened; otherwise, redundant connections are weakened or deleted. The optimized graph structure is input into the GNN for multi-level feature aggregation, and then the semantic embedding extracted by the LLM is dynamically fused with the structural embedding generated by the GNN through a cross-modal attention mechanism.

[0020] (3) The system design of the present invention is highly adaptable and scalable. Through modular design, the system can quickly adapt to the needs of different industrial scenarios: in the supply chain optimization scenario, LLM constructs a supplier relationship diagram by analyzing order logs and logistics records, and GNN predicts potential delivery risk nodes; in the energy consumption management scenario, LLM analyzes equipment energy efficiency reports to identify abnormal energy consumption patterns, and GNN establishes a correlation network between energy-consuming devices to locate energy-saving potential points. This flexible architecture allows enterprises to complete model deployment for new scenarios in a relatively short time by simply adjusting the prompt word template and task-specific header (TaskHead), saving implementation costs compared to traditional custom development solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of a method for monitoring marine oil and gas equipment data based on enhanced graph learning according to Example 1 of the present invention. DETAILED DESCRIPTION

[0022] The present invention will be further described in detail below with reference to the accompanying drawings.

[0023] Example 1 Reference Figure 1 , a marine oil and gas equipment data monitoring method based on enhanced graph learning in this embodiment includes: S1. Collect multi-source sensor data (numerical data, text, images, etc.) from marine oil and gas equipment systems and construct a dynamic graph structure, in which the devices serve as nodes and the physical connections, communication relationships, and semantic associations between devices serve as edges. Node features are fused with numerical signals, text logs, and spatiotemporal information to form a multimodal graph representation.

[0024] S2. Preprocess and textualize node attributes, normalize and encode numerical features (such as sensor readings), perform one-hot encoding on discrete features, and extract key semantic information from device logs to generate a structured-text hybrid node feature representation to enhance the model's ability to understand complex industrial data.

[0025] S3. Use a large language model (LLM) to parse device logs and status descriptions. Prompt templates are used to guide the LLM to output semantic information between nodes. Dynamic edge weights are calculated based on the original graph structure. A time-series sliding window mechanism is introduced, and an LSTM is used to dynamically update the adjacency matrix.

[0026] S4. On the semantic embedding side, a hierarchical encoding strategy is adopted to extract local time window features and global temporal features of device logs respectively, and adaptively fuse them through a gating mechanism. On the structural embedding side, a spatiotemporal graph convolutional network (TCN+GCN) is integrated to capture the spatiotemporal dependencies of device states, and residual connections are added to maintain gradient stability. Adversarial training is further introduced to force the alignment of semantic embedding and structural embedding in the latent space through a domain discriminator, enhancing the synergy of bimodal representation and providing more robust feature expression for downstream tasks.

[0027] S5. Design a multi-head attention mechanism to conduct cross-modal interaction between semantic embedding and structural embedding, dynamically calculate the attention weights of different modalities, achieve deep fusion of semantic and topological information, output a joint embedding representation, and enhance the model's modeling capabilities for complex industrial scenarios.

[0028] S6. Based on the requirements of downstream tasks (such as fault prediction and anomaly detection), the fused node features are converted into node embeddings through GCN to obtain device status representations. LLM and GNN parameters are optimized through end-to-end training to achieve real-time monitoring and intelligent decision-making of industrial IoT data.

[0029] The step S1 specifically includes: S1.1: In the marine oil and gas data twin intelligent monitoring system, by deploying a multi-source heterogeneous sensor network (including vibration sensors, temperature sensors, current transformers, etc.) to collect equipment operating status data in real time, a dynamic network structure diagram G=(V,A,X,Y) with spatiotemporal characteristics is constructed.

[0030] The node set V = {v_1,v_2,...,v_N} fully represents all types of device entities in the system. Each node v_i not only contains basic information such as device ID and type, but also records topological attributes such as the device's spatial coordinates and installation location. The adjacency matrix A∈R^(N×N) accurately captures the complex relationships between nodes along three dimensions: hard connections based on physical connections (such as cable connections and communication links), soft connections based on data correlations (such as covariance analysis), and semantic connections based on maintenance records (such as fault propagation paths). The node feature set X = {x_1,x_2,...,x_N} is a multidimensional spatiotemporal tensor. Each x_i∈R^(T×D) integrates D-dimensional features of the device over T time steps, including: 1) numerical sensor readings (such as time-series signals such as temperature and vibration amplitude), 2) textual maintenance logs (semantic vectors encoded using BERT), and 3) inherent device attributes (static parameters such as model and age). The device label set Y is flexibly defined according to the specific monitoring task. It can be a discrete fault type code (for classification tasks) or a continuous remaining life prediction value (for regression tasks), supporting multi-task joint learning.

[0031] S1.2: In order to address the core issues of spatiotemporal heterogeneity and multimodal fragmentation that traditional static graph models have difficulty in handling in the marine oil and gas equipment system scenario, we first intelligently expand the basic node set V and upgrade each device node vi∈V to an enhanced node entity with spatiotemporal awareness. Its data structure is defined as: vi={ ID: device unique identifier (device code), Device type: three-level classification system based on function (sensor / controller / actuator), Physical coordinates: (x, y, z) three-dimensional space position (accurate to millimeter level), Timestamp queue: {t1,...,tn} (records key state change moments), Dynamic properties: Sampling frequency: device-specific sampling rate (e.g. 10Hz vibration sensor), Communication protocol: Modbus / TCP and other industrial protocol identification, Health status: Status coding based on ISO standards } To address the non-uniform sampling data commonly found in industrial sites, we use interpolation to unify the time granularity: , Hierarchical labels are assigned based on device type (sensor / controller / actuator) for subsequent hierarchical graph attention. (1) Sensor nodes: labeled as L1 (data acquisition layer); (2) Controller nodes: labeled as L2 (logical control layer); (3) Actuator nodes: labeled as L3 (physical execution layer). Differentiated attention calculation strategies are set between each layer, such as time-sensitive attention for L1→L2 and causal constraint attention for L2→L3. This hierarchical design can capture the vertical information propagation characteristics in industrial control systems.

[0032] S1.3: Adjacency Matrix A Upgrade from static definition to spatiotemporal correlation matrix, where i and j are node numbers: , in It is a learnable parameter that dynamically adjusts the contribution of each mode through back propagation.

[0033] (1) Physical connection weight : Based on device communication links (such as Modbus / TCP connections): , in is the distance between the two devices.

[0034] (2) Semantic association weight : LLM is used to parse device logs to generate relationship descriptions and calculate cosine similarity. This weight reveals implicit associations between devices: : in 、 It is the information of sensors i and j at time t (the key time point of device status change).

[0035] (3) Causal temporal weight : Based on Granger causality test and DTW time series alignment: , When the Granger test p-value is less than 0.05, it is assumed to be 1 (causality exists); otherwise, it is assumed to be 0. This metric ensures that only statistically significant causal relationships are preserved. In actual industrial scenarios, lag order constraints are added to avoid long-time lag spurious causality. DTW time series alignment addresses the issue of inconsistent sampling rates between devices and calculates the optimal alignment path cost for the two time series.

[0036] S1.4: Perform multimodal fusion of node feature X: Feature Matrix X Integrate four types of modal features to form a unified representation: , Temporal alignment: Sliding windows and cubic spline interpolation are used to unify the timestamps of non-uniformly sampled data. Spatial alignment: Geohash-3D encoding is used to map device coordinates to a unified spatial grid (accuracy ±5cm). Semantic alignment: The BERT model fine-tuned in the industrial field is used to extract text features and construct an embedding space for device fault codes.

[0037] The step S2 specifically includes: S2.1: Preprocess the original node attributes to ensure that they are suitable for conversion into text. For continuous attributes, use the min-max normalization method to map them to the [0,1] interval: , in, represents the attribute value of the i-th node, and Respectively represent the minimum and maximum values of the attribute, Represents the normalized attribute value. For discrete attributes, we use one-hot encoding to convert them into binary vectors: , in, represents the discrete attribute vector of the i-th node, Indicates whether the node belongs to the kth category, and k represents the total number of categories of the attribute.

[0038] S2.2: Generate a text fragment for each node, which contains all the attribute information of the node and sorts and describes the attributes according to the weight of semantic association. Use the following template to generate text description: "The attributes of node [node number] include: [attribute 1 name]: [attribute 1 value], [attribute 2 name]: [attribute 2 value], ..., [attribute n name]: [attribute n value]." Among them, attribute names and attribute values are sorted according to the weight of semantic association, and the attributes with higher weights are ranked higher.

[0039] The step S3 specifically includes: S3.1: After obtaining the textual descriptions of node attributes, the graph structure is optimized using a large language model (LLM). LLM can understand the semantic information in the node attribute text and infer the potential relationships between nodes based on this information, thereby optimizing the graph structure.

[0040] First, we need to infer node relationships. We use LLM to analyze node attribute text and infer potential relationships between nodes. For each pair of nodes, we input their attribute text into LLM and obtain the relationship description output by LLM. We use the following template to generate the relationship description: "The relationship between node [node number 1] and node [node number 2] is: [relationship description]." S3.2 calculates the relationship weight between each pair of nodes using the following formula: , in, represents the relationship weight between node i and node j, and Represent the attribute text vectors of node i and node j respectively, represents the similarity between the attribute text vectors of node i and node j, and n represents the total number of nodes.

[0041] S3.3 analyzes the network topology of the offshore oil and gas production system and uses a graph calculation method based on the adjacency matrix to extract the topological characteristics of the underwater equipment nodes. For the submarine production control system (such as the submarine control module SCM, submarine valves, jumper pipes and other key equipment nodes), first calculate its degree centrality as the basic structural attribute: for each node , calculate its degree centrality as the basic structural property: , In deep-sea production networks, this indicator directly reflects the physical / data connection density of underwater equipment. It is typically expressed as: High-value nodes (degree centrality > 15) are usually underwater master control stations (MasterControlStation) or subsea distribution units (SDU); Medium-value nodes (3<degree centrality≤15) are mostly relay equipment such as underwater Christmas trees and manifolds; Low-value nodes (degree centrality ≤ 3) correspond to edge sensors (e.g., seabed pressure and temperature gauges, corrosion monitoring probes); When the adjacency matrix contains edge weights When the weighted degree centrality is based on the acoustic communication frequency or the umbilical cable data transmission volume, the weighted degree centrality is: , This improved metric can accurately identify: 1) high-frequency communication nodes (such as underwater acoustic modem hubs); 2) critical data relay equipment (such as fiber-optic communication nodes between jumper pipes); and 3) emergency communication paths (such as the burst communication load of the ROV docking interface).

[0042] At the same time, the Dynamic Weighted PageRank algorithm (DWPR) is used to more accurately assess the global importance of underwater device nodes. Based on the traditional PageRank algorithm, this algorithm introduces time-varying communication weights and environmental disturbance factors to better adapt to the complex working conditions of deep sea. The iterative calculation process is as follows: , Among them, N ( i） Represents the nodes constructed by the marine oil and gas equipment system you Neighbor set, damping coefficient α ∈(0,1) is optimized to 0.8 (the default value is 0.85) according to the characteristics of the ocean environment. N The total number of device nodes for all builds.

[0043] is the time-varying edge weight, reflecting the node you and vj At the moment t Real-time communication load (such as acoustic data transmission volume, hydraulic signal strength); is an environmental disturbance factor used to quantify the impact of ocean currents on underwater wireless communications (0.7-1.2), sensor signal attenuation caused by microbial attachment (0.8-1.0), and network topology changes during typhoon passage (0.5-1.5); Set to 0.8 to match the delay characteristics of underwater communication.

[0044] S3.3 addresses the unique challenges of marine oil and gas system data: (1) the "data island" problem unique to deep-sea equipment, (2) the triple uncertainty of the deep-sea environment (ocean current impact, biological adhesion, and communication delay). For example, when topological analysis detects a sudden increase in communication delay, integrating the semantic feature of "sudden increase in seabed turbidity" in the ROV log makes it difficult to distinguish whether it is environmental interference or equipment failure, and (3) traditional methods have hidden fault chains that are difficult to detect. This paper deeply integrates the equipment topological features extracted by GCN with the operation and maintenance semantic features analyzed by LLM to build a multi-dimensional cognitive system for equipment status.

[0045] First, the continuous signals collected by the sensor (such as temperature and vibration amplitude) are adaptively normalized according to the working conditions: , The sliding window W is dynamically adjusted according to the device type: Rapid response equipment (such as blowout preventers): 10-second short window; Slowly varying parameters (such as seabed soil temperature): 24-hour long window; To address the asynchronous sampling problem unique to the marine environment, a Kalman-particle filter hybrid algorithm is used to compensate for possible influences (such as acoustic communication delay, non-periodicity of ROV inspection data, and sensor drift caused by tidal cycles) and generate a time-aligned composite feature matrix: , , Where A is the state transition matrix for modeling the physical dynamics of underwater equipment, H is the observation matrix for observing the relationship between vibration, temperature, and acoustic channels, Q is the process noise covariance (the temperature term in the Q matrix can be adjusted to address the problem of asynchronous data alignment), and N is the number of particles that accounts for computational efficiency and the nonlinear characteristics of deep-sea data.

[0046] Use a sliding window (e.g., a 10-second window) to align sensor data with different sampling rates. After filling missing values with a Kalman filter, splice the data into a composite matrix along the feature dimension, and then perform the splicing operation: , Corresponding to discrete features, enhancement operations are performed: , Among them, the LLM embedded input dimension dLLM =768 (BERT-base model), and use attention weighting (weight α ∈[0.8,1.5]). Encoded into standard term vectors, the topological structure of the standard document is extracted through a graph encoder.

[0047] Then perform cross-modal gating fusion: , Among them, the topological features : Number of graph convolution layers L =3: covers the three-level topology of the underwater production system "wellhead-manifold-umbilical cable"; gating weight g: bias term b g Initialized to -0.5: Tends to prioritize trust in topological features (due to sparse ocean text data) Threshold adjustment after activation: Forced in typhoon mode g >0.7 (mainly based on topological stability).

[0048] S3.3 Edge adjustment using LLM: First use LLM to parse the node pairs ( you , vj ) (from S2.3), output relationship type (causal / cooperative / dependent) and confidence level: , Constructing a temporal relationship tensor , aggregate historical information through LSTM: , Then the topological features are dynamically fused with the semantic features: , The balance factor β Jointly learned from topological and semantic features: , S3.4 Adaptive graph structure optimization: Dynamically adjust the original adjacency matrix : , Learning rate λ Adaptive adjustment of variance through sliding window: , The updated adjacency matrix needs to be sparsely processed (pruning and adding edges) to make its information density higher, including removing ( value is set to 0.2), and for high confidence semantic relations And topologically sparse nodes Add a new edge. Then, you need to Apply a smooth constraint: , The step S4 specifically includes: S4.1 Semantic Embedding Learning (LLM Side) After adjusting the graph structure, semantic embedding and structural embedding need to be obtained for subsequent optimization. These two embeddings represent nodes from different perspectives and solve two key problems in graph data analysis: (1) Semantic embedding: By encoding the node attribute text through LLM, the semantic embedding captures the semantic information of the node attributes. (2) Structural embedding: By encoding the optimized graph structure through GNN, the structural embedding captures the topological structure information of the nodes in the graph. The former can help the model understand the deep semantic relationship between node attributes, especially when the attribute text is complex and diverse, and the latter can help the model capture the connection pattern between nodes and the global characteristics of the graph structure.

[0049] First, we perform hierarchical semantic encoding on the nodes in the offshore oil and gas system. For equipment operating at a specific time, we extract features from local time windows. We then partition the equipment logs into fixed time windows and input them into a lightweight LLM (DistilBERT is used in this article) to obtain local semantic representations: , A sliding window strategy (step size = 1) is used to cover all time series data.

[0050] For devices that are always working, global temporal feature extraction is performed. The entire device history log (24 hours) is input into the LLM to extract the global context representation: , in It is a keyword attention mask based on TF-IDF to strengthen the weight of fault-related words.

[0051] Design dynamic gating to integrate local and global features: , is the Sigmoid function, are learnable parameters.

[0052] S4.2 Structural Embedding Learning (GNN Side) In order to adapt to the needs of ocean scenes, this paper uses the spatiotemporal graph convolutional network (ST-GNN) to obtain information at the temporal and spatial levels.

[0053] First, spatial dependency modeling is performed, and a multi-scale graph attention network is used to aggregate the topological associations of the underwater production system: , According to the ocean characteristic parameters, the attention coefficient Introducing environmental disturbance factors, calculated by node similarity: , in is the edge weight, taking into account the signal communication quality (SNR≥20dB η =1.0, otherwise it decays linearly) and ocean current impact intensity (ADCP measured velocity > 1m / s η ← η ×0.8) two key parameters.

[0054] In view of the long-period characteristics of deepwater equipment signals, DilatedTCN is used to capture long-term timing patterns: , in l is the number of layers, k is the convolution kernel size.

[0055] Perform residual connection and layer normalization: , Used to align input and output dimensions.

[0056] S4.3 Adversarial Cross-Modal Alignment, In deep-sea oil and gas production systems, there are significant differences in the multimodal data (topological structure vs. semantic description) of device nodes. Topological features (GNN side): reflect the physical connection between underwater equipment (such as umbilical cable communication, hydraulic control link) and dynamic interaction (such as sensor data drift under ocean current disturbance). Semantic features (LLM side): include unstructured text descriptions such as maintenance logs (such as "underwater valve seal failure") and ROV inspection reports (such as "abnormal detachment of anode blocks"). Discriminator D It is necessary to distinguish the source of embedding (semantic / topological) while adapting to the dynamic disturbances of the marine environment.

[0057] Considering the above factors, the neighborhood discriminator is designed and the discriminator is constructed. D Distinguish embedding sources (semantic / structural): , Adversarial training objective, step 1: discriminator loss (maximize source classification accuracy): , Step 2: Generator loss (minimizing modality difference): , Compared with traditional GRL, the discriminator gradient is multiplied by λ =−1, adaptive weighting needs to be introduced in dynamic ocean environments: , The step S5 specifically includes: S5.1: Semantic embeddings and structural embeddings focus on the attribute semantics and topological structure of nodes, respectively. There is a certain degree of information isolation between the two. To fully leverage the complementary nature of these two embeddings, a multi-head attention mechanism is employed to fuse them. This multi-head attention mechanism dynamically balances the importance of semantic and structural information and adaptively adjusts the contributions of the two embeddings based on the needs of specific tasks, thereby generating more comprehensive and expressive node representations.

[0058] S5.2: We perform linear transformations on the semantic embedding and structural embedding to generate the query vector (Query), key vector (Key), and value vector (Value): , Among them, Qs, Ks and Vs represent the query vector, key vector and value vector of semantic embedding respectively. denote the query vector, key vector and value vector of the structure embedding respectively, They represent the weight matrices of linear transformation respectively.

[0059] S5.3: Next, we calculate the attention weight between the semantic embedding and the structural embedding to measure the importance of the two in the fusion process: , in, represents the attention weight between node i and node j, and denote the query vector of semantic embedding and the key vector of structural embedding respectively, and n denotes the total number of nodes.

[0060] Finally, we use the calculated attention weights to perform weighted fusion of semantic embedding and structural embedding to generate the final node representation: , in, represents the final representation of node i, represents the attention weight between node i and node j, A vector of values representing the embedding of the structure.

[0061] The step S6 specifically includes: S6.1: Use residual connections to fuse semantic information back to the original features to avoid information loss caused by direct overlay: , in, is a learnable linear transformation matrix used to align dimensions.

[0062] S6.2: Execute downstream tasks based on fusion embedding (such as equipment failure prediction and network anomaly detection, which depend on the specific task requirements and are not the focus of this article), and jointly optimize network parameters through end-to-end training to achieve real-time monitoring of industrial IoT node data.

[0063] This article takes fault prediction as an example: For the processed graph data, we first use GCN to aggregate node neighbor information through the normalized adjacency matrix to obtain the node embedding H: , in is the degree matrix, is a trainable parameter. Then the failure probability of marine oil and gas equipment is obtained through mapping: , The output range is [0,1], indicating the possibility of device failure.

[0064] Example 2 This embodiment provides a marine oil and gas equipment data monitoring system based on enhanced graph learning, including: The data acquisition module is configured as follows: A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, a method for monitoring marine oil and gas equipment data based on enhanced graph learning.

[0065] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded and executed by the processor to implement a method for monitoring marine oil and gas equipment data based on enhanced graph learning.

[0066] The above are all preferred embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for monitoring marine oil and gas equipment data based on enhanced graph learning, characterized in that: include: Acquire multi-source sensor data and construct a dynamic network structure graph with spatiotemporal characteristics; Preprocess and convert the original node attributes of the dynamic network structure graph into text; Use the large language model LLM to parse the converted dynamic network structure graph; Feature extraction is performed from the semantic embedding side and the structural embedding side respectively; Use multi-head attention mechanism to fuse semantic embedding and structural embedding features; Execute downstream tasks based on fusion embedding to achieve data monitoring.

2. The method for monitoring marine oil and gas equipment data based on enhanced graph learning according to claim 1, characterized in that: The method of acquiring multi-source sensor data and constructing a dynamic network structure diagram with spatiotemporal characteristics includes deploying a multi-source heterogeneous sensor network to collect device operation status data in real time, constructing a dynamic network structure diagram G=(V,A,X,Y) with spatiotemporal characteristics, and expanding the basic node set V, upgrading each device node vi∈V to an enhanced node entity with spatiotemporal perception capabilities, upgrading the adjacency matrix A from a static definition to a spatiotemporal correlation matrix, and integrating four types of modal features into the feature matrix X to form a unified representation.

3. The method for monitoring marine oil and gas equipment data based on enhanced graph learning according to claim 2, characterized in that: The original node attributes of the dynamic network structure diagram are preprocessed and converted into text, including mapping continuous attributes to the [0, 1] interval using the min-max normalization method; for discrete attributes, converting them into binary vectors using one-hot encoding; and generating a text segment for each node, which contains all the attribute information of the node, and sorting and describing the attributes according to the weight of the semantic association. The description is expressed as: The attributes of node [node number] include: [attribute 1 name]: [attribute 1 value], [attribute 2 name]: [attribute 2 value], ..., [attribute n name]: [attribute n value]; Among them, attribute names and attribute values are sorted according to the weight of semantic association, and the attributes with higher weights are ranked higher.

4. The method for monitoring marine oil and gas equipment data based on enhanced graph learning according to claim 3, characterized in that: The converted dynamic network structure diagram is parsed using the large language model LLM, including using the LLM to analyze the node attribute text, inferring the potential relationship between the nodes, for each pair of nodes, inputting their attribute text into the LLM, and obtaining the relationship description output by the LLM, and calculating the relationship weight between each pair of nodes based on the similarity of the attribute text vectors of the nodes, which is expressed as: ， in, represents the relationship weight between node i and node j, and Represent the attribute text vectors of node i and node j respectively, It represents the similarity between the attribute text vectors of node i and node j, n represents the total number of nodes, and k represents the number of nodes.

5. The method for monitoring marine oil and gas equipment data based on enhanced graph learning according to claim 4, characterized in that: The converted dynamic network structure graph is parsed using the large language model (LLM), and the topological features of the underwater equipment nodes are extracted using a graph calculation method based on an adjacency matrix. For the deep-sea production network, the degree centrality of each node is calculated as the basic structural attribute, and a dynamic weighted PageRank algorithm is used to accurately evaluate the global importance of the underwater equipment nodes. The time-varying communication weight and environmental disturbance factor are introduced to adapt to the complex working conditions of the deep sea. The iterative calculation process is expressed as follows: ， Among them, N ( i） Represents the nodes constructed by the marine oil and gas equipment system vi Neighbor set, damping coefficient α ∈(0,1) is optimized to 0.8 according to the characteristics of the ocean environment, N The total number of device nodes for all builds, is the time-varying edge weight, reflecting the node vi and vj At the moment t Real-time communication load, is the environmental disturbance factor, Set to 0.

8.

6. The method for monitoring marine oil and gas equipment data based on enhanced graph learning according to claim 5, characterized in that: The method of using the large language model LLM to parse the converted dynamic network structure graph also includes using the LLM to perform edge adjustment and adaptive graph structure optimization, wherein the LLM is first used to parse the node pair ( vi , vj ), output the relationship type and confidence and construct the time series relationship tensor , aggregate historical information through LSTM, and then dynamically fuse topological features with semantic features, expressed as: ， The balance factor β Learned jointly by topological and semantic features, .

7. The method for monitoring marine oil and gas equipment data based on enhanced graph learning according to claim 6, characterized in that: The method extracts features from both the semantic embedding and structural embedding sides, including performing hierarchical semantic encoding on the nodes of the marine oil and gas system on the semantic embedding side, dividing the features of the equipment working at a specific time into local time windows for feature extraction; dividing the equipment logs into blocks according to fixed time windows, inputting lightweight LLM to obtain local semantic representations, and using a sliding window strategy to cover all time series data; and performing global time series feature extraction on the equipment working continuously. Input the entire device history log into the LLM, extract the global context representation, design dynamic gating, and fuse local and global features, which can be expressed as: ， in, is the Sigmoid function, are learnable parameters.

8. The method for monitoring marine oil and gas equipment data based on enhanced graph learning according to claim 7, characterized in that: The feature extraction is performed from the semantic embedding side and the structural embedding side respectively, and also includes using a spatiotemporal graph convolutional network to obtain information at the temporal and spatial levels on the structural embedding side. First, spatial dependency modeling is performed, and a multi-scale graph attention network is used to aggregate the topological associations of the underwater production system. According to the ocean characteristic parameters, the environmental disturbance factor is introduced into the attention coefficient, and the probability of connection between nodes is calculated by node similarity. For the long-period characteristics of deep-water equipment signals, a void causal convolution is used to capture the long-term temporal pattern, and residual connection and hierarchical normalization are performed. The node similarity is expressed as: ， in is the edge weight.

9. The method for monitoring marine oil and gas equipment data based on enhanced graph learning according to claim 8, characterized in that: The multi-head attention mechanism is used to fuse semantic embedding and structural embedding features, including distinguishing the embedding sources based on adversarial cross-modal alignment, and performing linear transformations on the semantic embedding and structural embedding respectively to generate query vectors, key vectors and value vectors. The attention weights between the semantic embedding and the structural embedding are then calculated to measure their importance in the fusion process, and residual connections are used to fuse the semantic information back to the original features to avoid information loss caused by direct coverage. Finally, downstream tasks are performed based on the fused embedding, and LLM and GNN parameters are jointly optimized through end-to-end training to achieve real-time monitoring of marine oil and gas equipment node data.

10. A marine oil and gas equipment data monitoring system based on enhanced graph learning, characterized in that: include: The data acquisition module is configured to acquire multi-source sensor data and construct a dynamic network structure graph with spatiotemporal characteristics; The preprocessing module is configured to preprocess and convert the original node attributes of the dynamic network structure graph into text; The parsing module is configured to parse the converted dynamic network structure graph using a large language model (LLM); The feature extraction module is configured to extract features from the semantic embedding side and the structural embedding side respectively; The feature fusion module is configured to use the multi-head attention mechanism to fuse semantic embedding and structural embedding features; The monitoring module is configured to perform downstream tasks based on the fused embedding to achieve data monitoring.

Images